ui-svelte: fix histogram calculation (#695)

- Fix the histogram calculation to use server provided generation
tokens/second.
- Move histogram to Activities page where it can exist with the rest of
the token metrics

Fixes #681

.coderabbit.yaml

@@ -4,7 +4,7 @@ early_access: false
 reviews:
   profile: "chill"
   request_changes_workflow: false
-  high_level_summary: true
+  high_level_summary: false
   poem: false
   review_status: true
   collapse_walkthrough: false

.github/workflows/config-schema.yml

@@ -4,11 +4,15 @@ on:
   pull_request:
     paths:
       - "config-schema.json"
+      - "config.example.yaml"
+      - ".github/workflows/config-schema.yml"
   push:
     branches:
       - main
     paths:
       - "config-schema.json"
+      - "config.example.yaml"
+      - ".github/workflows/config-schema.yml"
 
   workflow_dispatch:
 
@@ -39,3 +43,14 @@ jobs:
           fi
 
           echo "✓ config-schema.json is valid"
+
+      - name: Set up Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: "3.x"
+
+      - name: Install check-jsonschema
+        run: pip install check-jsonschema
+
+      - name: Validate config.example.yaml against schema
+        run: check-jsonschema --schemafile config-schema.json config.example.yaml

.github/workflows/containers.yml

@@ -10,17 +10,44 @@ on:
   # Allows manual triggering of the workflow
   workflow_dispatch:
 
+  # Run on workflow file changes (without pushing)
+  push:
+    paths:
+      - '.github/workflows/containers.yml'
+      - 'docker/build-container.sh'
+      - 'docker/*.Containerfile'
+
+# grant permissions on GITHUB_TOKEN to publish packages
+# ref: https://docs.github.com/en/packages/managing-github-packages-using-github-actions-workflows/publishing-and-installing-a-package-with-github-actions#publishing-a-package-using-an-action
+permissions:
+  contents: read
+  packages: write
+  id-token: write
+
 jobs:
   build-and-push:
     runs-on: ubuntu-latest
     strategy:
       matrix:
-        platform: [intel, cuda, vulkan, cpu, musa]
+        platform: [intel, cuda, cuda13, vulkan, cpu, musa, rocm]
       fail-fast: false
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
 
+      - name: Free up disk space
+        if: matrix.platform == 'rocm'
+        run: |
+          echo "Before cleanup:"
+          df -h
+          sudo rm -rf /usr/share/dotnet
+          sudo rm -rf /usr/local/lib/android
+          sudo rm -rf /opt/ghc
+          sudo rm -rf /opt/hostedtoolcache/CodeQL
+          sudo docker system prune -af
+          echo "After cleanup:"
+          df -h
+
       - name: Log in to GitHub Container Registry
         uses: docker/login-action@v2
         with:
@@ -31,7 +58,7 @@ jobs:
       - name: Run build-container
         env:
           GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-        run: ./docker/build-container.sh ${{ matrix.platform }} true
+        run: ./docker/build-container.sh ${{ matrix.platform }} ${{ github.event_name != 'push' }}
 
   # note make sure mostlygeek/llama-swap has admin rights to the llama-swap package
   # see: https://github.com/actions/delete-package-versions/issues/74

.github/workflows/go-ci-windows.yml

@@ -3,9 +3,25 @@ name: Windows CI
 on:
   push:
     branches: [ "main" ]
+    # only run when backend source changes
+    # cmd/ is excluded because it contains utilities without tests
+    paths:
+      - '**/*.go'
+      - '!cmd/**'
+      - 'go.mod'
+      - 'go.sum'
+      - 'Makefile'
+      - '.github/workflows/go-ci-windows.yml'
 
   pull_request:
     branches: [ "main" ]
+    paths:
+      - '**/*.go'
+      - '!cmd/**'
+      - 'go.mod'
+      - 'go.sum'
+      - 'Makefile'
+      - '.github/workflows/go-ci-windows.yml'
 
   # Allows manual triggering of the workflow
   workflow_dispatch:
@@ -28,7 +44,7 @@ jobs:
       uses: actions/cache/restore@v4
       with:
         path: ./build
-        key: ${{ runner.os }}-simple-responder-${{ hashFiles('misc/simple-responder/simple-responder.go') }}
+        key: ${{ runner.os }}-simple-responder-${{ hashFiles('cmd/simple-responder/simple-responder.go') }}
 
     # necessary for testing proxy/Process swapping
     - name: Create simple-responder
@@ -43,7 +59,7 @@ jobs:
       uses: actions/cache/save@v4
       with:
         path: ./build
-        key: ${{ runner.os }}-simple-responder-${{ hashFiles('misc/simple-responder/simple-responder.go') }}
+        key: ${{ runner.os }}-simple-responder-${{ hashFiles('cmd/simple-responder/simple-responder.go') }}
 
     - name: Test all
       shell: bash

.github/workflows/go-ci.yml

@@ -2,53 +2,68 @@ name: Linux CI
 
 on:
   push:
-    branches: [ "main" ]
+    branches: ["main"]
+    # only run when backend source changes
+    # cmd/ is excluded because it contains utilities without tests
+    paths:
+      - "**/*.go"
+      - "!cmd/**"
+      - "go.mod"
+      - "go.sum"
+      - "Makefile"
+      - ".github/workflows/go-ci.yml"
 
   pull_request:
-    branches: [ "main" ]
+    branches: ["main"]
+    paths:
+      - "**/*.go"
+      - "!cmd/**"
+      - "go.mod"
+      - "go.sum"
+      - "Makefile"
+      - ".github/workflows/go-ci.yml"
 
   # Allows manual triggering of the workflow
   workflow_dispatch:
 
 jobs:
-
   run-tests:
     runs-on: ubuntu-latest
     steps:
-    - uses: actions/checkout@v4
+      - uses: actions/checkout@v4
 
-    - name: Set up Go
-      uses: actions/setup-go@v4
-      with:
-        go-version: '1.23'
+      - name: Set up Go
+        uses: actions/setup-go@v4
+        with:
+          go-version-file: go.mod
 
-    # Only run in this linux based runner
-    - name: Check Formatting
-      run: |
-        if [ "$(gofmt -l . | grep -v 'event/.*_test.go' | wc -l)" -gt 0 ]; then
-          gofmt -l . | grep -v 'event/.*_test.go'
-          exit 1
-        fi
-    # cache simple-responder to save the build time
-    - name: Restore Simple Responder
-      id: restore-simple-responder
-      uses: actions/cache/restore@v4
-      with:
-        path: ./build
-        key: ${{ runner.os }}-simple-responder-${{ hashFiles('misc/simple-responder/simple-responder.go') }}
+      # Only run in this linux based runner
+      - name: Check Formatting
+        run: |
+          if [ "$(gofmt -l . | grep -v 'event/.*_test.go' | wc -l)" -gt 0 ]; then
+            gofmt -l . | grep -v 'event/.*_test.go'
+            exit 1
+          fi
+      # cache simple-responder to save the build time
+      - name: Restore Simple Responder
+        id: restore-simple-responder
+        uses: actions/cache/restore@v4
+        with:
+          path: ./build
+          key: ${{ runner.os }}-simple-responder-${{ hashFiles('cmd/simple-responder/simple-responder.go') }}
 
-    # necessary for testing proxy/Process swapping
-    - name: Create simple-responder
-      run: make simple-responder
+      # necessary for testing proxy/Process swapping
+      - name: Create simple-responder
+        run: make simple-responder
 
-    - name: Save Simple Responder
-      # nothing new to save ... skip this step
-      if: steps.restore-simple-responder.outputs.cache-hit != 'true'
-      id: save-simple-responder
-      uses: actions/cache/save@v4
-      with:
-        path: ./build
-        key: ${{ runner.os }}-simple-responder-${{ hashFiles('misc/simple-responder/simple-responder.go') }}
+      - name: Save Simple Responder
+        # nothing new to save ... skip this step
+        if: steps.restore-simple-responder.outputs.cache-hit != 'true'
+        id: save-simple-responder
+        uses: actions/cache/save@v4
+        with:
+          path: ./build
+          key: ${{ runner.os }}-simple-responder-${{ hashFiles('cmd/simple-responder/simple-responder.go') }}
 
-    - name: Test all
-      run: make test-all
+      - name: Test all
+        run: make test-all

.github/workflows/release.yml

@@ -3,13 +3,13 @@ name: goreleaser
 on:
   push:
     tags:
-      - '*'
+      - "*"
 
   # Allows manual triggering of the workflow
   workflow_dispatch:
     inputs:
       tag:
-        description: 'Tag version to release (e.g. v144)'
+        description: "Tag version to release (e.g. v144)"
         required: true
 
 permissions:
@@ -19,35 +19,30 @@ jobs:
   goreleaser:
     runs-on: ubuntu-latest
     steps:
-      -
-        name: Checkout
+      - name: Checkout
         uses: actions/checkout@v4
         with:
           fetch-depth: 0
           ref: ${{ github.event.inputs.tag || github.ref }}
-      -
-        name: Set up Go
+      - name: Set up Go
         uses: actions/setup-go@v5
-      -
-        name: Set up Node.js
+      - name: Set up Node.js
         uses: actions/setup-node@v4
         with:
-          node-version: '23'
-      -
-        name: Install dependencies and build UI
+          node-version: "24"
+      - name: Install dependencies and build UI
         run: |
-          cd ui
+          cd ui-svelte
           npm ci
           npm run build
 
-      -
-        name: Run GoReleaser
+      - name: Run GoReleaser
         uses: goreleaser/goreleaser-action@v6
         with:
           # either 'goreleaser' (default) or 'goreleaser-pro'
           distribution: goreleaser
           # 'latest', 'nightly', or a semver
-          version: '~> v2'
+          version: "~> v2"
           args: release --clean
         env:
           GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
@@ -76,4 +71,4 @@ jobs:
               "release": {
                 "tag_name": "${{ steps.tag.outputs.tag }}"
               }
-            }
+            }

.github/workflows/ui-tests.yml

@@ -0,0 +1,33 @@
+name: UI Tests
+
+on:
+  push:
+    branches: [ "main" ]
+    paths:
+      - 'ui-svelte/**'
+      - '.github/workflows/ui-tests.yml'
+
+  pull_request:
+    branches: [ "main" ]
+    paths:
+      - 'ui-svelte/**'
+      - '.github/workflows/ui-tests.yml'
+
+  workflow_dispatch:
+
+jobs:
+
+  run-tests:
+    runs-on: ubuntu-latest
+    steps:
+    - uses: actions/checkout@v4
+
+    - name: Set up Node.js
+      uses: actions/setup-node@v4
+      with:
+        node-version: '24'
+        cache: 'npm'
+        cache-dependency-path: ui-svelte/package-lock.json
+
+    - name: Run UI tests
+      run: make test-ui

.github/workflows/unified-docker.yml

@@ -0,0 +1,136 @@
+name: Build Unified Docker Image
+
+on:
+  schedule:
+    - cron: "37 5 * * *"
+
+  workflow_dispatch:
+    inputs:
+      llama_cpp_ref:
+        description: "llama.cpp commit hash, tag, or branch"
+        required: false
+        default: "master"
+      whisper_ref:
+        description: "whisper.cpp commit hash, tag, or branch"
+        required: false
+        default: "master"
+      sd_ref:
+        description: "stable-diffusion.cpp commit hash, tag, or branch"
+        required: false
+        default: "master"
+      ik_llama_ref:
+        description: "ik_llama.cpp commit hash, tag, or branch (CUDA only)"
+        required: false
+        default: "main"
+      llama_swap_version:
+        description: "llama-swap version (e.g. v198, latest, main)"
+        required: false
+        default: "main"
+      build_cuda:
+        description: "Build CUDA image"
+        type: boolean
+        required: false
+        default: true
+      build_vulkan:
+        description: "Build Vulkan image"
+        type: boolean
+        required: false
+        default: true
+      push_to_ghcr:
+        description: "Push images to ghcr.io"
+        type: boolean
+        required: false
+        default: true
+
+permissions:
+  contents: read
+  packages: write
+
+jobs:
+  setup:
+    runs-on: ubuntu-latest
+    outputs:
+      matrix: ${{ steps.set-matrix.outputs.matrix }}
+    steps:
+      - id: set-matrix
+        run: |
+          backends=()
+          # schedule uses defaults (build both); workflow_dispatch respects inputs
+          if [[ "${{ github.event_name }}" == "schedule" ]] || [[ "${{ inputs.build_cuda }}" == "true" ]]; then
+            backends+=("cuda")
+          fi
+          if [[ "${{ github.event_name }}" == "schedule" ]] || [[ "${{ inputs.build_vulkan }}" == "true" ]]; then
+            backends+=("vulkan")
+          fi
+          matrix=$(printf '%s\n' "${backends[@]}" | jq -R . | jq -sc .)
+          echo "matrix=$matrix" >> $GITHUB_OUTPUT
+
+  build:
+    needs: setup
+    if: ${{ needs.setup.outputs.matrix != '[]' }}
+    runs-on: ubuntu-latest
+    strategy:
+      fail-fast: false
+      matrix:
+        backend: ${{ fromJSON(needs.setup.outputs.matrix) }}
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+
+      - name: Free up disk space
+        run: |
+          echo "Before cleanup:"
+          df -h
+          sudo rm -rf /usr/share/dotnet
+          sudo rm -rf /usr/local/lib/android
+          sudo rm -rf /opt/ghc
+          sudo rm -rf /opt/hostedtoolcache/CodeQL
+          sudo docker system prune -af
+          echo "After cleanup:"
+          df -h
+
+      # On GitHub Actions runners, create a fresh builder.
+      # When running locally under act, skip this and reuse the existing
+      # llama-swap-builder (which has ccache warm) to avoid exhausting disk.
+      - name: Set up Docker Buildx
+        if: ${{ !env.ACT }}
+        uses: docker/setup-buildx-action@v3
+
+      - name: Log in to GitHub Container Registry
+        if: ${{ !env.ACT }}
+        uses: docker/login-action@v3
+        with:
+          registry: ghcr.io
+          username: ${{ github.actor }}
+          password: ${{ secrets.GITHUB_TOKEN }}
+
+      - name: Build unified Docker image (${{ matrix.backend }})
+        env:
+          LLAMA_REF: ${{ inputs.llama_cpp_ref || 'master' }}
+          WHISPER_REF: ${{ inputs.whisper_ref || 'master' }}
+          SD_REF: ${{ inputs.sd_ref || 'master' }}
+          IK_LLAMA_REF: ${{ inputs.ik_llama_ref || 'main' }}
+          LS_VERSION: ${{ inputs.llama_swap_version || 'main' }}
+          DOCKER_IMAGE_TAG: ghcr.io/mostlygeek/llama-swap:unified-${{ matrix.backend }}
+          # When running under act, use the local builder that has warm ccache.
+          # On GitHub Actions, BUILDX_BUILDER is unset so docker uses the builder
+          # created by setup-buildx-action above.
+          BUILDX_BUILDER: ${{ env.ACT == 'true' && 'llama-swap-builder' || '' }}
+        run: |
+          chmod +x docker/unified/build-image.sh
+          docker/unified/build-image.sh --${{ matrix.backend }}
+
+      - name: Push to GitHub Container Registry
+        if: ${{ !env.ACT && (github.event_name == 'schedule' || inputs.push_to_ghcr == true) }}
+        run: |
+          BASE_TAG="ghcr.io/mostlygeek/llama-swap:unified-${{ matrix.backend }}"
+          DATE_TAG=$(date -u +%Y-%m-%d)
+
+          docker push "${BASE_TAG}"
+          docker tag "${BASE_TAG}" "${BASE_TAG}-${DATE_TAG}"
+          docker push "${BASE_TAG}-${DATE_TAG}"
+
+          ROOTLESS_TAG="${BASE_TAG}-rootless"
+          docker push "${ROOTLESS_TAG}"
+          docker tag "${ROOTLESS_TAG}" "${ROOTLESS_TAG}-${DATE_TAG}"
+          docker push "${ROOTLESS_TAG}-${DATE_TAG}"

AGENTS.md

@@ -0,0 +1,52 @@
+## Project Description:
+
+llama-swap is a light weight, transparent proxy server that provides automatic model swapping to llama.cpp's server.
+
+## Tech stack
+
+- golang
+- typescript, vite and svelt5 for UI (located in ui/)
+
+## Workflow Tasks
+
+- when summarizing changes only include details that require further action
+- just say "Done." when there is no further action
+- use the github CLI `gh` to create pull requests and work with github
+- Rules for creating pull requests:
+  - keep them short and focused on changes.
+  - never include a test plan
+  - write the summary using the same style rules as commit message
+
+## Testing
+
+- Follow test naming conventions like `TestProxyManager_<test name>`, `TestProcessGroup_<test name>`, etc.
+- Use `go test -v -run <name pattern for new tests>` to run any new tests you've written.
+- Run `gofmt -l .` before committing to verify formatting. Fix any reported files with `gofmt -w <file>`.
+- Use `make test-dev` after running new tests for a quick over all test run. This runs `go test` and `staticcheck`. Fix any static checking errors. Use this only when changes are made to any code under the `proxy/` directory
+- Use `make test-all` before completing work. This includes long running concurrency tests.
+- Use `make test-ui` after making changes to the UI in ui-svelte/
+
+### Commit message example format:
+
+```
+proxy: add new feature
+
+Add new feature that implements functionality X and Y.
+
+- key change 1
+- key change 2
+- key change 3
+
+fixes #123
+```
+
+## Code Reviews
+
+- use three levels High, Medium, Low severity
+- label each discovered issue with a label like H1, M2, L3 respectively
+- High severity are must fix issues (security, race conditions, critical bugs)
+- Medium severity are recommended improvements (coding style, missing functionality, inconsistencies)
+- Low severity are nice to have changes and nits
+- Include a suggestion with each discovered item
+- Limit your code review to three items with the highest priority first
+- Double check your discovered items and recommended remediations

CLAUDE.md

@@ -1,45 +1 @@
-# Project: llama-swap
-
-## Project Description:
-
-llama-swap is a light weight, transparent proxy server that provides automatic model swapping to llama.cpp's server.
-
-## Tech stack
-
-- golang
-- typescript, vite and react for UI (ui/)
-
-## Testing
-
-- Follow test naming conventions like `TestProxyManager_<test name>`, `TestProcessGroup_<test name>`, etc.
-- Use `go test -v -run <name pattern for new tests>` to run any new tests you've written.
-- Use `make test-dev` after running new tests for a quick over all test run. This runs `go test` and `staticcheck`. Fix any static checking errors. Use this only when changes are made to any code under the `proxy/` directory
-- Use `make test-all` before completing work. This includes long running concurrency tests.
-
-## Workflow Tasks
-
-### Plan Improvements
-
-Work plans are located in ai-plans/. Plans written by the user may be incomplete, contain inconsistencies or errors.
-
-When the user asks to improve a plan follow these guidelines for expanding and improving it.
-
-- Identify any inconsistencies.
-- Expand plans out to be detailed specification of requirements and changes to be made.
-- Plans should have at least these sections:
-  - Title - very short, describes changes
-  - Overview: A more detailed summary of goal and outcomes desired
-  - Design Requirements: Detailed descriptions of what needs to be done
-  - Testing Plan: Tests to be implemented
-  - Checklist: A detailed list of changes to be made
-
-Look for "plan expansion" as explicit instructions to improve a plan.
-
-### Implementation of plans
-
-When the user says "paint it", respond with "commencing automated assembly". Then implement the changes as described by the plan. Update the checklist as you complete items.
-
-## General Rules
-
-- when summarizing changes only include details that require further action (action items)
-- when there are no action items, just say "Done."
+@AGENTS.md

Makefile

@@ -36,11 +36,11 @@ test-all: proxy/ui_dist/placeholder.txt
 	go test -race -count=1 ./proxy/...
 
 ui/node_modules:
-	cd ui && npm install
+	cd ui-svelte && npm install
 
 # build react UI
 ui: ui/node_modules
-	cd ui && npm run build
+	cd ui-svelte && npm run build
 
 # Build OSX binary
 mac: ui
@@ -48,9 +48,14 @@ mac: ui
 	GOOS=darwin GOARCH=arm64 go build -ldflags="-X main.commit=${GIT_HASH} -X main.version=local_${GIT_HASH} -X main.date=${BUILD_DATE}" -o $(BUILD_DIR)/$(APP_NAME)-darwin-arm64
 
 # Build Linux binary
-linux: ui
-	@echo "Building Linux binary..."
+linux: linux-arm64 linux-amd64
+
+linux-amd64: ui
+	@echo "Building Linux AMD64 binary..."
 	GOOS=linux GOARCH=amd64 go build -ldflags="-X main.commit=${GIT_HASH} -X main.version=local_${GIT_HASH} -X main.date=${BUILD_DATE}" -o $(BUILD_DIR)/$(APP_NAME)-linux-amd64
+
+linux-arm64: ui
+	@echo "Building Linux ARM64 binary..."
 	GOOS=linux GOARCH=arm64 go build -ldflags="-X main.commit=${GIT_HASH} -X main.version=local_${GIT_HASH} -X main.date=${BUILD_DATE}" -o $(BUILD_DIR)/$(APP_NAME)-linux-arm64
 
 # Build Windows binary
@@ -92,5 +97,9 @@ wol-proxy: $(BUILD_DIR)
 	@echo "Building wol-proxy"
 	go build -o $(BUILD_DIR)/wol-proxy-$(GOOS)-$(GOARCH)-$(shell date +%Y-%m-%d) cmd/wol-proxy/wol-proxy.go
 
+test-ui:
+	cd ui-svelte && npm ci && npm run check && npm test
+
 # Phony targets
-.PHONY: all clean ui mac linux windows simple-responder simple-responder-windows test test-all test-dev wol-proxy
+.PHONY: all clean ui mac windows simple-responder simple-responder-windows test test-all test-dev test-ui wol-proxy
+.PHONE: linux linux-arm64 linux-amd64

README.md

@@ -5,7 +5,7 @@
 
 # llama-swap
 
-Run multiple LLM models on your machine and hot-swap between them as needed. llama-swap works with any OpenAI API-compatible server, giving you the flexibility to switch models without restarting your applications.
+Run multiple generative AI models on your machine and hot-swap between them on demand. llama-swap works with any OpenAI and Anthropic API compatible server and is used by thousands of people to power their local AI workflows.
 
 Built in Go for performance and simplicity, llama-swap has zero dependencies and is incredibly easy to set up. Get started in minutes - just one binary and one configuration file.
 
@@ -13,7 +13,7 @@ Built in Go for performance and simplicity, llama-swap has zero dependencies and
 
 - ✅ Easy to deploy and configure: one binary, one configuration file. no external dependencies
 - ✅ On-demand model switching
-- ✅ Use any local OpenAI compatible server (llama.cpp, vllm, tabbyAPI, etc.)
+- ✅ Use any local OpenAI compatible server (llama.cpp, vllm, tabbyAPI, stable-diffusion.cpp, etc.)
   - future proof, upgrade your inference servers at any time.
 - ✅ OpenAI API supported endpoints:
   - `v1/completions`
@@ -27,10 +27,15 @@ Built in Go for performance and simplicity, llama-swap has zero dependencies and
   - `v1/images/edits`
 - ✅ Anthropic API supported endpoints:
   - `v1/messages`
+  - `v1/messages/count_tokens`
 - ✅ llama-server (llama.cpp) supported endpoints
   - `v1/rerank`, `v1/reranking`, `/rerank`
   - `/infill` - for code infilling
   - `/completion` - for completion endpoint
+- ✅ SDAPI via [stable-diffusion.cpp's server](https://github.com/leejet/stable-diffusion.cpp/tree/master/examples/server)
+  - `/sdapi/v1/txt2img`
+  - `/sdapi/v1/img2img`
+  - `/sdapi/v1/loras` - requires `model` in request body to fetch the correct loras
 - ✅ llama-swap API
   - `/ui` - web UI
   - `/upstream/:model_id` - direct access to upstream server ([demo](https://github.com/mostlygeek/llama-swap/pull/31))
@@ -40,20 +45,32 @@ Built in Go for performance and simplicity, llama-swap has zero dependencies and
   - `/health` - just returns "OK"
 - ✅ API Key support - define keys to restrict access to API endpoints
 - ✅ Customizable
-  - Run multiple models at once with `Groups` ([#107](https://github.com/mostlygeek/llama-swap/issues/107))
+  - Run concurrent models with a custom DSL swap matrix ([#643](https://github.com/mostlygeek/llama-swap/issues/643))
   - Automatic unloading of models after timeout by setting a `ttl`
   - Reliable Docker and Podman support using `cmd` and `cmdStop` together
   - Preload models on startup with `hooks` ([#235](https://github.com/mostlygeek/llama-swap/pull/235))
 
 ### Web UI
 
-llama-swap includes a real time web interface for monitoring logs and controlling models:
+llama-swap includes a real time web interface with a playground for testing out all sorts of local models:
 
-<img width="1164" height="745" alt="image" src="https://github.com/user-attachments/assets/bacf3f9d-819f-430b-9ed2-1bfaa8d54579" />
+<img width="1125" height="876" alt="image" src="https://github.com/user-attachments/assets/8ee41947-97af-463d-b0f0-8e9c478fac07" />
 
-The Activity Page shows recent requests:
+View detailed token metrics:
 
-<img width="1360" height="963" alt="image" src="https://github.com/user-attachments/assets/5f3edee6-d03a-4ae5-ae06-b20ac1f135bd" />
+<img width="1111" height="515" alt="image" src="https://github.com/user-attachments/assets/64bfb280-d7a3-4126-971a-a128fd40410c" />
+
+Inspect request and responses:
+
+<img width="1111" height="720" alt="image" src="https://github.com/user-attachments/assets/24fe4aca-1448-4d7c-b9e8-a967589bda6c" />
+
+Manually load and unload models:
+
+<img width="1109" height="719" alt="image" src="https://github.com/user-attachments/assets/02b1e1f2-abd0-4050-84ae-facd66ff01c4" />
+
+Real time log streaming:
+
+<img width="1107" height="559" alt="image" src="https://github.com/user-attachments/assets/39669a10-cff2-409e-836a-5bad8bd0140c" />
 
 ## Installation
 
@@ -68,6 +85,7 @@ llama-swap can be installed in multiple ways
 ### Docker Install ([download images](https://github.com/mostlygeek/llama-swap/pkgs/container/llama-swap))
 
 Nightly container images with llama-swap and llama-server are built for multiple platforms (cuda, vulkan, intel, etc.) including [non-root variants with improved security](docs/container-security.md).
+The stable-diffusion.cpp server is also included for the musa and vulkan platforms.
 
 ```shell
 $ docker pull ghcr.io/mostlygeek/llama-swap:cuda
@@ -162,7 +180,7 @@ That's all you need to get started:
 Almost all configuration settings are optional and can be added one step at a time:
 
 - Advanced features
-  - `groups` to run multiple models at once
+  - `matrix` to run concurrent models with a custom swap logic DSL
   - `hooks` to run things on startup
   - `macros` reusable snippets
 - Model customization
@@ -180,7 +198,7 @@ See the [configuration documentation](docs/configuration.md) for all options.
 
 When a request is made to an OpenAI compatible endpoint, llama-swap will extract the `model` value and load the appropriate server configuration to serve it. If the wrong upstream server is running, it will be replaced with the correct one. This is where the "swap" part comes in. The upstream server is automatically swapped to handle the request correctly.
 
-In the most basic configuration llama-swap handles one model at a time. For more advanced use cases, the `groups` feature allows multiple models to be loaded at the same time. You have complete control over how your system resources are used.
+In the most basic configuration llama-swap handles one model at a time. For more advanced use cases, using a `matrix` allows multiple models to be loaded at the same time. You have complete control over how your system resources are used.
 
 ## Reverse Proxy Configuration (nginx)
 

ai-plans/improve-tests-655.md

@@ -0,0 +1,183 @@
+# Improve Testability (#655)
+
+## Current Pain Points
+
+1. **Tests bypass config loading** - ~80% of tests build `config.Config` structs directly, skipping YAML parsing, env var substitution, macro expansion, and `${PORT}` assignment. Config bugs in those paths go untested.
+
+2. **simple-responder is everywhere** - Every proxy/routing test launches a real subprocess, waits for health checks (~healthCheckTimeout: 15), and manages process lifecycle just to test HTTP routing. Most of that overhead is wasted.
+
+3. **Port counter is fragile** - A global `nextTestPort` counter starting at 12000 with a mutex. Parallel tests or leftover processes can collide.
+
+## Stages
+
+### Stage 1: YAML-based test config helper
+
+**Goal:** Tests go through the real `LoadConfigFromReader` path instead of hand-building structs.
+
+**Effort:** Low | **Impact:** Config bugs caught earlier | **Risk:** None
+
+Create a test helper in `proxy/helpers_test.go`:
+
+```go
+// testConfigFromYAML substitutes simple-responder paths and loads through
+// the real config pipeline (env vars, macros, port assignment, etc.)
+func testConfigFromYAML(t *testing.T, yamlTmpl string) config.Config {
+    t.Helper()
+    yamlStr := strings.ReplaceAll(yamlTmpl, "{{RESPONDER}}", filepath.ToSlash(simpleResponderPath))
+    cfg, err := config.LoadConfigFromReader(strings.NewReader(yamlStr))
+    require.NoError(t, err)
+    return cfg
+}
+```
+
+Tests would then look like:
+
+```go
+func TestProxyManager_SwapProcessCorrectly(t *testing.T) {
+    config := testConfigFromYAML(t, `
+healthCheckTimeout: 15
+logLevel: error
+models:
+  model1:
+    cmd: {{RESPONDER}} --port ${PORT} -silent -respond model1
+  model2:
+    cmd: {{RESPONDER}} --port ${PORT} -silent -respond model2
+`)
+    proxy := New(config)
+    // ... same assertions
+}
+```
+
+**Why this stage first:** Zero production code changes. Pure test-side refactoring. Can be done incrementally - migrate tests one at a time. Each migrated test now validates the full config pipeline.
+
+**Scope:** ~20-30 tests in `proxymanager_test.go`, `processgroup_test.go`, `peerproxy_test.go`.
+
+### Stage 2: Injected test handler (eliminate simple-responder for routing tests)
+
+**Goal:** Replace simple-responder subprocess launches with an injected `http.Handler` for tests that don't specifically test process lifecycle.
+
+**Effort:** Medium | **Impact:** 10-100x faster routing tests | **Risk:** Low (additive, no existing code broken)
+
+Add a `testHandler http.Handler` field to `Process`. When set, `ProxyRequest` delegates directly to this handler instead of going through the reverse proxy. No subprocess, no health checks, no TCP roundtrip.
+
+**2a. Add testHandler to Process:**
+
+```go
+// In Process struct (process.go):
+testHandler http.Handler  // set only in tests; bypasses subprocess and reverse proxy
+```
+
+In `Process.Start()`, skip subprocess + health check when handler is set:
+
+```go
+func (p *Process) start() error {
+    if p.testHandler != nil {
+        p.setState(StateReady)
+        return nil
+    }
+    // existing subprocess logic...
+}
+```
+
+In `Process.ProxyRequest()`, delegate directly to the handler:
+
+```go
+// Before the reverseProxy.ServeHTTP call:
+if p.testHandler != nil {
+    p.testHandler.ServeHTTP(w, r)
+    return
+}
+```
+
+**2b. Test helper to create the handler:**
+
+```go
+// newTestHandler returns an http.Handler that mimics llama.cpp's API
+// (same endpoints as simple-responder).
+func newTestHandler(respond string) http.Handler {
+    mux := http.NewServeMux()
+    mux.HandleFunc("/v1/chat/completions", func(w http.ResponseWriter, r *http.Request) { ... })
+    mux.HandleFunc("/health", func(w http.ResponseWriter, r *http.Request) { ... })
+    // ... other endpoints
+    return mux
+}
+```
+
+Tests for routing/auth/CORS/streaming then become:
+
+```go
+func TestProxyManager_AuthRequired(t *testing.T) {
+    handler := newTestHandler("model1")
+
+    config := testConfigFromYAML(t, `
+healthCheckTimeout: 15
+logLevel: error
+requiredAPIKeys: [test-key]
+models:
+  model1:
+    cmd: {{RESPONDER}} --port ${PORT} -silent -respond model1
+`)
+    pm := NewProxyManager(config)
+    // inject handler — skips subprocess, health check, port allocation
+    pm.processGroups["model1"].process.testHandler = handler
+}
+```
+
+**Why this matters:** The handler is called directly in-process. No subprocess spawn, no health check timeout, no port allocation, no TCP roundtrip, no reverse proxy overhead. Routing tests go from ~100ms each (process startup + health check) to ~1ms. Unlike an `httptest.Server` approach, there are zero network hops.
+
+**Why not blank-cmd + proxy URL:** A blank `cmd` with a `proxy` field pointing at `httptest.Server` still requires a real TCP roundtrip through the reverse proxy and introduces "external process" semantics to the config schema. Injecting the handler directly keeps it purely a test concern with no config changes.
+
+**Scope:** Most tests in `proxymanager_test.go` (auth, CORS, model listing, streaming, peer proxy), `peerproxy_test.go`, `metrics_monitor_test.go`.
+
+### Stage 3: Migrate tests incrementally
+
+**Goal:** Convert existing tests to use the Stage 1 + Stage 2 helpers.
+
+**Effort:** Medium | **Impact:** Cleaner, more reliable tests | **Risk:** None
+
+Priority order:
+1. `proxymanager_test.go` routing tests (highest count, most repetition)
+2. `peerproxy_test.go` (straightforward, all HTTP routing)
+3. `metrics_monitor_test.go` (capture logic doesn't need real processes)
+4. `processgroup_test.go` swap tests (keep simple-responder for actual swap lifecycle tests)
+
+Tests that **must keep simple-responder:**
+- Process lifecycle: start/stop, SIGKILL, SIGTERM, TTL expiry, health check failures, failed start counting
+- ProcessGroup swap concurrency (the port-collision test in `TestProcessGroup_ProxyRequestSwapIsTrueParallel`)
+
+**Scope:** ~60-70% of tests can drop simple-responder.
+
+### Stage 4 (optional): Process interface for ProcessGroup
+
+**Goal:** Enable pure unit tests of ProcessGroup's swap/exclusive/concurrency logic without any HTTP server at all.
+
+**Effort:** High | **Impact:** Pure unit tests possible | **Risk:** Medium (refactor core code)
+
+```go
+type ProcessController interface {
+    Start() error
+    Stop(StopStrategy)
+    ProxyRequest(http.ResponseWriter, *http.Request) error
+    CurrentState() ProcessState
+    ID() string
+    SetState(ProcessState)  // for test setup
+}
+```
+
+This requires:
+- Extracting the interface
+- A `MockProcess` implementation
+- Refactoring `ProcessGroup` to use the interface instead of `*Process`
+
+**Recommendation:** Only do this if ProcessGroup grows significantly more complex. Stages 1-3 give 80% of the benefit for 20% of the effort.
+
+## Effort/Impact Summary
+
+| Stage | Effort | Impact | Risk |
+|-------|--------|--------|------|
+| 1. YAML config helper | Low | Config bugs caught earlier | None |
+| 2. Injected test handler | Medium | 10-100x faster routing tests | Low |
+| 3. Migrate tests | Medium | Cleaner, more reliable tests | None |
+| 4. Process interface | High | Pure unit tests possible | Medium |
+
+**Recommended approach:** Do stages 1-3 in order. Each stage is independently valuable and can ship on its own. Stage 4 is deferred unless there's a specific need.

cmd/simple-responder/simple-responder.go

@@ -210,6 +210,11 @@ func main() {
 		})
 	})
 
+	r.GET("/v1/audio/voices", func(c *gin.Context) {
+		model := c.Query("model")
+		c.JSON(http.StatusOK, gin.H{"voices": []string{"voice1"}, "model": model})
+	})
+
 	r.GET("/slow-respond", func(c *gin.Context) {
 		echo := c.Query("echo")
 		delay := c.Query("delay")
@@ -269,6 +274,43 @@ func main() {
 		c.String(200, fmt.Sprintf("%s %s", c.Request.Method, c.Request.URL.Path))
 	})
 
+	// SD API endpoints
+	r.POST("/sdapi/v1/txt2img", func(c *gin.Context) {
+		body, err := io.ReadAll(c.Request.Body)
+		if err != nil {
+			c.JSON(http.StatusInternalServerError, gin.H{"error": "Failed to read request body"})
+			return
+		}
+		defer c.Request.Body.Close()
+
+		modelName := gjson.GetBytes(body, "model").String()
+		c.JSON(http.StatusOK, gin.H{
+			"model":  modelName,
+			"images": []string{},
+		})
+	})
+
+	r.POST("/sdapi/v1/img2img", func(c *gin.Context) {
+		body, err := io.ReadAll(c.Request.Body)
+		if err != nil {
+			c.JSON(http.StatusInternalServerError, gin.H{"error": "Failed to read request body"})
+			return
+		}
+		defer c.Request.Body.Close()
+
+		modelName := gjson.GetBytes(body, "model").String()
+		c.JSON(http.StatusOK, gin.H{
+			"model":  modelName,
+			"images": []string{},
+		})
+	})
+
+	r.GET("/sdapi/v1/loras", func(c *gin.Context) {
+		c.JSON(http.StatusOK, gin.H{
+			"loras": []string{},
+		})
+	})
+
 	address := "127.0.0.1:" + *port // Address with the specified port
 
 	srv := &http.Server{

config-schema.json

@@ -39,6 +39,49 @@
             },
             "default": {},
             "description": "A dictionary of string substitutions. Macros are reusable snippets used in model cmd, cmdStop, proxy, checkEndpoint, filters.stripParams. Macro names must be <64 chars, match ^[a-zA-Z0-9_-]+$, and not be PORT or MODEL_ID. Values can be string, number, or boolean. Macros can reference other macros defined before them."
+        },
+        "timeouts": {
+            "type": "object",
+            "properties": {
+                "connect": {
+                    "type": "integer",
+                    "minimum": 0,
+                    "default": 30,
+                    "description": "TCP connection timeout in seconds. Set to 0 to disable."
+                },
+                "keepalive": {
+                    "type": "integer",
+                    "minimum": 0,
+                    "default": 30,
+                    "description": "TCP keepalive timeout in seconds. Set to 0 to disable."
+                },
+                "responseHeader": {
+                    "type": "integer",
+                    "minimum": 0,
+                    "default": 0,
+                    "description": "Time to wait for response headers in seconds. Set to 0 to disable."
+                },
+                "tlsHandshake": {
+                    "type": "integer",
+                    "minimum": 0,
+                    "default": 10,
+                    "description": "TLS handshake timeout in seconds. Set to 0 to disable."
+                },
+                "expectContinue": {
+                    "type": "integer",
+                    "minimum": 0,
+                    "default": 1,
+                    "description": "Expect-Continue timeout in seconds. Set to 0 to disable."
+                },
+                "idleConn": {
+                    "type": "integer",
+                    "minimum": 0,
+                    "default": 90,
+                    "description": "Idle connection timeout in seconds. Set to 0 to disable."
+                }
+            },
+            "additionalProperties": false,
+            "description": "Timeout settings for proxy connections."
         }
     },
     "properties": {
@@ -48,6 +91,12 @@
             "default": 120,
             "description": "Number of seconds to wait for a model to be ready to serve requests."
         },
+        "globalTTL": {
+            "type": "integer",
+            "minimum": 0,
+            "default": 0,
+            "description": "Default TTL for all models in seconds, 0 means no TTL and models will never be automatically unloaded"
+        },
         "logLevel": {
             "type": "string",
             "enum": [
@@ -87,6 +136,12 @@
             "default": 1000,
             "description": "Maximum number of metrics to keep in memory. Controls how many metrics are stored before older ones are discarded."
         },
+        "captureBuffer": {
+            "type": "integer",
+            "minimum": 0,
+            "default": 5,
+            "description": "Size in megabytes of the buffer for storing request/response captures. Set to 0 to disable captures."
+        },
         "startPort": {
             "type": "integer",
             "default": 5800,
@@ -171,9 +226,9 @@
                     },
                     "ttl": {
                         "type": "integer",
-                        "minimum": 0,
-                        "default": 0,
-                        "description": "Automatically unload the model after ttl seconds. 0 disables unloading. Must be >0 to enable."
+                        "minimum": -1,
+                        "default": -1,
+                        "description": "Automatically unload the model after ttl seconds. -1 uses the global TTL value, 0 disables unloading. Must be >0 to enable."
                     },
                     "useModelName": {
                         "type": "string",
@@ -188,11 +243,26 @@
                                 "default": "",
                                 "pattern": "^[a-zA-Z0-9_, ]*$",
                                 "description": "Comma separated list of parameters to remove from the request. Used for server-side enforcement of sampling parameters."
+                            },
+                            "setParams": {
+                                "type": "object",
+                                "additionalProperties": true,
+                                "default": {},
+                                "description": "Dictionary of parameters to set/override in requests. Useful for enforcing specific parameter values. Protected params like 'model' cannot be overridden. Values can be strings, numbers, booleans, arrays, or objects."
+                            },
+                            "setParamsByID": {
+                                "type": "object",
+                                "additionalProperties": {
+                                    "type": "object",
+                                    "additionalProperties": true
+                                },
+                                "default": {},
+                                "description": "Dictionary mapping requested model IDs (or aliases) to parameters to set/override in requests. Applied after setParams and can override those values. Useful with aliases to vary behaviour depending on which alias the client used (e.g. different reasoning_effort per alias). Keys support ${MODEL_ID} macro substitution. Protected params like 'model' cannot be overridden."
                             }
                         },
                         "additionalProperties": false,
                         "default": {},
-                        "description": "Dictionary of filter settings. Only stripParams is supported."
+                        "description": "Dictionary of filter settings. Supports stripParams, setParams, and setParamsByID."
                     },
                     "metadata": {
                         "type": "object",
@@ -214,6 +284,9 @@
                         "type": "boolean",
                         "default": false,
                         "description": "If true the model will not show up in /v1/models responses. It can still be used as normal in API requests."
+                    },
+                    "timeouts": {
+                        "$ref": "#/definitions/timeouts"
                     }
                 }
             }
@@ -252,6 +325,44 @@
             },
             "description": "A dictionary of group settings. Provides advanced controls over model swapping behaviour. Model IDs must be defined in models. A model can only be a member of one group. Behaviour controlled via swap, exclusive, persistent."
         },
+        "matrix": {
+            "type": "object",
+            "description": "Solver-based alternative to groups. Declares valid combinations of concurrent models. The solver minimizes eviction cost when swapping. A config must use either groups or matrix, not both.",
+            "required": [
+                "vars",
+                "sets"
+            ],
+            "properties": {
+                "vars": {
+                    "type": "object",
+                    "description": "Short names for models. Keys must be alphanumeric, 1-8 characters. All sets and evict_costs must use these IDs.",
+                    "minProperties": 1,
+                    "additionalProperties": {
+                        "type": "string"
+                    },
+                    "propertyNames": {
+                        "pattern": "^[a-zA-Z0-9]{1,8}$"
+                    }
+                },
+                "evict_costs": {
+                    "type": "object",
+                    "description": "Relative cost of evicting a running model. Models not listed default to 1. Values must be positive integers.",
+                    "additionalProperties": {
+                        "type": "integer",
+                        "minimum": 1
+                    }
+                },
+                "sets": {
+                    "type": "object",
+                    "description": "Named sets of concurrent model combinations. Values are DSL strings using & (AND), | (OR), () (grouping), and +ref (inline another set). Definition order is used for tie-breaking.",
+                    "minProperties": 1,
+                    "additionalProperties": {
+                        "type": "string"
+                    }
+                }
+            },
+            "additionalProperties": false
+        },
         "hooks": {
             "type": "object",
             "properties": {
@@ -320,11 +431,90 @@
                             "minLength": 1
                         },
                         "description": "A list of models served by the peer."
+                    },
+                    "filters": {
+                        "type": "object",
+                        "properties": {
+                            "stripParams": {
+                                "type": "string",
+                                "default": "",
+                                "pattern": "^[a-zA-Z0-9_, ]*$",
+                                "description": "Comma separated list of parameters to remove from the request. Useful for removing parameters that the peer doesn't support."
+                            },
+                            "setParams": {
+                                "type": "object",
+                                "additionalProperties": true,
+                                "default": {},
+                                "description": "Dictionary of parameters to set/override in requests to this peer. Useful for injecting provider-specific settings. Protected params like 'model' cannot be overridden. Values can be strings, numbers, booleans, arrays, or objects."
+                            }
+                        },
+                        "additionalProperties": false,
+                        "default": {},
+                        "description": "Dictionary of filter settings for peer requests. Supports stripParams and setParams."
+                    },
+                    "timeouts": {
+                        "type": "object",
+                        "properties": {
+                            "connect": {
+                                "type": "integer",
+                                "minimum": 0,
+                                "default": 30,
+                                "description": "TCP connection timeout in seconds."
+                            },
+                            "keepalive": {
+                                "type": "integer",
+                                "minimum": 0,
+                                "default": 30,
+                                "description": "TCP keepalive connection timeout in seconds."
+                            },
+                            "responseHeader": {
+                                "type": "integer",
+                                "minimum": 0,
+                                "default": 0,
+                                "description": "Time to wait for response headers in seconds."
+                            },
+                            "tlsHandshake": {
+                                "type": "integer",
+                                "minimum": 0,
+                                "default": 10,
+                                "description": "TLS handshake timeout in seconds."
+                            },
+                            "idleConn": {
+                                "type": "integer",
+                                "minimum": 0,
+                                "default": 90,
+                                "description": "Idle connection timeout in seconds."
+                            }
+                        },
+                        "additionalProperties": false,
+                        "description": "Timeout settings for proxy connections to this peer."
                     }
                 }
             },
             "default": {},
             "description": "A dictionary of remote peers and models they provide. Peers can be another llama-swap or any server that provides the /v1/ generative API endpoints supported by llama-swap."
         }
-    }
-}
+    },
+    "allOf": [
+        {
+            "if": {
+                "required": ["groups"]
+            },
+            "then": {
+                "not": {
+                    "required": ["matrix"]
+                }
+            }
+        },
+        {
+            "if": {
+                "required": ["matrix"]
+            },
+            "then": {
+                "not": {
+                    "required": ["groups"]
+                }
+            }
+        }
+    ]
+}

config.example.yaml

@@ -50,6 +50,11 @@ logToStdout: "proxy"
 # - useful for limiting memory usage when processing large volumes of metrics
 metricsMaxInMemory: 1000
 
+# captureBuffer: how many MBs to allocate for storing request/response captures
+# - optional, default: 10
+# - set to 0 to disable
+captureBuffer: 15
+
 # startPort: sets the starting port number for the automatic ${PORT} macro.
 # - optional, default: 5800
 # - the ${PORT} macro can be used in model.cmd and model.proxy settings
@@ -70,15 +75,10 @@ sendLoadingState: true
 #   all fields except for Id so chat UIs can use the alias equivalent to the original.
 includeAliasesInList: false
 
-# apiKeys: require an API key when making requests to inference endpoints
-# - optional, default: []
-# - when empty (the default) authorization will not be checked as llama-swap is default-allow
-# - each key is a non-empty string
-apiKeys:
-  - "sk-hunter2"
-  # hint, one liner: printf "sk-%s\n" "$(head -c 48 /dev/urandom | base64 )"
-  - "sk-gyCPiKUcIfPlaM4OSMZekkprgijPx6+OsmQs8Rsg0xZ9qpy6gKWsIKqHOk+cgXVx"
-  - "sk-+QtIn0Zjj4UHjiaZYiZEnru4mrwKM9RzhmJeK5SobNXLl8QMFXxGz1/2lEuvQpkb"
+# globalTTL: the default TTL in seconds before unloading a model
+# - optional, default: 0 (never automatically unload)
+# - must be >= 0
+globalTTL: 0
 
 # macros: a dictionary of string substitutions
 # - optional, default: empty dictionary
@@ -90,6 +90,9 @@ apiKeys:
 # - macro names must not be a reserved name: PORT or MODEL_ID
 # - macro values can be numbers, bools, or strings
 # - macros can contain other macros, but they must be defined before they are used
+# - environment variables can be referenced with ${env.VAR_NAME} syntax
+#   - env macros are substituted first, before regular macros
+#   - if the env var is not set, config loading will fail with an error
 macros:
   # Example of a multi-line macro
   "latest-llama": >
@@ -102,6 +105,24 @@ macros:
   # but they must be previously declared.
   "default_args": "--ctx-size ${default_ctx}"
 
+  # Example of environment variable macros
+  # - ${env.VAR_NAME} pulls the value from the system environment
+  # - useful for paths, secrets, or machine-specific configuration
+  "models_dir": "${env.HOME}/models"
+
+# apiKeys: require an API key when making requests to inference endpoints
+# - optional, default: []
+# - when empty (the default) authorization will not be checked as llama-swap is default-allow
+# - each key is a non-empty string
+apiKeys:
+  - "sk-hunter2"
+  # tip, one liner: printf "sk-%s\n" "$(head -c 48 /dev/urandom | base64 )"
+  - "sk-gyCPiKUcIfPlaM4OSMZekkprgijPx6+OsmQs8Rsg0xZ9qpy6gKWsIKqHOk+cgXVx"
+
+  # use environment variable macros to keep secrets out of the config
+  - "${env.API_KEY_1}"
+  - "${env.API_KEY_2}"
+
 # models: a dictionary of model configurations
 # - required
 # - each key is the model's ID, used in API requests
@@ -110,7 +131,7 @@ macros:
 # - below are examples of the all the settings a model can have
 models:
   # keys are the model names used in API requests
-  "llama":
+  "gpt-oss-120b":
     # macros: a dictionary of string substitutions specific to this model
     # - optional, default: empty dictionary
     # - macros defined here override macros defined in the global macros section
@@ -127,7 +148,7 @@ models:
     cmd: |
       # ${latest-llama} is a macro that is defined above
       ${latest-llama}
-      --model path/to/llama-8B-Q4_K_M.gguf
+      --model path/to/gpt-oss-120B.gguf
       --ctx-size ${default_ctx}
       --temperature ${temp}
 
@@ -135,13 +156,13 @@ models:
     # - optional, default: empty string
     # - if set, it will be used in the v1/models API response
     # - if not set, it will be omitted in the JSON model record
-    name: "llama 3.1 8B"
+    name: "gpt-oss 120B"
 
     # description: a description for the model
     # - optional, default: empty string
     # - if set, it will be used in the v1/models API response
     # - if not set, it will be omitted in the JSON model record
-    description: "A small but capable model used for quick testing"
+    description: "A thinking model from OpenAI"
 
     # env: define an array of environment variables to inject into cmd's environment
     # - optional, default: empty array
@@ -156,14 +177,6 @@ models:
     # - if you use a custom port in cmd this *must* be set
     proxy: http://127.0.0.1:8999
 
-    # aliases: alternative model names that this model configuration is used for
-    # - optional, default: empty array
-    # - aliases must be unique globally
-    # - useful for impersonating a specific model
-    aliases:
-      - "gpt-4o-mini"
-      - "gpt-3.5-turbo"
-
     # checkEndpoint: URL path to check if the server is ready
     # - optional, default: /health
     # - endpoint is expected to return an HTTP 200 response
@@ -172,8 +185,10 @@ models:
     checkEndpoint: /custom-endpoint
 
     # ttl: automatically unload the model after ttl seconds
-    # - optional, default: 0
-    # - ttl values must be a value greater than 0
+    # - optional, default: -1 (use global default)
+    # - ttl values must be a value greater than or equal to 0
+    # - a ttl of -1 will use the global TTL value as the default
+    # - a ttl of 0 will mean never unload
     # - a value of 0 disables automatic unloading of the model
     ttl: 60
 
@@ -181,11 +196,11 @@ models:
     # - optional, default: ""
     # - useful for when the upstream server expects a specific model name that
     #   is different from the model's ID
-    useModelName: "qwen:qwq"
+    useModelName: "openai/gpt-oss-120B"
 
     # filters: a dictionary of filter settings
     # - optional, default: empty dictionary
-    # - only stripParams is currently supported
+    # - same capabilities as peer filters (stripParams, setParams)
     filters:
       # stripParams: a comma separated list of parameters to remove from the request
       # - optional, default: ""
@@ -195,6 +210,43 @@ models:
       # - recommended to stick to sampling parameters
       stripParams: "temperature, top_p, top_k"
 
+      # setParams: a dictionary of parameters to set/override in requests
+      # - optional, default: empty dictionary
+      # - useful for enforcing specific parameter values
+      # - protected params like "model" cannot be overridden
+      # - values can be strings, numbers, booleans, arrays, or objects
+      # - always runs for the model
+      setParams:
+        # Example: enforce specific sampling parameters
+        temperature: 0.7
+        top_p: 0.9
+
+      # setParamsByID: a dictionary of parameters to set based the model ID
+      # - optional, default: empty dictionary
+      # - combine with aliases to create variant behaviour without reloading the model
+      # - parameters are set in the request body JSON
+      # - run after setParams so it will override any settings
+      # - protected params like "model" cannot be overridden
+      # - values can be strings, numbers, booleans, arrays, or objects
+      # - model aliases will be automatically created for each key
+      setParamsByID:
+        "${MODEL_ID}":
+          chat_template_kwargs:
+            reasoning_effort: medium
+        "${MODEL_ID}:high":
+          chat_template_kwargs:
+            reasoning_effort: high
+        "${MODEL_ID}:low":
+          chat_template_kwargs:
+            reasoning_effort: low
+
+    # aliases: alternative model names that this model configuration is used for
+    # - optional, default: empty array
+    # - aliases must be unique globally
+    # - useful for impersonating a specific model
+    aliases:
+      - "gpt-4o-mini"
+
     # metadata: a dictionary of arbitrary values that are included in /v1/models
     # - optional, default: empty dictionary
     # - while metadata can contains complex types it is recommended to keep it simple
@@ -232,6 +284,22 @@ models:
     # - optional, default: undefined (use global setting)
     sendLoadingState: false
 
+    # timeouts: configure proxy connection timeouts for this model
+    # - optional, defaults shown below
+    # - useful for models running on slower hardware that need longer timeouts
+    # - connect: TCP dial connection timeout in seconds, default: 30 seconds
+    # - keepalive: TCP connection keepalive timeout, default: 30 seconds
+    # - responseHeader: time to wait for response headers in seconds, default: 0 (no timeout)
+    # - tlsHandshake: TLS handshake timeout in seconds, default: 10 seconds
+    # - idleConn: idle connection timeout in seconds, default: 90 seconds
+    # - set any value to 0 to disable that timeout (not recommended)
+    timeouts:
+      connect: 30
+      keepalive: 0
+      responseHeader: 60
+      tlsHandshake: 10
+      idleConn: 90
+
   # Unlisted model example:
   "qwen-unlisted":
     # unlisted: boolean, true or false
@@ -263,68 +331,83 @@ models:
     # - processes have 5 seconds to shutdown until forceful termination is attempted
     cmdStop: docker stop ${MODEL_ID}
 
-# groups: a dictionary of group settings
-# - optional, default: empty dictionary
-# - provides advanced controls over model swapping behaviour
-# - using groups some models can be kept loaded indefinitely, while others are swapped out
-# - model IDs must be defined in the Models section
-# - a model can only be a member of one group
-# - group behaviour is controlled via the `swap`, `exclusive` and `persistent` fields
-# - see issue #109 for details
+# =============================================================================
+# matrix: run concurrent models with a solver-based swap DSL
+# =============================================================================
 #
-# NOTE: the example below uses model names that are not defined above for demonstration purposes
-groups:
-  # group1 works the same as the default behaviour of llama-swap where only one model is allowed
-  # to run a time across the whole llama-swap instance
-  "group1":
-    # swap: controls the model swapping behaviour in within the group
-    # - optional, default: true
-    # - true : only one model is allowed to run at a time
-    # - false: all models can run together, no swapping
-    swap: true
+# Note:
+# A config must use either a matrix or legacy groups, not both. A configuration error
+# will occur if both are defined. Configuration examples for legacy Groups can be found:
+# https://github.com/mostlygeek/llama-swap/blob/40e39f7/config.example.yaml#L334-L396
+#
+# The matrix declares valid combinations of models that can run concurrently.
+# When a model is requested, the solver finds the cheapest way to make it
+# available by evicting as few (and least costly) running models as possible.
+#
+# Solver behavior:
+#   1. Request arrives for model X
+#   2. If X is already running, forward immediately. Done.
+#   3. Find all sets containing X
+#   4. For each candidate set, compute cost: sum of evict_costs for
+#      every running model NOT in that set
+#   5. Pick lowest cost candidate. Ties broken by definition order.
+#   6. Evict what needs to stop. Start X. Forward request.
+#
+# Subset semantics: a set [a, b, c] means any subset is valid.
+# Only the requested model is started — others are not preloaded.
+#
+# A model not appearing in any set can only run alone.
+#
+matrix:
+  # vars: short names for models (alphanumeric, 1-8 chars)
+  # - required for sets and evict_costs settings
+  # - each entry is a short name to a real model ID. Do not use an alias
+  # - used to keep set DSL logic short and easier to read
+  # - sets and evict_costs only use identifiers defined in vars
+  vars:
+    g: gemma-model
+    q: qwen-model
+    m: mistral-model
+    v: voxtral-model
+    e: reranker-model
+    L: llama-70B
+    sd: stable-diffusion
 
-    # exclusive: controls how the group affects other groups
-    # - optional, default: true
-    # - true: causes all other groups to unload when this group runs a model
-    # - false: does not affect other groups
-    exclusive: true
+  # evict_costs: relative cost of losing a running model (default: 1)
+  evict_costs:
+    v: 50 # vllm backend, slow cold start
+    L: 30 # 70B weights, slow to load
 
-    # members references the models defined above
-    # required
-    members:
-      - "llama"
-      - "qwen-unlisted"
+  # sets: named sets of concurrent model combinations
+  # Values are DSL strings with operators:
+  #   &     AND (models run together)
+  #   |     OR  (alternatives)
+  #   ()    grouping
+  #   +ref  inline another set's expression
+  #
+  # Expansion examples:
+  #   "L"                  → [L]
+  #   "a & b"              → [a, b]
+  #   "a | b"              → [a], [b]
+  #   "(a | b) & c"        → [a, c], [b, c]
+  #   "(a | b) & (c | d)"  → [a,c], [a,d], [b,c], [b,d]
+  #   "+llms & v"          → expands llms inline, then applies & v
+  sets:
+    # LLM + TTS: switching between g/q/m won't evict v
+    # expands to: [g,v], [q,v], [m,v]
+    standard: "(g | q | m) & v"
 
-  # Example:
-  # - in group2 all models can run at the same time
-  # - when a different group is loaded it causes all running models in this group to unload
-  "group2":
-    swap: false
+    # LLM + TTS + reranker
+    # expands to: [g,v,e], [q,v,e]
+    with_rerank: "(g | q) & v & e"
 
-    # exclusive: false does not unload other groups when a model in group2 is requested
-    # - the models in group2 will be loaded but will not unload any other groups
-    exclusive: false
-    members:
-      - "docker-llama"
-      - "modelA"
-      - "modelB"
+    # LLM + image generation, no TTS
+    # expands to: [g,sd], [q,sd]
+    creative: "(g | q) & sd"
 
-  # Example:
-  # - a persistent group, prevents other groups from unloading it
-  "forever":
-    # persistent: prevents over groups from unloading the models in this group
-    # - optional, default: false
-    # - does not affect individual model behaviour
-    persistent: true
-
-    # set swap/exclusive to false to prevent swapping inside the group
-    # and the unloading of other groups
-    swap: false
-    exclusive: false
-    members:
-      - "forever-modelA"
-      - "forever-modelB"
-      - "forever-modelc"
+    # 70B model uses all GPUs, can only run alone
+    # expands to: [L]
+    full: "L"
 
 # hooks: a dictionary of event triggers and actions
 # - optional, default: empty dictionary
@@ -365,7 +448,8 @@ peers:
     # - optional, default: ""
     # - if blank, no key will be added to the request
     # - key will be injected into headers: Authorization: Bearer <key> and x-api-key: <key>
-    apiKey: sk-your-openrouter-key
+    # - can be a string or a macro
+    apiKey: ${env.OPENROUTER_API_KEY}
     models:
       - meta-llama/llama-3.1-8b-instruct
       - qwen/qwen3-235b-a22b-2507
@@ -373,3 +457,34 @@ peers:
       - z-ai/glm-4.7
       - moonshotai/kimi-k2-0905
       - minimax/minimax-m2.1
+    # timeouts: configure proxy connection timeouts for this peer
+    # - optional, defaults shown below
+    # - useful when the peer runs on slower hardware
+    # - set any value to 0 to disable that timeout (not recommended)
+    timeouts:
+      connect: 30
+      keepalive: 30
+      responseHeader: 60
+      tlsHandshake: 10
+      idleConn: 90
+
+    # filters: a dictionary of filter settings for peer requests
+    # - optional, default: empty dictionary
+    # - same capabilities as model filters (stripParams, setParams)
+    filters:
+      # stripParams: a comma separated list of parameters to remove from the request
+      # - optional, default: ""
+      # - useful for removing parameters that the peer doesn't support
+      # - the `model` parameter can never be removed
+      stripParams: "temperature, top_p"
+
+      # setParams: a dictionary of parameters to set/override in requests to this peer
+      # - optional, default: empty dictionary
+      # - useful for injecting provider-specific settings like data retention policies
+      # - protected params like "model" cannot be overridden
+      # - values can be strings, numbers, booleans, arrays, or objects
+      setParams:
+        # Example: enforce zero-data-retention for OpenRouter
+        provider:
+          data_collection: "deny"
+          zdr: true

docker/build-container.sh

@@ -1,28 +1,50 @@
 #!/bin/bash
 
+set -euo pipefail
+
 cd $(dirname "$0")
 
+# use this to test locally, example:
+# GITHUB_TOKEN=$(gh auth token) LOG_DEBUG=1 DEBUG_ABORT_BUILD=1 ./docker/build-container.sh rocm
+# you need read:package scope on the token. Generate a personal access token with
+# the scopes: gist, read:org, repo, write:packages
+# then: gh auth login (and copy/paste the new token)
+
+LOG_DEBUG=${LOG_DEBUG:-0}
+DEBUG_ABORT_BUILD=${DEBUG_ABORT_BUILD:-}
+
+log_debug() {
+    if [ "$LOG_DEBUG" = "1" ]; then
+        echo "[DEBUG] $*"
+    fi
+}
+
+log_info() {
+    echo "[INFO] $*"
+}
+
 ARCH=$1
 PUSH_IMAGES=${2:-false}
 
 # List of allowed architectures
-ALLOWED_ARCHS=("intel" "vulkan" "musa" "cuda" "cpu")
+ALLOWED_ARCHS=("intel" "vulkan" "musa" "cuda" "cuda13" "cpu" "rocm")
 
 # Check if ARCH is in the allowed list
 if [[ ! " ${ALLOWED_ARCHS[@]} " =~ " ${ARCH} " ]]; then
-  echo "Error: ARCH must be one of the following: ${ALLOWED_ARCHS[@]}"
+  log_info "Error: ARCH must be one of the following: ${ALLOWED_ARCHS[@]}"
   exit 1
 fi
 
 # Check if GITHUB_TOKEN is set and not empty
-if [[ -z "$GITHUB_TOKEN" ]]; then
-  echo "Error: GITHUB_TOKEN is not set or is empty."
+if [[ -z "${GITHUB_TOKEN:-}" ]]; then
+  log_info "Error: GITHUB_TOKEN is not set or is empty."
   exit 1
 fi
 
 # Set llama.cpp base image, customizable using the BASE_LLAMACPP_IMAGE environment
 # variable, this permits testing with forked llama.cpp repositories
 BASE_IMAGE=${BASE_LLAMACPP_IMAGE:-ghcr.io/ggml-org/llama.cpp}
+SD_IMAGE=${BASE_SDCPP_IMAGE:-ghcr.io/leejet/stable-diffusion.cpp}
 
 # Set llama-swap repository, automatically uses GITHUB_REPOSITORY variable
 # to enable easy container builds on forked repos
@@ -32,25 +54,76 @@ LS_REPO=${GITHUB_REPOSITORY:-mostlygeek/llama-swap}
 # have to strip out the 'v' due to .tar.gz file naming
 LS_VER=$(curl -s https://api.github.com/repos/${LS_REPO}/releases/latest | jq -r .tag_name | sed 's/v//')
 
+# Fetches the most recent llama.cpp tag matching the given prefix
+# Handles pagination to search beyond the first 100 results
+# $1 - tag_prefix (e.g., "server" or "server-vulkan")
+# Returns: the version number extracted from the tag
+fetch_llama_tag() {
+    local tag_prefix=$1
+    local page=1
+    local per_page=100
+
+    while true; do
+        log_debug "Fetching page $page for tag prefix: $tag_prefix"
+
+        local response=$(curl -s -H "Authorization: Bearer $GITHUB_TOKEN" \
+            "https://api.github.com/users/ggml-org/packages/container/llama.cpp/versions?per_page=${per_page}&page=${page}")
+
+        # Check for API errors
+        if echo "$response" | jq -e '.message' > /dev/null 2>&1; then
+            local error_msg=$(echo "$response" | jq -r '.message')
+            log_info "GitHub API error: $error_msg"
+            return 1
+        fi
+
+        # Check if response is empty array (no more pages)
+        if [ "$(echo "$response" | jq 'length')" -eq 0 ]; then
+            log_debug "No more pages (empty response)"
+            return 1
+        fi
+
+        # Extract matching tag from this page
+        local found_tag=$(echo "$response" | jq -r \
+            ".[] | select(.metadata.container.tags[]? | startswith(\"$tag_prefix\")) | .metadata.container.tags[] | select(startswith(\"$tag_prefix\"))" \
+            | sort -r | head -n1)
+
+        if [ -n "$found_tag" ]; then
+            log_debug "Found tag: $found_tag on page $page"
+            echo "$found_tag" | awk -F '-' '{print $NF}'
+            return 0
+        fi
+
+        page=$((page + 1))
+
+        # Safety limit to prevent infinite loops
+        if [ $page -gt 50 ]; then
+            log_info "Reached pagination safety limit (50 pages)"
+            return 1
+        fi
+    done
+}
+
 if [ "$ARCH" == "cpu" ]; then
-    # cpu only containers just use the server tag
-    LCPP_TAG=$(curl -s -H "Authorization: Bearer $GITHUB_TOKEN" \
-        "https://api.github.com/users/ggml-org/packages/container/llama.cpp/versions" \
-        | jq -r '.[] | select(.metadata.container.tags[] | startswith("server")) | .metadata.container.tags[]' \
-        | sort -r | head -n1 | awk -F '-' '{print $3}')
+    LCPP_TAG=$(fetch_llama_tag "server")
     BASE_TAG=server-${LCPP_TAG}
 else
-    LCPP_TAG=$(curl -s -H "Authorization: Bearer $GITHUB_TOKEN" \
-        "https://api.github.com/users/ggml-org/packages/container/llama.cpp/versions" \
-        | jq -r --arg arch "$ARCH" '.[] | select(.metadata.container.tags[] | startswith("server-\($arch)")) | .metadata.container.tags[]' \
-        | sort -r | head -n1 | awk -F '-' '{print $3}')
+    LCPP_TAG=$(fetch_llama_tag "server-${ARCH}")
     BASE_TAG=server-${ARCH}-${LCPP_TAG}
 fi
 
+SD_TAG=master-${ARCH}
+
 # Abort if LCPP_TAG is empty.
 if [[ -z "$LCPP_TAG" ]]; then
-    echo "Abort: Could not find llama-server container for arch: $ARCH"
+    log_info "Abort: Could not find llama-server container for arch: $ARCH"
     exit 1
+else
+    log_info "LCPP_TAG: $LCPP_TAG"
+fi
+
+if [[ ! -z "$DEBUG_ABORT_BUILD" ]]; then
+    log_info "Abort: DEBUG_ABORT_BUILD set"
+    exit 0
 fi
 
 for CONTAINER_TYPE in non-root root; do
@@ -68,10 +141,22 @@ for CONTAINER_TYPE in non-root root; do
     USER_HOME=/app
   fi
 
-  echo "Building $CONTAINER_TYPE $CONTAINER_TAG $LS_VER"
-  docker build -f llama-swap.Containerfile --build-arg BASE_TAG=${BASE_TAG} --build-arg LS_VER=${LS_VER} --build-arg UID=${USER_UID} \
+  log_info "Building $CONTAINER_TYPE $CONTAINER_TAG $LS_VER"
+  docker build --provenance=false -f llama-swap.Containerfile --build-arg BASE_TAG=${BASE_TAG} --build-arg LS_VER=${LS_VER} --build-arg UID=${USER_UID} \
     --build-arg LS_REPO=${LS_REPO} --build-arg GID=${USER_GID} --build-arg USER_HOME=${USER_HOME} -t ${CONTAINER_TAG} -t ${CONTAINER_LATEST} \
     --build-arg BASE_IMAGE=${BASE_IMAGE} .
+
+  # For architectures with stable-diffusion.cpp support, layer sd-server on top
+  case "$ARCH" in
+    "musa" | "vulkan")
+      log_info "Adding sd-server to $CONTAINER_TAG"
+      docker build --provenance=false -f llama-swap-sd.Containerfile \
+        --build-arg BASE=${CONTAINER_TAG} \
+        --build-arg SD_IMAGE=${SD_IMAGE} --build-arg SD_TAG=${SD_TAG} \
+        --build-arg UID=${USER_UID} --build-arg GID=${USER_GID} \
+        -t ${CONTAINER_TAG} -t ${CONTAINER_LATEST} . ;;
+  esac
+
   if [ "$PUSH_IMAGES" == "true" ]; then
     docker push ${CONTAINER_TAG}
     docker push ${CONTAINER_LATEST}

docker/build-image.sh

@@ -0,0 +1,305 @@
+#!/bin/bash
+#
+# Build script for llama-swap-docker with commit hash pinning
+#
+# Usage:
+#   ./build-image.sh --cuda                    # Build CUDA image
+#   ./build-image.sh --vulkan                  # Build Vulkan image
+#   ./build-image.sh --cuda --no-cache         # Build CUDA image without cache
+#   LLAMA_COMMIT_HASH=abc123 ./build-image.sh --cuda      # Override llama.cpp commit
+#   LLAMA_COMMIT_HASH=b8429 ./build-image.sh --vulkan    # Override llama.cpp release tag (vulkan uses prebuilt binaries)
+#   WHISPER_COMMIT_HASH=def456 ./build-image.sh --vulkan  # Override whisper.cpp commit
+#   SD_COMMIT_HASH=ghi789 ./build-image.sh --cuda        # Override stable-diffusion.cpp commit
+#
+# Features:
+#   - Auto-detects latest commit hashes from git repos
+#   - Builds llama-swap from local source code
+#   - Allows environment variable overrides for reproducible builds
+#   - Cache-friendly: changing commit hash busts cache appropriately
+#   - Supports both CUDA and Vulkan backends (requires explicit flag)
+#
+
+set -euo pipefail
+
+# Parse command line arguments
+BACKEND=""
+NO_CACHE=false
+
+if [[ $# -eq 0 ]]; then
+    echo "Error: No backend specified. Please use --cuda or --vulkan."
+    echo ""
+    echo "Usage: ./build-image.sh --cuda|--vulkan [--no-cache]"
+    echo ""
+    echo "Options:"
+    echo "  --cuda      Build CUDA image (NVIDIA GPUs)"
+    echo "  --vulkan    Build Vulkan image (AMD GPUs and compatible hardware)"
+    echo "  --no-cache  Force rebuild without using Docker cache"
+    echo "  --help, -h  Show this help message"
+    echo ""
+    echo "Environment variables:"
+    echo "  DOCKER_IMAGE_TAG     Set custom image tag (default: llama-swap:cuda or llama-swap:vulkan)"
+    echo "  LLAMA_COMMIT_HASH    Override llama.cpp commit hash"
+    echo "  WHISPER_COMMIT_HASH  Override whisper.cpp commit hash"
+    echo "  SD_COMMIT_HASH       Override stable-diffusion.cpp commit hash"
+    exit 1
+fi
+
+for arg in "$@"; do
+    case $arg in
+        --cuda)
+            BACKEND="cuda"
+            ;;
+        --vulkan)
+            BACKEND="vulkan"
+            ;;
+        --no-cache)
+            NO_CACHE=true
+            ;;
+        --help|-h)
+            echo "Usage: ./build-image.sh --cuda|--vulkan [--no-cache]"
+            echo ""
+            echo "Options:"
+            echo "  --cuda      Build CUDA image (NVIDIA GPUs)"
+            echo "  --vulkan    Build Vulkan image (AMD GPUs and compatible hardware)"
+            echo "  --no-cache  Force rebuild without using Docker cache"
+            echo "  --help, -h  Show this help message"
+            echo ""
+            echo "Environment variables:"
+            echo "  DOCKER_IMAGE_TAG     Set custom image tag (default: llama-swap:cuda or llama-swap:vulkan)"
+            echo "  LLAMA_COMMIT_HASH    Override llama.cpp commit hash"
+            echo "  WHISPER_COMMIT_HASH  Override whisper.cpp commit hash"
+            echo "  SD_COMMIT_HASH       Override stable-diffusion.cpp commit hash"
+            exit 0
+            ;;
+    esac
+done
+
+# Validate backend selection
+if [[ -z "$BACKEND" ]]; then
+    echo "Error: No backend specified. Please use --cuda or --vulkan."
+    exit 1
+fi
+
+# Configuration
+if [[ -n "${DOCKER_IMAGE_TAG:-}" ]]; then
+    # User provided a custom tag, use it as-is
+    :
+elif [[ "$BACKEND" == "vulkan" ]]; then
+    DOCKER_IMAGE_TAG="llama-swap:vulkan"
+else
+    DOCKER_IMAGE_TAG="llama-swap:cuda"
+fi
+DOCKER_BUILDKIT="${DOCKER_BUILDKIT:-1}"
+
+# Single unified Dockerfile, backend selected via build arg
+DOCKERFILE="Dockerfile"
+if [[ "$BACKEND" == "vulkan" ]]; then
+    echo "Building for: Vulkan (AMD GPUs and compatible hardware)"
+else
+    echo "Building for: CUDA (NVIDIA GPUs)"
+fi
+
+# Git repository URLs
+LLAMA_REPO="https://github.com/ggml-org/llama.cpp.git"
+WHISPER_REPO="https://github.com/ggml-org/whisper.cpp.git"
+SD_REPO="https://github.com/leejet/stable-diffusion.cpp.git"
+
+# Function to get the latest commit hash from a git repo's default branch
+get_latest_commit() {
+    local repo_url="$1"
+    local branch="${2:-master}"
+
+    # Try to get the latest commit hash for the specified branch
+    git ls-remote --heads "${repo_url}" "${branch}" 2>/dev/null | head -1 | cut -f1
+}
+
+# Function to get the default branch name (master or main)
+get_default_branch() {
+    local repo_url="$1"
+
+    # Check for master first
+    if git ls-remote --heads "${repo_url}" master &>/dev/null; then
+        echo "master"
+    elif git ls-remote --heads "${repo_url}" main &>/dev/null; then
+        echo "main"
+    else
+        echo "master"  # fallback
+    fi
+}
+
+# Function to get the latest release tag from a GitHub repo
+get_latest_release_tag() {
+    local owner_repo="$1"
+    curl -fsSL "https://api.github.com/repos/${owner_repo}/releases/latest" \
+        | grep '"tag_name"' | head -1 | cut -d'"' -f4
+}
+
+echo "=========================================="
+echo "llama-swap-docker Build Script"
+echo "=========================================="
+echo ""
+
+# Determine commit hashes / release tags - use env vars or auto-detect
+# For vulkan builds, llama and sd use GitHub release tags (prebuilt binaries).
+# For cuda builds (or whisper on any backend), use git commit hashes.
+if [[ -n "${LLAMA_COMMIT_HASH:-}" ]]; then
+    LLAMA_HASH="${LLAMA_COMMIT_HASH}"
+    echo "llama.cpp: Using provided version: ${LLAMA_HASH}"
+elif [[ "$BACKEND" == "vulkan" ]]; then
+    LLAMA_HASH=$(get_latest_release_tag "ggml-org/llama.cpp")
+    if [[ -z "${LLAMA_HASH}" ]]; then
+        echo "ERROR: Could not determine latest release tag for llama.cpp" >&2
+        exit 1
+    fi
+    echo "llama.cpp: Auto-detected latest release tag: ${LLAMA_HASH}"
+else
+    LLAMA_BRANCH=$(get_default_branch "${LLAMA_REPO}")
+    LLAMA_HASH=$(get_latest_commit "${LLAMA_REPO}" "${LLAMA_BRANCH}")
+    if [[ -z "${LLAMA_HASH}" ]]; then
+        echo "ERROR: Could not determine latest commit for llama.cpp" >&2
+        exit 1
+    fi
+    echo "llama.cpp: Auto-detected latest commit (${LLAMA_BRANCH}): ${LLAMA_HASH}"
+fi
+
+if [[ -n "${WHISPER_COMMIT_HASH:-}" ]]; then
+    WHISPER_HASH="${WHISPER_COMMIT_HASH}"
+    echo "whisper.cpp: Using provided commit hash: ${WHISPER_HASH}"
+else
+    WHISPER_BRANCH=$(get_default_branch "${WHISPER_REPO}")
+    WHISPER_HASH=$(get_latest_commit "${WHISPER_REPO}" "${WHISPER_BRANCH}")
+    if [[ -z "${WHISPER_HASH}" ]]; then
+        echo "ERROR: Could not determine latest commit for whisper.cpp" >&2
+        exit 1
+    fi
+    echo "whisper.cpp: Auto-detected latest commit (${WHISPER_BRANCH}): ${WHISPER_HASH}"
+fi
+
+if [[ -n "${SD_COMMIT_HASH:-}" ]]; then
+    SD_HASH="${SD_COMMIT_HASH}"
+    echo "stable-diffusion.cpp: Using provided version: ${SD_HASH}"
+elif [[ "$BACKEND" == "vulkan" ]]; then
+    SD_HASH=$(get_latest_release_tag "leejet/stable-diffusion.cpp")
+    if [[ -z "${SD_HASH}" ]]; then
+        echo "ERROR: Could not determine latest release tag for stable-diffusion.cpp" >&2
+        exit 1
+    fi
+    echo "stable-diffusion.cpp: Auto-detected latest release tag: ${SD_HASH}"
+else
+    SD_BRANCH=$(get_default_branch "${SD_REPO}")
+    SD_HASH=$(get_latest_commit "${SD_REPO}" "${SD_BRANCH}")
+    if [[ -z "${SD_HASH}" ]]; then
+        echo "ERROR: Could not determine latest commit for stable-diffusion.cpp" >&2
+        exit 1
+    fi
+    echo "stable-diffusion.cpp: Auto-detected latest commit (${SD_BRANCH}): ${SD_HASH}"
+fi
+
+echo ""
+echo "=========================================="
+echo "Starting Docker build..."
+echo "=========================================="
+echo ""
+
+# Build the Docker image with commit hashes as build args
+# Build context is the repository root (..) so the Dockerfile can access Go source
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+REPO_ROOT="$(cd "${SCRIPT_DIR}/.." && pwd)"
+BUILD_ARGS=(
+    --build-arg "BACKEND=${BACKEND}"
+    --build-arg "LLAMA_COMMIT_HASH=${LLAMA_HASH}"
+    --build-arg "WHISPER_COMMIT_HASH=${WHISPER_HASH}"
+    --build-arg "SD_COMMIT_HASH=${SD_HASH}"
+    -t "${DOCKER_IMAGE_TAG}"
+    -f "${SCRIPT_DIR}/${DOCKERFILE}"
+)
+
+if [[ "$NO_CACHE" == true ]]; then
+    BUILD_ARGS+=(--no-cache)
+    echo "Note: Building without cache"
+fi
+
+# Use docker buildx with a custom builder for parallelism control
+# The legacy DOCKER_BUILDKIT=1 docker build doesn't respect BUILDKIT_MAX_PARALLELISM env var
+# We need to use a custom builder with a buildkitd.toml config file
+BUILDER_NAME="llama-swap-builder"
+
+# Check if our custom builder exists with the right config, create/update if needed
+if ! docker buildx inspect "$BUILDER_NAME" >/dev/null 2>&1; then
+    echo "Creating custom buildx builder with max-parallelism=1..."
+    
+    # Create buildkitd.toml config file
+    cat > buildkitd.toml << 'BUILDKIT_EOF'
+[worker.oci]
+  max-parallelism = 1
+BUILDKIT_EOF
+    
+    # Create the builder with the config
+    docker buildx create --name "$BUILDER_NAME" \
+        --driver docker-container \
+        --buildkitd-config buildkitd.toml \
+        --use
+else
+    # Switch to our builder
+    docker buildx use "$BUILDER_NAME"
+fi
+
+echo "Building with sequential stages (one at a time), each using all CPU cores..."
+echo "Using builder: $BUILDER_NAME"
+
+# Use docker buildx build with --load to load the image into Docker
+# The --builder flag ensures we use our custom builder with max-parallelism=1
+# Build context is the repository root so we can access Go source files
+docker buildx build --builder "$BUILDER_NAME" --load "${BUILD_ARGS[@]}" "${REPO_ROOT}"
+
+echo ""
+echo "=========================================="
+echo "Verifying build artifacts..."
+echo "=========================================="
+echo ""
+
+# Verify all expected binaries exist in the image
+MISSING_BINARIES=()
+
+for binary in llama-server llama-cli whisper-server whisper-cli sd-server sd-cli llama-swap; do
+    if ! docker run --rm "${DOCKER_IMAGE_TAG}" which "${binary}" >/dev/null 2>&1; then
+        MISSING_BINARIES+=("${binary}")
+    fi
+done
+
+if [[ ${#MISSING_BINARIES[@]} -gt 0 ]]; then
+    echo "ERROR: Build succeeded but the following binaries are missing from the image:"
+    for binary in "${MISSING_BINARIES[@]}"; do
+        echo "  - ${binary}"
+    done
+    echo ""
+    echo "This usually indicates a build stage failure. Try running with --no-cache flag:"
+    echo "  ./build-image.sh --vulkan --no-cache"
+    exit 1
+fi
+
+echo "All expected binaries verified: llama-server, llama-cli, whisper-server, whisper-cli, sd-server, sd-cli, llama-swap"
+
+echo ""
+echo "=========================================="
+echo "Build complete!"
+echo "=========================================="
+echo ""
+echo "Image tag: ${DOCKER_IMAGE_TAG}"
+echo ""
+echo "Built with:"
+echo "  llama.cpp:           ${LLAMA_HASH}"
+echo "  whisper.cpp:         ${WHISPER_HASH}"
+echo "  stable-diffusion.cpp: ${SD_HASH}"
+echo "  llama-swap:          $(docker run --rm "${DOCKER_IMAGE_TAG}" cat /versions.txt | grep llama-swap | cut -d' ' -f2-)"
+echo ""
+if [[ "$BACKEND" == "vulkan" ]]; then
+    echo "Run with:"
+    echo "  docker run -it --rm --device /dev/dri:/dev/dri ${DOCKER_IMAGE_TAG}"
+    echo ""
+    echo "Note: For AMD GPUs, you may also need to mount render devices:"
+    echo "  docker run -it --rm --device /dev/dri:/dev/dri --group-add video ${DOCKER_IMAGE_TAG}"
+else
+    echo "Run with:"
+    echo "  docker run -it --rm --gpus all ${DOCKER_IMAGE_TAG}"
+fi

docker/config.example.yaml

@@ -15,4 +15,19 @@ models:
     cmd: >
       /app/llama-server
       -hf bartowski/SmolLM2-135M-Instruct-GGUF:Q4_K_M
-      --port 9999
+      --port 9999
+
+  z-image:
+    checkEndpoint: /
+    cmd: |
+      /app/sd-server
+      --listen-port 9999
+      --diffusion-fa
+      --diffusion-model /models/z_image_turbo-Q8_0.gguf
+      --vae /models/ae.safetensors
+      --llm /models/qwen3-4b-instruct-2507-q8_0.gguf
+      --offload-to-cpu
+      --cfg-scale 1.0
+      --height 512 --width 512
+      --steps 8
+    aliases: [gpt-image-1,dall-e-2,dall-e-3,gpt-image-1-mini,gpt-image-1.5]

docker/llama-swap-sd.Containerfile

@@ -0,0 +1,11 @@
+ARG SD_IMAGE=ghcr.io/leejet/stable-diffusion.cpp
+ARG SD_TAG=master-vulkan
+ARG BASE=llama-swap:latest
+
+FROM ${SD_IMAGE}:${SD_TAG} AS sd-source
+FROM ${BASE}
+
+ARG UID=10001
+ARG GID=10001
+
+COPY --from=sd-source --chown=${UID}:${GID} /sd-server /app/sd-server

docker/unified/Dockerfile

@@ -0,0 +1,207 @@
+# Unified multi-stage Dockerfile for AI inference tools
+# Supports CUDA and Vulkan backends via BACKEND build arg
+#
+# Usage:
+#   docker buildx build --build-arg BACKEND=cuda -t llama-swap:unified-cuda .
+#   docker buildx build --build-arg BACKEND=vulkan -t llama-swap:unified-vulkan .
+#   docker buildx build --build-arg BACKEND=cuda --build-arg CMAKE_CUDA_ARCHITECTURES="86;89" -t llama-swap:unified-cuda .
+#
+# Each project has its own install script that handles cloning, building,
+# and installing binaries. Build stages are independent for cache efficiency.
+
+ARG BACKEND=cuda
+
+# ── Builder bases ──────────────────────────────────────────────────────
+
+FROM nvidia/cuda:12.9.1-devel-ubuntu24.04 AS builder-base-cuda
+
+ARG CMAKE_CUDA_ARCHITECTURES="60;61;75;86;89"
+ENV DEBIAN_FRONTEND=noninteractive
+ENV CMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES}
+ENV CCACHE_DIR=/ccache
+ENV CCACHE_MAXSIZE=2G
+ENV PATH="/usr/lib/ccache:${PATH}"
+
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    build-essential cmake git python3 python3-pip libssl-dev \
+    curl ca-certificates ccache make wget \
+    && rm -rf /var/lib/apt/lists/*
+
+WORKDIR /build
+
+# ──
+
+FROM ubuntu:24.04 AS builder-base-vulkan
+
+ENV DEBIAN_FRONTEND=noninteractive
+ENV CCACHE_DIR=/ccache
+ENV CCACHE_MAXSIZE=2G
+ENV PATH="/usr/lib/ccache:${PATH}"
+
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    build-essential cmake git python3 python3-pip libssl-dev \
+    curl ca-certificates ccache make wget software-properties-common \
+    libvulkan-dev glslang-tools spirv-tools vulkan-validationlayers glslc \
+    spirv-headers \
+    && rm -rf /var/lib/apt/lists/*
+
+WORKDIR /build
+
+# ── Select builder base by BACKEND ────────────────────────────────────
+
+FROM builder-base-${BACKEND} AS builder-base
+
+# ── Build whisper.cpp (fastest build, run first) ──────────────────────
+
+FROM builder-base AS whisper-build
+ARG BACKEND=cuda
+ARG WHISPER_COMMIT_HASH=master
+COPY install-whisper.sh /build/
+RUN --mount=type=cache,id=ccache-${BACKEND},target=/ccache \
+    --mount=type=cache,id=whisper-${BACKEND},target=/src/whisper.cpp/build \
+    BACKEND=${BACKEND} bash /build/install-whisper.sh "${WHISPER_COMMIT_HASH}"
+
+# ── Build stable-diffusion.cpp ────────────────────────────────────────
+
+FROM builder-base AS sd-build
+ARG BACKEND=cuda
+ARG SD_COMMIT_HASH=master
+COPY install-sd.sh /build/
+RUN --mount=type=cache,id=ccache-${BACKEND},target=/ccache \
+    --mount=type=cache,id=sd-${BACKEND},target=/src/stable-diffusion.cpp/build \
+    BACKEND=${BACKEND} bash /build/install-sd.sh "${SD_COMMIT_HASH}"
+
+# ── Build llama.cpp (slowest build, run last) ─────────────────────────
+
+FROM builder-base AS llama-build
+ARG BACKEND=cuda
+ARG LLAMA_COMMIT_HASH=master
+COPY install-llama.sh /build/
+RUN --mount=type=cache,id=ccache-${BACKEND},target=/ccache \
+    --mount=type=cache,id=llama-${BACKEND},target=/src/llama.cpp/build \
+    BACKEND=${BACKEND} bash /build/install-llama.sh "${LLAMA_COMMIT_HASH}"
+
+# ── Build ik_llama.cpp (CUDA only) ────────────────────────────────────
+#
+# Two named stages allow ARG BACKEND to select at build time:
+#   - ik-llama-cuda  : real build (from builder-base-cuda)
+#   - ik-llama-vulkan: no-op (empty /install/bin, skips CUDA pull entirely)
+# BuildKit only evaluates the selected branch, so vulkan builds never
+# pull nvidia/cuda:*-devel or compile ik_llama.cpp.
+
+FROM builder-base-vulkan AS ik-llama-vulkan
+RUN mkdir -p /install/bin
+
+FROM builder-base-cuda AS ik-llama-cuda
+ARG IK_LLAMA_COMMIT_HASH=main
+COPY install-ik-llama.sh /build/
+RUN --mount=type=cache,id=ccache-cuda,target=/ccache \
+    --mount=type=cache,id=ik-llama-cuda,target=/src/ik_llama.cpp/build \
+    bash /build/install-ik-llama.sh "${IK_LLAMA_COMMIT_HASH}"
+
+ARG BACKEND=cuda
+FROM ik-llama-${BACKEND} AS ik-llama-build
+
+# ── Download llama-swap release binary ────────────────────────────────
+
+FROM builder-base AS llama-swap-download
+ARG LS_VERSION=latest
+COPY install-llama-swap.sh /build/
+RUN bash /build/install-llama-swap.sh "${LS_VERSION}"
+
+# ── Runtime bases ─────────────────────────────────────────────────────
+
+FROM nvidia/cuda:12.9.1-runtime-ubuntu24.04 AS runtime-cuda
+
+ENV DEBIAN_FRONTEND=noninteractive
+ENV LD_LIBRARY_PATH="/usr/local/cuda/lib64:${LD_LIBRARY_PATH}"
+ENV PATH="/usr/local/bin:${PATH}"
+
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    libgomp1 python3 curl ca-certificates \
+    && rm -rf /var/lib/apt/lists/*
+
+# CUDA stub drivers for container compatibility
+COPY --from=builder-base-cuda /usr/local/cuda/lib64/stubs/libcuda.so /usr/local/cuda/lib64/stubs/libcuda.so
+COPY --from=builder-base-cuda /usr/local/cuda/lib64/stubs/libcuda.so /usr/local/cuda/lib64/stubs/libcuda.so.1
+
+# ──
+
+FROM ubuntu:24.04 AS runtime-vulkan
+
+ENV DEBIAN_FRONTEND=noninteractive
+ENV PATH="/usr/local/bin:${PATH}"
+
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    libgomp1 libvulkan1 mesa-vulkan-drivers \
+    python3 curl ca-certificates \
+    && rm -rf /var/lib/apt/lists/*
+
+# ── Select runtime base by BACKEND ────────────────────────────────────
+
+FROM runtime-${BACKEND} AS runtime
+
+ARG BACKEND=cuda
+ARG LLAMA_COMMIT_HASH=unknown
+ARG WHISPER_COMMIT_HASH=unknown
+ARG SD_COMMIT_HASH=unknown
+ARG IK_LLAMA_COMMIT_HASH=unknown
+ARG RUN_UID=0
+
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    python3-numpy python3-sentencepiece python3-pip \
+    && rm -rf /var/lib/apt/lists/*
+
+# Create non-root user when RUN_UID != 0
+RUN if [ "$RUN_UID" != "0" ]; then \
+      groupadd --system --gid $RUN_UID llama-swap && \
+      useradd --system --uid $RUN_UID --gid $RUN_UID \
+        --home /app --shell /sbin/nologin llama-swap; \
+    fi && \
+    mkdir -p /etc/llama-swap/config && \
+    chown -R ${RUN_UID}:${RUN_UID} /etc/llama-swap
+
+WORKDIR /app
+
+# Copy whisper.cpp binaries and libraries
+COPY --from=whisper-build /install/bin/whisper-server /usr/local/bin/
+COPY --from=whisper-build /install/bin/whisper-cli /usr/local/bin/
+COPY --from=whisper-build /install/lib/ /usr/local/lib/
+
+# Copy stable-diffusion.cpp binaries and libraries
+COPY --from=sd-build /install/bin/sd-server /usr/local/bin/
+COPY --from=sd-build /install/bin/sd-cli /usr/local/bin/
+COPY --from=sd-build /install/lib/ /usr/local/lib/
+
+# Copy llama.cpp binaries (statically linked)
+COPY --from=llama-build /install/bin/llama-server /usr/local/bin/
+COPY --from=llama-build /install/bin/llama-cli /usr/local/bin/
+
+# Copy ik-llama-server (CUDA only; empty copy for vulkan)
+COPY --from=ik-llama-build /install/bin/ /usr/local/bin/
+
+# Install uv
+RUN pip install uv --break-system-packages
+
+# Copy llama-swap binary
+COPY --from=llama-swap-download /install/bin/llama-swap /usr/local/bin/
+COPY --from=llama-swap-download /install/llama-swap-version /tmp/
+
+RUN ldconfig
+
+COPY config.example.yaml /etc/llama-swap/config/config.yaml
+
+# Version tracking
+RUN echo "llama.cpp: ${LLAMA_COMMIT_HASH}" > /versions.txt && \
+    echo "whisper.cpp: ${WHISPER_COMMIT_HASH}" >> /versions.txt && \
+    echo "stable-diffusion.cpp: ${SD_COMMIT_HASH}" >> /versions.txt && \
+    echo "ik_llama.cpp: ${IK_LLAMA_COMMIT_HASH}" >> /versions.txt && \
+    echo "llama-swap: $(cat /tmp/llama-swap-version)" >> /versions.txt && \
+    echo "backend: ${BACKEND}" >> /versions.txt && \
+    echo "build_timestamp: $(date -u +%Y-%m-%dT%H:%M:%SZ)" >> /versions.txt
+
+RUN mkdir -p /models && chown ${RUN_UID}:${RUN_UID} /models
+WORKDIR /models
+USER ${RUN_UID}
+ENTRYPOINT ["llama-swap"]
+CMD ["-config", "/etc/llama-swap/config/config.yaml", "-listen", "0.0.0.0:8080"]

docker/unified/README.md

@@ -0,0 +1,8 @@
+# Unified Docker Container
+
+These scripts create a custom llama-swap container that contains:
+
+- llama-server for LLMs, rerank and embedding model support
+- sd-server (stable-diffusion.cpp) for image generation
+- whisper.cpp for ASR
+

docker/unified/build-image.sh

@@ -0,0 +1,303 @@
+#!/bin/bash
+#
+# Build script for unified container with version pinning
+#
+# Usage:
+#   ./build-image.sh --cuda                              # Build CUDA image
+#   ./build-image.sh --vulkan                            # Build Vulkan image
+#   ./build-image.sh --cuda --no-cache                   # Build without cache
+#   LLAMA_REF=b1234 ./build-image.sh --vulkan            # Pin llama.cpp to a commit hash
+#   LLAMA_REF=v1.2.3 ./build-image.sh --cuda             # Pin llama.cpp to a tag
+#   WHISPER_REF=v1.0.0 ./build-image.sh --vulkan         # Pin whisper.cpp to a tag
+#   SD_REF=master ./build-image.sh --cuda                # Pin stable-diffusion.cpp to a branch
+#   LS_VERSION=170 ./build-image.sh --cuda               # Override llama-swap version
+#   IK_LLAMA_REF=main ./build-image.sh --cuda            # Pin ik_llama.cpp to main branch (CUDA only)
+#
+
+set -euo pipefail
+
+BACKEND=""
+NO_CACHE=false
+
+for arg in "$@"; do
+    case $arg in
+        --cuda)
+            BACKEND="cuda"
+            ;;
+        --vulkan)
+            BACKEND="vulkan"
+            ;;
+        --no-cache)
+            NO_CACHE=true
+            ;;
+        --help|-h)
+            echo "Usage: ./build-image.sh --cuda|--vulkan [--no-cache]"
+            echo ""
+            echo "Options:"
+            echo "  --cuda      Build CUDA image (NVIDIA GPUs)"
+            echo "  --vulkan    Build Vulkan image (AMD GPUs and compatible hardware)"
+            echo "  --no-cache  Force rebuild without using Docker cache"
+            echo "  --help, -h  Show this help message"
+            echo ""
+            echo "Environment variables:"
+            echo "  DOCKER_IMAGE_TAG     Set custom image tag (default: llama-swap:unified-cuda or llama-swap:unified-vulkan)"
+            echo "  LLAMA_REF            Pin llama.cpp to a commit, tag, or branch"
+            echo "  WHISPER_REF          Pin whisper.cpp to a commit, tag, or branch"
+            echo "  SD_REF               Pin stable-diffusion.cpp to a commit, tag, or branch"
+            echo "  IK_LLAMA_REF         Pin ik_llama.cpp to a commit, tag, or branch (CUDA only)"
+            echo "  LS_VERSION           Override llama-swap version (e.g., '170' or 'latest')"
+            exit 0
+            ;;
+    esac
+done
+
+if [[ -z "$BACKEND" ]]; then
+    echo "Error: No backend specified. Please use --cuda or --vulkan."
+    echo ""
+    echo "Usage: ./build-image.sh --cuda|--vulkan [--no-cache]"
+    exit 1
+fi
+
+DOCKER_IMAGE_TAG="${DOCKER_IMAGE_TAG:-llama-swap:unified-${BACKEND}}"
+
+# Git repository URLs
+LLAMA_REPO="https://github.com/ggml-org/llama.cpp.git"
+WHISPER_REPO="https://github.com/ggml-org/whisper.cpp.git"
+SD_REPO="https://github.com/leejet/stable-diffusion.cpp.git"
+LLAMA_SWAP_REPO="https://github.com/mostlygeek/llama-swap.git"
+IK_LLAMA_REPO="https://github.com/ikawrakow/ik_llama.cpp.git"
+
+# Resolve a git ref (commit hash, tag, or branch) to a full commit hash.
+# Requires only: git, network access to the remote.
+resolve_ref() {
+    local repo_url="$1"
+    local ref="$2"
+
+    # Full 40-char SHA — use as-is
+    if [[ "${ref}" =~ ^[0-9a-f]{40}$ ]]; then
+        echo "${ref}"
+        return
+    fi
+
+    # Try tag then branch (exact match)
+    local hash
+    hash=$(git ls-remote "${repo_url}" "refs/tags/${ref}" "refs/heads/${ref}" 2>/dev/null | head -1 | cut -f1)
+    if [[ -n "${hash}" ]]; then
+        echo "${hash}"
+        return
+    fi
+
+    # Short hash (7+ chars): scan all refs for a SHA with this prefix
+    if [[ "${ref}" =~ ^[0-9a-f]{7,}$ ]]; then
+        hash=$(git ls-remote "${repo_url}" 2>/dev/null | grep "^${ref}" | head -1 | cut -f1)
+        if [[ -n "${hash}" ]]; then
+            echo "${hash}"
+            return
+        fi
+    fi
+
+    echo "ERROR: Could not resolve ref '${ref}' for ${repo_url}" >&2
+    if [[ "${ref}" =~ ^[0-9a-f]+$ && ${#ref} -lt 7 ]]; then
+        echo "  Short hashes must be at least 7 characters (got ${#ref})." >&2
+    else
+        echo "  Tried: tag, branch, git ls-remote prefix match" >&2
+    fi
+    echo "  Use a full 40-char SHA, a tag name, a branch name, or a 7+ char short hash." >&2
+    return 1
+}
+
+# Resolve HEAD of a repo without needing to know the default branch name.
+get_latest_hash() {
+    git ls-remote "${1}" HEAD 2>/dev/null | head -1 | cut -f1
+}
+
+echo "=========================================="
+echo "llama-swap Unified Build (${BACKEND})"
+echo "=========================================="
+echo ""
+
+# Resolve llama.cpp ref
+if [[ -n "${LLAMA_REF:-}" ]]; then
+    LLAMA_HASH=$(resolve_ref "${LLAMA_REPO}" "${LLAMA_REF}") || exit 1
+    echo "llama.cpp: ${LLAMA_REF} -> ${LLAMA_HASH}"
+else
+    LLAMA_HASH=$(get_latest_hash "${LLAMA_REPO}")
+    if [[ -z "${LLAMA_HASH}" ]]; then
+        echo "ERROR: Could not determine latest commit for llama.cpp" >&2
+        exit 1
+    fi
+    echo "llama.cpp: latest HEAD: ${LLAMA_HASH}"
+fi
+
+# Resolve whisper.cpp ref
+if [[ -n "${WHISPER_REF:-}" ]]; then
+    WHISPER_HASH=$(resolve_ref "${WHISPER_REPO}" "${WHISPER_REF}") || exit 1
+    echo "whisper.cpp: ${WHISPER_REF} -> ${WHISPER_HASH}"
+else
+    WHISPER_HASH=$(get_latest_hash "${WHISPER_REPO}")
+    if [[ -z "${WHISPER_HASH}" ]]; then
+        echo "ERROR: Could not determine latest commit for whisper.cpp" >&2
+        exit 1
+    fi
+    echo "whisper.cpp: latest HEAD: ${WHISPER_HASH}"
+fi
+
+# Resolve stable-diffusion.cpp ref
+if [[ -n "${SD_REF:-}" ]]; then
+    SD_HASH=$(resolve_ref "${SD_REPO}" "${SD_REF}") || exit 1
+    echo "stable-diffusion.cpp: ${SD_REF} -> ${SD_HASH}"
+else
+    SD_HASH=$(get_latest_hash "${SD_REPO}")
+    if [[ -z "${SD_HASH}" ]]; then
+        echo "ERROR: Could not determine latest commit for stable-diffusion.cpp" >&2
+        exit 1
+    fi
+    echo "stable-diffusion.cpp: latest HEAD: ${SD_HASH}"
+fi
+
+# Resolve ik_llama.cpp ref (CUDA only)
+if [[ "$BACKEND" == "cuda" ]]; then
+    if [[ -n "${IK_LLAMA_REF:-}" ]]; then
+        IK_LLAMA_HASH=$(resolve_ref "${IK_LLAMA_REPO}" "${IK_LLAMA_REF}") || exit 1
+        echo "ik_llama.cpp: ${IK_LLAMA_REF} -> ${IK_LLAMA_HASH}"
+    else
+        IK_LLAMA_HASH=$(get_latest_hash "${IK_LLAMA_REPO}")
+        if [[ -z "${IK_LLAMA_HASH}" ]]; then
+            echo "ERROR: Could not determine latest commit for ik_llama.cpp" >&2
+            exit 1
+        fi
+        echo "ik_llama.cpp: latest HEAD: ${IK_LLAMA_HASH}"
+    fi
+else
+    IK_LLAMA_HASH="n/a"
+    echo "ik_llama.cpp: skipped (vulkan build)"
+fi
+
+# Resolve llama-swap ref
+if [[ -n "${LS_VERSION:-}" ]]; then
+    LS_HASH=$(resolve_ref "${LLAMA_SWAP_REPO}" "${LS_VERSION}") || exit 1
+    echo "llama-swap: ${LS_VERSION} -> ${LS_HASH}"
+else
+    LS_HASH=$(get_latest_hash "${LLAMA_SWAP_REPO}")
+    if [[ -z "${LS_HASH}" ]]; then
+        echo "ERROR: Could not determine latest commit for llama-swap" >&2
+        exit 1
+    fi
+    echo "llama-swap: latest HEAD: ${LS_HASH}"
+fi
+
+echo ""
+echo "=========================================="
+echo "Starting Docker build..."
+echo "=========================================="
+echo ""
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+
+BUILD_ARGS=(
+    --build-arg "BACKEND=${BACKEND}"
+    --build-arg "LLAMA_COMMIT_HASH=${LLAMA_HASH}"
+    --build-arg "WHISPER_COMMIT_HASH=${WHISPER_HASH}"
+    --build-arg "SD_COMMIT_HASH=${SD_HASH}"
+    --build-arg "IK_LLAMA_COMMIT_HASH=${IK_LLAMA_HASH}"
+    --build-arg "LS_VERSION=${LS_HASH}"
+    -t "${DOCKER_IMAGE_TAG}"
+    -f "${SCRIPT_DIR}/Dockerfile"
+)
+
+if [[ "$NO_CACHE" == true ]]; then
+    BUILD_ARGS+=(--no-cache)
+    echo "Note: Building without cache"
+elif [[ "${GITHUB_ACTIONS:-}" == "true" && "${ACT:-}" != "true" ]]; then
+    CACHE_REF="ghcr.io/mostlygeek/llama-swap:unified-${BACKEND}-cache"
+    BUILD_ARGS+=(
+        --cache-from "type=registry,ref=${CACHE_REF}"
+        --cache-to "type=registry,ref=${CACHE_REF},mode=max"
+    )
+    echo "Note: Using registry cache (${CACHE_REF})"
+fi
+
+DOCKER_BUILDKIT=1 docker buildx build --load "${BUILD_ARGS[@]}" "${SCRIPT_DIR}"
+
+echo ""
+echo "=========================================="
+echo "Verifying build artifacts..."
+echo "=========================================="
+echo ""
+
+EXPECTED_BINARIES=(llama-server llama-cli whisper-server whisper-cli sd-server sd-cli llama-swap)
+if [[ "$BACKEND" == "cuda" ]]; then
+    EXPECTED_BINARIES+=(ik-llama-server)
+fi
+
+MISSING_BINARIES=()
+for binary in "${EXPECTED_BINARIES[@]}"; do
+    if ! docker run --rm --entrypoint which "${DOCKER_IMAGE_TAG}" "${binary}" >/dev/null 2>&1; then
+        MISSING_BINARIES+=("${binary}")
+    fi
+done
+
+if [[ ${#MISSING_BINARIES[@]} -gt 0 ]]; then
+    echo "ERROR: Build succeeded but the following binaries are missing:"
+    for binary in "${MISSING_BINARIES[@]}"; do
+        echo "  - ${binary}"
+    done
+    echo ""
+    echo "Try running with --no-cache flag:"
+    echo "  ./build-image.sh --${BACKEND} --no-cache"
+    exit 1
+fi
+
+VERIFIED_LIST="llama-server, llama-cli, whisper-server, whisper-cli, sd-server, sd-cli, llama-swap"
+if [[ "$BACKEND" == "cuda" ]]; then
+    VERIFIED_LIST="${VERIFIED_LIST}, ik-llama-server"
+fi
+echo "All expected binaries verified: ${VERIFIED_LIST}"
+
+echo ""
+echo "=========================================="
+echo "Building rootless image..."
+echo "=========================================="
+echo ""
+
+ROOTLESS_TAG="${DOCKER_IMAGE_TAG}-rootless"
+docker buildx build --load -t "${ROOTLESS_TAG}" - <<EOF
+FROM ${DOCKER_IMAGE_TAG}
+USER root
+RUN groupadd --system --gid 10001 llama-swap && \\
+    useradd --system --uid 10001 --gid 10001 \\
+      --home /app --shell /sbin/nologin llama-swap && \\
+    chown -R 10001:10001 /etc/llama-swap /models
+USER 10001
+EOF
+
+echo "Rootless image built: ${ROOTLESS_TAG}"
+
+echo ""
+echo "=========================================="
+echo "Build complete!"
+echo "=========================================="
+echo ""
+echo "Image tags:"
+echo "  ${DOCKER_IMAGE_TAG}"
+echo "  ${ROOTLESS_TAG}"
+echo ""
+echo "Built with:"
+echo "  llama.cpp:            ${LLAMA_HASH}"
+echo "  whisper.cpp:          ${WHISPER_HASH}"
+echo "  stable-diffusion.cpp: ${SD_HASH}"
+if [[ "$BACKEND" == "cuda" ]]; then
+    echo "  ik_llama.cpp:         ${IK_LLAMA_HASH}"
+fi
+echo "  llama-swap:           $(docker run --rm --entrypoint cat "${DOCKER_IMAGE_TAG}" /versions.txt | grep llama-swap | cut -d' ' -f2-)"
+echo ""
+if [[ "$BACKEND" == "vulkan" ]]; then
+    echo "Run with:"
+    echo "  docker run -it --rm --device /dev/dri:/dev/dri ${DOCKER_IMAGE_TAG}"
+    echo ""
+    echo "Note: For AMD GPUs, you may also need:"
+    echo "  docker run -it --rm --device /dev/dri:/dev/dri --group-add video ${DOCKER_IMAGE_TAG}"
+else
+    echo "Run with:"
+    echo "  docker run -it --rm --gpus all ${DOCKER_IMAGE_TAG}"
+fi

docker/unified/config.example.yaml

@@ -0,0 +1,33 @@
+# placeholder example configuration
+healthCheckTimeout: 300
+logRequests: true
+
+models:
+  "llama":
+    cmd: >
+      llama-server
+      -hf bartowski/Qwen2.5-0.5B-Instruct-GGUF:Q4_K_M
+      --port ${PORT}
+
+  "whisper":
+    checkEndpoint: /v1/audio/transcriptions/
+    cmd: >
+      whisper-server
+      --port ${PORT}
+      --m /models/whisper.bin
+      --flash-attn
+      --request-path /v1/audio/transcriptions --inference-path ""
+
+  "image":
+    checkEndpoint: /
+    cmd: |
+      /app/sd-server
+      --listen-port 9999
+      --diffusion-fa
+      --diffusion-model /models/z_image_turbo-Q8_0.gguf
+      --vae /models/ae.safetensors
+      --llm /models/qwen3-4b-instruct-2507-q8_0.gguf
+      --offload-to-cpu
+      --cfg-scale 1.0
+      --height 512 --width 512
+      --steps 8

docker/unified/install-ik-llama.sh

@@ -0,0 +1,48 @@
+#!/bin/bash
+# Install ik_llama.cpp - clone, build, and install binaries
+# Usage: ./install-ik-llama.sh <commit_hash>
+# Note: CUDA only; always built against builder-base-cuda
+set -e
+
+COMMIT_HASH="${1:-main}"
+
+mkdir -p /install/bin
+
+# Clone and checkout (init-based so cache-mounted build dir doesn't break clone)
+echo "=== Cloning ik_llama.cpp at ${COMMIT_HASH} ==="
+mkdir -p /src/ik_llama.cpp
+cd /src/ik_llama.cpp
+if [ ! -d .git ]; then
+    git init
+    git remote add origin https://github.com/ikawrakow/ik_llama.cpp.git
+fi
+git fetch --depth=1 origin "${COMMIT_HASH}"
+git checkout FETCH_HEAD
+
+CMAKE_FLAGS=(
+    -DGGML_NATIVE=OFF
+    -DBUILD_SHARED_LIBS=OFF
+    -DCMAKE_BUILD_TYPE=Release
+    -DCMAKE_C_COMPILER_LAUNCHER=ccache
+    -DCMAKE_CXX_COMPILER_LAUNCHER=ccache
+    -DGGML_CUDA=ON
+    "-DCMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES:?CMAKE_CUDA_ARCHITECTURES must be set}"
+    "-DCMAKE_CUDA_FLAGS=-allow-unsupported-compiler"
+    "-DCMAKE_EXE_LINKER_FLAGS=-Wl,-rpath-link,/usr/local/cuda/lib64/stubs -lcuda -Wl,--allow-shlib-undefined"
+)
+
+rm -rf build/CMakeCache.txt build/CMakeFiles 2>/dev/null || true
+
+echo "=== Building ik_llama.cpp ==="
+cmake -B build "${CMAKE_FLAGS[@]}"
+cmake --build build --config Release -j"$(nproc)" --target llama-server
+
+if [ ! -f "build/bin/llama-server" ]; then
+    echo "FATAL: llama-server not found in build/bin/" >&2
+    exit 1
+fi
+
+# Install as ik-llama-server to avoid collision with llama.cpp's llama-server
+cp "build/bin/llama-server" "/install/bin/ik-llama-server"
+echo "=== ik_llama.cpp build complete ==="
+ls -la /install/bin/

docker/unified/install-llama-swap.sh

@@ -0,0 +1,59 @@
+#!/bin/bash
+# Install llama-swap - download latest release binary from GitHub
+# Usage: ./install-llama-swap.sh [version]
+#   version: release version number (e.g., "170") or "latest" (default)
+set -e
+
+VERSION="${1:-latest}"
+REPO="mostlygeek/llama-swap"
+
+mkdir -p /install/bin
+
+# If a full commit hash is given, find the release tag that points to it
+if echo "${VERSION}" | grep -qE '^[0-9a-f]{40}$'; then
+    echo "=== Resolving commit ${VERSION:0:7} to release tag ==="
+    TAG=$(git ls-remote --tags "https://github.com/${REPO}.git" 2>/dev/null \
+        | grep "^${VERSION}" | sed 's|.*refs/tags/||' | grep -v '\^{}' | head -1)
+    if [ -n "${TAG}" ]; then
+        echo "Resolved to tag: ${TAG}"
+        VERSION="${TAG#v}"
+    else
+        echo "No release tag found for commit ${VERSION:0:7}, using latest"
+        VERSION="latest"
+    fi
+fi
+
+# Strip leading 'v' prefix so both "198" and "v198" work
+VERSION="${VERSION#v}"
+
+# Resolve "latest" to actual version number
+if [ "$VERSION" = "latest" ]; then
+    echo "=== Resolving latest llama-swap release ==="
+    VERSION=$(curl -fsSL "https://api.github.com/repos/${REPO}/releases/latest" \
+        | grep '"tag_name"' | head -1 | cut -d'"' -f4 | sed 's/^v//')
+    if [ -z "$VERSION" ]; then
+        echo "FATAL: Could not determine latest release version" >&2
+        exit 1
+    fi
+    echo "Latest version: ${VERSION}"
+fi
+
+# Download and extract
+URL="https://github.com/${REPO}/releases/download/v${VERSION}/llama-swap_${VERSION}_linux_amd64.tar.gz"
+echo "=== Downloading llama-swap v${VERSION} ==="
+echo "URL: $URL"
+curl -fSL -o /tmp/llama-swap.tar.gz "$URL"
+tar -xzf /tmp/llama-swap.tar.gz -C /install/bin/
+rm /tmp/llama-swap.tar.gz
+
+# Validate
+if [ ! -x "/install/bin/llama-swap" ]; then
+    echo "FATAL: llama-swap binary not found or not executable" >&2
+    ls -la /install/bin/ >&2
+    exit 1
+fi
+
+echo "$VERSION" > /install/llama-swap-version
+
+echo "=== llama-swap v${VERSION} installed ==="
+ls -la /install/bin/llama-swap

docker/unified/install-llama.sh

@@ -0,0 +1,63 @@
+#!/bin/bash
+# Install llama.cpp - clone, build, and install binaries
+# Usage: BACKEND=cuda|vulkan ./install-llama.sh <commit_hash>
+set -e
+
+COMMIT_HASH="${1:-master}"
+BACKEND="${BACKEND:-cuda}"
+
+mkdir -p /install/bin
+
+# Clone and checkout (init-based so cache-mounted /src/llama.cpp/build dir doesn't break clone)
+echo "=== Cloning llama.cpp at ${COMMIT_HASH} ==="
+mkdir -p /src/llama.cpp
+cd /src/llama.cpp
+if [ ! -d .git ]; then
+    git init
+    git remote add origin https://github.com/ggml-org/llama.cpp.git
+fi
+git fetch --depth=1 origin "${COMMIT_HASH}"
+git checkout FETCH_HEAD
+
+# Common cmake flags
+CMAKE_FLAGS=(
+    -DGGML_NATIVE=OFF
+    -DBUILD_SHARED_LIBS=OFF
+    -DCMAKE_BUILD_TYPE=Release
+    -DCMAKE_C_COMPILER_LAUNCHER=ccache
+    -DCMAKE_CXX_COMPILER_LAUNCHER=ccache
+    -DLLAMA_BUILD_TESTS=OFF
+)
+
+if [ "$BACKEND" = "cuda" ]; then
+    CMAKE_FLAGS+=(
+        -DGGML_CUDA=ON
+        -DGGML_VULKAN=OFF
+        "-DCMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES:?CMAKE_CUDA_ARCHITECTURES must be set}"
+        "-DCMAKE_CUDA_FLAGS=-allow-unsupported-compiler"
+        "-DCMAKE_EXE_LINKER_FLAGS=-Wl,-rpath-link,/usr/local/cuda/lib64/stubs -lcuda"
+    )
+elif [ "$BACKEND" = "vulkan" ]; then
+    CMAKE_FLAGS+=(
+        -DGGML_CUDA=OFF
+        -DGGML_VULKAN=ON
+    )
+fi
+
+TARGETS=(llama-cli llama-server)
+
+rm -rf build/CMakeCache.txt build/CMakeFiles 2>/dev/null || true
+
+echo "=== Building llama.cpp for ${BACKEND} ==="
+cmake -B build "${CMAKE_FLAGS[@]}"
+cmake --build build --config Release -j"$(nproc)" --target "${TARGETS[@]}"
+
+for bin in "${TARGETS[@]}"; do
+    if [ ! -f "build/bin/$bin" ]; then
+        echo "FATAL: $bin not found in build/bin/" >&2
+        exit 1
+    fi
+    cp "build/bin/$bin" "/install/bin/"
+done
+echo "=== llama.cpp build complete ==="
+ls -la /install/bin/

docker/unified/install-sd.sh

@@ -0,0 +1,68 @@
+#!/bin/bash
+# Install stable-diffusion.cpp - clone, build, and install binaries and library
+# Usage: BACKEND=cuda|vulkan ./install-sd.sh <commit_hash>
+set -e
+
+COMMIT_HASH="${1:-master}"
+BACKEND="${BACKEND:-cuda}"
+
+mkdir -p /install/bin /install/lib
+
+# Clone and checkout (init-based so cache-mounted /src/stable-diffusion.cpp/build dir doesn't break clone)
+echo "=== Cloning stable-diffusion.cpp at ${COMMIT_HASH} ==="
+mkdir -p /src/stable-diffusion.cpp
+cd /src/stable-diffusion.cpp
+if [ ! -d .git ]; then
+    git init
+    git remote add origin https://github.com/leejet/stable-diffusion.cpp.git
+fi
+git fetch --depth=1 origin "${COMMIT_HASH}"
+git checkout FETCH_HEAD
+git submodule update --init --recursive --depth=1
+
+# Common cmake flags
+CMAKE_FLAGS=(
+    -DGGML_NATIVE=OFF
+    -DCMAKE_BUILD_TYPE=Release
+    -DCMAKE_C_COMPILER_LAUNCHER=ccache
+    -DCMAKE_CXX_COMPILER_LAUNCHER=ccache
+    -DSD_BUILD_EXAMPLES=ON
+)
+
+if [ "$BACKEND" = "cuda" ]; then
+    CMAKE_FLAGS+=(
+        -DGGML_CUDA=ON
+        -DGGML_VULKAN=OFF
+        "-DCMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES:?CMAKE_CUDA_ARCHITECTURES must be set}"
+        "-DCMAKE_CUDA_FLAGS=-allow-unsupported-compiler"
+        "-DCMAKE_EXE_LINKER_FLAGS=-Wl,-rpath-link,/usr/local/cuda/lib64/stubs -lcuda"
+        "-DCMAKE_SHARED_LINKER_FLAGS=-Wl,-rpath-link,/usr/local/cuda/lib64/stubs -lcuda"
+        -DSD_CUDA=ON
+    )
+elif [ "$BACKEND" = "vulkan" ]; then
+    CMAKE_FLAGS+=(
+        -DGGML_CUDA=OFF
+        -DGGML_VULKAN=ON
+        -DSD_VULKAN=ON
+    )
+fi
+
+TARGETS=(stable-diffusion sd-cli sd-server)
+
+rm -rf build/CMakeCache.txt build/CMakeFiles 2>/dev/null || true
+
+echo "=== Building stable-diffusion.cpp for ${BACKEND} ==="
+cmake -B build "${CMAKE_FLAGS[@]}"
+cmake --build build --config Release -j"$(nproc)" --target "${TARGETS[@]}"
+
+for bin in sd-cli sd-server; do
+    if [ ! -f "build/bin/$bin" ]; then
+        echo "FATAL: $bin not found in build/bin/" >&2
+        exit 1
+    fi
+    cp "build/bin/$bin" "/install/bin/"
+done
+find build -name "*.so*" -type f -exec cp {} /install/lib/ \;
+
+echo "=== stable-diffusion.cpp build complete ==="
+ls -la /install/bin/ /install/lib/

docker/unified/install-whisper.sh

@@ -0,0 +1,64 @@
+#!/bin/bash
+# Install whisper.cpp - clone, build, and install binaries
+# Usage: BACKEND=cuda|vulkan ./install-whisper.sh <commit_hash>
+set -e
+
+COMMIT_HASH="${1:-master}"
+BACKEND="${BACKEND:-cuda}"
+
+mkdir -p /install/bin /install/lib
+
+# Clone and checkout (init-based so cache-mounted /src/whisper.cpp/build dir doesn't break clone)
+echo "=== Cloning whisper.cpp at ${COMMIT_HASH} ==="
+mkdir -p /src/whisper.cpp
+cd /src/whisper.cpp
+if [ ! -d .git ]; then
+    git init
+    git remote add origin https://github.com/ggml-org/whisper.cpp.git
+fi
+git fetch --depth=1 origin "${COMMIT_HASH}"
+git checkout FETCH_HEAD
+
+# Common cmake flags
+CMAKE_FLAGS=(
+    -DGGML_NATIVE=OFF
+    -DCMAKE_BUILD_TYPE=Release
+    -DCMAKE_C_COMPILER_LAUNCHER=ccache
+    -DCMAKE_CXX_COMPILER_LAUNCHER=ccache
+)
+
+if [ "$BACKEND" = "cuda" ]; then
+    CMAKE_FLAGS+=(
+        -DGGML_CUDA=ON
+        -DGGML_VULKAN=OFF
+        "-DCMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES:?CMAKE_CUDA_ARCHITECTURES must be set}"
+        "-DCMAKE_CUDA_FLAGS=-allow-unsupported-compiler"
+        "-DCMAKE_EXE_LINKER_FLAGS=-Wl,-rpath-link,/usr/local/cuda/lib64/stubs -lcuda"
+        "-DCMAKE_SHARED_LINKER_FLAGS=-Wl,-rpath-link,/usr/local/cuda/lib64/stubs -lcuda"
+    )
+elif [ "$BACKEND" = "vulkan" ]; then
+    CMAKE_FLAGS+=(
+        -DGGML_CUDA=OFF
+        -DGGML_VULKAN=ON
+    )
+fi
+
+TARGETS=(whisper-cli whisper-server)
+
+rm -rf build/CMakeCache.txt build/CMakeFiles 2>/dev/null || true
+
+echo "=== Building whisper.cpp for ${BACKEND} ==="
+cmake -B build "${CMAKE_FLAGS[@]}"
+cmake --build build --config Release -j"$(nproc)" --target "${TARGETS[@]}"
+
+for bin in "${TARGETS[@]}"; do
+    if [ ! -f "build/bin/$bin" ]; then
+        echo "FATAL: $bin not found in build/bin/" >&2
+        exit 1
+    fi
+    cp "build/bin/$bin" "/install/bin/"
+done
+find build -name "*.so*" -type f -exec cp {} /install/lib/ \;
+
+echo "=== whisper.cpp build complete ==="
+ls -la /install/bin/

docs/configuration.md

@@ -22,7 +22,7 @@ models:
     cmd: llama-server --port ${PORT} -m /path/to/third_model.gguf
 ```
 
-With this configuration models will be hot swapped and loaded on demand. The special `${PORT}` macro provides a unique port per model. Useful if you want to run multiple models at the same time with the `groups` feature.
+With this configuration models will be hot swapped and loaded on demand. The special `${PORT}` macro provides a unique port per model which is useful if you want to run multiple models at the same time with the `matrix` feature.
 
 ## Advanced control with `cmd`
 
@@ -76,7 +76,7 @@ llama-swap supports many more features to customize how you want to manage your
 | --------- | ---------------------------------------------- |
 | `ttl`     | automatic unloading of models after a timeout  |
 | `macros`  | reusable snippets to use in configurations     |
-| `groups`  | run multiple models at a time                  |
+| `matrix`  | run multiple models at a time                  |
 | `hooks`   | event driven functionality                     |
 | `env`     | define environment variables per model         |
 | `aliases` | serve a model with different names             |
@@ -141,6 +141,11 @@ logToStdout: "proxy"
 # - useful for limiting memory usage when processing large volumes of metrics
 metricsMaxInMemory: 1000
 
+# captureBuffer: how many MBs to allocate for storing request/response captures
+# - optional, default: 10
+# - set to 0 to disable
+captureBuffer: 15
+
 # startPort: sets the starting port number for the automatic ${PORT} macro.
 # - optional, default: 5800
 # - the ${PORT} macro can be used in model.cmd and model.proxy settings
@@ -161,15 +166,10 @@ sendLoadingState: true
 #   all fields except for Id so chat UIs can use the alias equivalent to the original.
 includeAliasesInList: false
 
-# apiKeys: require an API key when making requests to inference endpoints
-# - optional, default: []
-# - when empty (the default) authorization will not be checked as llama-swap is default-allow
-# - each key is a non-empty string
-apiKeys:
-  - "sk-hunter2"
-  # hint, one liner: printf "sk-%s\n" "$(head -c 48 /dev/urandom | base64 )"
-  - "sk-gyCPiKUcIfPlaM4OSMZekkprgijPx6+OsmQs8Rsg0xZ9qpy6gKWsIKqHOk+cgXVx"
-  - "sk-+QtIn0Zjj4UHjiaZYiZEnru4mrwKM9RzhmJeK5SobNXLl8QMFXxGz1/2lEuvQpkb"
+# globalTTL: the default TTL in seconds before unloading a model
+# - optional, default: 0 (never automatically unload)
+# - must be >= 0
+globalTTL: 0
 
 # macros: a dictionary of string substitutions
 # - optional, default: empty dictionary
@@ -181,6 +181,9 @@ apiKeys:
 # - macro names must not be a reserved name: PORT or MODEL_ID
 # - macro values can be numbers, bools, or strings
 # - macros can contain other macros, but they must be defined before they are used
+# - environment variables can be referenced with ${env.VAR_NAME} syntax
+#   - env macros are substituted first, before regular macros
+#   - if the env var is not set, config loading will fail with an error
 macros:
   # Example of a multi-line macro
   "latest-llama": >
@@ -193,6 +196,24 @@ macros:
   # but they must be previously declared.
   "default_args": "--ctx-size ${default_ctx}"
 
+  # Example of environment variable macros
+  # - ${env.VAR_NAME} pulls the value from the system environment
+  # - useful for paths, secrets, or machine-specific configuration
+  "models_dir": "${env.HOME}/models"
+
+# apiKeys: require an API key when making requests to inference endpoints
+# - optional, default: []
+# - when empty (the default) authorization will not be checked as llama-swap is default-allow
+# - each key is a non-empty string
+apiKeys:
+  - "sk-hunter2"
+  # tip, one liner: printf "sk-%s\n" "$(head -c 48 /dev/urandom | base64 )"
+  - "sk-gyCPiKUcIfPlaM4OSMZekkprgijPx6+OsmQs8Rsg0xZ9qpy6gKWsIKqHOk+cgXVx"
+
+  # use environment variable macros to keep secrets out of the config
+  - "${env.API_KEY_1}"
+  - "${env.API_KEY_2}"
+
 # models: a dictionary of model configurations
 # - required
 # - each key is the model's ID, used in API requests
@@ -201,7 +222,7 @@ macros:
 # - below are examples of the all the settings a model can have
 models:
   # keys are the model names used in API requests
-  "llama":
+  "gpt-oss-120b":
     # macros: a dictionary of string substitutions specific to this model
     # - optional, default: empty dictionary
     # - macros defined here override macros defined in the global macros section
@@ -218,7 +239,7 @@ models:
     cmd: |
       # ${latest-llama} is a macro that is defined above
       ${latest-llama}
-      --model path/to/llama-8B-Q4_K_M.gguf
+      --model path/to/gpt-oss-120B.gguf
       --ctx-size ${default_ctx}
       --temperature ${temp}
 
@@ -226,13 +247,13 @@ models:
     # - optional, default: empty string
     # - if set, it will be used in the v1/models API response
     # - if not set, it will be omitted in the JSON model record
-    name: "llama 3.1 8B"
+    name: "gpt-oss 120B"
 
     # description: a description for the model
     # - optional, default: empty string
     # - if set, it will be used in the v1/models API response
     # - if not set, it will be omitted in the JSON model record
-    description: "A small but capable model used for quick testing"
+    description: "A thinking model from OpenAI"
 
     # env: define an array of environment variables to inject into cmd's environment
     # - optional, default: empty array
@@ -247,14 +268,6 @@ models:
     # - if you use a custom port in cmd this *must* be set
     proxy: http://127.0.0.1:8999
 
-    # aliases: alternative model names that this model configuration is used for
-    # - optional, default: empty array
-    # - aliases must be unique globally
-    # - useful for impersonating a specific model
-    aliases:
-      - "gpt-4o-mini"
-      - "gpt-3.5-turbo"
-
     # checkEndpoint: URL path to check if the server is ready
     # - optional, default: /health
     # - endpoint is expected to return an HTTP 200 response
@@ -263,8 +276,10 @@ models:
     checkEndpoint: /custom-endpoint
 
     # ttl: automatically unload the model after ttl seconds
-    # - optional, default: 0
-    # - ttl values must be a value greater than 0
+    # - optional, default: -1 (use global default)
+    # - ttl values must be a value greater than or equal to 0
+    # - a ttl of -1 will use the global TTL value as the default
+    # - a ttl of 0 will mean never unload
     # - a value of 0 disables automatic unloading of the model
     ttl: 60
 
@@ -272,11 +287,11 @@ models:
     # - optional, default: ""
     # - useful for when the upstream server expects a specific model name that
     #   is different from the model's ID
-    useModelName: "qwen:qwq"
+    useModelName: "openai/gpt-oss-120B"
 
     # filters: a dictionary of filter settings
     # - optional, default: empty dictionary
-    # - only stripParams is currently supported
+    # - same capabilities as peer filters (stripParams, setParams)
     filters:
       # stripParams: a comma separated list of parameters to remove from the request
       # - optional, default: ""
@@ -286,6 +301,43 @@ models:
       # - recommended to stick to sampling parameters
       stripParams: "temperature, top_p, top_k"
 
+      # setParams: a dictionary of parameters to set/override in requests
+      # - optional, default: empty dictionary
+      # - useful for enforcing specific parameter values
+      # - protected params like "model" cannot be overridden
+      # - values can be strings, numbers, booleans, arrays, or objects
+      # - always runs for the model
+      setParams:
+        # Example: enforce specific sampling parameters
+        temperature: 0.7
+        top_p: 0.9
+
+      # setParamsByID: a dictionary of parameters to set based the model ID
+      # - optional, default: empty dictionary
+      # - combine with aliases to create variant behaviour without reloading the model
+      # - parameters are set in the request body JSON
+      # - run after setParams so it will override any settings
+      # - protected params like "model" cannot be overridden
+      # - values can be strings, numbers, booleans, arrays, or objects
+      # - model aliases will be automatically created for each key
+      setParamsByID:
+        "${MODEL_ID}":
+          chat_template_kwargs:
+            reasoning_effort: medium
+        "${MODEL_ID}:high":
+          chat_template_kwargs:
+            reasoning_effort: high
+        "${MODEL_ID}:low":
+          chat_template_kwargs:
+            reasoning_effort: low
+
+    # aliases: alternative model names that this model configuration is used for
+    # - optional, default: empty array
+    # - aliases must be unique globally
+    # - useful for impersonating a specific model
+    aliases:
+      - "gpt-4o-mini"
+
     # metadata: a dictionary of arbitrary values that are included in /v1/models
     # - optional, default: empty dictionary
     # - while metadata can contains complex types it is recommended to keep it simple
@@ -323,6 +375,22 @@ models:
     # - optional, default: undefined (use global setting)
     sendLoadingState: false
 
+    # timeouts: configure proxy connection timeouts for this model
+    # - optional, defaults shown below
+    # - useful for models running on slower hardware that need longer timeouts
+    # - connect: TCP dial connection timeout in seconds, default: 30 seconds
+    # - keepalive: TCP connection keepalive timeout, default: 30 seconds
+    # - responseHeader: time to wait for response headers in seconds, default: 0 (no timeout)
+    # - tlsHandshake: TLS handshake timeout in seconds, default: 10 seconds
+    # - idleConn: idle connection timeout in seconds, default: 90 seconds
+    # - set any value to 0 to disable that timeout (not recommended)
+    timeouts:
+      connect: 30
+      keepalive: 0
+      responseHeader: 60
+      tlsHandshake: 10
+      idleConn: 90
+
   # Unlisted model example:
   "qwen-unlisted":
     # unlisted: boolean, true or false
@@ -354,68 +422,83 @@ models:
     # - processes have 5 seconds to shutdown until forceful termination is attempted
     cmdStop: docker stop ${MODEL_ID}
 
-# groups: a dictionary of group settings
-# - optional, default: empty dictionary
-# - provides advanced controls over model swapping behaviour
-# - using groups some models can be kept loaded indefinitely, while others are swapped out
-# - model IDs must be defined in the Models section
-# - a model can only be a member of one group
-# - group behaviour is controlled via the `swap`, `exclusive` and `persistent` fields
-# - see issue #109 for details
+# =============================================================================
+# matrix: run concurrent models with a solver-based swap DSL
+# =============================================================================
 #
-# NOTE: the example below uses model names that are not defined above for demonstration purposes
-groups:
-  # group1 works the same as the default behaviour of llama-swap where only one model is allowed
-  # to run a time across the whole llama-swap instance
-  "group1":
-    # swap: controls the model swapping behaviour in within the group
-    # - optional, default: true
-    # - true : only one model is allowed to run at a time
-    # - false: all models can run together, no swapping
-    swap: true
+# Note:
+# A config must use either a matrix or legacy groups, not both. A configuration error
+# will occur if both are defined. Configuration examples for legacy Groups can be found:
+# https://github.com/mostlygeek/llama-swap/blob/40e39f7/config.example.yaml#L334-L396
+#
+# The matrix declares valid combinations of models that can run concurrently.
+# When a model is requested, the solver finds the cheapest way to make it
+# available by evicting as few (and least costly) running models as possible.
+#
+# Solver behavior:
+#   1. Request arrives for model X
+#   2. If X is already running, forward immediately. Done.
+#   3. Find all sets containing X
+#   4. For each candidate set, compute cost: sum of evict_costs for
+#      every running model NOT in that set
+#   5. Pick lowest cost candidate. Ties broken by definition order.
+#   6. Evict what needs to stop. Start X. Forward request.
+#
+# Subset semantics: a set [a, b, c] means any subset is valid.
+# Only the requested model is started — others are not preloaded.
+#
+# A model not appearing in any set can only run alone.
+#
+matrix:
+  # vars: short names for models (alphanumeric, 1-8 chars)
+  # - required for sets and evict_costs settings
+  # - each entry is a short name to a real model ID. Do not use an alias
+  # - used to keep set DSL logic short and easier to read
+  # - sets and evict_costs only use identifiers defined in vars
+  vars:
+    g: gemma-model
+    q: qwen-model
+    m: mistral-model
+    v: voxtral-model
+    e: reranker-model
+    L: llama-70B
+    sd: stable-diffusion
 
-    # exclusive: controls how the group affects other groups
-    # - optional, default: true
-    # - true: causes all other groups to unload when this group runs a model
-    # - false: does not affect other groups
-    exclusive: true
+  # evict_costs: relative cost of losing a running model (default: 1)
+  evict_costs:
+    v: 50 # vllm backend, slow cold start
+    L: 30 # 70B weights, slow to load
 
-    # members references the models defined above
-    # required
-    members:
-      - "llama"
-      - "qwen-unlisted"
+  # sets: named sets of concurrent model combinations
+  # Values are DSL strings with operators:
+  #   &     AND (models run together)
+  #   |     OR  (alternatives)
+  #   ()    grouping
+  #   +ref  inline another set's expression
+  #
+  # Expansion examples:
+  #   "L"                  → [L]
+  #   "a & b"              → [a, b]
+  #   "a | b"              → [a], [b]
+  #   "(a | b) & c"        → [a, c], [b, c]
+  #   "(a | b) & (c | d)"  → [a,c], [a,d], [b,c], [b,d]
+  #   "+llms & v"          → expands llms inline, then applies & v
+  sets:
+    # LLM + TTS: switching between g/q/m won't evict v
+    # expands to: [g,v], [q,v], [m,v]
+    standard: "(g | q | m) & v"
 
-  # Example:
-  # - in group2 all models can run at the same time
-  # - when a different group is loaded it causes all running models in this group to unload
-  "group2":
-    swap: false
+    # LLM + TTS + reranker
+    # expands to: [g,v,e], [q,v,e]
+    with_rerank: "(g | q) & v & e"
 
-    # exclusive: false does not unload other groups when a model in group2 is requested
-    # - the models in group2 will be loaded but will not unload any other groups
-    exclusive: false
-    members:
-      - "docker-llama"
-      - "modelA"
-      - "modelB"
+    # LLM + image generation, no TTS
+    # expands to: [g,sd], [q,sd]
+    creative: "(g | q) & sd"
 
-  # Example:
-  # - a persistent group, prevents other groups from unloading it
-  "forever":
-    # persistent: prevents over groups from unloading the models in this group
-    # - optional, default: false
-    # - does not affect individual model behaviour
-    persistent: true
-
-    # set swap/exclusive to false to prevent swapping inside the group
-    # and the unloading of other groups
-    swap: false
-    exclusive: false
-    members:
-      - "forever-modelA"
-      - "forever-modelB"
-      - "forever-modelc"
+    # 70B model uses all GPUs, can only run alone
+    # expands to: [L]
+    full: "L"
 
 # hooks: a dictionary of event triggers and actions
 # - optional, default: empty dictionary
@@ -456,7 +539,8 @@ peers:
     # - optional, default: ""
     # - if blank, no key will be added to the request
     # - key will be injected into headers: Authorization: Bearer <key> and x-api-key: <key>
-    apiKey: sk-your-openrouter-key
+    # - can be a string or a macro
+    apiKey: ${env.OPENROUTER_API_KEY}
     models:
       - meta-llama/llama-3.1-8b-instruct
       - qwen/qwen3-235b-a22b-2507
@@ -464,4 +548,35 @@ peers:
       - z-ai/glm-4.7
       - moonshotai/kimi-k2-0905
       - minimax/minimax-m2.1
+    # timeouts: configure proxy connection timeouts for this peer
+    # - optional, defaults shown below
+    # - useful when the peer runs on slower hardware
+    # - set any value to 0 to disable that timeout (not recommended)
+    timeouts:
+      connect: 30
+      keepalive: 30
+      responseHeader: 60
+      tlsHandshake: 10
+      idleConn: 90
+
+    # filters: a dictionary of filter settings for peer requests
+    # - optional, default: empty dictionary
+    # - same capabilities as model filters (stripParams, setParams)
+    filters:
+      # stripParams: a comma separated list of parameters to remove from the request
+      # - optional, default: ""
+      # - useful for removing parameters that the peer doesn't support
+      # - the `model` parameter can never be removed
+      stripParams: "temperature, top_p"
+
+      # setParams: a dictionary of parameters to set/override in requests to this peer
+      # - optional, default: empty dictionary
+      # - useful for injecting provider-specific settings like data retention policies
+      # - protected params like "model" cannot be overridden
+      # - values can be strings, numbers, booleans, arrays, or objects
+      setParams:
+        # Example: enforce zero-data-retention for OpenRouter
+        provider:
+          data_collection: "deny"
+          zdr: true
 ```

go.mod

@@ -1,11 +1,11 @@
 module github.com/mostlygeek/llama-swap
 
-go 1.25.4
+go 1.26.1
 
 require (
 	github.com/billziss-gh/golib v0.2.0
-	github.com/fsnotify/fsnotify v1.9.0
 	github.com/gin-gonic/gin v1.10.0
+	github.com/klauspost/compress v1.18.5
 	github.com/stretchr/testify v1.9.0
 	github.com/tidwall/gjson v1.18.0
 	github.com/tidwall/sjson v1.2.5

go.sum

@@ -11,8 +11,6 @@ github.com/cloudwego/iasm v0.2.0/go.mod h1:8rXZaNYT2n95jn+zTI1sDr+IgcD2GVs0nlbbQ
 github.com/davecgh/go-spew v1.1.0/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38=
 github.com/davecgh/go-spew v1.1.1 h1:vj9j/u1bqnvCEfJOwUhtlOARqs3+rkHYY13jYWTU97c=
 github.com/davecgh/go-spew v1.1.1/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38=
-github.com/fsnotify/fsnotify v1.9.0 h1:2Ml+OJNzbYCTzsxtv8vKSFD9PbJjmhYF14k/jKC7S9k=
-github.com/fsnotify/fsnotify v1.9.0/go.mod h1:8jBTzvmWwFyi3Pb8djgCCO5IBqzKJ/Jwo8TRcHyHii0=
 github.com/gabriel-vasile/mimetype v1.4.3 h1:in2uUcidCuFcDKtdcBxlR0rJ1+fsokWf+uqxgUFjbI0=
 github.com/gabriel-vasile/mimetype v1.4.3/go.mod h1:d8uq/6HKRL6CGdk+aubisF/M5GcPfT7nKyLpA0lbSSk=
 github.com/gin-contrib/sse v0.1.0 h1:Y/yl/+YNO8GZSjAhjMsSuLt29uWRFHdHYUb5lYOV9qE=
@@ -34,6 +32,8 @@ github.com/google/go-cmp v0.5.5/go.mod h1:v8dTdLbMG2kIc/vJvl+f65V22dbkXbowE6jgT/
 github.com/google/gofuzz v1.0.0/go.mod h1:dBl0BpW6vV/+mYPU4Po3pmUjxk6FQPldtuIdl/M65Eg=
 github.com/json-iterator/go v1.1.12 h1:PV8peI4a0ysnczrg+LtxykD8LfKY9ML6u2jnxaEnrnM=
 github.com/json-iterator/go v1.1.12/go.mod h1:e30LSqwooZae/UwlEbR2852Gd8hjQvJoHmT4TnhNGBo=
+github.com/klauspost/compress v1.18.5 h1:/h1gH5Ce+VWNLSWqPzOVn6XBO+vJbCNGvjoaGBFW2IE=
+github.com/klauspost/compress v1.18.5/go.mod h1:cwPg85FWrGar70rWktvGQj8/hthj3wpl0PGDogxkrSQ=
 github.com/klauspost/cpuid/v2 v2.0.9/go.mod h1:FInQzS24/EEf25PyTYn52gqo7WaD8xa0213Md/qVLRg=
 github.com/klauspost/cpuid/v2 v2.2.7 h1:ZWSB3igEs+d0qvnxR/ZBzXVmxkgt8DdzP6m9pfuVLDM=
 github.com/klauspost/cpuid/v2 v2.2.7/go.mod h1:Lcz8mBdAVJIBVzewtcLocK12l3Y+JytZYpaMropDUws=

llama-swap.go

@@ -9,14 +9,15 @@ import (
 	"os"
 	"os/signal"
 	"path/filepath"
+	"runtime"
 	"syscall"
 	"time"
 
-	"github.com/fsnotify/fsnotify"
 	"github.com/gin-gonic/gin"
 	"github.com/mostlygeek/llama-swap/event"
 	"github.com/mostlygeek/llama-swap/proxy"
 	"github.com/mostlygeek/llama-swap/proxy/config"
+	"github.com/mostlygeek/llama-swap/proxy/configwatcher"
 )
 
 var (
@@ -79,6 +80,17 @@ func main() {
 	sigChan := make(chan os.Signal, 1)
 	signal.Notify(sigChan, syscall.SIGINT, syscall.SIGTERM)
 
+	// Reload signals (SIGHUP on POSIX, none on Windows — Windows does not
+	// deliver SIGHUP). Always wired up so `kill -HUP` works regardless of
+	// --watch-config.
+	reloadChan := make(chan os.Signal, 1)
+	if runtime.GOOS != "windows" {
+		signal.Notify(reloadChan, syscall.SIGHUP)
+	}
+
+	// Context that bounds the lifetime of background watcher goroutines.
+	watcherCtx, watcherCancel := context.WithCancel(context.Background())
+
 	// Create server with initial handler
 	srv := &http.Server{
 		Addr: *listenStr,
@@ -121,52 +133,45 @@ func main() {
 	// load the initial proxy manager
 	reloadProxyManager()
 	debouncedReload := debounce(time.Second, reloadProxyManager)
-	if *watchConfig {
-		defer event.On(func(e proxy.ConfigFileChangedEvent) {
-			if e.ReloadingState == proxy.ReloadingStateStart {
-				debouncedReload()
-			}
-		})()
 
-		fmt.Println("Watching Configuration for changes")
+	// Listen for ConfigFileChangedEvent unconditionally so SIGHUP and the
+	// poll-based watcher both feed the same debounced reload pipeline. The
+	// UI also listens for the matching ReloadingStateEnd emitted from
+	// reloadProxyManager.
+	defer event.On(func(e proxy.ConfigFileChangedEvent) {
+		if e.ReloadingState == proxy.ReloadingStateStart {
+			debouncedReload()
+		}
+	})()
+
+	// SIGHUP (or platform-equivalent) → reload. Back-to-back signals collapse
+	// to one reload via the debounce window, which is the desired behavior.
+	go func() {
+		for range reloadChan {
+			fmt.Println("Received reload signal, reloading configuration")
+			event.Emit(proxy.ConfigFileChangedEvent{
+				ReloadingState: proxy.ReloadingStateStart,
+			})
+		}
+	}()
+
+	if *watchConfig {
 		go func() {
 			absConfigPath, err := filepath.Abs(*configPath)
 			if err != nil {
 				fmt.Printf("Error getting absolute path for watching config file: %v\n", err)
 				return
 			}
-			watcher, err := fsnotify.NewWatcher()
-			if err != nil {
-				fmt.Printf("Error creating file watcher: %v. File watching disabled.\n", err)
-				return
-			}
-
-			configDir := filepath.Dir(absConfigPath)
-			err = watcher.Add(configDir)
-			if err != nil {
-				fmt.Printf("Error adding config path directory (%s) to watcher: %v. File watching disabled.", configDir, err)
-				return
-			}
-
-			defer watcher.Close()
-			for {
-				select {
-				case changeEvent := <-watcher.Events:
-					if changeEvent.Name == absConfigPath && (changeEvent.Has(fsnotify.Write) || changeEvent.Has(fsnotify.Create) || changeEvent.Has(fsnotify.Remove)) {
-						event.Emit(proxy.ConfigFileChangedEvent{
-							ReloadingState: proxy.ReloadingStateStart,
-						})
-					} else if changeEvent.Name == filepath.Join(configDir, "..data") && changeEvent.Has(fsnotify.Create) {
-						// the change for k8s configmap
-						event.Emit(proxy.ConfigFileChangedEvent{
-							ReloadingState: proxy.ReloadingStateStart,
-						})
-					}
-
-				case err := <-watcher.Errors:
-					log.Printf("File watcher error: %v", err)
-				}
-			}
+			fmt.Println("Watching configuration for changes (poll-based, 2s interval)")
+			(&configwatcher.Watcher{
+				Path:     absConfigPath,
+				Interval: configwatcher.DefaultInterval,
+				OnChange: func() {
+					event.Emit(proxy.ConfigFileChangedEvent{
+						ReloadingState: proxy.ReloadingStateStart,
+					})
+				},
+			}).Run(watcherCtx)
 		}()
 	}
 
@@ -174,6 +179,7 @@ func main() {
 	go func() {
 		sig := <-sigChan
 		fmt.Printf("Received signal %v, shutting down...\n", sig)
+		watcherCancel()
 		ctx, cancel := context.WithTimeout(context.Background(), time.Second*5)
 		defer cancel()
 

proxy/config/config.go

@@ -87,6 +87,7 @@ type GroupConfig struct {
 var (
 	macroNameRegex    = regexp.MustCompile(`^[a-zA-Z0-9_-]+$`)
 	macroPatternRegex = regexp.MustCompile(`\$\{([a-zA-Z0-9_-]+)\}`)
+	envMacroRegex     = regexp.MustCompile(`\$\{env\.([a-zA-Z_][a-zA-Z0-9_]*)\}`)
 )
 
 // set default values for GroupConfig
@@ -122,10 +123,18 @@ type Config struct {
 	LogTimeFormat      string                 `yaml:"logTimeFormat"`
 	LogToStdout        string                 `yaml:"logToStdout"`
 	MetricsMaxInMemory int                    `yaml:"metricsMaxInMemory"`
+	CaptureBuffer      int                    `yaml:"captureBuffer"`
+	GlobalTTL          int                    `yaml:"globalTTL"`
 	Models             map[string]ModelConfig `yaml:"models"` /* key is model ID */
 	Profiles           map[string][]string    `yaml:"profiles"`
 	Groups             map[string]GroupConfig `yaml:"groups"` /* key is group ID */
 
+	// swap matrix: solver-based alternative to groups
+	Matrix *MatrixConfig `yaml:"matrix"`
+
+	// populated during validation when matrix is configured
+	ExpandedSets []ExpandedSet `yaml:"-"`
+
 	// for key/value replacements in model's cmd, cmdStop, proxy, checkEndPoint
 	Macros MacroList `yaml:"macros"`
 
@@ -183,8 +192,16 @@ func LoadConfigFromReader(r io.Reader) (Config, error) {
 	if err != nil {
 		return Config{}, err
 	}
+	yamlStr := string(data)
 
-	// default configuration values
+	// Phase 1: Substitute all ${env.VAR} macros at string level
+	// This is safe because env values are simple strings without YAML formatting
+	yamlStr, err = substituteEnvMacros(yamlStr)
+	if err != nil {
+		return Config{}, err
+	}
+
+	// Unmarshal into full Config with defaults
 	config := Config{
 		HealthCheckTimeout: 120,
 		StartPort:          5800,
@@ -192,14 +209,14 @@ func LoadConfigFromReader(r io.Reader) (Config, error) {
 		LogTimeFormat:      "",
 		LogToStdout:        LogToStdoutProxy,
 		MetricsMaxInMemory: 1000,
+		CaptureBuffer:      5,
+		GlobalTTL:          0,
 	}
-	err = yaml.Unmarshal(data, &config)
-	if err != nil {
+	if err = yaml.Unmarshal([]byte(yamlStr), &config); err != nil {
 		return Config{}, err
 	}
 
 	if config.HealthCheckTimeout < 15 {
-		// set a minimum of 15 seconds
 		config.HealthCheckTimeout = 15
 	}
 
@@ -207,6 +224,10 @@ func LoadConfigFromReader(r io.Reader) (Config, error) {
 		return Config{}, fmt.Errorf("startPort must be greater than 1")
 	}
 
+	if config.GlobalTTL < 0 {
+		return Config{}, fmt.Errorf("globalTTL must be >= 0")
+	}
+
 	switch config.LogToStdout {
 	case LogToStdoutProxy, LogToStdoutUpstream, LogToStdoutBoth, LogToStdoutNone:
 	default:
@@ -224,55 +245,55 @@ func LoadConfigFromReader(r io.Reader) (Config, error) {
 		}
 	}
 
-	/* check macro constraint rules:
-
-	- name must fit the regex ^[a-zA-Z0-9_-]+$
-	- names must be less than 64 characters (no reason, just cause)
-	- name can not be any reserved macros: PORT, MODEL_ID
-	- macro values must be less than 1024 characters
-	*/
+	// Validate global macros
 	for _, macro := range config.Macros {
 		if err = validateMacro(macro.Name, macro.Value); err != nil {
 			return Config{}, err
 		}
 	}
 
-	// Get and sort all model IDs first, makes testing more consistent
+	// Get and sort all model IDs for consistent port assignment
 	modelIds := make([]string, 0, len(config.Models))
 	for modelId := range config.Models {
 		modelIds = append(modelIds, modelId)
 	}
-	sort.Strings(modelIds) // This guarantees stable iteration order
+	sort.Strings(modelIds)
 
 	nextPort := config.StartPort
 	for _, modelId := range modelIds {
 		modelConfig := config.Models[modelId]
 
-		// Strip comments from command fields before macro expansion
+		// Strip comments from command fields
 		modelConfig.Cmd = StripComments(modelConfig.Cmd)
 		modelConfig.CmdStop = StripComments(modelConfig.CmdStop)
 
-		// validate model macros
+		// set model TTL to globalTTL it is the default value
+		if modelConfig.UnloadAfter == MODEL_CONFIG_DEFAULT_TTL {
+			modelConfig.UnloadAfter = config.GlobalTTL
+		}
+
+		if modelConfig.UnloadAfter < 0 {
+			return Config{}, fmt.Errorf("model %s: invalid TTL value %d", modelId, modelConfig.UnloadAfter)
+		}
+
+		// Validate model macros
 		for _, macro := range modelConfig.Macros {
 			if err = validateMacro(macro.Name, macro.Value); err != nil {
 				return Config{}, fmt.Errorf("model %s: %s", modelId, err.Error())
 			}
 		}
 
-		// Merge global config and model macros. Model macros take precedence
-		mergedMacros := make(MacroList, 0, len(config.Macros)+len(modelConfig.Macros))
+		// Build merged macro list: MODEL_ID + global macros + model macros (model overrides global)
+		mergedMacros := make(MacroList, 0, len(config.Macros)+len(modelConfig.Macros)+1)
 		mergedMacros = append(mergedMacros, MacroEntry{Name: "MODEL_ID", Value: modelId})
-
-		// Add global macros first
 		mergedMacros = append(mergedMacros, config.Macros...)
 
-		// Add model macros (can override global)
+		// Add model macros (override globals with same name)
 		for _, entry := range modelConfig.Macros {
-			// Remove any existing global macro with same name
 			found := false
 			for i, existing := range mergedMacros {
 				if existing.Name == entry.Name {
-					mergedMacros[i] = entry // Override
+					mergedMacros[i] = entry
 					found = true
 					break
 				}
@@ -282,23 +303,40 @@ func LoadConfigFromReader(r io.Reader) (Config, error) {
 			}
 		}
 
-		// First pass: Substitute user-defined macros in reverse order (LIFO - last defined first)
-		// This allows later macros to reference earlier ones
+		// Substitute remaining macros in model fields (LIFO order)
 		for i := len(mergedMacros) - 1; i >= 0; i-- {
 			entry := mergedMacros[i]
 			macroSlug := fmt.Sprintf("${%s}", entry.Name)
 			macroStr := fmt.Sprintf("%v", entry.Value)
 
-			// Substitute in command fields
 			modelConfig.Cmd = strings.ReplaceAll(modelConfig.Cmd, macroSlug, macroStr)
 			modelConfig.CmdStop = strings.ReplaceAll(modelConfig.CmdStop, macroSlug, macroStr)
 			modelConfig.Proxy = strings.ReplaceAll(modelConfig.Proxy, macroSlug, macroStr)
 			modelConfig.CheckEndpoint = strings.ReplaceAll(modelConfig.CheckEndpoint, macroSlug, macroStr)
 			modelConfig.Filters.StripParams = strings.ReplaceAll(modelConfig.Filters.StripParams, macroSlug, macroStr)
+			modelConfig.Name = strings.ReplaceAll(modelConfig.Name, macroSlug, macroStr)
+			modelConfig.Description = strings.ReplaceAll(modelConfig.Description, macroSlug, macroStr)
 
-			// Substitute in metadata (recursive)
+			// Substitute macros in SetParamsByID keys and values
+			if len(modelConfig.Filters.SetParamsByID) > 0 {
+				newSetParamsByID := make(map[string]map[string]any, len(modelConfig.Filters.SetParamsByID))
+				for key, paramMap := range modelConfig.Filters.SetParamsByID {
+					newKey := strings.ReplaceAll(key, macroSlug, macroStr)
+					newValAny, err := substituteMacroInValue(any(paramMap), entry.Name, entry.Value)
+					if err != nil {
+						return Config{}, fmt.Errorf("model %s filters.setParamsByID: %s", modelId, err.Error())
+					}
+					newParamMap, ok := newValAny.(map[string]any)
+					if !ok {
+						return Config{}, fmt.Errorf("model %s filters.setParamsByID: unexpected type after macro substitution", modelId)
+					}
+					newSetParamsByID[newKey] = newParamMap
+				}
+				modelConfig.Filters.SetParamsByID = newSetParamsByID
+			}
+
+			// Substitute in metadata (type-preserving)
 			if len(modelConfig.Metadata) > 0 {
-				var err error
 				result, err := substituteMacroInValue(modelConfig.Metadata, entry.Name, entry.Value)
 				if err != nil {
 					return Config{}, fmt.Errorf("model %s metadata: %s", modelId, err.Error())
@@ -307,29 +345,25 @@ func LoadConfigFromReader(r io.Reader) (Config, error) {
 			}
 		}
 
-		// Final pass: check if PORT macro is needed after macro expansion
-		// ${PORT} is a resource on the local machine so a new port is only allocated
-		// if it is required in either cmd or proxy keys
+		// Handle PORT macro - only allocate if cmd uses it
 		cmdHasPort := strings.Contains(modelConfig.Cmd, "${PORT}")
 		proxyHasPort := strings.Contains(modelConfig.Proxy, "${PORT}")
-		if cmdHasPort || proxyHasPort { // either has it
-			if !cmdHasPort && proxyHasPort { // but both don't have it
+		if cmdHasPort || proxyHasPort {
+			if !cmdHasPort && proxyHasPort {
 				return Config{}, fmt.Errorf("model %s: proxy uses ${PORT} but cmd does not - ${PORT} is only available when used in cmd", modelId)
 			}
 
-			// Add PORT macro and substitute it
-			portEntry := MacroEntry{Name: "PORT", Value: nextPort}
 			macroSlug := "${PORT}"
 			macroStr := fmt.Sprintf("%v", nextPort)
 
 			modelConfig.Cmd = strings.ReplaceAll(modelConfig.Cmd, macroSlug, macroStr)
 			modelConfig.CmdStop = strings.ReplaceAll(modelConfig.CmdStop, macroSlug, macroStr)
 			modelConfig.Proxy = strings.ReplaceAll(modelConfig.Proxy, macroSlug, macroStr)
+			modelConfig.Name = strings.ReplaceAll(modelConfig.Name, macroSlug, macroStr)
+			modelConfig.Description = strings.ReplaceAll(modelConfig.Description, macroSlug, macroStr)
 
-			// Substitute PORT in metadata
 			if len(modelConfig.Metadata) > 0 {
-				var err error
-				result, err := substituteMacroInValue(modelConfig.Metadata, portEntry.Name, portEntry.Value)
+				result, err := substituteMacroInValue(modelConfig.Metadata, "PORT", nextPort)
 				if err != nil {
 					return Config{}, fmt.Errorf("model %s metadata: %s", modelId, err.Error())
 				}
@@ -339,13 +373,15 @@ func LoadConfigFromReader(r io.Reader) (Config, error) {
 			nextPort++
 		}
 
-		// make sure there are no unknown macros that have not been replaced
+		// Validate no unknown macros remain
 		fieldMap := map[string]string{
 			"cmd":                 modelConfig.Cmd,
 			"cmdStop":             modelConfig.CmdStop,
 			"proxy":               modelConfig.Proxy,
 			"checkEndpoint":       modelConfig.CheckEndpoint,
 			"filters.stripParams": modelConfig.Filters.StripParams,
+			"name":                modelConfig.Name,
+			"description":         modelConfig.Description,
 		}
 
 		for fieldName, fieldValue := range fieldMap {
@@ -353,62 +389,94 @@ func LoadConfigFromReader(r io.Reader) (Config, error) {
 			for _, match := range matches {
 				macroName := match[1]
 				if macroName == "PID" && fieldName == "cmdStop" {
-					continue // this is ok, has to be replaced by process later
+					continue // replaced at runtime
 				}
-				// Reserved macros are always valid (they should have been substituted already)
 				if macroName == "PORT" || macroName == "MODEL_ID" {
 					return Config{}, fmt.Errorf("macro '${%s}' should have been substituted in %s.%s", macroName, modelId, fieldName)
 				}
-				// Any other macro is unknown
 				return Config{}, fmt.Errorf("unknown macro '${%s}' found in %s.%s", macroName, modelId, fieldName)
 			}
 		}
 
-		// Check for unknown macros in metadata
 		if len(modelConfig.Metadata) > 0 {
-			if err := validateMetadataForUnknownMacros(modelConfig.Metadata, modelId); err != nil {
+			if err := validateNestedForUnknownMacros(modelConfig.Metadata, fmt.Sprintf("model %s metadata", modelId)); err != nil {
 				return Config{}, err
 			}
 		}
 
-		// Validate the proxy URL.
-		if _, err := url.Parse(modelConfig.Proxy); err != nil {
-			return Config{}, fmt.Errorf(
-				"model %s: invalid proxy URL: %w", modelId, err,
-			)
+		// Validate SetParamsByID keys and values
+		for key, paramMap := range modelConfig.Filters.SetParamsByID {
+			if matches := macroPatternRegex.FindAllStringSubmatch(key, -1); len(matches) > 0 {
+				return Config{}, fmt.Errorf("unknown macro '${%s}' found in model %s filters.setParamsByID key", matches[0][1], modelId)
+			}
+			if err := validateNestedForUnknownMacros(any(paramMap), fmt.Sprintf("model %s filters.setParamsByID[%s]", modelId, key)); err != nil {
+				return Config{}, err
+			}
+		}
+
+		// Auto-register setParamsByID keys as aliases (skip the model's own ID)
+		for key := range modelConfig.Filters.SetParamsByID {
+			if key == modelId {
+				continue
+			}
+			if _, exists := config.Models[key]; exists {
+				return Config{}, fmt.Errorf("model %s filters.setParamsByID: key '%s' conflicts with an existing model ID", modelId, key)
+			}
+			if existingModel, exists := config.aliases[key]; exists {
+				if existingModel != modelId {
+					return Config{}, fmt.Errorf("duplicate alias '%s' in model %s filters.setParamsByID, already used by model %s", key, modelId, existingModel)
+				}
+				continue // already registered as explicit alias for this model
+			}
+			config.aliases[key] = modelId
+			modelConfig.Aliases = append(modelConfig.Aliases, key)
+		}
+
+		if _, err := url.Parse(modelConfig.Proxy); err != nil {
+			return Config{}, fmt.Errorf("model %s: invalid proxy URL: %w", modelId, err)
 		}
 
-		// if sendLoadingState is nil, set it to the global config value
-		// see #366
 		if modelConfig.SendLoadingState == nil {
-			v := config.SendLoadingState // copy it
+			v := config.SendLoadingState
 			modelConfig.SendLoadingState = &v
 		}
 
 		config.Models[modelId] = modelConfig
 	}
 
-	config = AddDefaultGroupToConfig(config)
-	// check that members are all unique in the groups
-	memberUsage := make(map[string]string) // maps member to group it appears in
-	for groupID, groupConfig := range config.Groups {
-		prevSet := make(map[string]bool)
-		for _, member := range groupConfig.Members {
-			// Check for duplicates within this group
-			if _, found := prevSet[member]; found {
-				return Config{}, fmt.Errorf("duplicate model member %s found in group: %s", member, groupID)
-			}
-			prevSet[member] = true
+	// groups XOR matrix
+	if config.Matrix != nil && len(config.Groups) > 0 {
+		return Config{}, fmt.Errorf("config cannot use both 'groups' and 'matrix'")
+	}
 
-			// Check if member is used in another group
-			if existingGroup, exists := memberUsage[member]; exists {
-				return Config{}, fmt.Errorf("model member %s is used in multiple groups: %s and %s", member, existingGroup, groupID)
+	if config.Matrix != nil {
+		expandedSets, err := ValidateMatrix(*config.Matrix, config.Models)
+		if err != nil {
+			return Config{}, fmt.Errorf("matrix: %w", err)
+		}
+		config.ExpandedSets = expandedSets
+	} else {
+		config = AddDefaultGroupToConfig(config)
+
+		// Validate group members
+		memberUsage := make(map[string]string)
+		for groupID, groupConfig := range config.Groups {
+			prevSet := make(map[string]bool)
+			for _, member := range groupConfig.Members {
+				if _, found := prevSet[member]; found {
+					return Config{}, fmt.Errorf("duplicate model member %s found in group: %s", member, groupID)
+				}
+				prevSet[member] = true
+
+				if existingGroup, exists := memberUsage[member]; exists {
+					return Config{}, fmt.Errorf("model member %s is used in multiple groups: %s and %s", member, existingGroup, groupID)
+				}
+				memberUsage[member] = groupID
 			}
-			memberUsage[member] = groupID
 		}
 	}
 
-	// clean up hooks preload
+	// Clean up hooks preload
 	if len(config.Hooks.OnStartup.Preload) > 0 {
 		var toPreload []string
 		for _, modelID := range config.Hooks.OnStartup.Preload {
@@ -420,19 +488,54 @@ func LoadConfigFromReader(r io.Reader) (Config, error) {
 				toPreload = append(toPreload, real)
 			}
 		}
-
 		config.Hooks.OnStartup.Preload = toPreload
 	}
 
-	// check api keys validatity
-	for _, apikey := range config.RequiredAPIKeys {
+	// Validate API keys (env macros already substituted at string level)
+	for i, apikey := range config.RequiredAPIKeys {
 		if apikey == "" {
 			return Config{}, fmt.Errorf("empty api key found in apiKeys")
 		}
-
 		if strings.Contains(apikey, " ") {
 			return Config{}, fmt.Errorf("api key cannot contain spaces: `%s`", apikey)
 		}
+		config.RequiredAPIKeys[i] = apikey
+	}
+
+	// Process peers with global macro substitution
+	for peerName, peerConfig := range config.Peers {
+		// Substitute global macros (LIFO order)
+		for i := len(config.Macros) - 1; i >= 0; i-- {
+			entry := config.Macros[i]
+			macroSlug := fmt.Sprintf("${%s}", entry.Name)
+			macroStr := fmt.Sprintf("%v", entry.Value)
+
+			peerConfig.ApiKey = strings.ReplaceAll(peerConfig.ApiKey, macroSlug, macroStr)
+			peerConfig.Filters.StripParams = strings.ReplaceAll(peerConfig.Filters.StripParams, macroSlug, macroStr)
+
+			// Substitute in setParams (type-preserving)
+			if len(peerConfig.Filters.SetParams) > 0 {
+				result, err := substituteMacroInValue(peerConfig.Filters.SetParams, entry.Name, entry.Value)
+				if err != nil {
+					return Config{}, fmt.Errorf("peers.%s.filters.setParams: %w", peerName, err)
+				}
+				peerConfig.Filters.SetParams = result.(map[string]any)
+			}
+		}
+
+		// Validate no unknown macros remain
+		if matches := macroPatternRegex.FindAllStringSubmatch(peerConfig.ApiKey, -1); len(matches) > 0 {
+			return Config{}, fmt.Errorf("peers.%s.apiKey: unknown macro '${%s}'", peerName, matches[0][1])
+		}
+		if matches := macroPatternRegex.FindAllStringSubmatch(peerConfig.Filters.StripParams, -1); len(matches) > 0 {
+			return Config{}, fmt.Errorf("peers.%s.filters.stripParams: unknown macro '${%s}'", peerName, matches[0][1])
+		}
+		if len(peerConfig.Filters.SetParams) > 0 {
+			if err := validateNestedForUnknownMacros(peerConfig.Filters.SetParams, fmt.Sprintf("peers.%s.filters.setParams", peerName)); err != nil {
+				return Config{}, err
+			}
+		}
+		config.Peers[peerName] = peerConfig
 	}
 
 	return config, nil
@@ -565,20 +668,26 @@ func validateMacro(name string, value any) error {
 	return nil
 }
 
-// validateMetadataForUnknownMacros recursively checks for any remaining macro references in metadata
-func validateMetadataForUnknownMacros(value any, modelId string) error {
+// validateNestedForUnknownMacros recursively checks for any remaining macro references in nested structures
+func validateNestedForUnknownMacros(value any, context string) error {
 	switch v := value.(type) {
 	case string:
 		matches := macroPatternRegex.FindAllStringSubmatch(v, -1)
 		for _, match := range matches {
 			macroName := match[1]
-			return fmt.Errorf("model %s metadata: unknown macro '${%s}'", modelId, macroName)
+			return fmt.Errorf("%s: unknown macro '${%s}'", context, macroName)
+		}
+		// Check for unsubstituted env macros
+		envMatches := envMacroRegex.FindAllStringSubmatch(v, -1)
+		for _, match := range envMatches {
+			varName := match[1]
+			return fmt.Errorf("%s: environment variable '%s' not set", context, varName)
 		}
 		return nil
 
 	case map[string]any:
 		for _, val := range v {
-			if err := validateMetadataForUnknownMacros(val, modelId); err != nil {
+			if err := validateNestedForUnknownMacros(val, context); err != nil {
 				return err
 			}
 		}
@@ -586,7 +695,7 @@ func validateMetadataForUnknownMacros(value any, modelId string) error {
 
 	case []any:
 		for _, val := range v {
-			if err := validateMetadataForUnknownMacros(val, modelId); err != nil {
+			if err := validateNestedForUnknownMacros(val, context); err != nil {
 				return err
 			}
 		}
@@ -645,3 +754,67 @@ func substituteMacroInValue(value any, macroName string, macroValue any) (any, e
 		return value, nil
 	}
 }
+
+// substituteEnvMacros replaces ${env.VAR_NAME} with environment variable values.
+// Returns error if any referenced env var is not set or contains invalid characters.
+// Env macros inside YAML comments are ignored by unmarshalling the YAML first
+// (which strips comments) and only checking the comment-free version for macros.
+func substituteEnvMacros(s string) (string, error) {
+	// Unmarshal and remarshal to strip YAML comments
+	var raw any
+	if err := yaml.Unmarshal([]byte(s), &raw); err != nil {
+		// If YAML is invalid, fall back to scanning the original string
+		// so the user gets the env var error rather than a confusing YAML parse error
+		return substituteEnvMacrosInString(s, s)
+	}
+	clean, err := yaml.Marshal(raw)
+	if err != nil {
+		return substituteEnvMacrosInString(s, s)
+	}
+
+	return substituteEnvMacrosInString(s, string(clean))
+}
+
+// substituteEnvMacrosInString finds ${env.VAR} macros in scanStr and substitutes
+// them in target. This separation allows scanning comment-free YAML while
+// substituting in the original string.
+func substituteEnvMacrosInString(target, scanStr string) (string, error) {
+	result := target
+	matches := envMacroRegex.FindAllStringSubmatch(scanStr, -1)
+	for _, match := range matches {
+		fullMatch := match[0] // ${env.VAR_NAME}
+		varName := match[1]   // VAR_NAME
+
+		value, exists := os.LookupEnv(varName)
+		if !exists {
+			return "", fmt.Errorf("environment variable '%s' is not set", varName)
+		}
+
+		// Sanitize the value for safe YAML substitution
+		value, err := sanitizeEnvValueForYAML(value, varName)
+		if err != nil {
+			return "", err
+		}
+
+		result = strings.ReplaceAll(result, fullMatch, value)
+	}
+	return result, nil
+}
+
+// sanitizeEnvValueForYAML ensures an environment variable value is safe for YAML substitution.
+// It rejects values with characters that break YAML structure and escapes quotes/backslashes
+// for compatibility with double-quoted YAML strings.
+func sanitizeEnvValueForYAML(value, varName string) (string, error) {
+	// Reject values that would break YAML structure regardless of quoting context
+	if strings.ContainsAny(value, "\n\r\x00") {
+		return "", fmt.Errorf("environment variable '%s' contains newlines or null bytes which are not allowed in YAML substitution", varName)
+	}
+
+	// Escape backslashes and double quotes for safe use in double-quoted YAML strings.
+	// In unquoted contexts, these escapes appear literally (harmless for most use cases).
+	// In double-quoted contexts, they are interpreted correctly.
+	value = strings.ReplaceAll(value, `\`, `\\`)
+	value = strings.ReplaceAll(value, `"`, `\"`)
+
+	return value, nil
+}

proxy/config/config_posix_test.go

@@ -163,6 +163,15 @@ groups:
 
 	modelLoadingState := false
 
+	defaultTimeout := TimeoutsConfig{
+		Connect:        30,
+		KeepAlive:      30,
+		ResponseHeader: 0,
+		TLSHandshake:   10,
+		ExpectContinue: 1,
+		IdleConn:       90,
+	}
+
 	expected := Config{
 		LogLevel:      "info",
 		LogTimeFormat: "",
@@ -187,6 +196,7 @@ groups:
 				Name:             "Model 1",
 				Description:      "This is model 1",
 				SendLoadingState: &modelLoadingState,
+				Timeouts:         defaultTimeout,
 			},
 			"model2": {
 				Cmd:              "path/to/server --arg1 one",
@@ -195,6 +205,7 @@ groups:
 				Env:              []string{},
 				CheckEndpoint:    "/",
 				SendLoadingState: &modelLoadingState,
+				Timeouts:         defaultTimeout,
 			},
 			"model3": {
 				Cmd:              "path/to/cmd --arg1 one",
@@ -203,6 +214,7 @@ groups:
 				Env:              []string{},
 				CheckEndpoint:    "/",
 				SendLoadingState: &modelLoadingState,
+				Timeouts:         defaultTimeout,
 			},
 			"model4": {
 				Cmd:              "path/to/cmd --arg1 one",
@@ -211,10 +223,12 @@ groups:
 				Aliases:          []string{},
 				Env:              []string{},
 				SendLoadingState: &modelLoadingState,
+				Timeouts:         defaultTimeout,
 			},
 		},
 		HealthCheckTimeout: 15,
 		MetricsMaxInMemory: 1000,
+		CaptureBuffer:      5,
 		Profiles: map[string][]string{
 			"test": {"model1", "model2"},
 		},

proxy/config/config_test.go

@@ -6,6 +6,7 @@ import (
 	"testing"
 
 	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
 )
 
 func TestConfig_GroupMemberIsUnique(t *testing.T) {
@@ -809,3 +810,737 @@ func TestConfig_APIKeys_Invalid(t *testing.T) {
 		})
 	}
 }
+
+func TestConfig_APIKeys_EnvMacros(t *testing.T) {
+	t.Run("env substitution in apiKeys", func(t *testing.T) {
+		t.Setenv("TEST_API_KEY", "secret-key-123")
+
+		content := `apiKeys: ["${env.TEST_API_KEY}"]`
+		config, err := LoadConfigFromReader(strings.NewReader(content))
+		assert.NoError(t, err)
+		assert.Equal(t, []string{"secret-key-123"}, config.RequiredAPIKeys)
+	})
+
+	t.Run("multiple env substitutions in apiKeys", func(t *testing.T) {
+		t.Setenv("TEST_API_KEY_1", "key-one")
+		t.Setenv("TEST_API_KEY_2", "key-two")
+
+		content := `apiKeys: ["${env.TEST_API_KEY_1}", "${env.TEST_API_KEY_2}", "static-key"]`
+		config, err := LoadConfigFromReader(strings.NewReader(content))
+		assert.NoError(t, err)
+		assert.Equal(t, []string{"key-one", "key-two", "static-key"}, config.RequiredAPIKeys)
+	})
+
+	t.Run("missing env var in apiKeys", func(t *testing.T) {
+		content := `apiKeys: ["${env.NONEXISTENT_API_KEY}"]`
+		_, err := LoadConfigFromReader(strings.NewReader(content))
+		assert.Error(t, err)
+		// With string-level env substitution, error only includes var name
+		assert.Contains(t, err.Error(), "NONEXISTENT_API_KEY")
+	})
+
+	t.Run("env substitution results in empty key", func(t *testing.T) {
+		t.Setenv("TEST_EMPTY_KEY", "")
+
+		content := `apiKeys: ["${env.TEST_EMPTY_KEY}"]`
+		_, err := LoadConfigFromReader(strings.NewReader(content))
+		assert.Error(t, err)
+		assert.Equal(t, "empty api key found in apiKeys", err.Error())
+	})
+}
+
+func TestConfig_GlobalTTL(t *testing.T) {
+	t.Run("globalTTL sets default for models", func(t *testing.T) {
+		content := `
+globalTTL: 300
+models:
+  model1:
+    cmd: server --port ${PORT}
+`
+		config, err := LoadConfigFromReader(strings.NewReader(content))
+		assert.NoError(t, err)
+		assert.Equal(t, 300, config.GlobalTTL)
+		assert.Equal(t, 300, config.Models["model1"].UnloadAfter)
+	})
+
+	t.Run("model ttl=0 overrides globalTTL", func(t *testing.T) {
+		content := `
+globalTTL: 300
+models:
+  model1:
+    cmd: server --port ${PORT}
+    ttl: 0
+`
+		config, err := LoadConfigFromReader(strings.NewReader(content))
+		assert.NoError(t, err)
+		assert.Equal(t, 0, config.Models["model1"].UnloadAfter)
+	})
+
+	t.Run("model explicit ttl overrides globalTTL", func(t *testing.T) {
+		content := `
+globalTTL: 300
+models:
+  model1:
+    cmd: server --port ${PORT}
+    ttl: 600
+`
+		config, err := LoadConfigFromReader(strings.NewReader(content))
+		assert.NoError(t, err)
+		assert.Equal(t, 600, config.Models["model1"].UnloadAfter)
+	})
+
+	t.Run("globalTTL defaults to 0", func(t *testing.T) {
+		content := `
+models:
+  model1:
+    cmd: server --port ${PORT}
+`
+		config, err := LoadConfigFromReader(strings.NewReader(content))
+		assert.NoError(t, err)
+		assert.Equal(t, 0, config.GlobalTTL)
+		assert.Equal(t, 0, config.Models["model1"].UnloadAfter)
+	})
+
+	t.Run("negative globalTTL rejected", func(t *testing.T) {
+		content := `
+globalTTL: -1
+models:
+  model1:
+    cmd: server --port ${PORT}
+`
+		_, err := LoadConfigFromReader(strings.NewReader(content))
+		assert.Error(t, err)
+		assert.Contains(t, err.Error(), "globalTTL must be >= 0")
+	})
+}
+
+func TestConfig_EnvMacros(t *testing.T) {
+	t.Run("basic env substitution in cmd", func(t *testing.T) {
+		t.Setenv("TEST_MODEL_PATH", "/opt/models")
+
+		content := `
+models:
+  test:
+    cmd: "${env.TEST_MODEL_PATH}/llama-server"
+    proxy: "http://localhost:8080"
+`
+		config, err := LoadConfigFromReader(strings.NewReader(content))
+		assert.NoError(t, err)
+		assert.Equal(t, "/opt/models/llama-server", config.Models["test"].Cmd)
+	})
+
+	t.Run("env substitution in multiple fields", func(t *testing.T) {
+		t.Setenv("TEST_HOST", "myserver")
+		t.Setenv("TEST_PORT", "9999")
+
+		content := `
+models:
+  test:
+    cmd: "server --host ${env.TEST_HOST}"
+    proxy: "http://${env.TEST_HOST}:${env.TEST_PORT}"
+    checkEndpoint: "http://${env.TEST_HOST}/health"
+`
+		config, err := LoadConfigFromReader(strings.NewReader(content))
+		assert.NoError(t, err)
+		assert.Equal(t, "server --host myserver", config.Models["test"].Cmd)
+		assert.Equal(t, "http://myserver:9999", config.Models["test"].Proxy)
+		assert.Equal(t, "http://myserver/health", config.Models["test"].CheckEndpoint)
+	})
+
+	t.Run("env in global macro value", func(t *testing.T) {
+		t.Setenv("TEST_BASE_PATH", "/usr/local")
+
+		content := `
+macros:
+  SERVER_PATH: "${env.TEST_BASE_PATH}/bin/server"
+models:
+  test:
+    cmd: "${SERVER_PATH} --port 8080"
+    proxy: "http://localhost:8080"
+`
+		config, err := LoadConfigFromReader(strings.NewReader(content))
+		assert.NoError(t, err)
+		assert.Equal(t, "/usr/local/bin/server --port 8080", config.Models["test"].Cmd)
+	})
+
+	t.Run("env in model-level macro value", func(t *testing.T) {
+		t.Setenv("TEST_MODEL_DIR", "/models/llama")
+
+		content := `
+models:
+  test:
+    macros:
+      MODEL_FILE: "${env.TEST_MODEL_DIR}/model.gguf"
+    cmd: "server --model ${MODEL_FILE}"
+    proxy: "http://localhost:8080"
+`
+		config, err := LoadConfigFromReader(strings.NewReader(content))
+		assert.NoError(t, err)
+		assert.Equal(t, "server --model /models/llama/model.gguf", config.Models["test"].Cmd)
+	})
+
+	t.Run("env in metadata", func(t *testing.T) {
+		t.Setenv("TEST_API_KEY", "secret123")
+
+		content := `
+models:
+  test:
+    cmd: "server"
+    proxy: "http://localhost:8080"
+    metadata:
+      api_key: "${env.TEST_API_KEY}"
+      nested:
+        key: "${env.TEST_API_KEY}"
+`
+		config, err := LoadConfigFromReader(strings.NewReader(content))
+		assert.NoError(t, err)
+		assert.Equal(t, "secret123", config.Models["test"].Metadata["api_key"])
+		nested := config.Models["test"].Metadata["nested"].(map[string]any)
+		assert.Equal(t, "secret123", nested["key"])
+	})
+
+	t.Run("env in filters.stripParams", func(t *testing.T) {
+		t.Setenv("TEST_STRIP_PARAMS", "temperature,top_p")
+
+		content := `
+models:
+  test:
+    cmd: "server"
+    proxy: "http://localhost:8080"
+    filters:
+      stripParams: "${env.TEST_STRIP_PARAMS}"
+`
+		config, err := LoadConfigFromReader(strings.NewReader(content))
+		assert.NoError(t, err)
+		assert.Equal(t, "temperature,top_p", config.Models["test"].Filters.StripParams)
+	})
+
+	t.Run("env in cmdStop", func(t *testing.T) {
+		t.Setenv("TEST_KILL_SIGNAL", "SIGTERM")
+
+		content := `
+models:
+  test:
+    cmd: "server --port ${PORT}"
+    cmdStop: "kill -${env.TEST_KILL_SIGNAL} ${PID}"
+    proxy: "http://localhost:${PORT}"
+`
+		config, err := LoadConfigFromReader(strings.NewReader(content))
+		assert.NoError(t, err)
+		assert.Contains(t, config.Models["test"].CmdStop, "-SIGTERM")
+	})
+
+	t.Run("missing env var returns error", func(t *testing.T) {
+		content := `
+models:
+  test:
+    cmd: "${env.UNDEFINED_VAR_12345}/server"
+    proxy: "http://localhost:8080"
+`
+		_, err := LoadConfigFromReader(strings.NewReader(content))
+		if assert.Error(t, err) {
+			assert.Contains(t, err.Error(), "UNDEFINED_VAR_12345")
+			assert.Contains(t, err.Error(), "not set")
+		}
+	})
+
+	t.Run("missing env var in global macro", func(t *testing.T) {
+		content := `
+macros:
+  PATH: "${env.UNDEFINED_GLOBAL_VAR}"
+models:
+  test:
+    cmd: "server"
+    proxy: "http://localhost:8080"
+`
+		_, err := LoadConfigFromReader(strings.NewReader(content))
+		if assert.Error(t, err) {
+			assert.Contains(t, err.Error(), "UNDEFINED_GLOBAL_VAR")
+			assert.Contains(t, err.Error(), "not set")
+		}
+	})
+
+	t.Run("missing env var in model macro", func(t *testing.T) {
+		content := `
+models:
+  test:
+    macros:
+      MY_PATH: "${env.UNDEFINED_MODEL_VAR}"
+    cmd: "server"
+    proxy: "http://localhost:8080"
+`
+		_, err := LoadConfigFromReader(strings.NewReader(content))
+		if assert.Error(t, err) {
+			assert.Contains(t, err.Error(), "UNDEFINED_MODEL_VAR")
+			assert.Contains(t, err.Error(), "not set")
+		}
+	})
+
+	t.Run("missing env var in metadata", func(t *testing.T) {
+		content := `
+models:
+  test:
+    cmd: "server"
+    proxy: "http://localhost:8080"
+    metadata:
+      key: "${env.UNDEFINED_META_VAR}"
+`
+		_, err := LoadConfigFromReader(strings.NewReader(content))
+		if assert.Error(t, err) {
+			assert.Contains(t, err.Error(), "UNDEFINED_META_VAR")
+			assert.Contains(t, err.Error(), "not set")
+		}
+	})
+
+	t.Run("env combined with regular macros", func(t *testing.T) {
+		t.Setenv("TEST_ROOT", "/data")
+
+		content := `
+macros:
+  MODEL_BASE: "${env.TEST_ROOT}/models"
+models:
+  test:
+    cmd: "server --model ${MODEL_BASE}/${MODEL_ID}.gguf"
+    proxy: "http://localhost:8080"
+`
+		config, err := LoadConfigFromReader(strings.NewReader(content))
+		assert.NoError(t, err)
+		assert.Equal(t, "server --model /data/models/test.gguf", config.Models["test"].Cmd)
+	})
+
+	t.Run("multiple env vars in same string", func(t *testing.T) {
+		t.Setenv("TEST_USER", "admin")
+		t.Setenv("TEST_PASS", "secret")
+
+		content := `
+models:
+  test:
+    cmd: "server --auth ${env.TEST_USER}:${env.TEST_PASS}"
+    proxy: "http://localhost:8080"
+`
+		config, err := LoadConfigFromReader(strings.NewReader(content))
+		assert.NoError(t, err)
+		assert.Equal(t, "server --auth admin:secret", config.Models["test"].Cmd)
+	})
+
+	t.Run("env value with newline is rejected", func(t *testing.T) {
+		t.Setenv("TEST_MULTILINE", "line1\nline2")
+
+		content := `
+models:
+  test:
+    cmd: "server --config ${env.TEST_MULTILINE}"
+    proxy: "http://localhost:8080"
+`
+		_, err := LoadConfigFromReader(strings.NewReader(content))
+		if assert.Error(t, err) {
+			assert.Contains(t, err.Error(), "TEST_MULTILINE")
+			assert.Contains(t, err.Error(), "newlines")
+		}
+	})
+
+	t.Run("env value with carriage return is rejected", func(t *testing.T) {
+		t.Setenv("TEST_CR", "line1\rline2")
+
+		content := `
+models:
+  test:
+    cmd: "server --config ${env.TEST_CR}"
+    proxy: "http://localhost:8080"
+`
+		_, err := LoadConfigFromReader(strings.NewReader(content))
+		if assert.Error(t, err) {
+			assert.Contains(t, err.Error(), "TEST_CR")
+			assert.Contains(t, err.Error(), "newlines")
+		}
+	})
+
+	t.Run("env value with quotes is escaped for YAML", func(t *testing.T) {
+		t.Setenv("TEST_QUOTED", `value with "quotes"`)
+
+		content := `
+models:
+  test:
+    cmd: "server --arg \"${env.TEST_QUOTED}\""
+    proxy: "http://localhost:8080"
+`
+		config, err := LoadConfigFromReader(strings.NewReader(content))
+		assert.NoError(t, err)
+		// Quotes are escaped before YAML parsing, then YAML unescapes them
+		// Final result preserves the original value with quotes
+		assert.Contains(t, config.Models["test"].Cmd, `"quotes"`)
+	})
+
+	t.Run("env value with backslash is escaped for YAML", func(t *testing.T) {
+		t.Setenv("TEST_BACKSLASH", `path\to\file`)
+
+		content := `
+models:
+  test:
+    cmd: "server --path \"${env.TEST_BACKSLASH}\""
+    proxy: "http://localhost:8080"
+`
+		config, err := LoadConfigFromReader(strings.NewReader(content))
+		assert.NoError(t, err)
+		// Backslashes are escaped before YAML parsing, then YAML unescapes them
+		// Final result preserves the original value with backslashes
+		assert.Contains(t, config.Models["test"].Cmd, `path\to\file`)
+	})
+}
+
+func TestConfig_PeerApiKey_EnvMacros(t *testing.T) {
+	t.Run("env substitution in peer apiKey", func(t *testing.T) {
+		t.Setenv("TEST_PEER_API_KEY", "sk-peer-secret-123")
+
+		content := `
+peers:
+  openrouter:
+    proxy: https://openrouter.ai/api
+    apiKey: "${env.TEST_PEER_API_KEY}"
+    models:
+      - llama-3.1-8b
+`
+		config, err := LoadConfigFromReader(strings.NewReader(content))
+		assert.NoError(t, err)
+		assert.Equal(t, "sk-peer-secret-123", config.Peers["openrouter"].ApiKey)
+	})
+
+	t.Run("missing env var in peer apiKey", func(t *testing.T) {
+		content := `
+peers:
+  openrouter:
+    proxy: https://openrouter.ai/api
+    apiKey: "${env.NONEXISTENT_PEER_KEY}"
+    models:
+      - llama-3.1-8b
+`
+		_, err := LoadConfigFromReader(strings.NewReader(content))
+		assert.Error(t, err)
+		// With string-level env substitution, error only includes var name
+		assert.Contains(t, err.Error(), "NONEXISTENT_PEER_KEY")
+	})
+
+	t.Run("static apiKey unchanged", func(t *testing.T) {
+		content := `
+peers:
+  openrouter:
+    proxy: https://openrouter.ai/api
+    apiKey: sk-static-key
+    models:
+      - llama-3.1-8b
+`
+		config, err := LoadConfigFromReader(strings.NewReader(content))
+		assert.NoError(t, err)
+		assert.Equal(t, "sk-static-key", config.Peers["openrouter"].ApiKey)
+	})
+
+	t.Run("multiple peers with env apiKeys", func(t *testing.T) {
+		t.Setenv("TEST_PEER_KEY_1", "key-one")
+		t.Setenv("TEST_PEER_KEY_2", "key-two")
+
+		content := `
+peers:
+  peer1:
+    proxy: https://peer1.example.com
+    apiKey: "${env.TEST_PEER_KEY_1}"
+    models:
+      - model-a
+  peer2:
+    proxy: https://peer2.example.com
+    apiKey: "${env.TEST_PEER_KEY_2}"
+    models:
+      - model-b
+`
+		config, err := LoadConfigFromReader(strings.NewReader(content))
+		assert.NoError(t, err)
+		assert.Equal(t, "key-one", config.Peers["peer1"].ApiKey)
+		assert.Equal(t, "key-two", config.Peers["peer2"].ApiKey)
+	})
+
+	t.Run("global macro substitution in peer apiKey", func(t *testing.T) {
+		content := `
+macros:
+  API_KEY: sk-from-global-macro
+peers:
+  openrouter:
+    proxy: https://openrouter.ai/api
+    apiKey: "${API_KEY}"
+    models:
+      - llama-3.1-8b
+`
+		config, err := LoadConfigFromReader(strings.NewReader(content))
+		assert.NoError(t, err)
+		assert.Equal(t, "sk-from-global-macro", config.Peers["openrouter"].ApiKey)
+	})
+
+	t.Run("global macro in peer filters.stripParams", func(t *testing.T) {
+		content := `
+macros:
+  STRIP_LIST: "temperature, top_p"
+peers:
+  openrouter:
+    proxy: https://openrouter.ai/api
+    models:
+      - llama-3.1-8b
+    filters:
+      stripParams: "${STRIP_LIST}"
+`
+		config, err := LoadConfigFromReader(strings.NewReader(content))
+		assert.NoError(t, err)
+		assert.Equal(t, "temperature, top_p", config.Peers["openrouter"].Filters.StripParams)
+	})
+
+	t.Run("global macro in peer filters.setParams", func(t *testing.T) {
+		content := `
+macros:
+  MAX_TOKENS: 4096
+peers:
+  openrouter:
+    proxy: https://openrouter.ai/api
+    models:
+      - llama-3.1-8b
+    filters:
+      setParams:
+        max_tokens: "${MAX_TOKENS}"
+`
+		config, err := LoadConfigFromReader(strings.NewReader(content))
+		assert.NoError(t, err)
+		assert.Equal(t, 4096, config.Peers["openrouter"].Filters.SetParams["max_tokens"])
+	})
+
+	t.Run("env macro in peer filters.setParams", func(t *testing.T) {
+		t.Setenv("TEST_RETENTION_POLICY", "deny")
+
+		content := `
+peers:
+  openrouter:
+    proxy: https://openrouter.ai/api
+    models:
+      - llama-3.1-8b
+    filters:
+      setParams:
+        data_collection: "${env.TEST_RETENTION_POLICY}"
+`
+		config, err := LoadConfigFromReader(strings.NewReader(content))
+		assert.NoError(t, err)
+		assert.Equal(t, "deny", config.Peers["openrouter"].Filters.SetParams["data_collection"])
+	})
+
+	t.Run("env macro in peer filters.stripParams", func(t *testing.T) {
+		t.Setenv("TEST_STRIP_PARAMS", "frequency_penalty, presence_penalty")
+
+		content := `
+peers:
+  openrouter:
+    proxy: https://openrouter.ai/api
+    models:
+      - llama-3.1-8b
+    filters:
+      stripParams: "${env.TEST_STRIP_PARAMS}"
+`
+		config, err := LoadConfigFromReader(strings.NewReader(content))
+		assert.NoError(t, err)
+		assert.Equal(t, "frequency_penalty, presence_penalty", config.Peers["openrouter"].Filters.StripParams)
+	})
+
+	t.Run("unknown macro in peer apiKey fails", func(t *testing.T) {
+		content := `
+peers:
+  openrouter:
+    proxy: https://openrouter.ai/api
+    apiKey: "${UNDEFINED_MACRO}"
+    models:
+      - llama-3.1-8b
+`
+		_, err := LoadConfigFromReader(strings.NewReader(content))
+		assert.Error(t, err)
+		assert.Contains(t, err.Error(), "peers.openrouter.apiKey")
+		assert.Contains(t, err.Error(), "unknown macro")
+	})
+
+	t.Run("unknown macro in peer filters.setParams fails", func(t *testing.T) {
+		content := `
+peers:
+  openrouter:
+    proxy: https://openrouter.ai/api
+    models:
+      - llama-3.1-8b
+    filters:
+      setParams:
+        value: "${UNDEFINED_MACRO}"
+`
+		_, err := LoadConfigFromReader(strings.NewReader(content))
+		assert.Error(t, err)
+		assert.Contains(t, err.Error(), "peers.openrouter.filters.setParams")
+		assert.Contains(t, err.Error(), "unknown macro")
+	})
+
+	t.Run("env macros in comments are ignored", func(t *testing.T) {
+		content := `
+# apiKeys:
+#   - "${env.COMMENTED_OUT_KEY_1}"
+#   - "${env.COMMENTED_OUT_KEY_2}"
+models:
+  test:
+    cmd: "server"
+    proxy: "http://localhost:8080"
+`
+		// These env vars are NOT set, but should not cause an error
+		// because they only appear in comment lines
+		config, err := LoadConfigFromReader(strings.NewReader(content))
+		assert.NoError(t, err)
+		assert.Empty(t, config.RequiredAPIKeys)
+	})
+
+	t.Run("env macros in comments ignored while active ones resolve", func(t *testing.T) {
+		t.Setenv("TEST_ACTIVE_KEY", "active-key-value")
+
+		content := `
+# apiKeys: ["${env.COMMENTED_OUT_KEY}"]
+apiKeys: ["${env.TEST_ACTIVE_KEY}"]
+models:
+  test:
+    cmd: "server"
+    proxy: "http://localhost:8080"
+`
+		config, err := LoadConfigFromReader(strings.NewReader(content))
+		assert.NoError(t, err)
+		assert.Equal(t, []string{"active-key-value"}, config.RequiredAPIKeys)
+	})
+
+	t.Run("env macros in indented comments are ignored", func(t *testing.T) {
+		content := `
+models:
+  test:
+    cmd: |
+      server
+      --port 8080
+    proxy: "http://localhost:8080"
+    # metadata:
+    #   api_key: "${env.SOME_UNSET_KEY}"
+`
+		_, err := LoadConfigFromReader(strings.NewReader(content))
+		assert.NoError(t, err)
+	})
+
+	t.Run("env macros in inline comments are ignored", func(t *testing.T) {
+		t.Setenv("TEST_INLINE_KEY", "real-value")
+
+		content := `
+apiKeys: ["${env.TEST_INLINE_KEY}"] # TODO: add ${env.FUTURE_KEY} later
+models:
+  test:
+    cmd: "server"
+    proxy: "http://localhost:8080"
+`
+		config, err := LoadConfigFromReader(strings.NewReader(content))
+		assert.NoError(t, err)
+		assert.Equal(t, []string{"real-value"}, config.RequiredAPIKeys)
+	})
+
+}
+
+func TestConfig_TimeoutsParsing(t *testing.T) {
+	configYaml := `
+models:
+  model1:
+    cmd: test-server --port ${PORT}
+    timeouts:
+      connect: 45
+      responseHeader: 120
+`
+
+	config, err := LoadConfigFromReader(strings.NewReader(configYaml))
+	require.NoError(t, err)
+
+	modelConfig, found := config.Models["model1"]
+	require.True(t, found, "model1 should exist in config")
+
+	assert.Equal(t, 45, modelConfig.Timeouts.Connect)
+	assert.Equal(t, 120, modelConfig.Timeouts.ResponseHeader)
+}
+
+func TestConfig_TimeoutsDefaults(t *testing.T) {
+	configYaml := `
+models:
+  model1:
+    cmd: test-server --port ${PORT}
+`
+
+	config, err := LoadConfigFromReader(strings.NewReader(configYaml))
+	require.NoError(t, err)
+
+	modelConfig, found := config.Models["model1"]
+	require.True(t, found, "model1 should exist in config")
+
+	// Default values should be set during unmarshaling
+	assert.Equal(t, 30, modelConfig.Timeouts.Connect)
+	assert.Equal(t, 0, modelConfig.Timeouts.ResponseHeader)
+	assert.Equal(t, 10, modelConfig.Timeouts.TLSHandshake)
+	assert.Equal(t, 1, modelConfig.Timeouts.ExpectContinue)
+	assert.Equal(t, 90, modelConfig.Timeouts.IdleConn)
+}
+
+func TestConfig_TimeoutsZeroAllowed(t *testing.T) {
+	configYaml := `
+models:
+  model1:
+    cmd: test-server --port ${PORT}
+    timeouts:
+      connect: 0
+      responseHeader: 0
+`
+
+	config, err := LoadConfigFromReader(strings.NewReader(configYaml))
+	require.NoError(t, err)
+
+	modelConfig, found := config.Models["model1"]
+	require.True(t, found, "model1 should exist in config")
+
+	// Explicit 0 should be preserved (disables timeout)
+	assert.Equal(t, 0, modelConfig.Timeouts.Connect)
+	assert.Equal(t, 0, modelConfig.Timeouts.ResponseHeader)
+}
+
+func TestConfig_PeerTimeoutsParsing(t *testing.T) {
+	configYaml := `
+peers:
+  peer1:
+    proxy: http://example.com
+    models: [model1]
+    timeouts:
+      connect: 45
+      responseHeader: 120
+`
+
+	config, err := LoadConfigFromReader(strings.NewReader(configYaml))
+	require.NoError(t, err)
+
+	peerConfig, found := config.Peers["peer1"]
+	require.True(t, found, "peer1 should exist in config")
+
+	assert.Equal(t, 45, peerConfig.Timeouts.Connect)
+	assert.Equal(t, 120, peerConfig.Timeouts.ResponseHeader)
+}
+
+func TestConfig_PeerTimeoutsDefaults(t *testing.T) {
+	configYaml := `
+peers:
+  peer1:
+    proxy: http://example.com
+    models: [model1]
+`
+
+	config, err := LoadConfigFromReader(strings.NewReader(configYaml))
+	require.NoError(t, err)
+
+	peerConfig, found := config.Peers["peer1"]
+	require.True(t, found, "peer1 should exist in config")
+
+	// Default values should be set during unmarshaling
+	assert.Equal(t, 30, peerConfig.Timeouts.Connect)
+	assert.Equal(t, 60, peerConfig.Timeouts.ResponseHeader)
+	assert.Equal(t, 10, peerConfig.Timeouts.TLSHandshake)
+	assert.Equal(t, 1, peerConfig.Timeouts.ExpectContinue)
+	assert.Equal(t, 90, peerConfig.Timeouts.IdleConn)
+}

proxy/config/config_windows_test.go

@@ -155,6 +155,15 @@ groups:
 
 	modelLoadingState := false
 
+	defaultTimeout := TimeoutsConfig{
+		Connect:        30,
+		KeepAlive:      30,
+		ResponseHeader: 0,
+		TLSHandshake:   10,
+		ExpectContinue: 1,
+		IdleConn:       90,
+	}
+
 	expected := Config{
 		LogLevel:      "info",
 		LogTimeFormat: "",
@@ -173,6 +182,7 @@ groups:
 				Env:              []string{"VAR1=value1", "VAR2=value2"},
 				CheckEndpoint:    "/health",
 				SendLoadingState: &modelLoadingState,
+				Timeouts:         defaultTimeout,
 			},
 			"model2": {
 				Cmd:              "path/to/server --arg1 one",
@@ -182,6 +192,7 @@ groups:
 				Env:              []string{},
 				CheckEndpoint:    "/",
 				SendLoadingState: &modelLoadingState,
+				Timeouts:         defaultTimeout,
 			},
 			"model3": {
 				Cmd:              "path/to/cmd --arg1 one",
@@ -191,6 +202,7 @@ groups:
 				Env:              []string{},
 				CheckEndpoint:    "/",
 				SendLoadingState: &modelLoadingState,
+				Timeouts:         defaultTimeout,
 			},
 			"model4": {
 				Cmd:              "path/to/cmd --arg1 one",
@@ -200,10 +212,12 @@ groups:
 				Aliases:          []string{},
 				Env:              []string{},
 				SendLoadingState: &modelLoadingState,
+				Timeouts:         defaultTimeout,
 			},
 		},
 		HealthCheckTimeout: 15,
 		MetricsMaxInMemory: 1000,
+		CaptureBuffer:      5,
 		Profiles: map[string][]string{
 			"test": {"model1", "model2"},
 		},

proxy/config/filters.go

@@ -0,0 +1,114 @@
+package config
+
+import (
+	"slices"
+	"sort"
+	"strings"
+)
+
+// ProtectedParams is a list of parameters that cannot be set or stripped via filters
+// These are protected to prevent breaking the proxy's ability to route requests correctly
+var ProtectedParams = []string{"model"}
+
+// Filters contains filter settings for modifying request parameters
+// Used by both models and peers
+type Filters struct {
+	// StripParams is a comma-separated list of parameters to remove from requests
+	// The "model" parameter can never be removed
+	StripParams string `yaml:"stripParams"`
+
+	// SetParams is a dictionary of parameters to set/override in requests
+	// Protected params (like "model") cannot be set
+	SetParams map[string]any `yaml:"setParams"`
+
+	// SetParamsByID maps requested model IDs to parameters to set/override in requests.
+	// Useful with aliases: a single loaded model can behave differently depending on
+	// which alias the client used. Applied after SetParams, so it can override those values.
+	// Protected params (like "model") cannot be set.
+	SetParamsByID map[string]map[string]any `yaml:"setParamsByID"`
+}
+
+// SanitizedStripParams returns a sorted list of parameters to strip,
+// with duplicates, empty strings, and protected params removed
+func (f Filters) SanitizedStripParams() []string {
+	if f.StripParams == "" {
+		return nil
+	}
+
+	params := strings.Split(f.StripParams, ",")
+	cleaned := make([]string, 0, len(params))
+	seen := make(map[string]bool)
+
+	for _, param := range params {
+		trimmed := strings.TrimSpace(param)
+		// Skip protected params, empty strings, and duplicates
+		if slices.Contains(ProtectedParams, trimmed) || trimmed == "" || seen[trimmed] {
+			continue
+		}
+		seen[trimmed] = true
+		cleaned = append(cleaned, trimmed)
+	}
+
+	if len(cleaned) == 0 {
+		return nil
+	}
+
+	slices.Sort(cleaned)
+	return cleaned
+}
+
+// SanitizedSetParamsByID returns the params to set for the given requestedModelID,
+// with protected params removed and keys sorted for consistent iteration order.
+// Returns nil if the ID has no entry or all its params are protected.
+func (f Filters) SanitizedSetParamsByID(requestedModelID string) (map[string]any, []string) {
+	if len(f.SetParamsByID) == 0 {
+		return nil, nil
+	}
+	params, found := f.SetParamsByID[requestedModelID]
+	if !found || len(params) == 0 {
+		return nil, nil
+	}
+	result := make(map[string]any, len(params))
+	keys := make([]string, 0, len(params))
+	for key, value := range params {
+		if slices.Contains(ProtectedParams, key) {
+			continue
+		}
+		result[key] = value
+		keys = append(keys, key)
+	}
+	sort.Strings(keys)
+	if len(result) == 0 {
+		return nil, nil
+	}
+	return result, keys
+}
+
+// SanitizedSetParams returns a copy of SetParams with protected params removed
+// and keys sorted for consistent iteration order
+func (f Filters) SanitizedSetParams() (map[string]any, []string) {
+	if len(f.SetParams) == 0 {
+		return nil, nil
+	}
+
+	result := make(map[string]any, len(f.SetParams))
+	keys := make([]string, 0, len(f.SetParams))
+
+	for key, value := range f.SetParams {
+		// Skip protected params
+		if slices.Contains(ProtectedParams, key) {
+			continue
+		}
+		result[key] = value
+		keys = append(keys, key)
+	}
+
+	// Sort keys for consistent ordering
+	sort.Strings(keys)
+
+	if len(result) == 0 {
+		return nil, nil
+	}
+
+	return result, keys
+}

proxy/config/filters_test.go

@@ -0,0 +1,285 @@
+package config
+
+import (
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+)
+
+func TestFilters_SanitizedStripParams(t *testing.T) {
+	tests := []struct {
+		name        string
+		stripParams string
+		want        []string
+	}{
+		{
+			name:        "empty string",
+			stripParams: "",
+			want:        nil,
+		},
+		{
+			name:        "single param",
+			stripParams: "temperature",
+			want:        []string{"temperature"},
+		},
+		{
+			name:        "multiple params",
+			stripParams: "temperature, top_p, top_k",
+			want:        []string{"temperature", "top_k", "top_p"}, // sorted
+		},
+		{
+			name:        "model param filtered",
+			stripParams: "model, temperature, top_p",
+			want:        []string{"temperature", "top_p"},
+		},
+		{
+			name:        "only model param",
+			stripParams: "model",
+			want:        nil,
+		},
+		{
+			name:        "duplicates removed",
+			stripParams: "temperature, top_p, temperature",
+			want:        []string{"temperature", "top_p"},
+		},
+		{
+			name:        "extra whitespace",
+			stripParams: "  temperature  ,  top_p  ",
+			want:        []string{"temperature", "top_p"},
+		},
+		{
+			name:        "empty values filtered",
+			stripParams: "temperature,,top_p,",
+			want:        []string{"temperature", "top_p"},
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			f := Filters{StripParams: tt.stripParams}
+			got := f.SanitizedStripParams()
+			assert.Equal(t, tt.want, got)
+		})
+	}
+}
+
+func TestFilters_SanitizedSetParams(t *testing.T) {
+	tests := []struct {
+		name       string
+		setParams  map[string]any
+		wantParams map[string]any
+		wantKeys   []string
+	}{
+		{
+			name:       "empty setParams",
+			setParams:  nil,
+			wantParams: nil,
+			wantKeys:   nil,
+		},
+		{
+			name:       "empty map",
+			setParams:  map[string]any{},
+			wantParams: nil,
+			wantKeys:   nil,
+		},
+		{
+			name: "normal params",
+			setParams: map[string]any{
+				"temperature": 0.7,
+				"top_p":       0.9,
+			},
+			wantParams: map[string]any{
+				"temperature": 0.7,
+				"top_p":       0.9,
+			},
+			wantKeys: []string{"temperature", "top_p"},
+		},
+		{
+			name: "protected model param filtered",
+			setParams: map[string]any{
+				"model":       "should-be-filtered",
+				"temperature": 0.7,
+			},
+			wantParams: map[string]any{
+				"temperature": 0.7,
+			},
+			wantKeys: []string{"temperature"},
+		},
+		{
+			name: "only protected param",
+			setParams: map[string]any{
+				"model": "should-be-filtered",
+			},
+			wantParams: nil,
+			wantKeys:   nil,
+		},
+		{
+			name: "complex nested values",
+			setParams: map[string]any{
+				"provider": map[string]any{
+					"data_collection": "deny",
+					"allow_fallbacks": false,
+				},
+				"transforms": []string{"middle-out"},
+			},
+			wantParams: map[string]any{
+				"provider": map[string]any{
+					"data_collection": "deny",
+					"allow_fallbacks": false,
+				},
+				"transforms": []string{"middle-out"},
+			},
+			wantKeys: []string{"provider", "transforms"},
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			f := Filters{SetParams: tt.setParams}
+			gotParams, gotKeys := f.SanitizedSetParams()
+
+			assert.Equal(t, len(tt.wantKeys), len(gotKeys), "keys length mismatch")
+			for i, key := range gotKeys {
+				assert.Equal(t, tt.wantKeys[i], key, "key mismatch at %d", i)
+			}
+
+			if tt.wantParams == nil {
+				assert.Nil(t, gotParams, "expected nil params")
+				return
+			}
+
+			assert.Equal(t, len(tt.wantParams), len(gotParams), "params length mismatch")
+			for key, wantValue := range tt.wantParams {
+				gotValue, exists := gotParams[key]
+				assert.True(t, exists, "missing key: %s", key)
+				// Simple comparison for basic types
+				switch v := wantValue.(type) {
+				case string, int, float64, bool:
+					assert.Equal(t, v, gotValue, "value mismatch for key %s", key)
+				}
+			}
+		})
+	}
+}
+
+func TestFilters_SanitizedSetParamsByID(t *testing.T) {
+	tests := []struct {
+		name             string
+		setParamsByID    map[string]map[string]any
+		requestedModelID string
+		wantParams       map[string]any
+		wantKeys         []string
+	}{
+		{
+			name:             "empty SetParamsByID returns nil",
+			setParamsByID:    nil,
+			requestedModelID: "model1",
+			wantParams:       nil,
+			wantKeys:         nil,
+		},
+		{
+			name:             "empty map returns nil",
+			setParamsByID:    map[string]map[string]any{},
+			requestedModelID: "model1",
+			wantParams:       nil,
+			wantKeys:         nil,
+		},
+		{
+			name: "non-matching model ID returns nil",
+			setParamsByID: map[string]map[string]any{
+				"model2": {"temperature": 0.9},
+			},
+			requestedModelID: "model1",
+			wantParams:       nil,
+			wantKeys:         nil,
+		},
+		{
+			name: "matching model ID returns correct params",
+			setParamsByID: map[string]map[string]any{
+				"model1": {"temperature": 0.7, "top_p": 0.9},
+				"model2": {"temperature": 0.5},
+			},
+			requestedModelID: "model1",
+			wantParams: map[string]any{
+				"temperature": 0.7,
+				"top_p":       0.9,
+			},
+			wantKeys: []string{"temperature", "top_p"},
+		},
+		{
+			name: "protected param model is filtered out",
+			setParamsByID: map[string]map[string]any{
+				"model1": {
+					"model":       "should-be-filtered",
+					"temperature": 0.7,
+				},
+			},
+			requestedModelID: "model1",
+			wantParams: map[string]any{
+				"temperature": 0.7,
+			},
+			wantKeys: []string{"temperature"},
+		},
+		{
+			name: "only protected param returns nil",
+			setParamsByID: map[string]map[string]any{
+				"model1": {
+					"model": "should-be-filtered",
+				},
+			},
+			requestedModelID: "model1",
+			wantParams:       nil,
+			wantKeys:         nil,
+		},
+		{
+			name: "keys are sorted",
+			setParamsByID: map[string]map[string]any{
+				"model1": {
+					"z_param": "z",
+					"a_param": "a",
+					"m_param": "m",
+				},
+			},
+			requestedModelID: "model1",
+			wantParams: map[string]any{
+				"z_param": "z",
+				"a_param": "a",
+				"m_param": "m",
+			},
+			wantKeys: []string{"a_param", "m_param", "z_param"},
+		},
+		{
+			name: "alias style key lookup",
+			setParamsByID: map[string]map[string]any{
+				"model1:high": {"reasoning_effort": "high"},
+				"model1:low":  {"reasoning_effort": "low"},
+			},
+			requestedModelID: "model1:high",
+			wantParams: map[string]any{
+				"reasoning_effort": "high",
+			},
+			wantKeys: []string{"reasoning_effort"},
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			f := Filters{SetParamsByID: tt.setParamsByID}
+			gotParams, gotKeys := f.SanitizedSetParamsByID(tt.requestedModelID)
+
+			if tt.wantParams == nil {
+				assert.Nil(t, gotParams)
+				assert.Nil(t, gotKeys)
+				return
+			}
+
+			assert.Equal(t, tt.wantKeys, gotKeys)
+			assert.Equal(t, tt.wantParams, gotParams)
+		})
+	}
+}
+
+func TestProtectedParams(t *testing.T) {
+	// Verify that "model" is protected
+	assert.Contains(t, ProtectedParams, "model")
+}

proxy/config/macro_in_macro_test.go

@@ -104,6 +104,62 @@ models:
 	assert.Contains(t, err.Error(), "self-reference")
 }
 
+// Test macro substitution in name and description fields
+func TestConfig_MacroInNameAndDescription(t *testing.T) {
+	content := `
+startPort: 10000
+macros:
+  "VARIANT": "Q4_K_M"
+  "FAMILY": "llama"
+
+models:
+  my-model:
+    cmd: echo ok
+    proxy: http://localhost:8080
+    name: "${FAMILY} ${VARIANT}"
+    description: "A ${FAMILY} model in ${VARIANT} format"
+`
+
+	config, err := LoadConfigFromReader(strings.NewReader(content))
+	assert.NoError(t, err)
+	assert.Equal(t, "llama Q4_K_M", config.Models["my-model"].Name)
+	assert.Equal(t, "A llama model in Q4_K_M format", config.Models["my-model"].Description)
+}
+
+// Test MODEL_ID macro in name and description fields
+func TestConfig_ModelIDInNameAndDescription(t *testing.T) {
+	content := `
+startPort: 10000
+models:
+  llama-3b:
+    cmd: echo ok
+    proxy: http://localhost:8080
+    name: "Model: ${MODEL_ID}"
+    description: "Running ${MODEL_ID}"
+`
+
+	config, err := LoadConfigFromReader(strings.NewReader(content))
+	assert.NoError(t, err)
+	assert.Equal(t, "Model: llama-3b", config.Models["llama-3b"].Name)
+	assert.Equal(t, "Running llama-3b", config.Models["llama-3b"].Description)
+}
+
+// Test unknown macro in name or description returns an error
+func TestConfig_UnknownMacroInNameDescription(t *testing.T) {
+	content := `
+startPort: 10000
+models:
+  test:
+    cmd: echo ok
+    proxy: http://localhost:8080
+    name: "Model ${UNDEFINED}"
+`
+
+	_, err := LoadConfigFromReader(strings.NewReader(content))
+	assert.Error(t, err)
+	assert.Contains(t, err.Error(), "UNDEFINED")
+}
+
 // Test undefined macro reference error
 func TestConfig_UndefinedMacroReference(t *testing.T) {
 	content := `

proxy/config/matrix.go

@@ -0,0 +1,226 @@
+package config
+
+import (
+	"fmt"
+	"regexp"
+	"sort"
+
+	"gopkg.in/yaml.v3"
+)
+
+var varKeyPattern = regexp.MustCompile(`^[a-zA-Z0-9]{1,8}$`)
+
+// MatrixConfig represents the swap matrix configuration block.
+type MatrixConfig struct {
+	Var        map[string]string `yaml:"vars"`
+	EvictCosts map[string]int    `yaml:"evict_costs"`
+	Sets       OrderedSets       `yaml:"sets"`
+}
+
+// SetEntry is a single named set with its DSL expression.
+type SetEntry struct {
+	Name string
+	DSL  string
+}
+
+// OrderedSets preserves YAML definition order of sets (used for tie-breaking).
+type OrderedSets []SetEntry
+
+func (os *OrderedSets) UnmarshalYAML(value *yaml.Node) error {
+	if value.Kind != yaml.MappingNode {
+		return fmt.Errorf("sets must be a mapping")
+	}
+
+	entries := make([]SetEntry, 0, len(value.Content)/2)
+	for i := 0; i < len(value.Content); i += 2 {
+		keyNode := value.Content[i]
+		valueNode := value.Content[i+1]
+
+		var name string
+		if err := keyNode.Decode(&name); err != nil {
+			return fmt.Errorf("failed to decode set name: %w", err)
+		}
+
+		var dsl string
+		if err := valueNode.Decode(&dsl); err != nil {
+			return fmt.Errorf("failed to decode DSL for set %q: %w", name, err)
+		}
+
+		entries = append(entries, SetEntry{Name: name, DSL: dsl})
+	}
+
+	*os = entries
+	return nil
+}
+
+// ExpandedSet is one valid combination of concurrent models (real model names).
+type ExpandedSet struct {
+	SetName string
+	DSL     string
+	Models  []string // real model names, sorted
+}
+
+// ValidateMatrix validates the matrix config and returns all expanded sets.
+func ValidateMatrix(matrix MatrixConfig, models map[string]ModelConfig) ([]ExpandedSet, error) {
+	if len(matrix.Sets) == 0 {
+		return nil, fmt.Errorf("matrix must define at least one set")
+	}
+
+	if len(matrix.Var) == 0 {
+		return nil, fmt.Errorf("matrix must define at least one var")
+	}
+
+	// Validate var entries
+	if matrix.Var != nil {
+		for id, modelName := range matrix.Var {
+			if !varKeyPattern.MatchString(id) {
+				return nil, fmt.Errorf("var key %q must be alphanumeric and 1-8 characters", id)
+			}
+			if _, exists := models[modelName]; !exists {
+				return nil, fmt.Errorf("var key %q references unknown model %q", id, modelName)
+			}
+		}
+	}
+
+	// Validate evict_costs
+	if matrix.EvictCosts != nil {
+		for key, cost := range matrix.EvictCosts {
+			if cost <= 0 {
+				return nil, fmt.Errorf("evict_cost for %q must be a positive integer, got %d", key, cost)
+			}
+			if _, ok := matrix.Var[key]; !ok {
+				return nil, fmt.Errorf("evict_costs: unknown var ID %q", key)
+			}
+		}
+	}
+
+	// Build dependency graph for +ref topological sort
+	setNames := make(map[string]bool)
+	for _, entry := range matrix.Sets {
+		setNames[entry.Name] = true
+	}
+
+	deps := make(map[string][]string) // setName -> set names it depends on
+	for _, entry := range matrix.Sets {
+		refs, err := extractRefs(entry.DSL)
+		if err != nil {
+			return nil, fmt.Errorf("set %q: %w", entry.Name, err)
+		}
+		for _, ref := range refs {
+			if !setNames[ref] {
+				return nil, fmt.Errorf("set %q references undefined set %q", entry.Name, ref)
+			}
+		}
+		deps[entry.Name] = refs
+	}
+
+	// Topological sort with cycle detection
+	order, err := topologicalSort(matrix.Sets, deps)
+	if err != nil {
+		return nil, err
+	}
+
+	// Expand sets in topological order
+	resolvedRefs := make(map[string][][]string) // set name -> expanded alias-level combos
+	var allExpanded []ExpandedSet
+	totalCombinations := 0
+
+	// Build ordered map for efficient lookup
+	setDSL := make(map[string]string)
+	for _, entry := range matrix.Sets {
+		setDSL[entry.Name] = entry.DSL
+	}
+
+	for _, name := range order {
+		dsl := setDSL[name]
+		combos, err := ParseAndExpandDSL(dsl, resolvedRefs)
+		if err != nil {
+			return nil, fmt.Errorf("set %q: %w", name, err)
+		}
+
+		resolvedRefs[name] = combos
+
+		// Resolve var IDs to real model names
+		for _, combo := range combos {
+			resolved := make([]string, len(combo))
+			for i, ident := range combo {
+				realName, ok := matrix.Var[ident]
+				if !ok {
+					return nil, fmt.Errorf("set %q: unknown var ID %q", name, ident)
+				}
+				resolved[i] = realName
+			}
+			sort.Strings(resolved)
+			allExpanded = append(allExpanded, ExpandedSet{
+				SetName: name,
+				DSL:     dsl,
+				Models:  resolved,
+			})
+		}
+
+		totalCombinations += len(combos)
+		if totalCombinations > maxDSLExpansions {
+			return nil, fmt.Errorf("total expanded combinations (%d) exceed limit of %d", totalCombinations, maxDSLExpansions)
+		}
+	}
+
+	return allExpanded, nil
+}
+
+// topologicalSort returns set names in dependency order.
+// Returns an error if a cycle is detected.
+func topologicalSort(sets OrderedSets, deps map[string][]string) ([]string, error) {
+	// States: 0 = unvisited, 1 = visiting, 2 = visited
+	state := make(map[string]int)
+	var order []string
+
+	var visit func(name string) error
+	visit = func(name string) error {
+		switch state[name] {
+		case 1:
+			return fmt.Errorf("circular reference detected involving set %q", name)
+		case 2:
+			return nil
+		}
+		state[name] = 1
+
+		for _, dep := range deps[name] {
+			if err := visit(dep); err != nil {
+				return err
+			}
+		}
+
+		state[name] = 2
+		order = append(order, name)
+		return nil
+	}
+
+	// Visit in definition order for deterministic output
+	for _, entry := range sets {
+		if state[entry.Name] == 0 {
+			if err := visit(entry.Name); err != nil {
+				return nil, err
+			}
+		}
+	}
+
+	return order, nil
+}
+
+// ResolvedEvictCosts returns a map of real model name -> evict cost,
+// resolving var IDs. Models not listed default to 1.
+func (m *MatrixConfig) ResolvedEvictCosts() map[string]int {
+	costs := make(map[string]int)
+	if m.EvictCosts == nil {
+		return costs
+	}
+	for key, cost := range m.EvictCosts {
+		// Resolve var ID if present
+		if realName, ok := m.Var[key]; ok {
+			costs[realName] = cost
+		} else {
+			costs[key] = cost
+		}
+	}
+	return costs
+}

proxy/config/matrix_dsl.go

@@ -0,0 +1,376 @@
+package config
+
+import (
+	"fmt"
+	"sort"
+	"strings"
+	"unicode"
+)
+
+const maxDSLExpansions = 1000
+
+// Token types for the DSL lexer
+type tokenType int
+
+const (
+	tokIdent  tokenType = iota // model alias or name
+	tokAnd                     // &
+	tokOr                      // |
+	tokLParen                  // (
+	tokRParen                  // )
+	tokRef                     // +setName
+	tokEOF
+)
+
+type token struct {
+	typ tokenType
+	val string
+}
+
+// tokenize splits a DSL string into tokens.
+func tokenize(input string) ([]token, error) {
+	var tokens []token
+	i := 0
+	runes := []rune(input)
+
+	for i < len(runes) {
+		ch := runes[i]
+
+		// skip whitespace
+		if unicode.IsSpace(ch) {
+			i++
+			continue
+		}
+
+		switch ch {
+		case '&':
+			tokens = append(tokens, token{tokAnd, "&"})
+			i++
+		case '|':
+			tokens = append(tokens, token{tokOr, "|"})
+			i++
+		case '(':
+			tokens = append(tokens, token{tokLParen, "("})
+			i++
+		case ')':
+			tokens = append(tokens, token{tokRParen, ")"})
+			i++
+		case '+':
+			// +ref: read the identifier that follows
+			i++
+			start := i
+			for i < len(runes) && isIdentChar(runes[i]) {
+				i++
+			}
+			if i == start {
+				return nil, fmt.Errorf("expected set name after '+' at position %d", start)
+			}
+			tokens = append(tokens, token{tokRef, string(runes[start:i])})
+		default:
+			if isIdentChar(ch) {
+				start := i
+				for i < len(runes) && isIdentChar(runes[i]) {
+					i++
+				}
+				tokens = append(tokens, token{tokIdent, string(runes[start:i])})
+			} else {
+				return nil, fmt.Errorf("unexpected character %q at position %d", ch, i)
+			}
+		}
+	}
+
+	tokens = append(tokens, token{tokEOF, ""})
+	return tokens, nil
+}
+
+func isIdentChar(ch rune) bool {
+	return unicode.IsLetter(ch) || unicode.IsDigit(ch) || ch == '_' || ch == '-' || ch == '.'
+}
+
+// AST node types
+type dslNode interface {
+	dslNode()
+}
+
+type andNode struct {
+	children []dslNode
+}
+
+type orNode struct {
+	children []dslNode
+}
+
+type leafNode struct {
+	name string
+}
+
+type refNode struct {
+	setName string
+}
+
+func (andNode) dslNode()  {}
+func (orNode) dslNode()   {}
+func (leafNode) dslNode() {}
+func (refNode) dslNode()  {}
+
+// parser holds state for recursive-descent parsing.
+type parser struct {
+	tokens []token
+	pos    int
+}
+
+func (p *parser) peek() token {
+	if p.pos < len(p.tokens) {
+		return p.tokens[p.pos]
+	}
+	return token{tokEOF, ""}
+}
+
+func (p *parser) next() token {
+	t := p.peek()
+	if t.typ != tokEOF {
+		p.pos++
+	}
+	return t
+}
+
+func (p *parser) expect(typ tokenType) (token, error) {
+	t := p.next()
+	if t.typ != typ {
+		return t, fmt.Errorf("expected token type %d, got %q", typ, t.val)
+	}
+	return t, nil
+}
+
+// Grammar:
+//
+//	expr    = andExpr
+//	andExpr = orExpr ('&' orExpr)*
+//	orExpr  = atom ('|' atom)*
+//	atom    = ident | '+' ident | '(' expr ')'
+//
+// & binds tighter than |, so "a | b & c" means "a | (b & c)"
+func parse(tokens []token) (dslNode, error) {
+	p := &parser{tokens: tokens}
+	node, err := p.parseExpr()
+	if err != nil {
+		return nil, err
+	}
+	if p.peek().typ != tokEOF {
+		return nil, fmt.Errorf("unexpected token %q after expression", p.peek().val)
+	}
+	return node, nil
+}
+
+func (p *parser) parseExpr() (dslNode, error) {
+	return p.parseOrExpr()
+}
+
+func (p *parser) parseOrExpr() (dslNode, error) {
+	left, err := p.parseAndExpr()
+	if err != nil {
+		return nil, err
+	}
+
+	if p.peek().typ == tokOr {
+		children := []dslNode{left}
+		for p.peek().typ == tokOr {
+			p.next() // consume |
+			right, err := p.parseAndExpr()
+			if err != nil {
+				return nil, err
+			}
+			children = append(children, right)
+		}
+		return orNode{children: children}, nil
+	}
+
+	return left, nil
+}
+
+func (p *parser) parseAndExpr() (dslNode, error) {
+	left, err := p.parseAtom()
+	if err != nil {
+		return nil, err
+	}
+
+	if p.peek().typ == tokAnd {
+		children := []dslNode{left}
+		for p.peek().typ == tokAnd {
+			p.next() // consume &
+			right, err := p.parseAtom()
+			if err != nil {
+				return nil, err
+			}
+			children = append(children, right)
+		}
+		return andNode{children: children}, nil
+	}
+
+	return left, nil
+}
+
+func (p *parser) parseAtom() (dslNode, error) {
+	t := p.peek()
+
+	switch t.typ {
+	case tokIdent:
+		p.next()
+		return leafNode{name: t.val}, nil
+
+	case tokRef:
+		p.next()
+		return refNode{setName: t.val}, nil
+
+	case tokLParen:
+		p.next() // consume (
+		node, err := p.parseExpr()
+		if err != nil {
+			return nil, err
+		}
+		if _, err := p.expect(tokRParen); err != nil {
+			return nil, fmt.Errorf("missing closing parenthesis")
+		}
+		return node, nil
+
+	default:
+		return nil, fmt.Errorf("unexpected token %q", t.val)
+	}
+}
+
+// expand walks the AST and produces all combinations.
+// resolvedRefs contains previously expanded sets for +ref resolution.
+func expand(node dslNode, resolvedRefs map[string][][]string) ([][]string, error) {
+	switch n := node.(type) {
+	case leafNode:
+		return [][]string{{n.name}}, nil
+
+	case refNode:
+		expanded, ok := resolvedRefs[n.setName]
+		if !ok {
+			return nil, fmt.Errorf("unknown set reference +%s", n.setName)
+		}
+		// Return a copy
+		result := make([][]string, len(expanded))
+		for i, combo := range expanded {
+			result[i] = make([]string, len(combo))
+			copy(result[i], combo)
+		}
+		return result, nil
+
+	case orNode:
+		// Union of all children's expansions
+		var result [][]string
+		for _, child := range n.children {
+			childResult, err := expand(child, resolvedRefs)
+			if err != nil {
+				return nil, err
+			}
+			result = append(result, childResult...)
+			if len(result) > maxDSLExpansions {
+				return nil, fmt.Errorf("DSL expansion exceeded %d combinations", maxDSLExpansions)
+			}
+		}
+		return result, nil
+
+	case andNode:
+		// Cartesian product across children
+		result := [][]string{{}} // start with one empty combo
+		for _, child := range n.children {
+			childResult, err := expand(child, resolvedRefs)
+			if err != nil {
+				return nil, err
+			}
+			result, err = cartesianProduct(result, childResult, maxDSLExpansions)
+			if err != nil {
+				return nil, err
+			}
+		}
+		return result, nil
+
+	default:
+		return nil, fmt.Errorf("unknown node type %T", node)
+	}
+}
+
+// cartesianProduct computes the cartesian product of two sets of combinations.
+// It returns an error if the product would exceed cap.
+func cartesianProduct(left, right [][]string, cap int) ([][]string, error) {
+	if int64(len(left))*int64(len(right)) > int64(cap) {
+		return nil, fmt.Errorf("DSL expansion exceeded %d combinations", cap)
+	}
+	result := make([][]string, 0, len(left)*len(right))
+	for _, l := range left {
+		for _, r := range right {
+			combo := make([]string, 0, len(l)+len(r))
+			combo = append(combo, l...)
+			combo = append(combo, r...)
+			result = append(result, combo)
+		}
+	}
+	return result, nil
+}
+
+// ParseAndExpandDSL tokenizes, parses, and expands a DSL string.
+// resolvedRefs contains previously expanded sets for +ref inlining.
+func ParseAndExpandDSL(dsl string, resolvedRefs map[string][][]string) ([][]string, error) {
+	dsl = strings.TrimSpace(dsl)
+	if dsl == "" {
+		return nil, fmt.Errorf("empty DSL expression")
+	}
+
+	tokens, err := tokenize(dsl)
+	if err != nil {
+		return nil, fmt.Errorf("tokenize: %w", err)
+	}
+
+	tree, err := parse(tokens)
+	if err != nil {
+		return nil, fmt.Errorf("parse: %w", err)
+	}
+
+	result, err := expand(tree, resolvedRefs)
+	if err != nil {
+		return nil, err
+	}
+
+	// Deduplicate models within each combination and sort for consistency
+	for i, combo := range result {
+		result[i] = dedupAndSort(combo)
+	}
+
+	return result, nil
+}
+
+// dedupAndSort removes duplicate entries and sorts alphabetically.
+func dedupAndSort(items []string) []string {
+	seen := make(map[string]bool, len(items))
+	var unique []string
+	for _, item := range items {
+		if !seen[item] {
+			seen[item] = true
+			unique = append(unique, item)
+		}
+	}
+	sort.Strings(unique)
+	return unique
+}
+
+// extractRefs scans a DSL string for +ref tokens without full parsing.
+// Used for building the dependency graph for topological sorting.
+func extractRefs(dsl string) ([]string, error) {
+	tokens, err := tokenize(dsl)
+	if err != nil {
+		return nil, err
+	}
+
+	var refs []string
+	seen := make(map[string]bool)
+	for _, t := range tokens {
+		if t.typ == tokRef && !seen[t.val] {
+			seen[t.val] = true
+			refs = append(refs, t.val)
+		}
+	}
+	return refs, nil
+}

proxy/config/matrix_dsl_test.go

@@ -0,0 +1,300 @@
+package config
+
+import (
+	"fmt"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+func TestDSL_Tokenize(t *testing.T) {
+	tests := []struct {
+		name   string
+		input  string
+		expect []token
+		errMsg string
+	}{
+		{
+			name:  "single identifier",
+			input: "abc",
+			expect: []token{
+				{tokIdent, "abc"},
+				{tokEOF, ""},
+			},
+		},
+		{
+			name:  "identifier with hyphens and dots",
+			input: "model-name.v2",
+			expect: []token{
+				{tokIdent, "model-name.v2"},
+				{tokEOF, ""},
+			},
+		},
+		{
+			name:  "and expression",
+			input: "a & b",
+			expect: []token{
+				{tokIdent, "a"},
+				{tokAnd, "&"},
+				{tokIdent, "b"},
+				{tokEOF, ""},
+			},
+		},
+		{
+			name:  "or expression",
+			input: "a | b",
+			expect: []token{
+				{tokIdent, "a"},
+				{tokOr, "|"},
+				{tokIdent, "b"},
+				{tokEOF, ""},
+			},
+		},
+		{
+			name:  "parentheses",
+			input: "(a | b) & c",
+			expect: []token{
+				{tokLParen, "("},
+				{tokIdent, "a"},
+				{tokOr, "|"},
+				{tokIdent, "b"},
+				{tokRParen, ")"},
+				{tokAnd, "&"},
+				{tokIdent, "c"},
+				{tokEOF, ""},
+			},
+		},
+		{
+			name:  "ref token",
+			input: "+llms & v",
+			expect: []token{
+				{tokRef, "llms"},
+				{tokAnd, "&"},
+				{tokIdent, "v"},
+				{tokEOF, ""},
+			},
+		},
+		{
+			name:  "no whitespace",
+			input: "(a|b)&c",
+			expect: []token{
+				{tokLParen, "("},
+				{tokIdent, "a"},
+				{tokOr, "|"},
+				{tokIdent, "b"},
+				{tokRParen, ")"},
+				{tokAnd, "&"},
+				{tokIdent, "c"},
+				{tokEOF, ""},
+			},
+		},
+		{
+			name:   "empty ref",
+			input:  "+",
+			errMsg: "expected set name after '+'",
+		},
+		{
+			name:   "invalid character",
+			input:  "a @ b",
+			errMsg: "unexpected character",
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			tokens, err := tokenize(tt.input)
+			if tt.errMsg != "" {
+				require.Error(t, err)
+				assert.Contains(t, err.Error(), tt.errMsg)
+			} else {
+				require.NoError(t, err)
+				assert.Equal(t, tt.expect, tokens)
+			}
+		})
+	}
+}
+
+func TestDSL_ParseAndExpand(t *testing.T) {
+	tests := []struct {
+		name   string
+		dsl    string
+		refs   map[string][][]string
+		expect [][]string
+		errMsg string
+	}{
+		{
+			name:   "single model",
+			dsl:    "L",
+			expect: [][]string{{"L"}},
+		},
+		{
+			name:   "two models with AND",
+			dsl:    "a & b",
+			expect: [][]string{{"a", "b"}},
+		},
+		{
+			name:   "two models with OR",
+			dsl:    "a | b",
+			expect: [][]string{{"a"}, {"b"}},
+		},
+		{
+			name:   "three models with OR",
+			dsl:    "a | b | c",
+			expect: [][]string{{"a"}, {"b"}, {"c"}},
+		},
+		{
+			name: "cartesian product (a|b) & (c|d)",
+			dsl:  "(a | b) & (c | d)",
+			expect: [][]string{
+				{"a", "c"},
+				{"a", "d"},
+				{"b", "c"},
+				{"b", "d"},
+			},
+		},
+		{
+			name: "three-way AND",
+			dsl:  "a & b & c",
+			expect: [][]string{
+				{"a", "b", "c"},
+			},
+		},
+		{
+			name: "(g | q | m) & v",
+			dsl:  "(g | q | m) & v",
+			expect: [][]string{
+				{"g", "v"},
+				{"q", "v"},
+				{"m", "v"},
+			},
+		},
+		{
+			name: "(g | q) & v & e",
+			dsl:  "(g | q) & v & e",
+			expect: [][]string{
+				{"e", "g", "v"},
+				{"e", "q", "v"},
+			},
+		},
+		{
+			name: "precedence: a | b & c means a | (b & c)",
+			dsl:  "a | b & c",
+			expect: [][]string{
+				{"a"},
+				{"b", "c"},
+			},
+		},
+		{
+			name: "+ref inlining",
+			dsl:  "+llms & v",
+			refs: map[string][][]string{
+				"llms": {{"g"}, {"q"}, {"m"}},
+			},
+			expect: [][]string{
+				{"g", "v"},
+				{"q", "v"},
+				{"m", "v"},
+			},
+		},
+		{
+			name: "+ref chained",
+			dsl:  "+with_tts & e",
+			refs: map[string][][]string{
+				"with_tts": {{"g", "v"}, {"q", "v"}, {"m", "v"}},
+			},
+			expect: [][]string{
+				{"e", "g", "v"},
+				{"e", "q", "v"},
+				{"e", "m", "v"},
+			},
+		},
+		{
+			name: "dedup within combination",
+			dsl:  "a & a",
+			expect: [][]string{
+				{"a"},
+			},
+		},
+		{
+			name:   "empty expression",
+			dsl:    "",
+			errMsg: "empty DSL expression",
+		},
+		{
+			name:   "unmatched open paren",
+			dsl:    "(a | b",
+			errMsg: "missing closing parenthesis",
+		},
+		{
+			name:   "unmatched close paren",
+			dsl:    "a | b)",
+			errMsg: "unexpected token",
+		},
+		{
+			name:   "unknown ref",
+			dsl:    "+unknown",
+			errMsg: "unknown set reference +unknown",
+		},
+		{
+			name:   "empty parens",
+			dsl:    "()",
+			errMsg: "unexpected token",
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			refs := tt.refs
+			if refs == nil {
+				refs = map[string][][]string{}
+			}
+			result, err := ParseAndExpandDSL(tt.dsl, refs)
+			if tt.errMsg != "" {
+				require.Error(t, err)
+				assert.Contains(t, err.Error(), tt.errMsg)
+			} else {
+				require.NoError(t, err)
+				assert.Equal(t, tt.expect, result)
+			}
+		})
+	}
+}
+
+func TestDSL_ExpansionCap(t *testing.T) {
+	// Build an expression that would exceed 1000 combinations:
+	// (a1|a2|...|a32) & (b1|b2|...|b32) = 1024 combos
+	var aItems, bItems []string
+	for i := 0; i < 32; i++ {
+		aItems = append(aItems, fmt.Sprintf("a%d", i))
+		bItems = append(bItems, fmt.Sprintf("b%d", i))
+	}
+	dsl := fmt.Sprintf("(%s) & (%s)",
+		join(aItems, " | "),
+		join(bItems, " | "),
+	)
+	_, err := ParseAndExpandDSL(dsl, map[string][][]string{})
+	require.Error(t, err)
+	assert.Contains(t, err.Error(), "exceeded")
+}
+
+func TestDSL_ExtractRefs(t *testing.T) {
+	refs, err := extractRefs("+llms & v & +other")
+	require.NoError(t, err)
+	assert.Equal(t, []string{"llms", "other"}, refs)
+
+	refs, err = extractRefs("a & b")
+	require.NoError(t, err)
+	assert.Empty(t, refs)
+}
+
+func join(items []string, sep string) string {
+	result := ""
+	for i, item := range items {
+		if i > 0 {
+			result += sep
+		}
+		result += item
+	}
+	return result
+}

proxy/config/matrix_test.go

@@ -0,0 +1,305 @@
+package config
+
+import (
+	"strings"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+func makeModels(names ...string) map[string]ModelConfig {
+	m := make(map[string]ModelConfig)
+	for _, name := range names {
+		m[name] = ModelConfig{Cmd: "echo " + name}
+	}
+	return m
+}
+
+func TestValidateMatrix_Basic(t *testing.T) {
+	models := makeModels("gemma", "qwen", "mistral", "voxtral", "llama70B")
+
+	matrix := MatrixConfig{
+		Var: map[string]string{
+			"g": "gemma",
+			"q": "qwen",
+			"m": "mistral",
+			"v": "voxtral",
+			"L": "llama70B",
+		},
+		EvictCosts: map[string]int{
+			"L": 30,
+			"v": 50,
+		},
+		Sets: OrderedSets{
+			{Name: "standard", DSL: "(g | q | m) & v"},
+			{Name: "full", DSL: "L"},
+		},
+	}
+
+	expanded, err := ValidateMatrix(matrix, models)
+	require.NoError(t, err)
+
+	// standard expands to [gemma,voxtral], [qwen,voxtral], [mistral,voxtral]
+	// full expands to [llama70B]
+	assert.Len(t, expanded, 4)
+
+	assert.Equal(t, "standard", expanded[0].SetName)
+	assert.Equal(t, []string{"gemma", "voxtral"}, expanded[0].Models)
+
+	assert.Equal(t, "standard", expanded[1].SetName)
+	assert.Equal(t, []string{"qwen", "voxtral"}, expanded[1].Models)
+
+	assert.Equal(t, "standard", expanded[2].SetName)
+	assert.Equal(t, []string{"mistral", "voxtral"}, expanded[2].Models)
+
+	assert.Equal(t, "full", expanded[3].SetName)
+	assert.Equal(t, []string{"llama70B"}, expanded[3].Models)
+}
+
+func TestValidateMatrix_WithRef(t *testing.T) {
+	models := makeModels("gemma", "qwen", "mistral", "voxtral", "reranker")
+
+	matrix := MatrixConfig{
+		Var: map[string]string{
+			"g": "gemma",
+			"q": "qwen",
+			"m": "mistral",
+			"v": "voxtral",
+			"e": "reranker",
+		},
+		Sets: OrderedSets{
+			{Name: "llms", DSL: "g | q | m"},
+			{Name: "with_tts", DSL: "+llms & v"},
+			{Name: "mega", DSL: "+with_tts & e"},
+		},
+	}
+
+	expanded, err := ValidateMatrix(matrix, models)
+	require.NoError(t, err)
+
+	// llms: [gemma], [qwen], [mistral]
+	// with_tts: [gemma,voxtral], [qwen,voxtral], [mistral,voxtral]
+	// mega: [gemma,reranker,voxtral], [qwen,reranker,voxtral], [mistral,reranker,voxtral]
+	assert.Len(t, expanded, 9)
+
+	// Check mega entries
+	megaEntries := filterBySetName(expanded, "mega")
+	assert.Len(t, megaEntries, 3)
+	assert.Equal(t, []string{"gemma", "reranker", "voxtral"}, megaEntries[0].Models)
+}
+
+func TestValidateMatrix_MapIDRequired(t *testing.T) {
+	// DSL cannot use real model names directly — must use var IDs
+	models := makeModels("gemma", "voxtral")
+
+	matrix := MatrixConfig{
+		Var: map[string]string{"g": "gemma"},
+		Sets: OrderedSets{
+			{Name: "combo", DSL: "g & voxtral"},
+		},
+	}
+
+	_, err := ValidateMatrix(matrix, models)
+	require.Error(t, err)
+	assert.Contains(t, err.Error(), "unknown var ID")
+}
+
+func TestValidateMatrix_InvalidAliasKey(t *testing.T) {
+	models := makeModels("gemma")
+
+	tests := []struct {
+		name   string
+		alias  string
+		errMsg string
+	}{
+		{"too long", "abcdefghi", "alphanumeric and 1-8 characters"},
+		{"has underscore", "a_b", "alphanumeric and 1-8 characters"},
+		{"has hyphen", "a-b", "alphanumeric and 1-8 characters"},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			matrix := MatrixConfig{
+				Var:  map[string]string{tt.alias: "gemma"},
+				Sets: OrderedSets{{Name: "s", DSL: tt.alias}},
+			}
+			_, err := ValidateMatrix(matrix, models)
+			require.Error(t, err)
+			assert.Contains(t, err.Error(), tt.errMsg)
+		})
+	}
+}
+
+func TestValidateMatrix_AliasReferencesUnknownModel(t *testing.T) {
+	models := makeModels("gemma")
+
+	matrix := MatrixConfig{
+		Var:  map[string]string{"x": "nonexistent"},
+		Sets: OrderedSets{{Name: "s", DSL: "x"}},
+	}
+
+	_, err := ValidateMatrix(matrix, models)
+	require.Error(t, err)
+	assert.Contains(t, err.Error(), "unknown model")
+}
+
+func TestValidateMatrix_EvictCostInvalid(t *testing.T) {
+	models := makeModels("gemma")
+
+	t.Run("zero cost", func(t *testing.T) {
+		matrix := MatrixConfig{
+			Var:        map[string]string{"g": "gemma"},
+			EvictCosts: map[string]int{"g": 0},
+			Sets:       OrderedSets{{Name: "s", DSL: "g"}},
+		}
+		_, err := ValidateMatrix(matrix, models)
+		require.Error(t, err)
+		assert.Contains(t, err.Error(), "positive integer")
+	})
+
+	t.Run("negative cost", func(t *testing.T) {
+		matrix := MatrixConfig{
+			Var:        map[string]string{"g": "gemma"},
+			EvictCosts: map[string]int{"g": -1},
+			Sets:       OrderedSets{{Name: "s", DSL: "g"}},
+		}
+		_, err := ValidateMatrix(matrix, models)
+		require.Error(t, err)
+		assert.Contains(t, err.Error(), "positive integer")
+	})
+
+	t.Run("unknown var ID in evict_costs", func(t *testing.T) {
+		matrix := MatrixConfig{
+			Var:        map[string]string{"g": "gemma"},
+			EvictCosts: map[string]int{"unknown": 5},
+			Sets:       OrderedSets{{Name: "s", DSL: "g"}},
+		}
+		_, err := ValidateMatrix(matrix, models)
+		require.Error(t, err)
+		assert.Contains(t, err.Error(), "unknown var ID")
+	})
+}
+
+func TestValidateMatrix_CycleDetection(t *testing.T) {
+	models := makeModels("gemma")
+
+	matrix := MatrixConfig{
+		Var: map[string]string{"g": "gemma"},
+		Sets: OrderedSets{
+			{Name: "a", DSL: "+b"},
+			{Name: "b", DSL: "+a"},
+		},
+	}
+
+	_, err := ValidateMatrix(matrix, models)
+	require.Error(t, err)
+	assert.Contains(t, err.Error(), "circular reference")
+}
+
+func TestValidateMatrix_UndefinedRefTarget(t *testing.T) {
+	models := makeModels("gemma")
+
+	matrix := MatrixConfig{
+		Var: map[string]string{"g": "gemma"},
+		Sets: OrderedSets{
+			{Name: "a", DSL: "+nonexistent"},
+		},
+	}
+
+	_, err := ValidateMatrix(matrix, models)
+	require.Error(t, err)
+	assert.Contains(t, err.Error(), "references undefined set")
+}
+
+func TestValidateMatrix_NoSets(t *testing.T) {
+	_, err := ValidateMatrix(MatrixConfig{}, makeModels("gemma"))
+	require.Error(t, err)
+	assert.Contains(t, err.Error(), "at least one set")
+}
+
+func TestValidateMatrix_UnknownMapIDInDSL(t *testing.T) {
+	models := makeModels("gemma")
+
+	matrix := MatrixConfig{
+		Var: map[string]string{"g": "gemma"},
+		Sets: OrderedSets{
+			{Name: "s", DSL: "g & nonexistent"},
+		},
+	}
+
+	_, err := ValidateMatrix(matrix, models)
+	require.Error(t, err)
+	assert.Contains(t, err.Error(), "unknown var ID")
+}
+
+func TestValidateMatrix_ResolvedEvictCosts(t *testing.T) {
+	mc := &MatrixConfig{
+		Var: map[string]string{
+			"g": "gemma",
+			"L": "llama70B",
+		},
+		EvictCosts: map[string]int{
+			"L": 30,
+			"g": 5,
+		},
+	}
+
+	costs := mc.ResolvedEvictCosts()
+	assert.Equal(t, 30, costs["llama70B"])
+	assert.Equal(t, 5, costs["gemma"])
+}
+
+func TestValidateMatrix_ConfigXOR(t *testing.T) {
+	// groups and matrix both defined
+	yaml := `
+models:
+  model1:
+    cmd: echo model1
+    proxy: http://localhost:8080
+groups:
+  group1:
+    members:
+      - model1
+matrix:
+  sets:
+    s: "model1"
+`
+	_, err := LoadConfigFromReader(strings.NewReader(yaml))
+	require.Error(t, err)
+	assert.Contains(t, err.Error(), "cannot use both")
+}
+
+func TestValidateMatrix_ConfigMatrixOnly(t *testing.T) {
+	yaml := `
+models:
+  gemma:
+    cmd: echo gemma
+    proxy: http://localhost:8080
+  qwen:
+    cmd: echo qwen
+    proxy: http://localhost:8081
+matrix:
+  vars:
+    g: gemma
+    q: qwen
+  sets:
+    combo: "g | q"
+`
+	cfg, err := LoadConfigFromReader(strings.NewReader(yaml))
+	require.NoError(t, err)
+	assert.NotNil(t, cfg.Matrix)
+	assert.Len(t, cfg.ExpandedSets, 2)
+	// Groups should be empty when matrix is used
+	assert.Empty(t, cfg.Groups)
+}
+
+func filterBySetName(sets []ExpandedSet, name string) []ExpandedSet {
+	var result []ExpandedSet
+	for _, s := range sets {
+		if s.SetName == name {
+			result = append(result, s)
+		}
+	}
+	return result
+}

proxy/config/model_config.go

@@ -3,10 +3,23 @@ package config
 import (
 	"errors"
 	"runtime"
-	"slices"
-	"strings"
 )
 
+const (
+	MODEL_CONFIG_DEFAULT_TTL = -1
+)
+
+// TimeoutsConfig holds timeout settings for proxy connections
+// 0 = no timeout
+type TimeoutsConfig struct {
+	Connect        int `yaml:"connect"`
+	KeepAlive      int `yaml:"keepalive"`
+	ResponseHeader int `yaml:"responseHeader"`
+	TLSHandshake   int `yaml:"tlsHandshake"`
+	ExpectContinue int `yaml:"expectContinue"`
+	IdleConn       int `yaml:"idleConn"`
+}
+
 type ModelConfig struct {
 	Cmd           string   `yaml:"cmd"`
 	CmdStop       string   `yaml:"cmdStop"`
@@ -38,6 +51,9 @@ type ModelConfig struct {
 
 	// override global setting
 	SendLoadingState *bool `yaml:"sendLoadingState"`
+
+	// Timeout settings for proxy connections
+	Timeouts TimeoutsConfig `yaml:"timeouts"`
 }
 
 func (m *ModelConfig) UnmarshalYAML(unmarshal func(interface{}) error) error {
@@ -49,12 +65,22 @@ func (m *ModelConfig) UnmarshalYAML(unmarshal func(interface{}) error) error {
 		Aliases:          []string{},
 		Env:              []string{},
 		CheckEndpoint:    "/health",
-		UnloadAfter:      0,
+		UnloadAfter:      MODEL_CONFIG_DEFAULT_TTL, // use GlobalTTL
 		Unlisted:         false,
 		UseModelName:     "",
 		ConcurrencyLimit: 0,
 		Name:             "",
 		Description:      "",
+
+		// matches http.DefaultTransport
+		Timeouts: TimeoutsConfig{
+			Connect:        30,
+			KeepAlive:      30,
+			ResponseHeader: 0,
+			TLSHandshake:   10,
+			ExpectContinue: 1,
+			IdleConn:       90,
+		},
 	}
 
 	// the default cmdStop to taskkill /f /t /pid ${PID}
@@ -74,16 +100,15 @@ func (m *ModelConfig) SanitizedCommand() ([]string, error) {
 	return SanitizeCommand(m.Cmd)
 }
 
-// ModelFilters see issue #174
+// ModelFilters embeds Filters and adds legacy support for strip_params field
+// See issue #174
 type ModelFilters struct {
-	StripParams string `yaml:"stripParams"`
+	Filters `yaml:",inline"`
 }
 
 func (m *ModelFilters) UnmarshalYAML(unmarshal func(interface{}) error) error {
 	type rawModelFilters ModelFilters
-	defaults := rawModelFilters{
-		StripParams: "",
-	}
+	defaults := rawModelFilters{}
 
 	if err := unmarshal(&defaults); err != nil {
 		return err
@@ -104,25 +129,8 @@ func (m *ModelFilters) UnmarshalYAML(unmarshal func(interface{}) error) error {
 	return nil
 }
 
+// SanitizedStripParams wraps Filters.SanitizedStripParams for backwards compatibility
+// Returns ([]string, error) to match existing API
 func (f ModelFilters) SanitizedStripParams() ([]string, error) {
-	if f.StripParams == "" {
-		return nil, nil
-	}
-
-	params := strings.Split(f.StripParams, ",")
-	cleaned := make([]string, 0, len(params))
-	seen := make(map[string]bool)
-
-	for _, param := range params {
-		trimmed := strings.TrimSpace(param)
-		if trimmed == "model" || trimmed == "" || seen[trimmed] {
-			continue
-		}
-		seen[trimmed] = true
-		cleaned = append(cleaned, trimmed)
-	}
-
-	// sort cleaned
-	slices.Sort(cleaned)
-	return cleaned, nil
+	return f.Filters.SanitizedStripParams(), nil
 }

proxy/config/model_config_test.go

@@ -72,3 +72,101 @@ models:
 		assert.True(t, *config.Models["model2"].SendLoadingState)
 	}
 }
+
+func TestConfig_SetParamsByIDAutoAlias(t *testing.T) {
+	content := `
+models:
+  model1:
+    cmd: path/to/cmd --port ${PORT}
+    filters:
+      setParamsByID:
+        "${MODEL_ID}:high":
+          reasoning_effort: high
+        "${MODEL_ID}:low":
+          reasoning_effort: low
+`
+	cfg, err := LoadConfigFromReader(strings.NewReader(content))
+	assert.NoError(t, err)
+
+	// Keys (other than the model's own ID) should be registered as aliases
+	realName, found := cfg.RealModelName("model1:high")
+	assert.True(t, found, "model1:high should be an auto-registered alias")
+	assert.Equal(t, "model1", realName)
+
+	realName, found = cfg.RealModelName("model1:low")
+	assert.True(t, found, "model1:low should be an auto-registered alias")
+	assert.Equal(t, "model1", realName)
+
+	// Auto-aliases should also appear in modelConfig.Aliases
+	aliases := cfg.Models["model1"].Aliases
+	assert.Contains(t, aliases, "model1:high")
+	assert.Contains(t, aliases, "model1:low")
+}
+
+func TestConfig_SetParamsByIDAutoAliasConflictWithModelID(t *testing.T) {
+	content := `
+models:
+  model1:
+    cmd: path/to/cmd --port ${PORT}
+    filters:
+      setParamsByID:
+        model2:
+          reasoning_effort: high
+  model2:
+    cmd: path/to/cmd --port ${PORT}
+`
+	_, err := LoadConfigFromReader(strings.NewReader(content))
+	assert.ErrorContains(t, err, "conflicts with an existing model ID")
+}
+
+func TestConfig_SetParamsByIDAutoAliasConflictWithOtherModel(t *testing.T) {
+	content := `
+models:
+  model1:
+    cmd: path/to/cmd --port ${PORT}
+    filters:
+      setParamsByID:
+        "shared-alias":
+          reasoning_effort: high
+  model2:
+    cmd: path/to/cmd --port ${PORT}
+    filters:
+      setParamsByID:
+        "shared-alias":
+          reasoning_effort: low
+`
+	_, err := LoadConfigFromReader(strings.NewReader(content))
+	assert.ErrorContains(t, err, "duplicate alias")
+}
+
+func TestConfig_ModelFiltersWithSetParams(t *testing.T) {
+	content := `
+models:
+  model1:
+    cmd: path/to/cmd --port ${PORT}
+    filters:
+      stripParams: "top_k"
+      setParams:
+        temperature: 0.7
+        top_p: 0.9
+        stop:
+          - "<|end|>"
+          - "<|stop|>"
+`
+	config, err := LoadConfigFromReader(strings.NewReader(content))
+	assert.NoError(t, err)
+
+	modelConfig := config.Models["model1"]
+
+	// Check stripParams
+	stripParams, err := modelConfig.Filters.SanitizedStripParams()
+	assert.NoError(t, err)
+	assert.Equal(t, []string{"top_k"}, stripParams)
+
+	// Check setParams
+	setParams, keys := modelConfig.Filters.SanitizedSetParams()
+	assert.NotNil(t, setParams)
+	assert.Equal(t, []string{"stop", "temperature", "top_p"}, keys)
+	assert.Equal(t, 0.7, setParams["temperature"])
+	assert.Equal(t, 0.9, setParams["top_p"])
+}

proxy/config/peer.go

@@ -11,14 +11,30 @@ type PeerConfig struct {
 	ProxyURL *url.URL `yaml:"-"`
 	ApiKey   string   `yaml:"apiKey"`
 	Models   []string `yaml:"models"`
+	Filters  Filters  `yaml:"filters"`
+
+	// Timeout settings for proxy connections
+	Timeouts TimeoutsConfig `yaml:"timeouts"`
 }
 
 func (c *PeerConfig) UnmarshalYAML(unmarshal func(interface{}) error) error {
 	type rawPeerConfig PeerConfig
 	defaults := rawPeerConfig{
-		Proxy:  "",
-		ApiKey: "",
-		Models: []string{},
+		Proxy:   "",
+		ApiKey:  "",
+		Models:  []string{},
+		Filters: Filters{},
+
+		// mostly matches http.DefaultTransport but with a 60s ResponseHeader timeout
+		// to match the pre PR #619 functionality
+		Timeouts: TimeoutsConfig{
+			Connect:        30,
+			KeepAlive:      30,
+			ResponseHeader: 60,
+			TLSHandshake:   10,
+			ExpectContinue: 1,
+			IdleConn:       90,
+		},
 	}
 
 	if err := unmarshal(&defaults); err != nil {

proxy/config/peer_test.go

@@ -137,3 +137,73 @@ func searchSubstring(s, substr string) bool {
 	}
 	return false
 }
+
+func TestPeerConfig_WithFilters(t *testing.T) {
+	yamlData := `
+proxy: https://openrouter.ai/api
+apiKey: sk-test
+models:
+  - model_a
+filters:
+  setParams:
+    temperature: 0.7
+    provider:
+      data_collection: deny
+`
+	var config PeerConfig
+	err := yaml.Unmarshal([]byte(yamlData), &config)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	if config.Filters.SetParams == nil {
+		t.Fatal("Filters.SetParams should not be nil")
+	}
+
+	if config.Filters.SetParams["temperature"] != 0.7 {
+		t.Errorf("expected temperature 0.7, got %v", config.Filters.SetParams["temperature"])
+	}
+
+	provider, ok := config.Filters.SetParams["provider"].(map[string]any)
+	if !ok {
+		t.Fatal("provider should be a map")
+	}
+	if provider["data_collection"] != "deny" {
+		t.Errorf("expected data_collection deny, got %v", provider["data_collection"])
+	}
+}
+
+func TestPeerConfig_WithBothFilters(t *testing.T) {
+	yamlData := `
+proxy: https://openrouter.ai/api
+apiKey: sk-test
+models:
+  - model_a
+filters:
+  stripParams: "temperature, top_p"
+  setParams:
+    max_tokens: 1000
+`
+	var config PeerConfig
+	err := yaml.Unmarshal([]byte(yamlData), &config)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	// Check stripParams
+	stripParams := config.Filters.SanitizedStripParams()
+	if len(stripParams) != 2 {
+		t.Errorf("expected 2 strip params, got %d", len(stripParams))
+	}
+	if stripParams[0] != "temperature" || stripParams[1] != "top_p" {
+		t.Errorf("unexpected strip params: %v", stripParams)
+	}
+
+	// Check setParams
+	if config.Filters.SetParams == nil {
+		t.Fatal("Filters.SetParams should not be nil")
+	}
+	if config.Filters.SetParams["max_tokens"] != 1000 {
+		t.Errorf("expected max_tokens 1000, got %v", config.Filters.SetParams["max_tokens"])
+	}
+}

proxy/configwatcher/watcher.go

@@ -0,0 +1,85 @@
+// Package configwatcher provides a simple cross-platform file watcher based
+// on os.Stat polling. It works correctly inside Docker containers where the
+// config file is bind-mounted as an individual file, and for k8s ConfigMap
+// projections (which present the file as a symlink to an atomically swapped
+// target) — both cases where inotify-based watchers are unreliable.
+package configwatcher
+
+import (
+	"context"
+	"errors"
+	"io/fs"
+	"log"
+	"os"
+	"time"
+)
+
+const DefaultInterval = 2 * time.Second
+
+type Watcher struct {
+	Path     string
+	Interval time.Duration
+	OnChange func()
+}
+
+type snapshot struct {
+	exists  bool
+	modTime time.Time
+	size    int64
+}
+
+// Run blocks until ctx is canceled. It polls Path on Interval and invokes
+// OnChange whenever the file's modification time or size changes, or when
+// the file reappears after being missing. The baseline poll establishes
+// initial state and does not fire OnChange.
+func (w *Watcher) Run(ctx context.Context) {
+	interval := w.Interval
+	if interval <= 0 {
+		interval = DefaultInterval
+	}
+
+	prev := stat(w.Path)
+
+	ticker := time.NewTicker(interval)
+	defer ticker.Stop()
+
+	for {
+		select {
+		case <-ctx.Done():
+			return
+		case <-ticker.C:
+			cur := stat(w.Path)
+			if changed(prev, cur) && w.OnChange != nil {
+				w.OnChange()
+			}
+			prev = cur
+		}
+	}
+}
+
+func stat(path string) snapshot {
+	fi, err := os.Stat(path)
+	if err != nil {
+		if !errors.Is(err, fs.ErrNotExist) {
+			log.Printf("configwatcher: stat %s: %v", path, err)
+		}
+		return snapshot{}
+	}
+	return snapshot{
+		exists:  true,
+		modTime: fi.ModTime(),
+		size:    fi.Size(),
+	}
+}
+
+func changed(prev, cur snapshot) bool {
+	// Present → missing: stay quiet (likely a transient rename-style write).
+	// Missing → present: fire so we reload as soon as the file comes back.
+	if !cur.exists {
+		return false
+	}
+	if !prev.exists {
+		return true
+	}
+	return !prev.modTime.Equal(cur.modTime) || prev.size != cur.size
+}

proxy/configwatcher/watcher_test.go

@@ -0,0 +1,191 @@
+package configwatcher
+
+import (
+	"context"
+	"os"
+	"path/filepath"
+	"runtime"
+	"sync/atomic"
+	"testing"
+	"time"
+
+	"github.com/stretchr/testify/require"
+)
+
+const testInterval = 25 * time.Millisecond
+
+// startWatcher launches w.Run in a goroutine and returns a function that
+// cancels the context and waits for Run to return.
+func startWatcher(t *testing.T, w *Watcher) func() {
+	t.Helper()
+	ctx, cancel := context.WithCancel(context.Background())
+	done := make(chan struct{})
+	go func() {
+		w.Run(ctx)
+		close(done)
+	}()
+	return func() {
+		cancel()
+		select {
+		case <-done:
+		case <-time.After(2 * time.Second):
+			t.Fatal("watcher did not stop within 2s of cancel")
+		}
+	}
+}
+
+// waitForCount blocks until counter reaches want or timeout elapses.
+func waitForCount(t *testing.T, counter *int64, want int64, timeout time.Duration) bool {
+	t.Helper()
+	deadline := time.Now().Add(timeout)
+	for time.Now().Before(deadline) {
+		if atomic.LoadInt64(counter) >= want {
+			return true
+		}
+		time.Sleep(5 * time.Millisecond)
+	}
+	return false
+}
+
+func TestWatcher_NoFireOnBaseline(t *testing.T) {
+	dir := t.TempDir()
+	path := filepath.Join(dir, "config.yaml")
+	require.NoError(t, os.WriteFile(path, []byte("a"), 0o644))
+
+	var n int64
+	stop := startWatcher(t, &Watcher{
+		Path:     path,
+		Interval: testInterval,
+		OnChange: func() { atomic.AddInt64(&n, 1) },
+	})
+	defer stop()
+
+	time.Sleep(testInterval * 5)
+	require.Equal(t, int64(0), atomic.LoadInt64(&n), "baseline poll must not fire")
+}
+
+func TestWatcher_DetectsModTimeChange(t *testing.T) {
+	dir := t.TempDir()
+	path := filepath.Join(dir, "config.yaml")
+	require.NoError(t, os.WriteFile(path, []byte("a"), 0o644))
+
+	// Force a known baseline mtime.
+	base := time.Now().Add(-1 * time.Hour).Truncate(time.Second)
+	require.NoError(t, os.Chtimes(path, base, base))
+
+	var n int64
+	stop := startWatcher(t, &Watcher{
+		Path:     path,
+		Interval: testInterval,
+		OnChange: func() { atomic.AddInt64(&n, 1) },
+	})
+	defer stop()
+
+	// Let the baseline settle.
+	time.Sleep(testInterval * 2)
+
+	// Bump mtime well above the baseline so low-resolution filesystems still notice.
+	require.NoError(t, os.Chtimes(path, base.Add(10*time.Second), base.Add(10*time.Second)))
+
+	require.True(t, waitForCount(t, &n, 1, time.Second), "callback should fire after mtime change")
+}
+
+func TestWatcher_DetectsSizeChangeWithSameModTime(t *testing.T) {
+	dir := t.TempDir()
+	path := filepath.Join(dir, "config.yaml")
+	require.NoError(t, os.WriteFile(path, []byte("a"), 0o644))
+
+	fi, err := os.Stat(path)
+	require.NoError(t, err)
+	originalMtime := fi.ModTime()
+
+	var n int64
+	stop := startWatcher(t, &Watcher{
+		Path:     path,
+		Interval: testInterval,
+		OnChange: func() { atomic.AddInt64(&n, 1) },
+	})
+	defer stop()
+	time.Sleep(testInterval * 2)
+
+	require.NoError(t, os.WriteFile(path, []byte("aaaaa"), 0o644))
+	// Reset mtime back to the original so size is the only signal.
+	require.NoError(t, os.Chtimes(path, originalMtime, originalMtime))
+
+	require.True(t, waitForCount(t, &n, 1, time.Second), "callback should fire on size change")
+}
+
+func TestWatcher_SymlinkTargetSwap(t *testing.T) {
+	dir := t.TempDir()
+	targetA := filepath.Join(dir, "targetA")
+	targetB := filepath.Join(dir, "targetB")
+	link := filepath.Join(dir, "config.yaml")
+
+	require.NoError(t, os.WriteFile(targetA, []byte("AAAA"), 0o644))
+	require.NoError(t, os.WriteFile(targetB, []byte("BBBBBBBB"), 0o644))
+
+	if err := os.Symlink(targetA, link); err != nil {
+		if runtime.GOOS == "windows" {
+			t.Skipf("symlink creation requires privilege on Windows: %v", err)
+		}
+		t.Fatalf("os.Symlink: %v", err)
+	}
+
+	var n int64
+	stop := startWatcher(t, &Watcher{
+		Path:     link,
+		Interval: testInterval,
+		OnChange: func() { atomic.AddInt64(&n, 1) },
+	})
+	defer stop()
+	time.Sleep(testInterval * 2)
+
+	// Atomic symlink swap (k8s ConfigMap pattern): create new symlink at a
+	// temp name, then rename over the existing one.
+	tmpLink := filepath.Join(dir, "config.yaml.tmp")
+	require.NoError(t, os.Symlink(targetB, tmpLink))
+	require.NoError(t, os.Rename(tmpLink, link))
+
+	require.True(t, waitForCount(t, &n, 1, time.Second), "callback should fire after symlink target swap")
+}
+
+func TestWatcher_FileMissingThenReturns(t *testing.T) {
+	dir := t.TempDir()
+	path := filepath.Join(dir, "config.yaml")
+	require.NoError(t, os.WriteFile(path, []byte("a"), 0o644))
+
+	var n int64
+	stop := startWatcher(t, &Watcher{
+		Path:     path,
+		Interval: testInterval,
+		OnChange: func() { atomic.AddInt64(&n, 1) },
+	})
+	defer stop()
+	time.Sleep(testInterval * 2)
+
+	require.NoError(t, os.Remove(path))
+	time.Sleep(testInterval * 3)
+	require.Equal(t, int64(0), atomic.LoadInt64(&n), "removal alone must not fire")
+
+	require.NoError(t, os.WriteFile(path, []byte("b"), 0o644))
+	require.True(t, waitForCount(t, &n, 1, time.Second), "callback should fire when file returns")
+}
+
+func TestWatcher_ContextCancelStopsRun(t *testing.T) {
+	dir := t.TempDir()
+	path := filepath.Join(dir, "config.yaml")
+	require.NoError(t, os.WriteFile(path, []byte("a"), 0o644))
+
+	w := &Watcher{Path: path, Interval: testInterval}
+	ctx, cancel := context.WithCancel(context.Background())
+	done := make(chan struct{})
+	go func() { w.Run(ctx); close(done) }()
+
+	time.Sleep(testInterval * 2)
+	cancel()
+	select {
+	case <-done:
+	case <-time.After(2 * time.Second):
+		t.Fatal("Run did not return within 2s of cancel")
+	}
+}

proxy/events.go

@@ -8,6 +8,7 @@ const ConfigFileChangedEventID = 0x03
 const LogDataEventID = 0x04
 const TokenMetricsEventID = 0x05
 const ModelPreloadedEventID = 0x06
+const InFlightRequestsEventID = 0x07
 
 type ProcessStateChangeEvent struct {
 	ProcessName string
@@ -58,3 +59,11 @@ type ModelPreloadedEvent struct {
 func (e ModelPreloadedEvent) Type() uint32 {
 	return ModelPreloadedEventID
 }
+
+type InFlightRequestsEvent struct {
+	Total int
+}
+
+func (e InFlightRequestsEvent) Type() uint32 {
+	return InFlightRequestsEventID
+}

proxy/helpers_test.go

@@ -1,15 +1,22 @@
 package proxy
 
 import (
+	"encoding/json"
 	"fmt"
+	"io"
+	"net/http"
 	"os"
 	"path/filepath"
 	"runtime"
+	"strings"
 	"sync"
 	"testing"
+	"time"
 
 	"github.com/gin-gonic/gin"
 	"github.com/mostlygeek/llama-swap/proxy/config"
+	"github.com/stretchr/testify/require"
+	"github.com/tidwall/gjson"
 	"gopkg.in/yaml.v3"
 )
 
@@ -66,16 +73,30 @@ func getTestPort() int {
 	return port
 }
 
+// testConfigFromYAML substitutes {{RESPONDER}} with the simple-responder path and
+// loads through the real config pipeline (env vars, macros, port assignment, etc.)
+func testConfigFromYAML(t *testing.T, yamlTmpl string) config.Config {
+	t.Helper()
+	yamlStr := strings.ReplaceAll(yamlTmpl, "{{RESPONDER}}", filepath.ToSlash(simpleResponderPath))
+	cfg, err := config.LoadConfigFromReader(strings.NewReader(yamlStr))
+	require.NoError(t, err)
+	return cfg
+}
+
 func getTestSimpleResponderConfig(expectedMessage string) config.ModelConfig {
 	return getTestSimpleResponderConfigPort(expectedMessage, getTestPort())
 }
 
 func getTestSimpleResponderConfigPort(expectedMessage string, port int) config.ModelConfig {
+	// Convert path to forward slashes for cross-platform compatibility
+	// Windows handles forward slashes in paths correctly
+	cmdPath := filepath.ToSlash(simpleResponderPath)
+
 	// Create a YAML string with just the values we want to set
 	yamlStr := fmt.Sprintf(`
 cmd: '%s --port %d --silent --respond %s'
 proxy: "http://127.0.0.1:%d"
-`, simpleResponderPath, port, expectedMessage, port)
+`, cmdPath, port, expectedMessage, port)
 
 	var cfg config.ModelConfig
 	if err := yaml.Unmarshal([]byte(yamlStr), &cfg); err != nil {
@@ -84,3 +105,188 @@ proxy: "http://127.0.0.1:%d"
 
 	return cfg
 }
+
+// injectTestHandlers sets a testHandler on every Process in every ProcessGroup
+// of the given ProxyManager, bypassing subprocess launches. modelResponses maps
+// model IDs to their respond strings; if a model ID is not in the map, the model
+// ID itself is used.
+func injectTestHandlers(pm *ProxyManager, modelResponses map[string]string) {
+	for _, pg := range pm.processGroups {
+		for modelID, process := range pg.processes {
+			respond := modelID
+			if r, ok := modelResponses[modelID]; ok {
+				respond = r
+			}
+			process.testHandler = newTestHandler(respond)
+		}
+	}
+}
+
+// newTestHandler returns an http.Handler that mimics simple-responder's API.
+// It supports the endpoints that routing tests depend on, without launching
+// any subprocess or binding any port.
+func newTestHandler(respond string) http.Handler {
+	mux := http.NewServeMux()
+
+	mux.HandleFunc("/v1/chat/completions", func(w http.ResponseWriter, r *http.Request) {
+		bodyBytes, _ := io.ReadAll(r.Body)
+		isStreaming := r.URL.Query().Get("stream") == "true"
+
+		if wait := r.URL.Query().Get("wait"); wait != "" {
+			if d, err := time.ParseDuration(wait); err == nil {
+				time.Sleep(d)
+			}
+		}
+
+		if isStreaming {
+			w.Header().Set("Content-Type", "text/event-stream")
+			w.Header().Set("Cache-Control", "no-cache")
+			w.Header().Set("Connection", "keep-alive")
+			flusher := w.(http.Flusher)
+
+			for i := 0; i < 10; i++ {
+				data, _ := json.Marshal(map[string]any{
+					"created": time.Now().Unix(),
+					"choices": []map[string]any{
+						{"index": 0, "delta": map[string]any{"content": "asdf"}, "finish_reason": nil},
+					},
+				})
+				fmt.Fprintf(w, "event: message\ndata: %s\n\n", data)
+				flusher.Flush()
+			}
+
+			finalData, _ := json.Marshal(map[string]any{
+				"usage": map[string]any{
+					"completion_tokens": 10, "prompt_tokens": 25, "total_tokens": 35,
+				},
+				"timings": map[string]any{
+					"prompt_n": 25, "prompt_ms": 13, "predicted_n": 10,
+					"predicted_ms": 17, "predicted_per_second": 10,
+				},
+			})
+			fmt.Fprintf(w, "event: message\ndata: %s\n\n", finalData)
+			flusher.Flush()
+
+			fmt.Fprintf(w, "event: message\ndata: [DONE]\n\n")
+			flusher.Flush()
+		} else {
+			w.Header().Set("Content-Type", "application/json")
+			json.NewEncoder(w).Encode(map[string]any{
+				"responseMessage":  respond,
+				"h_content_length": r.Header.Get("Content-Length"),
+				"request_body":     string(bodyBytes),
+				"usage": map[string]any{
+					"completion_tokens": 10, "prompt_tokens": 25, "total_tokens": 35,
+				},
+				"timings": map[string]any{
+					"prompt_n": 25, "prompt_ms": 13, "predicted_n": 10,
+					"predicted_ms": 17, "predicted_per_second": 10,
+				},
+			})
+		}
+	})
+
+	mux.HandleFunc("/v1/audio/speech", func(w http.ResponseWriter, r *http.Request) {
+		body, _ := io.ReadAll(r.Body)
+		modelName := gjson.GetBytes(body, "model").String()
+		if modelName != respond {
+			w.WriteHeader(http.StatusBadRequest)
+			json.NewEncoder(w).Encode(map[string]string{"error": fmt.Sprintf("Invalid model: %s, expected: %s", modelName, respond)})
+			return
+		}
+		json.NewEncoder(w).Encode(map[string]string{"message": "ok"})
+	})
+
+	mux.HandleFunc("/v1/completions", func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		json.NewEncoder(w).Encode(map[string]any{
+			"responseMessage": respond,
+			"usage": map[string]any{
+				"completion_tokens": 10, "prompt_tokens": 25, "total_tokens": 35,
+			},
+		})
+	})
+
+	mux.HandleFunc("/completion", func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		json.NewEncoder(w).Encode(map[string]any{
+			"responseMessage": respond,
+			"usage": map[string]any{
+				"completion_tokens": 10, "prompt_tokens": 25, "total_tokens": 35,
+			},
+		})
+	})
+
+	mux.HandleFunc("/v1/audio/transcriptions", func(w http.ResponseWriter, r *http.Request) {
+		if err := r.ParseMultipartForm(10 << 20); err != nil {
+			w.WriteHeader(http.StatusBadRequest)
+			json.NewEncoder(w).Encode(map[string]string{"error": fmt.Sprintf("Error parsing multipart form: %s", err)})
+			return
+		}
+		model := r.FormValue("model")
+		if model == "" {
+			w.WriteHeader(http.StatusBadRequest)
+			json.NewEncoder(w).Encode(map[string]string{"error": "Missing model parameter"})
+			return
+		}
+		file, _, err := r.FormFile("file")
+		if err != nil {
+			w.WriteHeader(http.StatusBadRequest)
+			json.NewEncoder(w).Encode(map[string]string{"error": fmt.Sprintf("Error getting file: %s", err)})
+			return
+		}
+		fileBytes, _ := io.ReadAll(file)
+		file.Close()
+		json.NewEncoder(w).Encode(map[string]any{
+			"text":             fmt.Sprintf("The length of the file is %d bytes", len(fileBytes)),
+			"model":            model,
+			"h_content_type":   r.Header.Get("Content-Type"),
+			"h_content_length": r.Header.Get("Content-Length"),
+		})
+	})
+
+	mux.HandleFunc("/v1/audio/voices", func(w http.ResponseWriter, r *http.Request) {
+		model := r.URL.Query().Get("model")
+		json.NewEncoder(w).Encode(map[string]any{
+			"voices": []string{"voice1"}, "model": model,
+		})
+	})
+
+	mux.HandleFunc("/test", func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "text/plain")
+		fmt.Fprint(w, respond)
+	})
+
+	mux.HandleFunc("/", func(w http.ResponseWriter, r *http.Request) {
+		if r.URL.Path != "/" {
+			http.NotFound(w, r)
+			return
+		}
+		w.Header().Set("Content-Type", "text/plain")
+		fmt.Fprintf(w, "%s %s", r.Method, r.URL.Path)
+	})
+
+	mux.HandleFunc("/sdapi/v1/txt2img", func(w http.ResponseWriter, r *http.Request) {
+		body, _ := io.ReadAll(r.Body)
+		modelName := gjson.GetBytes(body, "model").String()
+		json.NewEncoder(w).Encode(map[string]any{
+			"model": modelName, "images": []string{},
+		})
+	})
+
+	mux.HandleFunc("/sdapi/v1/img2img", func(w http.ResponseWriter, r *http.Request) {
+		body, _ := io.ReadAll(r.Body)
+		modelName := gjson.GetBytes(body, "model").String()
+		json.NewEncoder(w).Encode(map[string]any{
+			"model": modelName, "images": []string{},
+		})
+	})
+
+	mux.HandleFunc("/sdapi/v1/loras", func(w http.ResponseWriter, r *http.Request) {
+		json.NewEncoder(w).Encode(map[string]any{
+			"loras": []string{},
+		})
+	})
+
+	return mux
+}

proxy/matrix.go

@@ -0,0 +1,329 @@
+package proxy
+
+import (
+	"fmt"
+	"net/http"
+	"slices"
+	"sort"
+	"sync"
+
+	"github.com/mostlygeek/llama-swap/proxy/config"
+)
+
+// MatrixSolver contains pure swap-decision logic with no Process dependencies.
+// It is safe for concurrent reads after construction.
+type MatrixSolver struct {
+	expandedSets []config.ExpandedSet // all valid model combinations
+	evictCosts   map[string]int       // real model name -> eviction cost (default 1)
+	modelToSets  map[string][]int     // model name -> indices into expandedSets
+}
+
+// NewMatrixSolver builds a solver from expanded sets and eviction costs.
+func NewMatrixSolver(expandedSets []config.ExpandedSet, evictCosts map[string]int) *MatrixSolver {
+	modelToSets := make(map[string][]int)
+	for i, es := range expandedSets {
+		for _, model := range es.Models {
+			modelToSets[model] = append(modelToSets[model], i)
+		}
+	}
+
+	return &MatrixSolver{
+		expandedSets: expandedSets,
+		evictCosts:   evictCosts,
+		modelToSets:  modelToSets,
+	}
+}
+
+// SolveResult describes what the solver decided.
+type SolveResult struct {
+	Evict     []string // running models that must be stopped
+	TargetSet []string // the chosen set of models (for informational purposes)
+	SetName   string   // name of the chosen set
+	DSL       string   // original DSL expression for the chosen set
+	TotalCost int      // total eviction cost
+}
+
+// Solve determines which models to evict when a model is requested.
+//
+// Algorithm:
+//  1. If requestedModel is already running, no eviction needed.
+//  2. Find all sets containing requestedModel.
+//  3. If no sets found, the model runs alone; evict all running models.
+//  4. For each candidate set, compute cost = sum of evict_costs for running
+//     models NOT in that set.
+//  5. Pick lowest cost. Ties broken by definition order (index in expandedSets).
+//  6. Return models to evict and the chosen set.
+func (s *MatrixSolver) Solve(requestedModel string, runningModels []string) (SolveResult, error) {
+	// If already running, nothing to do (but fill in set info for logging)
+	if slices.Contains(runningModels, requestedModel) {
+		setName, dsl := s.findMatchingSet(requestedModel, runningModels)
+		return SolveResult{
+			TargetSet: runningModels,
+			SetName:   setName,
+			DSL:       dsl,
+		}, nil
+	}
+
+	candidateIndices := s.modelToSets[requestedModel]
+
+	// Model not in any set: runs alone, evict everything
+	if len(candidateIndices) == 0 {
+		evict := make([]string, len(runningModels))
+		copy(evict, runningModels)
+		return SolveResult{
+			Evict:     evict,
+			TargetSet: []string{requestedModel},
+		}, nil
+	}
+
+	// Find the cheapest candidate set
+	bestCost := -1
+	bestIdx := -1
+
+	for _, idx := range candidateIndices {
+		setModels := s.expandedSets[idx].Models
+		cost := 0
+		for _, running := range runningModels {
+			if !slices.Contains(setModels, running) {
+				cost += s.evictCost(running)
+			}
+		}
+
+		if bestCost < 0 || cost < bestCost || (cost == bestCost && idx < bestIdx) {
+			bestCost = cost
+			bestIdx = idx
+		}
+	}
+
+	// Determine which running models to evict
+	chosen := s.expandedSets[bestIdx]
+	var evict []string
+	for _, running := range runningModels {
+		if !slices.Contains(chosen.Models, running) {
+			evict = append(evict, running)
+		}
+	}
+
+	return SolveResult{
+		Evict:     evict,
+		TargetSet: chosen.Models,
+		SetName:   chosen.SetName,
+		DSL:       chosen.DSL,
+		TotalCost: bestCost,
+	}, nil
+}
+
+// findMatchingSet finds the expanded set that contains all running models.
+// Returns the set name and DSL, or empty strings if no match.
+func (s *MatrixSolver) findMatchingSet(requestedModel string, runningModels []string) (string, string) {
+	for _, idx := range s.modelToSets[requestedModel] {
+		set := s.expandedSets[idx]
+		allInSet := true
+		for _, m := range runningModels {
+			if !slices.Contains(set.Models, m) {
+				allInSet = false
+				break
+			}
+		}
+		if allInSet {
+			return set.SetName, set.DSL
+		}
+	}
+	return "", ""
+}
+
+func (s *MatrixSolver) evictCost(model string) int {
+	if cost, ok := s.evictCosts[model]; ok {
+		return cost
+	}
+	return 1
+}
+
+// Matrix manages processes using solver-based swap logic.
+type Matrix struct {
+	sync.Mutex
+	solver         *MatrixSolver
+	processes      map[string]*Process // all processes keyed by real model name
+	config         config.Config
+	proxyLogger    *LogMonitor
+	upstreamLogger *LogMonitor
+
+	// inflight tracks ProxyRequest calls that have released m.Lock but may
+	// not yet have incremented Process.inFlightRequests. A concurrent
+	// request that needs to evict models waits for inflight to drain under
+	// m.Lock before stopping anything. Without this, a request that
+	// released m.Lock but has not yet reached Process.inFlightRequests.Add(1)
+	// races with Stop()'s Wait() and can be killed mid-request.
+	inflight sync.WaitGroup
+
+	// testDelayFastPath is a test-only hook invoked in the no-eviction path
+	// after m.Lock is released but before the request is dispatched to
+	// Process.ProxyRequest. Tests use it to park a request at the exact
+	// race window to deterministically reproduce the race.
+	testDelayFastPath func()
+}
+
+// NewMatrix creates a Matrix from config. It creates a Process for every
+// model defined in the config (any model can run alone even if not in a set).
+func NewMatrix(cfg config.Config, proxyLogger, upstreamLogger *LogMonitor) *Matrix {
+	processes := make(map[string]*Process)
+	for modelID, modelConfig := range cfg.Models {
+		processLogger := NewLogMonitorWriter(upstreamLogger)
+		process := NewProcess(modelID, cfg.HealthCheckTimeout, modelConfig, processLogger, proxyLogger)
+		processes[modelID] = process
+	}
+
+	evictCosts := cfg.Matrix.ResolvedEvictCosts()
+
+	return &Matrix{
+		solver:         NewMatrixSolver(cfg.ExpandedSets, evictCosts),
+		processes:      processes,
+		config:         cfg,
+		proxyLogger:    proxyLogger,
+		upstreamLogger: upstreamLogger,
+	}
+}
+
+// ProxyRequest handles the swap logic and proxies the request to the model.
+func (m *Matrix) ProxyRequest(modelID string, w http.ResponseWriter, r *http.Request) error {
+	process, ok := m.processes[modelID]
+	if !ok {
+		return fmt.Errorf("model %s not found in matrix", modelID)
+	}
+
+	m.Lock()
+	running := m.runningModels()
+	result, err := m.solver.Solve(modelID, running)
+	if err != nil {
+		m.Unlock()
+		return fmt.Errorf("matrix solver error: %w", err)
+	}
+
+	// Log solver decision
+	if len(result.Evict) > 0 {
+		m.proxyLogger.Infof("Matrix: model=%s set=%s dsl=%q evict=%v target=%v cost=%d",
+			modelID, result.SetName, result.DSL, result.Evict, result.TargetSet, result.TotalCost)
+	} else if len(running) == 0 {
+		m.proxyLogger.Infof("Matrix: model=%s starting (no models running)", modelID)
+	} else {
+		m.proxyLogger.Debugf("Matrix: model=%s already running in set=%s dsl=%q", modelID, result.SetName, result.DSL)
+	}
+
+	// Evict models that need to be stopped
+	if len(result.Evict) > 0 {
+		// Wait for any in-flight ProxyRequest calls to register on their
+		// Process before stopping anything. Without this, a request that
+		// released m.Lock but has not yet incremented
+		// Process.inFlightRequests races with Stop() and can be killed
+		// mid-request.
+		m.inflight.Wait()
+
+		var wg sync.WaitGroup
+		for _, evictModel := range result.Evict {
+			if p, exists := m.processes[evictModel]; exists {
+				wg.Add(1)
+				go func(p *Process) {
+					defer wg.Done()
+					p.Stop()
+				}(p)
+			}
+		}
+		wg.Wait()
+	}
+
+	// Register this request in inflight before releasing m.Lock so a
+	// concurrent eviction will wait for it to complete.
+	m.inflight.Add(1)
+	defer m.inflight.Done()
+	isFastPath := len(result.Evict) == 0
+	m.Unlock()
+
+	if isFastPath && m.testDelayFastPath != nil {
+		m.testDelayFastPath()
+	}
+
+	// Proxy the request (Process handles on-demand start)
+	process.ProxyRequest(w, r)
+	return nil
+}
+
+// StopProcesses stops all running processes.
+func (m *Matrix) StopProcesses(strategy StopStrategy) {
+	m.Lock()
+	defer m.Unlock()
+
+	var wg sync.WaitGroup
+	for _, process := range m.processes {
+		wg.Add(1)
+		go func(p *Process) {
+			defer wg.Done()
+			switch strategy {
+			case StopImmediately:
+				p.StopImmediately()
+			default:
+				p.Stop()
+			}
+		}(process)
+	}
+	wg.Wait()
+}
+
+// StopProcess stops a single process by model ID.
+func (m *Matrix) StopProcess(modelID string, strategy StopStrategy) error {
+	process, ok := m.processes[modelID]
+	if !ok {
+		return fmt.Errorf("process not found for %s", modelID)
+	}
+
+	switch strategy {
+	case StopImmediately:
+		process.StopImmediately()
+	default:
+		process.Stop()
+	}
+	return nil
+}
+
+// Shutdown shuts down all processes.
+func (m *Matrix) Shutdown() {
+	var wg sync.WaitGroup
+	for _, process := range m.processes {
+		wg.Add(1)
+		go func(p *Process) {
+			defer wg.Done()
+			p.Shutdown()
+		}(process)
+	}
+	wg.Wait()
+}
+
+// RunningModels returns model names currently in an active (non-stopped) state.
+func (m *Matrix) RunningModels() []string {
+	m.Lock()
+	defer m.Unlock()
+	return m.runningModels()
+}
+
+// runningModels returns running model names (caller must hold lock).
+func (m *Matrix) runningModels() []string {
+	var running []string
+	for id, process := range m.processes {
+		if process.CurrentState() != StateStopped && process.CurrentState() != StateShutdown {
+			running = append(running, id)
+		}
+	}
+	sort.Strings(running)
+	return running
+}
+
+// GetProcess returns the Process for a model.
+func (m *Matrix) GetProcess(modelID string) (*Process, bool) {
+	p, ok := m.processes[modelID]
+	return p, ok
+}
+
+// HasModel returns true if the model is managed by this matrix.
+func (m *Matrix) HasModel(modelID string) bool {
+	_, ok := m.processes[modelID]
+	return ok
+}

proxy/matrix_test.go

@@ -0,0 +1,349 @@
+package proxy
+
+import (
+	"net/http"
+	"net/http/httptest"
+	"runtime"
+	"testing"
+	"time"
+
+	"github.com/mostlygeek/llama-swap/proxy/config"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+// Helper to build expanded sets for solver tests
+func makeExpandedSets(sets ...struct {
+	name   string
+	models []string
+}) []config.ExpandedSet {
+	var result []config.ExpandedSet
+	for _, s := range sets {
+		result = append(result, config.ExpandedSet{
+			SetName: s.name,
+			Models:  s.models,
+		})
+	}
+	return result
+}
+
+func es(name string, models ...string) struct {
+	name   string
+	models []string
+} {
+	return struct {
+		name   string
+		models []string
+	}{name, models}
+}
+
+func TestMatrixSolver_AlreadyRunning(t *testing.T) {
+	solver := NewMatrixSolver(
+		makeExpandedSets(es("s1", "a", "b")),
+		nil,
+	)
+
+	result, err := solver.Solve("a", []string{"a"})
+	require.NoError(t, err)
+	assert.Empty(t, result.Evict)
+	assert.Equal(t, []string{"a"}, result.TargetSet)
+	assert.Equal(t, "s1", result.SetName)
+}
+
+func TestMatrixSolver_NotInAnySet_RunsAlone(t *testing.T) {
+	solver := NewMatrixSolver(
+		makeExpandedSets(es("s1", "a", "b")),
+		nil,
+	)
+
+	// Model "c" not in any set
+	result, err := solver.Solve("c", []string{"a", "b"})
+	require.NoError(t, err)
+	assert.ElementsMatch(t, []string{"a", "b"}, result.Evict)
+	assert.Equal(t, []string{"c"}, result.TargetSet)
+}
+
+func TestMatrixSolver_NotInAnySet_NothingRunning(t *testing.T) {
+	solver := NewMatrixSolver(
+		makeExpandedSets(es("s1", "a", "b")),
+		nil,
+	)
+
+	result, err := solver.Solve("c", []string{})
+	require.NoError(t, err)
+	assert.Empty(t, result.Evict)
+	assert.Equal(t, []string{"c"}, result.TargetSet)
+}
+
+func TestMatrixSolver_SingleSet_EvictsNonMembers(t *testing.T) {
+	// Set: [a, b]. Request a when b and c are running.
+	solver := NewMatrixSolver(
+		makeExpandedSets(es("s1", "a", "b")),
+		nil,
+	)
+
+	result, err := solver.Solve("a", []string{"b", "c"})
+	require.NoError(t, err)
+	// c is not in the set, so it gets evicted. b is in the set, so it stays.
+	assert.Equal(t, []string{"c"}, result.Evict)
+	assert.Equal(t, []string{"a", "b"}, result.TargetSet)
+}
+
+func TestMatrixSolver_PicksLowestCost(t *testing.T) {
+	// Two sets containing model "a":
+	// s1: [a, v]  — if v is running, cost=0; if L is running, cost=30
+	// s2: [a, L]  — if L is running, cost=0; if v is running, cost=50
+	solver := NewMatrixSolver(
+		makeExpandedSets(
+			es("s1", "a", "v"),
+			es("s2", "a", "L"),
+		),
+		map[string]int{"v": 50, "L": 30},
+	)
+
+	// v is running. Switching to a:
+	// s1 cost: v is in s1, so 0
+	// s2 cost: v is NOT in s2, so 50
+	// => pick s1
+	result, err := solver.Solve("a", []string{"v"})
+	require.NoError(t, err)
+	assert.Empty(t, result.Evict)
+	assert.Equal(t, []string{"a", "v"}, result.TargetSet)
+
+	// L is running. Switching to a:
+	// s1 cost: L is NOT in s1, so 30
+	// s2 cost: L is in s2, so 0
+	// => pick s2
+	result, err = solver.Solve("a", []string{"L"})
+	require.NoError(t, err)
+	assert.Empty(t, result.Evict)
+	assert.Equal(t, []string{"a", "L"}, result.TargetSet)
+}
+
+func TestMatrixSolver_TieBreakingByDefinitionOrder(t *testing.T) {
+	// Two sets with identical cost. Definition order should win.
+	solver := NewMatrixSolver(
+		makeExpandedSets(
+			es("s1", "a", "x"),
+			es("s2", "a", "y"),
+		),
+		nil,
+	)
+
+	// Nothing running, both sets cost 0. s1 is first.
+	result, err := solver.Solve("a", []string{})
+	require.NoError(t, err)
+	assert.Empty(t, result.Evict)
+	assert.Equal(t, []string{"a", "x"}, result.TargetSet)
+}
+
+func TestMatrixSolver_EvictCostPreservesExpensive(t *testing.T) {
+	// Model "v" costs 50 to evict, "m" costs 1 (default).
+	// Sets: [g,v], [g,m]
+	// Running: v, m. Request g.
+	// s1=[g,v]: evict m (cost 1), keep v
+	// s2=[g,m]: evict v (cost 50), keep m
+	// => pick s1
+	solver := NewMatrixSolver(
+		makeExpandedSets(
+			es("s1", "g", "v"),
+			es("s2", "g", "m"),
+		),
+		map[string]int{"v": 50},
+	)
+
+	result, err := solver.Solve("g", []string{"v", "m"})
+	require.NoError(t, err)
+	assert.Equal(t, []string{"m"}, result.Evict)
+	assert.Equal(t, []string{"g", "v"}, result.TargetSet)
+}
+
+func TestMatrixSolver_NothingRunning(t *testing.T) {
+	solver := NewMatrixSolver(
+		makeExpandedSets(
+			es("s1", "g", "v"),
+			es("s2", "q", "v"),
+		),
+		nil,
+	)
+
+	result, err := solver.Solve("g", []string{})
+	require.NoError(t, err)
+	assert.Empty(t, result.Evict)
+	assert.Equal(t, []string{"g", "v"}, result.TargetSet)
+}
+
+// TestMatrix_ProxyRequestSwapRaceAgainstFastPath verifies that an eviction
+// cannot stop a process while an in-flight ProxyRequest for that process is
+// still in the [m.Unlock, Process.inFlightRequests.Add(1)] window. Without
+// matrix-level inflight tracking, the eviction's Stop() races with the
+// pending request and kills it mid-start.
+func TestMatrix_ProxyRequestSwapRaceAgainstFastPath(t *testing.T) {
+	cfg := config.Config{
+		HealthCheckTimeout: 15,
+		Models: map[string]config.ModelConfig{
+			"model1": getTestSimpleResponderConfig("model1"),
+			"model2": getTestSimpleResponderConfig("model2"),
+		},
+		ExpandedSets: []config.ExpandedSet{
+			{SetName: "s1", Models: []string{"model1"}},
+			{SetName: "s2", Models: []string{"model2"}},
+		},
+		Matrix: &config.MatrixConfig{},
+	}
+
+	m := NewMatrix(cfg, testLogger, testLogger)
+	defer m.StopProcesses(StopImmediately)
+
+	// Bypass real subprocesses so the test is fast and deterministic.
+	m.processes["model1"].testHandler = newTestHandler("model1")
+	m.processes["model2"].testHandler = newTestHandler("model2")
+
+	// Prime: run a request through model1 so it reaches StateReady and
+	// subsequent requests take the no-eviction path.
+	primeReq := httptest.NewRequest("POST", "/v1/chat/completions", nil)
+	primeW := httptest.NewRecorder()
+	require.NoError(t, m.ProxyRequest("model1", primeW, primeReq))
+	require.Equal(t, http.StatusOK, primeW.Code)
+	require.Equal(t, StateReady, m.processes["model1"].CurrentState())
+	require.Equal(t, StateStopped, m.processes["model2"].CurrentState())
+
+	// Install fast-path hook that signals arrival and waits for release.
+	// This parks R2 at the race window — after m.Lock is released but
+	// before Process.inFlightRequests.Add(1).
+	r2Reached := make(chan struct{})
+	r2Release := make(chan struct{})
+	m.testDelayFastPath = func() {
+		close(r2Reached)
+		<-r2Release
+	}
+
+	// R2: no-eviction request for model1. Will pause at the hook.
+	r2Done := make(chan struct{})
+	w2 := httptest.NewRecorder()
+	go func() {
+		defer close(r2Done)
+		req := httptest.NewRequest("POST", "/v1/chat/completions", nil)
+		assert.NoError(t, m.ProxyRequest("model1", w2, req))
+	}()
+
+	// Deterministically wait for R2 to reach the race window.
+	<-r2Reached
+
+	// R3: request for model2 which requires evicting model1. Must wait for
+	// R2 to finish before touching model1.
+	r3Done := make(chan struct{})
+	w3 := httptest.NewRecorder()
+	go func() {
+		defer close(r3Done)
+		req := httptest.NewRequest("POST", "/v1/chat/completions", nil)
+		assert.NoError(t, m.ProxyRequest("model2", w3, req))
+	}()
+
+	// Spin until R3 has acquired m.Lock and entered the eviction path. In
+	// the fixed code, R3 then blocks on m.inflight.Wait() while still
+	// holding the lock, so TryLock keeps failing.
+	for m.TryLock() {
+		m.Unlock()
+		runtime.Gosched()
+	}
+
+	// Bounded poll: give R3 a chance to demonstrate the bug by mutating
+	// state. In the fixed code R3 is blocked and nothing changes; in the
+	// buggy code R3 will Stop() model1 and start model2 within microseconds.
+	deadline := time.Now().Add(100 * time.Millisecond)
+	for time.Now().Before(deadline) {
+		if m.processes["model1"].CurrentState() != StateReady ||
+			m.processes["model2"].CurrentState() != StateStopped {
+			break
+		}
+		done := false
+		select {
+		case <-r3Done:
+			done = true
+		default:
+		}
+		if done {
+			break
+		}
+		runtime.Gosched()
+	}
+
+	// Invariant: R3 must be blocked while R2 is still in flight.
+	select {
+	case <-r3Done:
+		t.Fatal("eviction completed while in-flight request was still pending — race not prevented")
+	default:
+	}
+	assert.Equal(t, StateReady, m.processes["model1"].CurrentState(),
+		"model1 must stay Ready while an in-flight request is pending")
+	assert.Equal(t, StateStopped, m.processes["model2"].CurrentState(),
+		"model2 must not be started until R2 finishes and model1 is evicted")
+
+	// Release R2 and let both requests finish.
+	close(r2Release)
+	<-r2Done
+	<-r3Done
+
+	assert.Equal(t, http.StatusOK, w2.Code)
+	assert.Contains(t, w2.Body.String(), "model1")
+	assert.Equal(t, http.StatusOK, w3.Code)
+	assert.Contains(t, w3.Body.String(), "model2")
+}
+
+func TestMatrixSolver_FullScenario(t *testing.T) {
+	// Simulates the example config:
+	// standard: [g,v], [q,v], [m,v]
+	// with_rerank: [g,v,e], [q,v,e]
+	// creative: [g,sd], [q,sd]
+	// full: [L]
+	solver := NewMatrixSolver(
+		makeExpandedSets(
+			es("standard", "g", "v"),
+			es("standard", "q", "v"),
+			es("standard", "m", "v"),
+			es("with_rerank", "e", "g", "v"),
+			es("with_rerank", "e", "q", "v"),
+			es("creative", "g", "sd"),
+			es("creative", "q", "sd"),
+			es("full", "L"),
+		),
+		map[string]int{"v": 50, "L": 30, "whisper": 10},
+	)
+
+	// Running: g, v. Request q.
+	// standard[q,v]: evict g (cost 1), keep v. Total: 1.
+	// with_rerank[q,v,e]: evict g (cost 1), keep v. Total: 1.
+	// => tie, pick first by definition order = standard[q,v]
+	result, err := solver.Solve("q", []string{"g", "v"})
+	require.NoError(t, err)
+	assert.Equal(t, []string{"g"}, result.Evict)
+	assert.Equal(t, []string{"q", "v"}, result.TargetSet)
+
+	// Running: g, v. Request L.
+	// full[L]: evict g (cost 1) + v (cost 50). Total: 51.
+	// Only one set contains L, so pick it.
+	result, err = solver.Solve("L", []string{"g", "v"})
+	require.NoError(t, err)
+	assert.ElementsMatch(t, []string{"g", "v"}, result.Evict)
+	assert.Equal(t, []string{"L"}, result.TargetSet)
+
+	// Running: g, v. Request sd.
+	// creative[g,sd]: evict v (cost 50). Total: 50.
+	// creative[q,sd]: evict g (cost 1) + v (cost 50). Total: 51.
+	// => pick creative[g,sd]
+	result, err = solver.Solve("sd", []string{"g", "v"})
+	require.NoError(t, err)
+	assert.Equal(t, []string{"v"}, result.Evict)
+	assert.Equal(t, []string{"g", "sd"}, result.TargetSet)
+
+	// Running: q, v, e. Request g.
+	// standard[g,v]: evict q (1) + e (1). Total: 2.
+	// with_rerank[g,v,e]: evict q (1). Total: 1.
+	// creative[g,sd]: evict q (1) + v (50) + e (1). Total: 52.
+	// => pick with_rerank[g,v,e]
+	result, err = solver.Solve("g", []string{"e", "q", "v"})
+	require.NoError(t, err)
+	assert.Equal(t, []string{"q"}, result.Evict)
+	assert.Equal(t, []string{"e", "g", "v"}, result.TargetSet)
+}

proxy/metrics_monitor.go

@@ -13,10 +13,54 @@ import (
 	"time"
 
 	"github.com/gin-gonic/gin"
+	"github.com/klauspost/compress/zstd"
 	"github.com/mostlygeek/llama-swap/event"
 	"github.com/tidwall/gjson"
 )
 
+// zstdEncOptions are the shared zstd encoder options for maximum compression.
+var zstdEncOptions = []zstd.EOption{
+	zstd.WithEncoderLevel(zstd.SpeedBetterCompression),
+}
+
+// zstdDecOptions are the shared zstd decoder options.
+var zstdDecOptions = []zstd.DOption{}
+
+// zstdEncPool pools zstd.Encoder instances to reduce allocations.
+var zstdEncPool = &sync.Pool{
+	New: func() interface{} {
+		enc, _ := zstd.NewWriter(nil, zstdEncOptions...)
+		return enc
+	},
+}
+
+// zstdDecPool pools zstd.Decoder instances to reduce allocations.
+var zstdDecPool = &sync.Pool{
+	New: func() interface{} {
+		dec, _ := zstd.NewReader(nil, zstdDecOptions...)
+		return dec
+	},
+}
+
+// compressCapture marshals a ReqRespCapture to JSON and compresses it with zstd.
+// Returns compressed bytes and the original JSON byte count for logging.
+func compressCapture(c *ReqRespCapture) ([]byte, int, error) {
+	jsonBytes, err := json.Marshal(c)
+	if err != nil {
+		return nil, 0, fmt.Errorf("marshal capture: %w", err)
+	}
+	enc := zstdEncPool.Get().(*zstd.Encoder)
+	defer zstdEncPool.Put(enc)
+	return enc.EncodeAll(jsonBytes, nil), len(jsonBytes), nil
+}
+
+// decompressCapture decompresses zstd-compressed JSON and returns it.
+func decompressCapture(data []byte) ([]byte, error) {
+	dec := zstdDecPool.Get().(*zstd.Decoder)
+	defer zstdDecPool.Put(dec)
+	return dec.DecodeAll(data, nil)
+}
+
 // TokenMetrics represents parsed token statistics from llama-server logs
 type TokenMetrics struct {
 	ID              int       `json:"id"`
@@ -28,6 +72,16 @@ type TokenMetrics struct {
 	PromptPerSecond float64   `json:"prompt_per_second"`
 	TokensPerSecond float64   `json:"tokens_per_second"`
 	DurationMs      int       `json:"duration_ms"`
+	HasCapture      bool      `json:"has_capture"`
+}
+
+type ReqRespCapture struct {
+	ID          int               `json:"id"`
+	ReqPath     string            `json:"req_path"`
+	ReqHeaders  map[string]string `json:"req_headers"`
+	ReqBody     []byte            `json:"req_body"`
+	RespHeaders map[string]string `json:"resp_headers"`
+	RespBody    []byte            `json:"resp_body"`
 }
 
 // TokenMetricsEvent represents a token metrics event
@@ -46,19 +100,32 @@ type metricsMonitor struct {
 	maxMetrics int
 	nextID     int
 	logger     *LogMonitor
+
+	// capture fields
+	enableCaptures bool
+	captures       map[int][]byte // zstd-compressed JSON of ReqRespCapture
+	captureOrder   []int          // track insertion order for FIFO eviction
+	captureSize    int            // current total compressed size in bytes
+	maxCaptureSize int            // max bytes for captures (uncompressed)
 }
 
-func newMetricsMonitor(logger *LogMonitor, maxMetrics int) *metricsMonitor {
-	mp := &metricsMonitor{
-		logger:     logger,
-		maxMetrics: maxMetrics,
+// newMetricsMonitor creates a new metricsMonitor. captureBufferMB is the
+// capture buffer size in megabytes; 0 disables captures.
+func newMetricsMonitor(logger *LogMonitor, maxMetrics int, captureBufferMB int) *metricsMonitor {
+	return &metricsMonitor{
+		logger:         logger,
+		maxMetrics:     maxMetrics,
+		enableCaptures: captureBufferMB > 0,
+		captures:       make(map[int][]byte),
+		captureOrder:   make([]int, 0),
+		captureSize:    0,
+		maxCaptureSize: captureBufferMB * 1024 * 1024,
 	}
-
-	return mp
 }
 
-// addMetrics adds a new metric to the collection and publishes an event
-func (mp *metricsMonitor) addMetrics(metric TokenMetrics) {
+// addMetrics adds a new metric to the collection and publishes an event.
+// Returns the assigned metric ID.
+func (mp *metricsMonitor) addMetrics(metric TokenMetrics) int {
 	mp.mu.Lock()
 	defer mp.mu.Unlock()
 
@@ -69,6 +136,84 @@ func (mp *metricsMonitor) addMetrics(metric TokenMetrics) {
 		mp.metrics = mp.metrics[len(mp.metrics)-mp.maxMetrics:]
 	}
 	event.Emit(TokenMetricsEvent{Metrics: metric})
+	return metric.ID
+}
+
+// addCapture adds a new capture to the buffer with size-based eviction.
+// Captures are skipped if enableCaptures is false or if compressed data exceeds maxCaptureSize.
+func (mp *metricsMonitor) addCapture(capture ReqRespCapture) {
+	if !mp.enableCaptures {
+		return
+	}
+
+	compressed, uncompressedBytes, err := compressCapture(&capture)
+	if err != nil {
+		mp.logger.Warnf("failed to compress capture: %v, skipping", err)
+		return
+	}
+
+	captureSize := len(compressed)
+	if captureSize > mp.maxCaptureSize {
+		mp.logger.Warnf("compressed capture size %d exceeds max %d, skipping", captureSize, mp.maxCaptureSize)
+		return
+	}
+
+	compressionRatio := (1 - float64(captureSize)/float64(uncompressedBytes)) * 100
+
+	mp.mu.Lock()
+	defer mp.mu.Unlock()
+
+	// Evict oldest (FIFO) until room available for the compressed data
+	for mp.captureSize+captureSize > mp.maxCaptureSize && len(mp.captureOrder) > 0 {
+		oldestID := mp.captureOrder[0]
+		mp.captureOrder = mp.captureOrder[1:]
+		if evicted, exists := mp.captures[oldestID]; exists {
+			l := len(evicted)
+			mp.captureSize -= l
+			delete(mp.captures, oldestID)
+			mp.logger.Debugf("Capture %d evicted to make space: %d bytes", oldestID, l)
+		}
+	}
+
+	mp.captures[capture.ID] = compressed
+	mp.captureOrder = append(mp.captureOrder, capture.ID)
+	mp.captureSize += captureSize
+
+	mp.logger.Debugf("Capture %d compressed and saved: %d bytes -> %d bytes (%.1f%% compression)", capture.ID, uncompressedBytes, len(compressed), compressionRatio)
+}
+
+// getCompressedBytes returns the raw compressed bytes for a capture by ID.
+func (mp *metricsMonitor) getCompressedBytes(id int) ([]byte, bool) {
+	mp.mu.RLock()
+	defer mp.mu.RUnlock()
+
+	data, exists := mp.captures[id]
+	return data, exists
+}
+
+// getCaptureByID returns decompressed capture bytes if found and decompress=true.
+// If decompress=false, returns the raw zstd-compressed bytes.
+// Returns nil if the capture is not found.
+func (mp *metricsMonitor) getCaptureByID(id int, decompress bool) []byte {
+	mp.mu.RLock()
+	defer mp.mu.RUnlock()
+
+	data, exists := mp.captures[id]
+	if !exists {
+		return nil
+	}
+
+	if !decompress {
+		return data
+	}
+
+	decompressed, err := decompressCapture(data)
+	if err != nil {
+		mp.logger.Warnf("failed to decompress capture %d: %v", id, err)
+		return nil
+	}
+
+	return decompressed
 }
 
 // getMetrics returns a copy of the current metrics
@@ -97,6 +242,28 @@ func (mp *metricsMonitor) wrapHandler(
 	request *http.Request,
 	next func(modelID string, w http.ResponseWriter, r *http.Request) error,
 ) error {
+	// Capture request body and headers if captures enabled
+	var reqBody []byte
+	var reqHeaders map[string]string
+	if mp.enableCaptures {
+		if request.Body != nil {
+			var err error
+			reqBody, err = io.ReadAll(request.Body)
+			if err != nil {
+				return fmt.Errorf("failed to read request body for capture: %w", err)
+			}
+			request.Body.Close()
+			request.Body = io.NopCloser(bytes.NewBuffer(reqBody))
+		}
+		reqHeaders = make(map[string]string)
+		for key, values := range request.Header {
+			if len(values) > 0 {
+				reqHeaders[key] = values[0]
+			}
+		}
+		redactHeaders(reqHeaders)
+	}
+
 	recorder := newBodyCopier(writer)
 
 	// Filter Accept-Encoding to only include encodings we can decompress for metrics
@@ -140,7 +307,6 @@ func (mp *metricsMonitor) wrapHandler(
 			return nil
 		}
 	}
-
 	if strings.Contains(recorder.Header().Get("Content-Type"), "text/event-stream") {
 		if parsed, err := processStreamingResponse(modelID, recorder.StartTime(), body); err != nil {
 			mp.logger.Warnf("error processing streaming response: %v, path=%s, recording minimal metrics", err, request.URL.Path)
@@ -153,6 +319,14 @@ func (mp *metricsMonitor) wrapHandler(
 			usage := parsed.Get("usage")
 			timings := parsed.Get("timings")
 
+			// extract timings for infill - response is an array, timings are in the last element
+			// see #463
+			if strings.HasPrefix(request.URL.Path, "/infill") {
+				if arr := parsed.Array(); len(arr) > 0 {
+					timings = arr[len(arr)-1].Get("timings")
+				}
+			}
+
 			if usage.Exists() || timings.Exists() {
 				if parsedMetrics, err := parseMetrics(modelID, recorder.StartTime(), usage, timings); err != nil {
 					mp.logger.Warnf("error parsing metrics: %v, path=%s, recording minimal metrics", err, request.URL.Path)
@@ -165,7 +339,38 @@ func (mp *metricsMonitor) wrapHandler(
 		}
 	}
 
-	mp.addMetrics(tm)
+	// Build capture if enabled and determine if it will be stored
+	var capture *ReqRespCapture
+	if mp.enableCaptures {
+		respHeaders := make(map[string]string)
+		for key, values := range recorder.Header() {
+			if len(values) > 0 {
+				respHeaders[key] = values[0]
+			}
+		}
+		redactHeaders(respHeaders)
+		delete(respHeaders, "Content-Encoding")
+		capture = &ReqRespCapture{
+			ReqPath:     request.URL.Path,
+			ReqHeaders:  reqHeaders,
+			ReqBody:     reqBody,
+			RespHeaders: respHeaders,
+			RespBody:    body,
+		}
+		compressed, _, err := compressCapture(capture)
+		if err == nil && len(compressed) <= mp.maxCaptureSize {
+			tm.HasCapture = true
+		}
+	}
+
+	metricID := mp.addMetrics(tm)
+
+	// Store capture if enabled
+	if capture != nil {
+		capture.ID = metricID
+		mp.addCapture(*capture)
+	}
+
 	return nil
 }
 
@@ -212,6 +417,11 @@ func processStreamingResponse(modelID string, start time.Time, body []byte) (Tok
 			usage := parsed.Get("usage")
 			timings := parsed.Get("timings")
 
+			// v1/responses format nests usage under response.usage
+			if !usage.Exists() {
+				usage = parsed.Get("response.usage")
+			}
+
 			if usage.Exists() || timings.Exists() {
 				return parseMetrics(modelID, start, usage, timings)
 			}
@@ -222,6 +432,8 @@ func processStreamingResponse(modelID string, start time.Time, body []byte) (Tok
 }
 
 func parseMetrics(modelID string, start time.Time, usage, timings gjson.Result) (TokenMetrics, error) {
+	wallDurationMs := int(time.Since(start).Milliseconds())
+
 	// default values
 	cachedTokens := -1 // unknown or missing data
 	outputTokens := 0
@@ -230,7 +442,7 @@ func parseMetrics(modelID string, start time.Time, usage, timings gjson.Result)
 	// timings data
 	tokensPerSecond := -1.0
 	promptPerSecond := -1.0
-	durationMs := int(time.Since(start).Milliseconds())
+	durationMs := wallDurationMs
 
 	if usage.Exists() {
 		if pt := usage.Get("prompt_tokens"); pt.Exists() {
@@ -259,7 +471,10 @@ func parseMetrics(modelID string, start time.Time, usage, timings gjson.Result)
 		outputTokens = int(timings.Get("predicted_n").Int())
 		promptPerSecond = timings.Get("prompt_per_second").Float()
 		tokensPerSecond = timings.Get("predicted_per_second").Float()
-		durationMs = int(timings.Get("prompt_ms").Float() + timings.Get("predicted_ms").Float())
+		timingsDurationMs := int(timings.Get("prompt_ms").Float() + timings.Get("predicted_ms").Float())
+		if timingsDurationMs > durationMs {
+			durationMs = timingsDurationMs
+		}
 
 		if cachedValue := timings.Get("cache_n"); cachedValue.Exists() {
 			cachedTokens = int(cachedValue.Int())
@@ -336,6 +551,24 @@ func (w *responseBodyCopier) StartTime() time.Time {
 	return w.start
 }
 
+// sensitiveHeaders lists headers that should be redacted in captures
+var sensitiveHeaders = map[string]bool{
+	"authorization":       true,
+	"proxy-authorization": true,
+	"cookie":              true,
+	"set-cookie":          true,
+	"x-api-key":           true,
+}
+
+// redactHeaders replaces sensitive header values in-place with "[REDACTED]"
+func redactHeaders(headers map[string]string) {
+	for key := range headers {
+		if sensitiveHeaders[strings.ToLower(key)] {
+			headers[key] = "[REDACTED]"
+		}
+	}
+}
+
 // filterAcceptEncoding filters the Accept-Encoding header to only include
 // encodings we can decompress (gzip, deflate). This respects the client's
 // preferences while ensuring we can parse response bodies for metrics.
@@ -347,9 +580,9 @@ func filterAcceptEncoding(acceptEncoding string) string {
 	supported := map[string]bool{"gzip": true, "deflate": true}
 	var filtered []string
 
-	for _, part := range strings.Split(acceptEncoding, ",") {
+	for part := range strings.SplitSeq(acceptEncoding, ",") {
 		// Parse encoding and optional quality value (e.g., "gzip;q=1.0")
-		encoding := strings.TrimSpace(strings.Split(part, ";")[0])
+		encoding, _, _ := strings.Cut(strings.TrimSpace(part), ";")
 		if supported[strings.ToLower(encoding)] {
 			filtered = append(filtered, strings.TrimSpace(part))
 		}

proxy/metrics_monitor_test.go

@@ -5,6 +5,7 @@ import (
 	"compress/flate"
 	"compress/gzip"
 	"encoding/json"
+	"math/rand"
 	"net/http"
 	"net/http/httptest"
 	"sync"
@@ -14,11 +15,12 @@ import (
 	"github.com/gin-gonic/gin"
 	"github.com/mostlygeek/llama-swap/event"
 	"github.com/stretchr/testify/assert"
+	"github.com/tidwall/gjson"
 )
 
 func TestMetricsMonitor_AddMetrics(t *testing.T) {
 	t.Run("adds metrics and assigns ID", func(t *testing.T) {
-		mm := newMetricsMonitor(testLogger, 10)
+		mm := newMetricsMonitor(testLogger, 10, 0)
 
 		metric := TokenMetrics{
 			Model:        "test-model",
@@ -37,7 +39,7 @@ func TestMetricsMonitor_AddMetrics(t *testing.T) {
 	})
 
 	t.Run("increments ID for each metric", func(t *testing.T) {
-		mm := newMetricsMonitor(testLogger, 10)
+		mm := newMetricsMonitor(testLogger, 10, 0)
 
 		for i := 0; i < 5; i++ {
 			mm.addMetrics(TokenMetrics{Model: "model"})
@@ -51,7 +53,7 @@ func TestMetricsMonitor_AddMetrics(t *testing.T) {
 	})
 
 	t.Run("respects max metrics limit", func(t *testing.T) {
-		mm := newMetricsMonitor(testLogger, 3)
+		mm := newMetricsMonitor(testLogger, 3, 0)
 
 		// Add 5 metrics
 		for i := 0; i < 5; i++ {
@@ -71,7 +73,7 @@ func TestMetricsMonitor_AddMetrics(t *testing.T) {
 	})
 
 	t.Run("emits TokenMetricsEvent", func(t *testing.T) {
-		mm := newMetricsMonitor(testLogger, 10)
+		mm := newMetricsMonitor(testLogger, 10, 0)
 
 		receivedEvent := make(chan TokenMetricsEvent, 1)
 		cancel := event.On(func(e TokenMetricsEvent) {
@@ -101,14 +103,14 @@ func TestMetricsMonitor_AddMetrics(t *testing.T) {
 
 func TestMetricsMonitor_GetMetrics(t *testing.T) {
 	t.Run("returns empty slice when no metrics", func(t *testing.T) {
-		mm := newMetricsMonitor(testLogger, 10)
+		mm := newMetricsMonitor(testLogger, 10, 0)
 		metrics := mm.getMetrics()
 		assert.NotNil(t, metrics)
 		assert.Equal(t, 0, len(metrics))
 	})
 
 	t.Run("returns copy of metrics", func(t *testing.T) {
-		mm := newMetricsMonitor(testLogger, 10)
+		mm := newMetricsMonitor(testLogger, 10, 0)
 		mm.addMetrics(TokenMetrics{Model: "model1"})
 		mm.addMetrics(TokenMetrics{Model: "model2"})
 
@@ -128,7 +130,7 @@ func TestMetricsMonitor_GetMetrics(t *testing.T) {
 
 func TestMetricsMonitor_GetMetricsJSON(t *testing.T) {
 	t.Run("returns valid JSON for empty metrics", func(t *testing.T) {
-		mm := newMetricsMonitor(testLogger, 10)
+		mm := newMetricsMonitor(testLogger, 10, 0)
 		jsonData, err := mm.getMetricsJSON()
 		assert.NoError(t, err)
 		assert.NotNil(t, jsonData)
@@ -140,7 +142,7 @@ func TestMetricsMonitor_GetMetricsJSON(t *testing.T) {
 	})
 
 	t.Run("returns valid JSON with metrics", func(t *testing.T) {
-		mm := newMetricsMonitor(testLogger, 10)
+		mm := newMetricsMonitor(testLogger, 10, 0)
 		mm.addMetrics(TokenMetrics{
 			Model:           "model1",
 			InputTokens:     100,
@@ -168,7 +170,7 @@ func TestMetricsMonitor_GetMetricsJSON(t *testing.T) {
 
 func TestMetricsMonitor_WrapHandler(t *testing.T) {
 	t.Run("successful non-streaming request with usage data", func(t *testing.T) {
-		mm := newMetricsMonitor(testLogger, 10)
+		mm := newMetricsMonitor(testLogger, 10, 0)
 
 		responseBody := `{
 			"usage": {
@@ -199,7 +201,7 @@ func TestMetricsMonitor_WrapHandler(t *testing.T) {
 	})
 
 	t.Run("successful request with timings data", func(t *testing.T) {
-		mm := newMetricsMonitor(testLogger, 10)
+		mm := newMetricsMonitor(testLogger, 10, 0)
 
 		responseBody := `{
 			"timings": {
@@ -239,7 +241,7 @@ func TestMetricsMonitor_WrapHandler(t *testing.T) {
 	})
 
 	t.Run("streaming request with SSE format", func(t *testing.T) {
-		mm := newMetricsMonitor(testLogger, 10)
+		mm := newMetricsMonitor(testLogger, 10, 0)
 
 		// Note: SSE format requires proper line breaks - each data line followed by blank line
 		responseBody := `data: {"choices":[{"text":"Hello"}]}
@@ -275,7 +277,7 @@ data: [DONE]
 	})
 
 	t.Run("non-OK status code does not record metrics", func(t *testing.T) {
-		mm := newMetricsMonitor(testLogger, 10)
+		mm := newMetricsMonitor(testLogger, 10, 0)
 
 		nextHandler := func(modelID string, w http.ResponseWriter, r *http.Request) error {
 			w.WriteHeader(http.StatusBadRequest)
@@ -295,7 +297,7 @@ data: [DONE]
 	})
 
 	t.Run("empty response body records minimal metrics", func(t *testing.T) {
-		mm := newMetricsMonitor(testLogger, 10)
+		mm := newMetricsMonitor(testLogger, 10, 0)
 
 		nextHandler := func(modelID string, w http.ResponseWriter, r *http.Request) error {
 			w.WriteHeader(http.StatusOK)
@@ -317,7 +319,7 @@ data: [DONE]
 	})
 
 	t.Run("invalid JSON records minimal metrics", func(t *testing.T) {
-		mm := newMetricsMonitor(testLogger, 10)
+		mm := newMetricsMonitor(testLogger, 10, 0)
 
 		nextHandler := func(modelID string, w http.ResponseWriter, r *http.Request) error {
 			w.Header().Set("Content-Type", "application/json")
@@ -341,7 +343,7 @@ data: [DONE]
 	})
 
 	t.Run("next handler error is propagated", func(t *testing.T) {
-		mm := newMetricsMonitor(testLogger, 10)
+		mm := newMetricsMonitor(testLogger, 10, 0)
 
 		expectedErr := assert.AnError
 		nextHandler := func(modelID string, w http.ResponseWriter, r *http.Request) error {
@@ -360,7 +362,7 @@ data: [DONE]
 	})
 
 	t.Run("response without usage or timings records minimal metrics", func(t *testing.T) {
-		mm := newMetricsMonitor(testLogger, 10)
+		mm := newMetricsMonitor(testLogger, 10, 0)
 
 		responseBody := `{"result": "ok"}`
 
@@ -384,6 +386,75 @@ data: [DONE]
 		assert.Equal(t, 0, metrics[0].InputTokens)
 		assert.Equal(t, 0, metrics[0].OutputTokens)
 	})
+
+	t.Run("infill request extracts timings from last array element", func(t *testing.T) {
+		mm := newMetricsMonitor(testLogger, 10, 0)
+
+		// Infill response is an array with timings in the last element
+		responseBody := `[
+			{"content": "first chunk"},
+			{"content": "second chunk"},
+			{"content": "final", "timings": {
+				"prompt_n": 150,
+				"predicted_n": 75,
+				"prompt_per_second": 200.5,
+				"predicted_per_second": 35.5,
+				"prompt_ms": 600.0,
+				"predicted_ms": 1800.0,
+				"cache_n": 30
+			}}
+		]`
+
+		nextHandler := func(modelID string, w http.ResponseWriter, r *http.Request) error {
+			w.Header().Set("Content-Type", "application/json")
+			w.WriteHeader(http.StatusOK)
+			w.Write([]byte(responseBody))
+			return nil
+		}
+
+		req := httptest.NewRequest("POST", "/infill", nil)
+		rec := httptest.NewRecorder()
+		ginCtx, _ := gin.CreateTestContext(rec)
+
+		err := mm.wrapHandler("test-model", ginCtx.Writer, req, nextHandler)
+		assert.NoError(t, err)
+
+		metrics := mm.getMetrics()
+		assert.Equal(t, 1, len(metrics))
+		assert.Equal(t, "test-model", metrics[0].Model)
+		assert.Equal(t, 150, metrics[0].InputTokens)
+		assert.Equal(t, 75, metrics[0].OutputTokens)
+		assert.Equal(t, 30, metrics[0].CachedTokens)
+		assert.Equal(t, 200.5, metrics[0].PromptPerSecond)
+		assert.Equal(t, 35.5, metrics[0].TokensPerSecond)
+		assert.Equal(t, 2400, metrics[0].DurationMs) // 600 + 1800
+	})
+
+	t.Run("infill request with empty array records minimal metrics", func(t *testing.T) {
+		mm := newMetricsMonitor(testLogger, 10, 0)
+
+		responseBody := `[]`
+
+		nextHandler := func(modelID string, w http.ResponseWriter, r *http.Request) error {
+			w.Header().Set("Content-Type", "application/json")
+			w.WriteHeader(http.StatusOK)
+			w.Write([]byte(responseBody))
+			return nil
+		}
+
+		req := httptest.NewRequest("POST", "/infill", nil)
+		rec := httptest.NewRecorder()
+		ginCtx, _ := gin.CreateTestContext(rec)
+
+		err := mm.wrapHandler("test-model", ginCtx.Writer, req, nextHandler)
+		assert.NoError(t, err)
+
+		metrics := mm.getMetrics()
+		assert.Equal(t, 1, len(metrics))
+		assert.Equal(t, "test-model", metrics[0].Model)
+		assert.Equal(t, 0, metrics[0].InputTokens)
+		assert.Equal(t, 0, metrics[0].OutputTokens)
+	})
 }
 
 func TestMetricsMonitor_ResponseBodyCopier(t *testing.T) {
@@ -437,7 +508,7 @@ func TestMetricsMonitor_ResponseBodyCopier(t *testing.T) {
 
 func TestMetricsMonitor_Concurrent(t *testing.T) {
 	t.Run("concurrent addMetrics is safe", func(t *testing.T) {
-		mm := newMetricsMonitor(testLogger, 1000)
+		mm := newMetricsMonitor(testLogger, 1000, 0)
 
 		var wg sync.WaitGroup
 		numGoroutines := 10
@@ -464,7 +535,7 @@ func TestMetricsMonitor_Concurrent(t *testing.T) {
 	})
 
 	t.Run("concurrent reads and writes are safe", func(t *testing.T) {
-		mm := newMetricsMonitor(testLogger, 100)
+		mm := newMetricsMonitor(testLogger, 100, 0)
 
 		done := make(chan bool)
 
@@ -501,8 +572,29 @@ func TestMetricsMonitor_Concurrent(t *testing.T) {
 }
 
 func TestMetricsMonitor_ParseMetrics(t *testing.T) {
+	t.Run("keeps wall clock duration when timings underreport request time", func(t *testing.T) {
+		start := time.Now().Add(-5 * time.Second)
+		usage := gjson.Parse(`{"prompt_tokens": 5, "completion_tokens": 1}`)
+		timings := gjson.Parse(`{
+			"prompt_n": 5,
+			"predicted_n": 1,
+			"prompt_per_second": 10.0,
+			"predicted_per_second": 2.0,
+			"prompt_ms": 5.0,
+			"predicted_ms": 15.0
+		}`)
+
+		metrics, err := parseMetrics("test-model", start, usage, timings)
+		assert.NoError(t, err)
+		assert.Equal(t, 5, metrics.InputTokens)
+		assert.Equal(t, 1, metrics.OutputTokens)
+		assert.Equal(t, 10.0, metrics.PromptPerSecond)
+		assert.Equal(t, 2.0, metrics.TokensPerSecond)
+		assert.GreaterOrEqual(t, metrics.DurationMs, 5000)
+	})
+
 	t.Run("prefers timings over usage data", func(t *testing.T) {
-		mm := newMetricsMonitor(testLogger, 10)
+		mm := newMetricsMonitor(testLogger, 10, 0)
 
 		// Timings should take precedence over usage
 		responseBody := `{
@@ -542,7 +634,7 @@ func TestMetricsMonitor_ParseMetrics(t *testing.T) {
 	})
 
 	t.Run("handles missing cache_n in timings", func(t *testing.T) {
-		mm := newMetricsMonitor(testLogger, 10)
+		mm := newMetricsMonitor(testLogger, 10, 0)
 
 		responseBody := `{
 			"timings": {
@@ -577,7 +669,7 @@ func TestMetricsMonitor_ParseMetrics(t *testing.T) {
 
 func TestMetricsMonitor_StreamingResponse(t *testing.T) {
 	t.Run("finds metrics in last valid SSE data", func(t *testing.T) {
-		mm := newMetricsMonitor(testLogger, 10)
+		mm := newMetricsMonitor(testLogger, 10, 0)
 
 		// Metrics should be found in the last data line before [DONE]
 		responseBody := `data: {"choices":[{"text":"First"}]}
@@ -611,7 +703,7 @@ data: [DONE]
 	})
 
 	t.Run("handles streaming with no valid JSON records minimal metrics", func(t *testing.T) {
-		mm := newMetricsMonitor(testLogger, 10)
+		mm := newMetricsMonitor(testLogger, 10, 0)
 
 		responseBody := `data: not json
 
@@ -640,8 +732,37 @@ data: [DONE]
 		assert.Equal(t, 0, metrics[0].OutputTokens)
 	})
 
+	t.Run("v1/responses format with nested response.usage", func(t *testing.T) {
+		mm := newMetricsMonitor(testLogger, 10, 0)
+
+		// v1/responses SSE format: usage is nested under response.usage
+		responseBody := "event: response.completed\n" +
+			`data: {"type":"response.completed","response":{"id":"resp_abc","object":"response","created_at":1773416985,"status":"completed","model":"test-model","output":[],"usage":{"input_tokens":17,"output_tokens":23,"total_tokens":40}}}` +
+			"\n\n"
+
+		nextHandler := func(modelID string, w http.ResponseWriter, r *http.Request) error {
+			w.Header().Set("Content-Type", "text/event-stream")
+			w.WriteHeader(http.StatusOK)
+			w.Write([]byte(responseBody))
+			return nil
+		}
+
+		req := httptest.NewRequest("POST", "/v1/responses", nil)
+		rec := httptest.NewRecorder()
+		ginCtx, _ := gin.CreateTestContext(rec)
+
+		err := mm.wrapHandler("test-model", ginCtx.Writer, req, nextHandler)
+		assert.NoError(t, err)
+
+		metrics := mm.getMetrics()
+		assert.Equal(t, 1, len(metrics))
+		assert.Equal(t, "test-model", metrics[0].Model)
+		assert.Equal(t, 17, metrics[0].InputTokens)
+		assert.Equal(t, 23, metrics[0].OutputTokens)
+	})
+
 	t.Run("handles empty streaming response records minimal metrics", func(t *testing.T) {
-		mm := newMetricsMonitor(testLogger, 10)
+		mm := newMetricsMonitor(testLogger, 10, 0)
 
 		responseBody := ``
 
@@ -669,7 +790,7 @@ data: [DONE]
 
 // Benchmark tests
 func BenchmarkMetricsMonitor_AddMetrics(b *testing.B) {
-	mm := newMetricsMonitor(testLogger, 1000)
+	mm := newMetricsMonitor(testLogger, 1000, 0)
 
 	metric := TokenMetrics{
 		Model:           "test-model",
@@ -690,7 +811,7 @@ func BenchmarkMetricsMonitor_AddMetrics(b *testing.B) {
 
 func BenchmarkMetricsMonitor_AddMetrics_SmallBuffer(b *testing.B) {
 	// Test performance with a smaller buffer where wrapping occurs more frequently
-	mm := newMetricsMonitor(testLogger, 100)
+	mm := newMetricsMonitor(testLogger, 100, 0)
 
 	metric := TokenMetrics{
 		Model:           "test-model",
@@ -711,7 +832,7 @@ func BenchmarkMetricsMonitor_AddMetrics_SmallBuffer(b *testing.B) {
 
 func TestMetricsMonitor_WrapHandler_Compression(t *testing.T) {
 	t.Run("gzip encoded response", func(t *testing.T) {
-		mm := newMetricsMonitor(testLogger, 10)
+		mm := newMetricsMonitor(testLogger, 10, 0)
 
 		responseBody := `{"usage": {"prompt_tokens": 100, "completion_tokens": 50}}`
 
@@ -745,7 +866,7 @@ func TestMetricsMonitor_WrapHandler_Compression(t *testing.T) {
 	})
 
 	t.Run("deflate encoded response", func(t *testing.T) {
-		mm := newMetricsMonitor(testLogger, 10)
+		mm := newMetricsMonitor(testLogger, 10, 0)
 
 		responseBody := `{"usage": {"prompt_tokens": 200, "completion_tokens": 75}}`
 
@@ -779,7 +900,7 @@ func TestMetricsMonitor_WrapHandler_Compression(t *testing.T) {
 	})
 
 	t.Run("invalid gzip data records minimal metrics", func(t *testing.T) {
-		mm := newMetricsMonitor(testLogger, 10)
+		mm := newMetricsMonitor(testLogger, 10, 0)
 
 		// Invalid compressed data
 		invalidData := []byte("this is not gzip data")
@@ -807,7 +928,7 @@ func TestMetricsMonitor_WrapHandler_Compression(t *testing.T) {
 	})
 
 	t.Run("unknown encoding treated as uncompressed", func(t *testing.T) {
-		mm := newMetricsMonitor(testLogger, 10)
+		mm := newMetricsMonitor(testLogger, 10, 0)
 
 		responseBody := `{"usage": {"prompt_tokens": 300, "completion_tokens": 100}}`
 
@@ -832,3 +953,268 @@ func TestMetricsMonitor_WrapHandler_Compression(t *testing.T) {
 		assert.Equal(t, 100, metrics[0].OutputTokens)
 	})
 }
+
+func TestReqRespCapture_CompressedSize(t *testing.T) {
+	t.Run("compressed size is smaller than uncompressed", func(t *testing.T) {
+		capture := ReqRespCapture{
+			ID:       1,
+			ReqPath:  "/v1/chat/completions",
+			ReqBody:  []byte(`{"model":"test","prompt":"hello world this is a test request body that is reasonably long"}`),
+			RespBody: []byte(`{"id":"resp-123","object":"chat.completion","created":1234567890,"model":"test-model","choices":[{"index":0,"message":{"role":"assistant","content":"This is a test response body with some meaningful content to compress"}},{"index":1,"message":{"role":"user","content":"Another message here"}}]}`),
+		}
+
+		compressed, uncompressed, err := compressCapture(&capture)
+		assert.NoError(t, err)
+		assert.Greater(t, uncompressed, 0)
+		assert.True(t, len(compressed) < uncompressed, "compressed (%d bytes) should be smaller than uncompressed JSON (%d bytes)", len(compressed), uncompressed)
+	})
+
+	t.Run("empty capture produces compressed output", func(t *testing.T) {
+		capture := ReqRespCapture{}
+		compressed, _, err := compressCapture(&capture)
+		assert.NoError(t, err)
+		assert.NotNil(t, compressed)
+		assert.True(t, len(compressed) > 0)
+	})
+}
+
+func TestMetricsMonitor_AddCapture(t *testing.T) {
+	t.Run("does nothing when captures disabled", func(t *testing.T) {
+		mm := newMetricsMonitor(testLogger, 10, 0)
+
+		capture := ReqRespCapture{
+			ID:      0,
+			ReqBody: []byte("test"),
+		}
+		mm.addCapture(capture)
+
+		// Should not store capture
+		assert.Nil(t, mm.getCaptureByID(0, false))
+	})
+
+	t.Run("adds capture when enabled", func(t *testing.T) {
+		mm := newMetricsMonitor(testLogger, 10, 5)
+
+		capture := ReqRespCapture{
+			ID:       0,
+			ReqBody:  []byte("test request"),
+			RespBody: []byte("test response"),
+		}
+		mm.addCapture(capture)
+
+		retrieved := mm.getCaptureByID(0, true)
+		assert.NotNil(t, retrieved)
+
+		var decoded ReqRespCapture
+		err := json.Unmarshal(retrieved, &decoded)
+		assert.NoError(t, err)
+		assert.Equal(t, 0, decoded.ID)
+		assert.Equal(t, []byte("test request"), decoded.ReqBody)
+		assert.Equal(t, []byte("test response"), decoded.RespBody)
+	})
+
+	t.Run("evicts oldest when exceeding max size", func(t *testing.T) {
+		mm := newMetricsMonitor(testLogger, 10, 5)
+		// Each full ReqRespCapture with 80 bytes random data compresses to ~185 bytes.
+		// 2 captures = ~370 bytes, 3 captures = ~555 bytes. Set limit so only 2 fit.
+		mm.maxCaptureSize = 450
+
+		// Use random-looking data that doesn't compress well with zstd
+		rng := rand.New(rand.NewSource(42))
+		capture1 := ReqRespCapture{ID: 0, ReqBody: make([]byte, 80)}
+		rng.Read(capture1.ReqBody)
+		capture2 := ReqRespCapture{ID: 1, ReqBody: make([]byte, 80)}
+		rng.Read(capture2.ReqBody)
+		capture3 := ReqRespCapture{ID: 2, ReqBody: make([]byte, 80)}
+		rng.Read(capture3.ReqBody)
+
+		mm.addCapture(capture1)
+		mm.addCapture(capture2)
+		// Adding capture3 should evict capture1
+		mm.addCapture(capture3)
+
+		assert.Nil(t, mm.getCaptureByID(0, true), "capture 0 should be evicted")
+		retrieved := mm.getCaptureByID(1, true)
+		assert.NotNil(t, retrieved, "capture 1 should exist")
+		retrieved = mm.getCaptureByID(2, true)
+		assert.NotNil(t, retrieved, "capture 2 should exist")
+	})
+
+	t.Run("skips capture larger than max size", func(t *testing.T) {
+		mm := newMetricsMonitor(testLogger, 10, 5)
+		mm.maxCaptureSize = 100
+
+		// Use random data that doesn't compress well to create an oversized capture
+		rng := rand.New(rand.NewSource(99))
+		largeCapture := ReqRespCapture{ID: 0, ReqBody: make([]byte, 300)}
+		rng.Read(largeCapture.ReqBody)
+		mm.addCapture(largeCapture)
+
+		assert.Nil(t, mm.getCaptureByID(0, false), "oversized capture should not be stored")
+	})
+}
+
+func TestMetricsMonitor_GetCaptureByID(t *testing.T) {
+	t.Run("returns nil for non-existent ID", func(t *testing.T) {
+		mm := newMetricsMonitor(testLogger, 10, 5)
+
+		assert.Nil(t, mm.getCaptureByID(999, false))
+	})
+
+	t.Run("returns decompressed capture by ID", func(t *testing.T) {
+		mm := newMetricsMonitor(testLogger, 10, 5)
+
+		capture := ReqRespCapture{
+			ID:       42,
+			ReqBody:  []byte("test request"),
+			RespBody: []byte("test response"),
+		}
+		mm.addCapture(capture)
+
+		retrieved := mm.getCaptureByID(42, true)
+		assert.NotNil(t, retrieved)
+
+		var decoded ReqRespCapture
+		err := json.Unmarshal(retrieved, &decoded)
+		assert.NoError(t, err)
+		assert.Equal(t, 42, decoded.ID)
+		assert.Equal(t, []byte("test request"), decoded.ReqBody)
+		assert.Equal(t, []byte("test response"), decoded.RespBody)
+	})
+
+	t.Run("returns compressed bytes when decompress=false", func(t *testing.T) {
+		mm := newMetricsMonitor(testLogger, 10, 5)
+
+		capture := ReqRespCapture{
+			ID:       42,
+			ReqBody:  []byte("test request body"),
+			RespBody: []byte("test response body"),
+		}
+		mm.addCapture(capture)
+
+		compressed := mm.getCaptureByID(42, false)
+		assert.NotNil(t, compressed)
+		// Compressed data should not be valid JSON (it's zstd-compressed)
+		assert.False(t, gjson.ValidBytes(compressed))
+	})
+}
+
+func TestRedactHeaders(t *testing.T) {
+	t.Run("redacts sensitive headers", func(t *testing.T) {
+		headers := map[string]string{
+			"Authorization":       "Bearer secret-token",
+			"Proxy-Authorization": "Basic creds",
+			"Cookie":              "session=abc123",
+			"Set-Cookie":          "session=xyz789",
+			"X-Api-Key":           "sk-12345",
+			"Content-Type":        "application/json",
+			"X-Custom":            "safe-value",
+		}
+
+		redactHeaders(headers)
+
+		assert.Equal(t, "[REDACTED]", headers["Authorization"])
+		assert.Equal(t, "[REDACTED]", headers["Proxy-Authorization"])
+		assert.Equal(t, "[REDACTED]", headers["Cookie"])
+		assert.Equal(t, "[REDACTED]", headers["Set-Cookie"])
+		assert.Equal(t, "[REDACTED]", headers["X-Api-Key"])
+		assert.Equal(t, "application/json", headers["Content-Type"])
+		assert.Equal(t, "safe-value", headers["X-Custom"])
+	})
+
+	t.Run("handles mixed case header names", func(t *testing.T) {
+		headers := map[string]string{
+			"authorization": "Bearer token",
+			"COOKIE":        "session=abc",
+			"x-api-key":     "key123",
+		}
+
+		redactHeaders(headers)
+
+		assert.Equal(t, "[REDACTED]", headers["authorization"])
+		assert.Equal(t, "[REDACTED]", headers["COOKIE"])
+		assert.Equal(t, "[REDACTED]", headers["x-api-key"])
+	})
+
+	t.Run("handles empty headers", func(t *testing.T) {
+		headers := map[string]string{}
+		redactHeaders(headers)
+		assert.Empty(t, headers)
+	})
+}
+
+func TestMetricsMonitor_WrapHandler_Capture(t *testing.T) {
+	t.Run("captures request and response when enabled", func(t *testing.T) {
+		mm := newMetricsMonitor(testLogger, 10, 5)
+
+		requestBody := `{"model": "test", "prompt": "hello"}`
+		responseBody := `{"usage": {"prompt_tokens": 100, "completion_tokens": 50}}`
+
+		nextHandler := func(modelID string, w http.ResponseWriter, r *http.Request) error {
+			w.Header().Set("Content-Type", "application/json")
+			w.Header().Set("X-Custom", "header-value")
+			w.WriteHeader(http.StatusOK)
+			w.Write([]byte(responseBody))
+			return nil
+		}
+
+		req := httptest.NewRequest("POST", "/test", bytes.NewBufferString(requestBody))
+		req.Header.Set("Content-Type", "application/json")
+		req.Header.Set("Authorization", "Bearer secret")
+		rec := httptest.NewRecorder()
+		ginCtx, _ := gin.CreateTestContext(rec)
+
+		err := mm.wrapHandler("test-model", ginCtx.Writer, req, nextHandler)
+		assert.NoError(t, err)
+
+		// Check metric was recorded
+		metrics := mm.getMetrics()
+		assert.Equal(t, 1, len(metrics))
+		metricID := metrics[0].ID
+
+		// Check capture was stored with same ID (decompressed)
+		captureData := mm.getCaptureByID(metricID, true)
+		assert.NotNil(t, captureData)
+
+		var capture ReqRespCapture
+		err = json.Unmarshal(captureData, &capture)
+		assert.NoError(t, err)
+		assert.Equal(t, metricID, capture.ID)
+		assert.Equal(t, []byte(requestBody), capture.ReqBody)
+		assert.Equal(t, []byte(responseBody), capture.RespBody)
+		assert.Equal(t, "/test", capture.ReqPath)
+		assert.Equal(t, "application/json", capture.ReqHeaders["Content-Type"])
+		assert.Equal(t, "[REDACTED]", capture.ReqHeaders["Authorization"])
+		assert.Equal(t, "application/json", capture.RespHeaders["Content-Type"])
+		assert.Equal(t, "header-value", capture.RespHeaders["X-Custom"])
+	})
+
+	t.Run("does not capture when disabled", func(t *testing.T) {
+		mm := newMetricsMonitor(testLogger, 10, 0)
+
+		requestBody := `{"model": "test"}`
+		responseBody := `{"usage": {"prompt_tokens": 100, "completion_tokens": 50}}`
+
+		nextHandler := func(modelID string, w http.ResponseWriter, r *http.Request) error {
+			w.Header().Set("Content-Type", "application/json")
+			w.WriteHeader(http.StatusOK)
+			w.Write([]byte(responseBody))
+			return nil
+		}
+
+		req := httptest.NewRequest("POST", "/test", bytes.NewBufferString(requestBody))
+		rec := httptest.NewRecorder()
+		ginCtx, _ := gin.CreateTestContext(rec)
+
+		err := mm.wrapHandler("test-model", ginCtx.Writer, req, nextHandler)
+		assert.NoError(t, err)
+
+		// Metrics should still be recorded
+		metrics := mm.getMetrics()
+		assert.Equal(t, 1, len(metrics))
+
+		// But no capture
+		capture := mm.getCaptureByID(metrics[0].ID, false)
+		assert.Nil(t, capture)
+	})
+}

proxy/peerproxy.go

@@ -34,23 +34,25 @@ func NewPeerProxy(peers config.PeerDictionaryConfig, proxyLogger *LogMonitor) (*
 	}
 	sort.Strings(peerIDs)
 
-	// Create a shared transport with reasonable timeouts for peer connections
-	// these can be tuned with feedback later
-	peerTransport := &http.Transport{
-		DialContext: (&net.Dialer{
-			Timeout:   30 * time.Second, // Connection timeout
-			KeepAlive: 30 * time.Second,
-		}).DialContext,
-		TLSHandshakeTimeout:   10 * time.Second,
-		ResponseHeaderTimeout: 60 * time.Second, // Time to wait for response headers
-		ExpectContinueTimeout: 1 * time.Second,
-		MaxIdleConns:          100,
-		MaxIdleConnsPerHost:   10,
-		IdleConnTimeout:       90 * time.Second,
-	}
-
 	for _, peerID := range peerIDs {
 		peer := peers[peerID]
+
+		// Create a transport with per-peer timeout configuration
+		peerTransport := &http.Transport{
+			Proxy: http.ProxyFromEnvironment,
+			DialContext: (&net.Dialer{
+				Timeout:   time.Duration(peer.Timeouts.Connect) * time.Second,
+				KeepAlive: time.Duration(peer.Timeouts.KeepAlive) * time.Second,
+			}).DialContext,
+			TLSHandshakeTimeout:   time.Duration(peer.Timeouts.TLSHandshake) * time.Second,
+			ResponseHeaderTimeout: time.Duration(peer.Timeouts.ResponseHeader) * time.Second,
+			ExpectContinueTimeout: time.Duration(peer.Timeouts.ExpectContinue) * time.Second,
+			ForceAttemptHTTP2:     true,
+			MaxIdleConns:          100,
+			MaxIdleConnsPerHost:   10,
+			IdleConnTimeout:       time.Duration(peer.Timeouts.IdleConn) * time.Second,
+		}
+
 		// Create reverse proxy for this peer
 		reverseProxy := httputil.NewSingleHostReverseProxy(peer.ProxyURL)
 		reverseProxy.Transport = peerTransport
@@ -106,6 +108,20 @@ func (p *PeerProxy) HasPeerModel(modelID string) bool {
 	return found
 }
 
+// GetPeerFilters returns the filters for a peer model, or empty filters if not found
+func (p *PeerProxy) GetPeerFilters(modelID string) config.Filters {
+	pp, found := p.proxyMap[modelID]
+	if !found {
+		return config.Filters{}
+	}
+	// Get the peer config using the peerID
+	peer, found := p.peers[pp.peerID]
+	if !found {
+		return config.Filters{}
+	}
+	return peer.Filters
+}
+
 func (p *PeerProxy) ListPeers() config.PeerDictionaryConfig {
 	return p.peers
 }

proxy/peerproxy_test.go

@@ -6,6 +6,7 @@ import (
 	"net/url"
 	"strings"
 	"testing"
+	"time"
 
 	"github.com/mostlygeek/llama-swap/proxy/config"
 	"github.com/stretchr/testify/assert"
@@ -266,3 +267,45 @@ func TestProxyRequest_SSEHeaderModification(t *testing.T) {
 	// The X-Accel-Buffering header should be set to "no" for SSE
 	assert.Equal(t, "no", w.Header().Get("X-Accel-Buffering"))
 }
+
+func TestNewPeerProxy_CustomTimeouts(t *testing.T) {
+	proxyURL, _ := url.Parse("http://localhost:8080")
+
+	peers := config.PeerDictionaryConfig{
+		"test-peer": config.PeerConfig{
+			Proxy:    "http://localhost:8080",
+			ProxyURL: proxyURL,
+			Models:   []string{"model1"},
+			Timeouts: config.TimeoutsConfig{
+				Connect:        45,
+				ResponseHeader: 300,
+				TLSHandshake:   15,
+				ExpectContinue: 2,
+				IdleConn:       120,
+			},
+		},
+	}
+
+	peerProxy, err := NewPeerProxy(peers, testLogger)
+
+	assert.NoError(t, err)
+	assert.NotNil(t, peerProxy)
+	assert.True(t, peerProxy.HasPeerModel("model1"))
+
+	// Verify the timeout values are actually applied to the transport
+	member, found := peerProxy.proxyMap["model1"]
+	require.True(t, found, "model1 should exist in proxyMap")
+	assert.NotNil(t, member.reverseProxy)
+	assert.NotNil(t, member.reverseProxy.Transport)
+
+	transport, ok := member.reverseProxy.Transport.(*http.Transport)
+	require.True(t, ok, "Transport should be *http.Transport")
+
+	// Verify all timeout values are correctly applied
+	assert.Equal(t, 300*time.Second, transport.ResponseHeaderTimeout)
+	assert.Equal(t, 15*time.Second, transport.TLSHandshakeTimeout)
+	assert.Equal(t, 2*time.Second, transport.ExpectContinueTimeout)
+	assert.Equal(t, 120*time.Second, transport.IdleConnTimeout)
+	// ForceAttemptHTTP2 should be enabled
+	assert.True(t, transport.ForceAttemptHTTP2)
+}

proxy/process.go

@@ -77,6 +77,9 @@ type Process struct {
 	// used for testing to override the default value
 	gracefulStopTimeout time.Duration
 
+	// used for testing to bypass subprocess and reverse proxy
+	testHandler http.Handler
+
 	// track the number of failed starts
 	failedStartCount int
 }
@@ -96,6 +99,24 @@ func NewProcess(ID string, healthCheckTimeout int, config config.ModelConfig, pr
 	var reverseProxy *httputil.ReverseProxy
 	if proxyURL != nil {
 		reverseProxy = httputil.NewSingleHostReverseProxy(proxyURL)
+
+		// Create custom transport with configured timeouts
+		transport := &http.Transport{
+			Proxy: http.ProxyFromEnvironment,
+			DialContext: (&net.Dialer{
+				Timeout:   time.Duration(config.Timeouts.Connect) * time.Second,
+				KeepAlive: time.Duration(config.Timeouts.KeepAlive) * time.Second,
+			}).DialContext,
+			TLSHandshakeTimeout:   time.Duration(config.Timeouts.TLSHandshake) * time.Second,
+			ResponseHeaderTimeout: time.Duration(config.Timeouts.ResponseHeader) * time.Second,
+			ExpectContinueTimeout: time.Duration(config.Timeouts.ExpectContinue) * time.Second,
+			ForceAttemptHTTP2:     true,
+			MaxIdleConns:          100,
+			MaxIdleConnsPerHost:   10,
+			IdleConnTimeout:       time.Duration(config.Timeouts.IdleConn) * time.Second,
+		}
+		reverseProxy.Transport = transport
+
 		reverseProxy.ModifyResponse = func(resp *http.Response) error {
 			// prevent nginx from buffering streaming responses (e.g., SSE)
 			if strings.Contains(strings.ToLower(resp.Header.Get("Content-Type")), "text/event-stream") {
@@ -218,6 +239,49 @@ func (p *Process) forceState(newState ProcessState) {
 // at any time.
 func (p *Process) start() error {
 
+	// test-only fast path: skip subprocess, health check, and TTL goroutine
+	if p.testHandler != nil {
+		if curState, err := p.swapState(StateStopped, StateStarting); err != nil {
+			if err == ErrExpectedStateMismatch {
+				if curState == StateStarting {
+					p.waitStarting.Wait()
+					curState = p.CurrentState()
+					if curState == StateReady {
+						return nil
+					}
+					return fmt.Errorf("process was already starting but wound up in state %v", curState)
+				}
+				return fmt.Errorf("process was in state %v when start() was called", curState)
+			}
+			return fmt.Errorf("failed to set Process state to starting: current state: %v, error: %v", curState, err)
+		}
+		defer p.waitStarting.Done()
+
+		// Mimic the real stop path: cancelUpstream transitions
+		// StateStopping -> StateStopped and closes cmdWaitChan,
+		// matching what waitForCmd does for real subprocesses.
+		ch := make(chan struct{})
+		p.cmdMutex.Lock()
+		p.cancelUpstream = func() {
+			if curState := p.CurrentState(); curState == StateStopping {
+				if _, err := p.swapState(StateStopping, StateStopped); err != nil {
+					p.forceState(StateStopped)
+				}
+			} else {
+				p.forceState(StateStopped)
+			}
+			close(ch)
+		}
+		p.cmdWaitChan = ch
+		p.cmdMutex.Unlock()
+
+		if curState, err := p.swapState(StateStarting, StateReady); err != nil {
+			return fmt.Errorf("failed to set Process state to ready: current state: %v, error: %v", curState, err)
+		}
+		p.failedStartCount = 0
+		return nil
+	}
+
 	if p.config.Proxy == "" {
 		return fmt.Errorf("can not start(), upstream proxy missing")
 	}
@@ -368,7 +432,10 @@ func (p *Process) start() error {
 
 // Stop will wait for inflight requests to complete before stopping the process.
 func (p *Process) Stop() {
+
+	// guard to prevent multiple goroutines from stopping
 	if !isValidTransition(p.CurrentState(), StateStopping) {
+		p.proxyLogger.Debugf("<%s> Stop() suppressing invalid transition from %s to StateStopping", p.ID, p.CurrentState())
 		return
 	}
 
@@ -381,13 +448,17 @@ func (p *Process) Stop() {
 // StopImmediately will transition the process to the stopping state and stop the process with a SIGTERM.
 // If the process does not stop within the specified timeout, it will be forcefully stopped with a SIGKILL.
 func (p *Process) StopImmediately() {
-	if !isValidTransition(p.CurrentState(), StateStopping) {
+
+	// guard to prevent multiple goroutines from stopping the process
+	enterState := p.CurrentState()
+	if !isValidTransition(enterState, StateStopping) {
+		p.proxyLogger.Debugf("<%s> StopImmediate() suppressing invalid transition from %s to StateStopping", p.ID, p.CurrentState())
 		return
 	}
 
-	p.proxyLogger.Debugf("<%s> Stopping process, current state: %s", p.ID, p.CurrentState())
-	if curState, err := p.swapState(StateReady, StateStopping); err != nil {
-		p.proxyLogger.Infof("<%s> Stop() Ready -> StateStopping err: %v, current state: %v", p.ID, err, curState)
+	p.proxyLogger.Debugf("<%s> Stopping process, enter state: %s", p.ID, enterState)
+	if curState, err := p.swapState(enterState, StateStopping); err != nil {
+		p.proxyLogger.Infof("<%s> Stop() %s -> StateStopping err: %v, current state: %v", p.ID, enterState, err, curState)
 		return
 	}
 
@@ -559,6 +630,11 @@ func (p *Process) ProxyRequest(w http.ResponseWriter, r *http.Request) {
 		if !srw.waitForCompletion(completionTimeout) {
 			p.proxyLogger.Warnf("<%s> status updates goroutine did not complete within %v, proceeding with proxy request", p.ID, completionTimeout)
 		}
+	}
+
+	if p.testHandler != nil {
+		p.testHandler.ServeHTTP(w, r)
+	} else if srw != nil {
 		p.reverseProxy.ServeHTTP(srw, r)
 	} else {
 		p.reverseProxy.ServeHTTP(w, r)

proxy/process_test.go

@@ -2,6 +2,7 @@ package proxy
 
 import (
 	"fmt"
+	"io"
 	"net/http"
 	"net/http/httptest"
 	"os"
@@ -117,12 +118,12 @@ func TestProcess_UnloadAfterTTL(t *testing.T) {
 	}
 
 	expectedMessage := "I_sense_imminent_danger"
-	config := getTestSimpleResponderConfig(expectedMessage)
-	assert.Equal(t, 0, config.UnloadAfter)
-	config.UnloadAfter = 3 // seconds
-	assert.Equal(t, 3, config.UnloadAfter)
+	conf := getTestSimpleResponderConfig(expectedMessage)
+	assert.Equal(t, config.MODEL_CONFIG_DEFAULT_TTL, conf.UnloadAfter)
+	conf.UnloadAfter = 3 // seconds
+	assert.Equal(t, 3, conf.UnloadAfter)
 
-	process := NewProcess("ttl_test", 2, config, debugLogger, debugLogger)
+	process := NewProcess("ttl_test", 2, conf, debugLogger, debugLogger)
 	defer process.Stop()
 
 	// this should take 4 seconds
@@ -159,12 +160,12 @@ func TestProcess_LowTTLValue(t *testing.T) {
 		t.Skip("skipping test, edit process_test.go to run it ")
 	}
 
-	config := getTestSimpleResponderConfig("fast_ttl")
-	assert.Equal(t, 0, config.UnloadAfter)
-	config.UnloadAfter = 1 // second
-	assert.Equal(t, 1, config.UnloadAfter)
+	conf := getTestSimpleResponderConfig("fast_ttl")
+	assert.Equal(t, config.MODEL_CONFIG_DEFAULT_TTL, conf.UnloadAfter)
+	conf.UnloadAfter = 1 // second
+	assert.Equal(t, 1, conf.UnloadAfter)
 
-	process := NewProcess("ttl", 2, config, debugLogger, debugLogger)
+	process := NewProcess("ttl", 2, conf, debugLogger, debugLogger)
 	defer process.Stop()
 
 	for i := 0; i < 100; i++ {
@@ -569,3 +570,39 @@ func (w *panicOnWriteResponseWriter) Write(b []byte) (int, error) {
 	}
 	return w.ResponseRecorder.Write(b)
 }
+
+func TestProcess_CustomTimeouts(t *testing.T) {
+	modelConfig := config.ModelConfig{
+		Cmd:           "echo test",
+		Proxy:         "http://localhost:8080",
+		CheckEndpoint: "/health",
+		Timeouts: config.TimeoutsConfig{
+			Connect:        45,
+			ResponseHeader: 120,
+			TLSHandshake:   15,
+			ExpectContinue: 2,
+			IdleConn:       120,
+		},
+	}
+
+	debugLogger := NewLogMonitorWriter(io.Discard)
+	process := NewProcess("test-model", 30, modelConfig, debugLogger, debugLogger)
+
+	// Verify the process was created successfully
+	assert.NotNil(t, process)
+	assert.Equal(t, "test-model", process.ID)
+	assert.NotNil(t, process.reverseProxy)
+	assert.NotNil(t, process.reverseProxy.Transport)
+
+	// Verify it's using http.Transport (not some other type)
+	transport, ok := process.reverseProxy.Transport.(*http.Transport)
+	assert.True(t, ok, "Transport should be *http.Transport")
+	assert.NotNil(t, transport)
+
+	// Verify the timeouts are correctly applied
+	assert.Equal(t, 120*time.Second, transport.ResponseHeaderTimeout)
+	assert.Equal(t, 15*time.Second, transport.TLSHandshakeTimeout)
+	assert.Equal(t, 2*time.Second, transport.ExpectContinueTimeout)
+	assert.Equal(t, 120*time.Second, transport.IdleConnTimeout)
+	assert.True(t, transport.ForceAttemptHTTP2)
+}

proxy/processgroup.go

@@ -24,6 +24,22 @@ type ProcessGroup struct {
 	// map of current processes
 	processes       map[string]*Process
 	lastUsedProcess string
+
+	// inflight tracks fast-path requests (requests for the already-selected
+	// model in a swap group). Fast-path requests Add(1) while holding pg.Lock
+	// and Done() on completion; a concurrent swap request calls inflight.Wait()
+	// under pg.Lock before stopping the current process. Without this tracking,
+	// a fast-path request that has released pg.Lock but has not yet called
+	// Process.inFlightRequests.Add(1) races with Stop()'s Wait() and can be
+	// killed mid-request.
+	inflight sync.WaitGroup
+
+	// testDelayFastPath is a test-only hook that, when non-nil, is invoked in
+	// the fast path after pg.Lock is released but before the request is
+	// dispatched to Process.ProxyRequest. Tests use it to park a fast-path
+	// request at the exact race window to deterministically reproduce the
+	// fast-path vs swap race.
+	testDelayFastPath func()
 }
 
 func NewProcessGroup(id string, config config.Config, proxyLogger *LogMonitor, upstreamLogger *LogMonitor) *ProcessGroup {
@@ -64,6 +80,13 @@ func (pg *ProcessGroup) ProxyRequest(modelID string, writer http.ResponseWriter,
 		pg.Lock()
 		if pg.lastUsedProcess != modelID {
 
+			// Wait for in-flight fast-path requests to drain before stopping
+			// the previous process. Without this, a fast-path request that has
+			// released pg.Lock but has not yet incremented
+			// Process.inFlightRequests races with Stop() and can be killed
+			// mid-request.
+			pg.inflight.Wait()
+
 			// is there something already running?
 			if pg.lastUsedProcess != "" {
 				pg.processes[pg.lastUsedProcess].Stop()
@@ -78,7 +101,16 @@ func (pg *ProcessGroup) ProxyRequest(modelID string, writer http.ResponseWriter,
 			pg.Unlock()
 			return nil
 		}
+
+		// Fast path: register this request in inflight before releasing
+		// pg.Lock so a concurrent swap will wait for it to complete.
+		pg.inflight.Add(1)
+		defer pg.inflight.Done()
 		pg.Unlock()
+
+		if pg.testDelayFastPath != nil {
+			pg.testDelayFastPath()
+		}
 	}
 
 	pg.processes[modelID].ProxyRequest(writer, request)
@@ -123,6 +155,10 @@ func (pg *ProcessGroup) StopProcesses(strategy StopStrategy) {
 	pg.Lock()
 	defer pg.Unlock()
 
+	if strategy != StopImmediately {
+		pg.inflight.Wait()
+	}
+
 	if len(pg.processes) == 0 {
 		return
 	}

proxy/processgroup_test.go

@@ -4,11 +4,14 @@ import (
 	"bytes"
 	"net/http"
 	"net/http/httptest"
+	"runtime"
 	"sync"
 	"testing"
+	"time"
 
 	"github.com/mostlygeek/llama-swap/proxy/config"
 	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
 )
 
 var processGroupTestConfig = config.AddDefaultGroupToConfig(config.Config{
@@ -95,6 +98,229 @@ func TestProcessGroup_ProxyRequestSwapIsTrueParallel(t *testing.T) {
 	wg.Wait()
 }
 
+// TestProcessGroup_ProxyRequestSwapRaceAgainstFastPath verifies that a swap
+// request cannot stop the current process while a fast-path request (for the
+// already-selected model) is in flight. Without ProcessGroup-level inflight
+// tracking, a fast-path request that has released pg.Lock but has not yet
+// incremented Process.inFlightRequests races with Stop()'s Wait() and the
+// process is killed mid-request.
+func TestProcessGroup_ProxyRequestSwapRaceAgainstFastPath(t *testing.T) {
+	cfg := config.AddDefaultGroupToConfig(config.Config{
+		HealthCheckTimeout: 15,
+		Models: map[string]config.ModelConfig{
+			"model1": getTestSimpleResponderConfig("model1"),
+			"model2": getTestSimpleResponderConfig("model2"),
+		},
+		Groups: map[string]config.GroupConfig{
+			"G1": {
+				Swap:    true,
+				Members: []string{"model1", "model2"},
+			},
+		},
+	})
+
+	pg := NewProcessGroup("G1", cfg, testLogger, testLogger)
+	defer pg.StopProcesses(StopImmediately)
+
+	// Bypass real subprocesses so the test is fast and deterministic.
+	pg.processes["model1"].testHandler = newTestHandler("model1")
+	pg.processes["model2"].testHandler = newTestHandler("model2")
+
+	// Prime: run a request through model1 via the swap path so that
+	// lastUsedProcess == "model1" and subsequent model1 requests take the
+	// fast path.
+	primeReq := httptest.NewRequest("POST", "/v1/chat/completions", nil)
+	primeW := httptest.NewRecorder()
+	require.NoError(t, pg.ProxyRequest("model1", primeW, primeReq))
+	require.Equal(t, http.StatusOK, primeW.Code)
+	require.Equal(t, StateReady, pg.processes["model1"].CurrentState())
+	require.Equal(t, StateStopped, pg.processes["model2"].CurrentState())
+
+	// Fast-path hook: signal arrival at the race window, then wait for
+	// release. This parks R2 deterministically at the point where pg.Lock
+	// has been released but Process.inFlightRequests has not yet been
+	// incremented — the exact window the race exploits.
+	r2Reached := make(chan struct{})
+	r2Release := make(chan struct{})
+	pg.testDelayFastPath = func() {
+		close(r2Reached)
+		<-r2Release
+	}
+
+	// R2: fast-path request for model1. Will pause at the test hook.
+	r2Done := make(chan struct{})
+	w2 := httptest.NewRecorder()
+	go func() {
+		defer close(r2Done)
+		req := httptest.NewRequest("POST", "/v1/chat/completions", nil)
+		assert.NoError(t, pg.ProxyRequest("model1", w2, req))
+	}()
+
+	// Deterministically wait for R2 to reach the race window.
+	<-r2Reached
+
+	// R3: swap request for model2. Must wait for R2 to finish before touching
+	// model1, otherwise model1 gets killed mid-request.
+	r3Done := make(chan struct{})
+	w3 := httptest.NewRecorder()
+	go func() {
+		defer close(r3Done)
+		req := httptest.NewRequest("POST", "/v1/chat/completions", nil)
+		assert.NoError(t, pg.ProxyRequest("model2", w3, req))
+	}()
+
+	// Spin until R3 has acquired pg.Lock and entered the swap critical
+	// section. In the fixed code, R3 then blocks on pg.inflight.Wait() while
+	// still holding the lock, so TryLock keeps failing.
+	for pg.TryLock() {
+		pg.Unlock()
+		runtime.Gosched()
+	}
+
+	// Bounded poll: give R3 a chance to demonstrate the bug by mutating
+	// state. In the fixed code, R3 is blocked on pg.inflight.Wait() and
+	// nothing changes, so we wait the full window. In the buggy code, R3
+	// will Stop() model1 and start serving via model2 within microseconds —
+	// we exit early once the mutation is observable.
+	deadline := time.Now().Add(100 * time.Millisecond)
+	for time.Now().Before(deadline) {
+		if pg.processes["model1"].CurrentState() != StateReady ||
+			pg.processes["model2"].CurrentState() != StateStopped {
+			break
+		}
+		done := false
+		select {
+		case <-r3Done:
+			done = true
+		default:
+		}
+		if done {
+			break
+		}
+		runtime.Gosched()
+	}
+
+	// Invariant: R3 must be blocked while R2 is still in flight.
+	select {
+	case <-r3Done:
+		t.Fatal("swap completed while fast-path request was still in flight — race not prevented")
+	default:
+	}
+	assert.Equal(t, StateReady, pg.processes["model1"].CurrentState(),
+		"model1 must stay Ready while a fast-path request is in flight")
+	assert.Equal(t, StateStopped, pg.processes["model2"].CurrentState(),
+		"model2 must not be started until R2 finishes and model1 is swapped out")
+
+	// Release R2 and let both requests finish.
+	close(r2Release)
+	<-r2Done
+	<-r3Done
+
+	assert.Equal(t, http.StatusOK, w2.Code)
+	assert.Contains(t, w2.Body.String(), "model1")
+	assert.Equal(t, http.StatusOK, w3.Code)
+	assert.Contains(t, w3.Body.String(), "model2")
+}
+
+// TestProcessGroup_StopProcessesWaitsForInflight verifies that StopProcesses
+// (called externally, e.g. from ProxyManager.swapProcessGroup) cannot stop a
+// process while a fast-path ProxyRequest is in the [pg.Unlock,
+// Process.inFlightRequests.Add(1)] window. Without pg.inflight.Wait() in
+// StopProcesses, the external caller bypasses the inflight guard and kills the
+// process mid-request.
+func TestProcessGroup_StopProcessesWaitsForInflight(t *testing.T) {
+	cfg := config.AddDefaultGroupToConfig(config.Config{
+		HealthCheckTimeout: 15,
+		Models: map[string]config.ModelConfig{
+			"model1": getTestSimpleResponderConfig("model1"),
+			"model2": getTestSimpleResponderConfig("model2"),
+		},
+		Groups: map[string]config.GroupConfig{
+			"G1": {
+				Swap:    true,
+				Members: []string{"model1", "model2"},
+			},
+		},
+	})
+
+	pg := NewProcessGroup("G1", cfg, testLogger, testLogger)
+	defer pg.StopProcesses(StopImmediately)
+
+	pg.processes["model1"].testHandler = newTestHandler("model1")
+	pg.processes["model2"].testHandler = newTestHandler("model2")
+
+	// Prime: model1 is active so subsequent model1 requests take the fast path.
+	primeReq := httptest.NewRequest("POST", "/v1/chat/completions", nil)
+	primeW := httptest.NewRecorder()
+	require.NoError(t, pg.ProxyRequest("model1", primeW, primeReq))
+	require.Equal(t, http.StatusOK, primeW.Code)
+	require.Equal(t, StateReady, pg.processes["model1"].CurrentState())
+
+	// Park a fast-path request at the race window.
+	r2Reached := make(chan struct{})
+	r2Release := make(chan struct{})
+	pg.testDelayFastPath = func() {
+		close(r2Reached)
+		<-r2Release
+	}
+
+	r2Done := make(chan struct{})
+	w2 := httptest.NewRecorder()
+	go func() {
+		defer close(r2Done)
+		req := httptest.NewRequest("POST", "/v1/chat/completions", nil)
+		assert.NoError(t, pg.ProxyRequest("model1", w2, req))
+	}()
+
+	<-r2Reached
+
+	// Simulate an external caller (e.g. ProxyManager.swapProcessGroup) stopping
+	// the group while a fast-path request is in flight.
+	r3Done := make(chan struct{})
+	go func() {
+		defer close(r3Done)
+		pg.StopProcesses(StopWaitForInflightRequest)
+	}()
+
+	// Spin until StopProcesses has acquired pg.Lock.
+	for pg.TryLock() {
+		pg.Unlock()
+		runtime.Gosched()
+	}
+
+	// Bounded poll: in the fixed code StopProcesses blocks on pg.inflight.Wait()
+	// and model1 stays Ready. In the buggy code it proceeds immediately and
+	// kills model1.
+	deadline := time.Now().Add(100 * time.Millisecond)
+	for time.Now().Before(deadline) {
+		if pg.processes["model1"].CurrentState() != StateReady {
+			break
+		}
+		select {
+		case <-r3Done:
+			goto done
+		default:
+		}
+		runtime.Gosched()
+	}
+done:
+
+	select {
+	case <-r3Done:
+		t.Fatal("StopProcesses completed while a fast-path request was still in flight — race not prevented")
+	default:
+	}
+	assert.Equal(t, StateReady, pg.processes["model1"].CurrentState(),
+		"model1 must stay Ready while a fast-path request is in flight")
+
+	close(r2Release)
+	<-r2Done
+	<-r3Done
+
+	assert.Equal(t, http.StatusOK, w2.Code)
+	assert.Contains(t, w2.Body.String(), "model1")
+}
+
 func TestProcessGroup_ProxyRequestSwapIsFalse(t *testing.T) {
 	pg := NewProcessGroup("G2", processGroupTestConfig, testLogger, testLogger)
 	defer pg.StopProcesses(StopWaitForInflightRequest)

proxy/proxymanager.go

@@ -28,6 +28,40 @@ const (
 
 type proxyCtxKey string
 
+type InflightCounter struct {
+	mu    sync.Mutex
+	total int
+}
+
+func newInflightCounter() *InflightCounter {
+	return &InflightCounter{}
+}
+
+func (ic *InflightCounter) Current() int {
+	ic.mu.Lock()
+	total := ic.total
+	ic.mu.Unlock()
+	return total
+}
+
+func (ic *InflightCounter) Increment() int {
+	ic.mu.Lock()
+	ic.total++
+	total := ic.total
+	ic.mu.Unlock()
+	return total
+}
+
+func (ic *InflightCounter) Decrement() int {
+	ic.mu.Lock()
+	if ic.total > 0 {
+		ic.total--
+	}
+	total := ic.total
+	ic.mu.Unlock()
+	return total
+}
+
 type ProxyManager struct {
 	sync.Mutex
 
@@ -43,6 +77,11 @@ type ProxyManager struct {
 
 	processGroups map[string]*ProcessGroup
 
+	// matrix-based swap (mutually exclusive with processGroups)
+	matrix *Matrix
+
+	inFlightCounter *InflightCounter
+
 	// shutdown signaling
 	shutdownCtx    context.Context
 	shutdownCancel context.CancelFunc
@@ -151,10 +190,12 @@ func New(proxyConfig config.Config) *ProxyManager {
 		muxLogger:      muxLogger,
 		upstreamLogger: upstreamLogger,
 
-		metricsMonitor: newMetricsMonitor(proxyLogger, maxMetrics),
+		metricsMonitor: newMetricsMonitor(proxyLogger, maxMetrics, proxyConfig.CaptureBuffer),
 
 		processGroups: make(map[string]*ProcessGroup),
 
+		inFlightCounter: newInflightCounter(),
+
 		shutdownCtx:    shutdownCtx,
 		shutdownCancel: shutdownCancel,
 
@@ -165,10 +206,14 @@ func New(proxyConfig config.Config) *ProxyManager {
 		peerProxy: peerProxy,
 	}
 
-	// create the process groups
-	for groupID := range proxyConfig.Groups {
-		processGroup := NewProcessGroup(groupID, proxyConfig, proxyLogger, upstreamLogger)
-		pm.processGroups[groupID] = processGroup
+	// create either matrix or process groups (mutually exclusive)
+	if proxyConfig.Matrix != nil {
+		pm.matrix = NewMatrix(proxyConfig, proxyLogger, upstreamLogger)
+	} else {
+		for groupID := range proxyConfig.Groups {
+			processGroup := NewProcessGroup(groupID, proxyConfig, proxyLogger, upstreamLogger)
+			pm.processGroups[groupID] = processGroup
+		}
 	}
 
 	pm.setupGinEngine()
@@ -187,18 +232,29 @@ func New(proxyConfig config.Config) *ProxyManager {
 				}
 
 				proxyLogger.Infof("Preloading model: %s", modelID)
-				processGroup, err := pm.swapProcessGroup(modelID)
 
-				if err != nil {
+				var preloadErr error
+				req, _ := http.NewRequest("GET", "/", nil)
+
+				if pm.matrix != nil {
+					preloadErr = pm.matrix.ProxyRequest(modelID, discardWriter, req)
+				} else {
+					processGroup, err := pm.swapProcessGroup(modelID)
+					if err != nil {
+						preloadErr = err
+					} else {
+						preloadErr = processGroup.ProxyRequest(modelID, discardWriter, req)
+					}
+				}
+
+				if preloadErr != nil {
 					event.Emit(ModelPreloadedEvent{
 						ModelName: modelID,
 						Success:   false,
 					})
-					proxyLogger.Errorf("Failed to preload model %s: %v", modelID, err)
+					proxyLogger.Errorf("Failed to preload model %s: %v", modelID, preloadErr)
 					continue
 				} else {
-					req, _ := http.NewRequest("GET", "/", nil)
-					processGroup.ProxyRequest(modelID, discardWriter, req)
 					event.Emit(ModelPreloadedEvent{
 						ModelName: modelID,
 						Success:   true,
@@ -276,34 +332,42 @@ func (pm *ProxyManager) setupGinEngine() {
 
 	// Set up routes using the Gin engine
 	// Protected routes use pm.apiKeyAuth() middleware
-	pm.ginEngine.POST("/v1/chat/completions", pm.apiKeyAuth(), pm.proxyInferenceHandler)
-	pm.ginEngine.POST("/v1/responses", pm.apiKeyAuth(), pm.proxyInferenceHandler)
+	pm.ginEngine.POST("/v1/chat/completions", pm.apiKeyAuth(), pm.trackInflight(), pm.proxyInferenceHandler)
+	pm.ginEngine.POST("/v1/responses", pm.apiKeyAuth(), pm.trackInflight(), pm.proxyInferenceHandler)
 	// Support legacy /v1/completions api, see issue #12
-	pm.ginEngine.POST("/v1/completions", pm.apiKeyAuth(), pm.proxyInferenceHandler)
+	pm.ginEngine.POST("/v1/completions", pm.apiKeyAuth(), pm.trackInflight(), pm.proxyInferenceHandler)
 	// Support anthropic /v1/messages (added https://github.com/ggml-org/llama.cpp/pull/17570)
-	pm.ginEngine.POST("/v1/messages", pm.apiKeyAuth(), pm.proxyInferenceHandler)
+	pm.ginEngine.POST("/v1/messages", pm.apiKeyAuth(), pm.trackInflight(), pm.proxyInferenceHandler)
+	// Support anthropic count_tokens API (Also added in the above PR)
+	pm.ginEngine.POST("/v1/messages/count_tokens", pm.apiKeyAuth(), pm.trackInflight(), pm.proxyInferenceHandler)
 
 	// Support embeddings and reranking
-	pm.ginEngine.POST("/v1/embeddings", pm.apiKeyAuth(), pm.proxyInferenceHandler)
+	pm.ginEngine.POST("/v1/embeddings", pm.apiKeyAuth(), pm.trackInflight(), pm.proxyInferenceHandler)
 
 	// llama-server's /reranking endpoint + aliases
-	pm.ginEngine.POST("/reranking", pm.apiKeyAuth(), pm.proxyInferenceHandler)
-	pm.ginEngine.POST("/rerank", pm.apiKeyAuth(), pm.proxyInferenceHandler)
-	pm.ginEngine.POST("/v1/rerank", pm.apiKeyAuth(), pm.proxyInferenceHandler)
-	pm.ginEngine.POST("/v1/reranking", pm.apiKeyAuth(), pm.proxyInferenceHandler)
+	pm.ginEngine.POST("/reranking", pm.apiKeyAuth(), pm.trackInflight(), pm.proxyInferenceHandler)
+	pm.ginEngine.POST("/rerank", pm.apiKeyAuth(), pm.trackInflight(), pm.proxyInferenceHandler)
+	pm.ginEngine.POST("/v1/rerank", pm.apiKeyAuth(), pm.trackInflight(), pm.proxyInferenceHandler)
+	pm.ginEngine.POST("/v1/reranking", pm.apiKeyAuth(), pm.trackInflight(), pm.proxyInferenceHandler)
 
 	// llama-server's /infill endpoint for code infilling
-	pm.ginEngine.POST("/infill", pm.apiKeyAuth(), pm.proxyInferenceHandler)
+	pm.ginEngine.POST("/infill", pm.apiKeyAuth(), pm.trackInflight(), pm.proxyInferenceHandler)
 
 	// llama-server's /completion endpoint
-	pm.ginEngine.POST("/completion", pm.apiKeyAuth(), pm.proxyInferenceHandler)
+	pm.ginEngine.POST("/completion", pm.apiKeyAuth(), pm.trackInflight(), pm.proxyInferenceHandler)
 
 	// Support audio/speech endpoint
-	pm.ginEngine.POST("/v1/audio/speech", pm.apiKeyAuth(), pm.proxyInferenceHandler)
-	pm.ginEngine.POST("/v1/audio/voices", pm.apiKeyAuth(), pm.proxyInferenceHandler)
-	pm.ginEngine.POST("/v1/audio/transcriptions", pm.apiKeyAuth(), pm.proxyOAIPostFormHandler)
-	pm.ginEngine.POST("/v1/images/generations", pm.apiKeyAuth(), pm.proxyInferenceHandler)
-	pm.ginEngine.POST("/v1/images/edits", pm.apiKeyAuth(), pm.proxyOAIPostFormHandler)
+	pm.ginEngine.POST("/v1/audio/speech", pm.apiKeyAuth(), pm.trackInflight(), pm.proxyInferenceHandler)
+	pm.ginEngine.POST("/v1/audio/voices", pm.apiKeyAuth(), pm.trackInflight(), pm.proxyInferenceHandler)
+	pm.ginEngine.GET("/v1/audio/voices", pm.apiKeyAuth(), pm.trackInflight(), pm.proxyGETModelHandler)
+	pm.ginEngine.POST("/v1/audio/transcriptions", pm.apiKeyAuth(), pm.trackInflight(), pm.proxyOAIPostFormHandler)
+	pm.ginEngine.POST("/v1/images/generations", pm.apiKeyAuth(), pm.trackInflight(), pm.proxyInferenceHandler)
+	pm.ginEngine.POST("/v1/images/edits", pm.apiKeyAuth(), pm.trackInflight(), pm.proxyOAIPostFormHandler)
+
+	// sd.cpp /sdapi/v1 endpoints
+	pm.ginEngine.POST("/sdapi/v1/txt2img", pm.apiKeyAuth(), pm.trackInflight(), pm.proxyInferenceHandler)
+	pm.ginEngine.POST("/sdapi/v1/img2img", pm.apiKeyAuth(), pm.trackInflight(), pm.proxyInferenceHandler)
+	pm.ginEngine.GET("/sdapi/v1/loras", pm.apiKeyAuth(), pm.trackInflight(), pm.proxyGETModelHandler)
 
 	pm.ginEngine.GET("/v1/models", pm.apiKeyAuth(), pm.listModelsHandler)
 
@@ -322,7 +386,7 @@ func (pm *ProxyManager) setupGinEngine() {
 	pm.ginEngine.GET("/upstream", func(c *gin.Context) {
 		c.Redirect(http.StatusFound, "/ui/models")
 	})
-	pm.ginEngine.Any("/upstream/*upstreamPath", pm.apiKeyAuth(), pm.proxyToUpstream)
+	pm.ginEngine.Any("/upstream/*upstreamPath", pm.apiKeyAuth(), pm.trackInflight(), pm.proxyToUpstream)
 	pm.ginEngine.GET("/unload", pm.apiKeyAuth(), pm.unloadAllModelsHandler)
 	pm.ginEngine.GET("/running", pm.apiKeyAuth(), pm.listRunningProcessesHandler)
 	pm.ginEngine.GET("/health", func(c *gin.Context) {
@@ -346,25 +410,35 @@ func (pm *ProxyManager) setupGinEngine() {
 	if err != nil {
 		pm.proxyLogger.Errorf("Failed to load React filesystem: %v", err)
 	} else {
+		// Serve files with compression support under /ui/*
+		// This handler checks for pre-compressed .br and .gz files
+		pm.ginEngine.GET("/ui/*filepath", func(c *gin.Context) {
+			filepath := strings.TrimPrefix(c.Param("filepath"), "/")
+			// Default to index.html for directory-like paths
+			if filepath == "" {
+				filepath = "index.html"
+			}
 
-		// serve files that exist under /ui/*
-		pm.ginEngine.StaticFS("/ui", reactFS)
+			ServeCompressedFile(reactFS, c.Writer, c.Request, filepath)
+		})
 
-		// server SPA for UI under /ui/*
+		// Serve SPA for UI under /ui/* - fallback to index.html for client-side routing
 		pm.ginEngine.NoRoute(func(c *gin.Context) {
 			if !strings.HasPrefix(c.Request.URL.Path, "/ui") {
 				c.AbortWithStatus(http.StatusNotFound)
 				return
 			}
 
-			file, err := reactFS.Open("index.html")
-			if err != nil {
-				c.String(http.StatusInternalServerError, err.Error())
+			// Check if this looks like a file request (has extension)
+			path := c.Request.URL.Path
+			if strings.Contains(path, ".") && !strings.HasSuffix(path, "/") {
+				// This was likely a file request that wasn't found
+				c.AbortWithStatus(http.StatusNotFound)
 				return
 			}
-			defer file.Close()
-			http.ServeContent(c.Writer, c.Request, "index.html", time.Now(), file)
 
+			// Serve index.html for SPA routing
+			ServeCompressedFile(reactFS, c.Writer, c.Request, "index.html")
 		})
 	}
 
@@ -376,6 +450,14 @@ func (pm *ProxyManager) setupGinEngine() {
 	gin.DisableConsoleColor()
 }
 
+func (pm *ProxyManager) trackInflight() gin.HandlerFunc {
+	return func(c *gin.Context) {
+		event.Emit(InFlightRequestsEvent{Total: pm.inFlightCounter.Increment()})
+		defer event.Emit(InFlightRequestsEvent{Total: pm.inFlightCounter.Decrement()})
+		c.Next()
+	}
+}
+
 // ServeHTTP implements http.Handler interface
 func (pm *ProxyManager) ServeHTTP(w http.ResponseWriter, r *http.Request) {
 	pm.ginEngine.ServeHTTP(w, r)
@@ -389,6 +471,11 @@ func (pm *ProxyManager) StopProcesses(strategy StopStrategy) {
 	pm.Lock()
 	defer pm.Unlock()
 
+	if pm.matrix != nil {
+		pm.matrix.StopProcesses(strategy)
+		return
+	}
+
 	// stop Processes in parallel
 	var wg sync.WaitGroup
 	for _, processGroup := range pm.processGroups {
@@ -409,6 +496,12 @@ func (pm *ProxyManager) Shutdown() {
 
 	pm.proxyLogger.Debug("Shutdown() called in proxy manager")
 
+	if pm.matrix != nil {
+		pm.matrix.Shutdown()
+		pm.shutdownCancel()
+		return
+	}
+
 	var wg sync.WaitGroup
 	// Send shutdown signal to all process in groups
 	for _, processGroup := range pm.processGroups {
@@ -575,10 +668,16 @@ func (pm *ProxyManager) proxyToUpstream(c *gin.Context) {
 		return
 	}
 
-	processGroup, err := pm.swapProcessGroup(modelID)
-	if err != nil {
-		pm.sendErrorResponse(c, http.StatusInternalServerError, fmt.Sprintf("error swapping process group: %s", err.Error()))
-		return
+	var handler func(string, http.ResponseWriter, *http.Request) error
+	if pm.matrix != nil {
+		handler = pm.matrix.ProxyRequest
+	} else {
+		processGroup, err := pm.swapProcessGroup(modelID)
+		if err != nil {
+			pm.sendErrorResponse(c, http.StatusInternalServerError, fmt.Sprintf("error swapping process group: %s", err.Error()))
+			return
+		}
+		handler = processGroup.ProxyRequest
 	}
 
 	// rewrite the path
@@ -587,13 +686,13 @@ func (pm *ProxyManager) proxyToUpstream(c *gin.Context) {
 
 	// attempt to record metrics if it is a POST request
 	if pm.metricsMonitor != nil && c.Request.Method == "POST" {
-		if err := pm.metricsMonitor.wrapHandler(modelID, c.Writer, c.Request, processGroup.ProxyRequest); err != nil {
+		if err := pm.metricsMonitor.wrapHandler(modelID, c.Writer, c.Request, handler); err != nil {
 			pm.sendErrorResponse(c, http.StatusInternalServerError, fmt.Sprintf("error proxying metrics wrapped request: %s", err.Error()))
 			pm.proxyLogger.Errorf("Error proxying wrapped upstream request for model %s, path=%s", modelID, originalPath)
 			return
 		}
 	} else {
-		if err := processGroup.ProxyRequest(modelID, c.Writer, c.Request); err != nil {
+		if err := handler(modelID, c.Writer, c.Request); err != nil {
 			pm.sendErrorResponse(c, http.StatusInternalServerError, fmt.Sprintf("error proxying request: %s", err.Error()))
 			pm.proxyLogger.Errorf("Error proxying upstream request for model %s, path=%s", modelID, originalPath)
 			return
@@ -619,10 +718,16 @@ func (pm *ProxyManager) proxyInferenceHandler(c *gin.Context) {
 
 	modelID, found := pm.config.RealModelName(requestedModel)
 	if found {
-		processGroup, err := pm.swapProcessGroup(modelID)
-		if err != nil {
-			pm.sendErrorResponse(c, http.StatusInternalServerError, fmt.Sprintf("error swapping process group: %s", err.Error()))
-			return
+		var localHandler func(string, http.ResponseWriter, *http.Request) error
+		if pm.matrix != nil {
+			localHandler = pm.matrix.ProxyRequest
+		} else {
+			processGroup, err := pm.swapProcessGroup(modelID)
+			if err != nil {
+				pm.sendErrorResponse(c, http.StatusInternalServerError, fmt.Sprintf("error swapping process group: %s", err.Error()))
+				return
+			}
+			localHandler = processGroup.ProxyRequest
 		}
 
 		// issue #69 allow custom model names to be sent to upstream
@@ -650,13 +755,60 @@ func (pm *ProxyManager) proxyInferenceHandler(c *gin.Context) {
 			}
 		}
 
+		// issue #453 set/override parameters in the JSON body
+		setParams, setParamKeys := pm.config.Models[modelID].Filters.SanitizedSetParams()
+		for _, key := range setParamKeys {
+			pm.proxyLogger.Debugf("<%s> setting param: %s", modelID, key)
+			bodyBytes, err = sjson.SetBytes(bodyBytes, key, setParams[key])
+			if err != nil {
+				pm.sendErrorResponse(c, http.StatusInternalServerError, fmt.Sprintf("error setting parameter %s in request", key))
+				return
+			}
+		}
+
+		// setParamsByID: set params based on the requested model ID (runs after setParams, can override it)
+		setParamsByIDParams, setParamsByIDKeys := pm.config.Models[modelID].Filters.SanitizedSetParamsByID(requestedModel)
+		for _, key := range setParamsByIDKeys {
+			pm.proxyLogger.Debugf("<%s> setting param by id: %s", requestedModel, key)
+			bodyBytes, err = sjson.SetBytes(bodyBytes, key, setParamsByIDParams[key])
+			if err != nil {
+				pm.sendErrorResponse(c, http.StatusInternalServerError, fmt.Sprintf("error setting parameter %s in request", key))
+				return
+			}
+		}
+
 		pm.proxyLogger.Debugf("ProxyManager using local Process for model: %s", requestedModel)
-		nextHandler = processGroup.ProxyRequest
+		nextHandler = localHandler
 	} else if pm.peerProxy != nil && pm.peerProxy.HasPeerModel(requestedModel) {
 		pm.proxyLogger.Debugf("ProxyManager using ProxyPeer for model: %s", requestedModel)
 		modelID = requestedModel
-		nextHandler = pm.peerProxy.ProxyRequest
 
+		// issue #453 apply filters for peer requests
+		peerFilters := pm.peerProxy.GetPeerFilters(requestedModel)
+
+		// Apply stripParams - remove specified parameters from request
+		stripParams := peerFilters.SanitizedStripParams()
+		for _, param := range stripParams {
+			pm.proxyLogger.Debugf("<%s> stripping param: %s", requestedModel, param)
+			bodyBytes, err = sjson.DeleteBytes(bodyBytes, param)
+			if err != nil {
+				pm.sendErrorResponse(c, http.StatusInternalServerError, fmt.Sprintf("error stripping parameter %s from request", param))
+				return
+			}
+		}
+
+		// Apply setParams - set/override specified parameters in request
+		setParams, setParamKeys := peerFilters.SanitizedSetParams()
+		for _, key := range setParamKeys {
+			pm.proxyLogger.Debugf("<%s> setting param: %s", requestedModel, key)
+			bodyBytes, err = sjson.SetBytes(bodyBytes, key, setParams[key])
+			if err != nil {
+				pm.sendErrorResponse(c, http.StatusInternalServerError, fmt.Sprintf("error setting parameter %s in request", key))
+				return
+			}
+		}
+
+		nextHandler = pm.peerProxy.ProxyRequest
 	}
 
 	if nextHandler == nil {
@@ -706,15 +858,33 @@ func (pm *ProxyManager) proxyOAIPostFormHandler(c *gin.Context) {
 		return
 	}
 
+	// Look for a matching local model first, then check peers
+	var nextHandler func(modelID string, w http.ResponseWriter, r *http.Request) error
+	var useModelName string
+
 	modelID, found := pm.config.RealModelName(requestedModel)
-	if !found {
-		pm.sendErrorResponse(c, http.StatusBadRequest, fmt.Sprintf("could not find real modelID for %s", requestedModel))
-		return
+	if found {
+		if pm.matrix != nil {
+			nextHandler = pm.matrix.ProxyRequest
+		} else {
+			processGroup, err := pm.swapProcessGroup(modelID)
+			if err != nil {
+				pm.sendErrorResponse(c, http.StatusInternalServerError, fmt.Sprintf("error swapping process group: %s", err.Error()))
+				return
+			}
+			nextHandler = processGroup.ProxyRequest
+		}
+
+		useModelName = pm.config.Models[modelID].UseModelName
+		pm.proxyLogger.Debugf("ProxyManager using local Process for model: %s", requestedModel)
+	} else if pm.peerProxy != nil && pm.peerProxy.HasPeerModel(requestedModel) {
+		pm.proxyLogger.Debugf("ProxyManager using ProxyPeer for model: %s", requestedModel)
+		modelID = requestedModel
+		nextHandler = pm.peerProxy.ProxyRequest
 	}
 
-	processGroup, err := pm.swapProcessGroup(modelID)
-	if err != nil {
-		pm.sendErrorResponse(c, http.StatusInternalServerError, fmt.Sprintf("error swapping process group: %s", err.Error()))
+	if nextHandler == nil {
+		pm.sendErrorResponse(c, http.StatusBadRequest, fmt.Sprintf("could not find suitable handler for %s", requestedModel))
 		return
 	}
 
@@ -730,8 +900,6 @@ func (pm *ProxyManager) proxyOAIPostFormHandler(c *gin.Context) {
 			// If this is the model field and we have a profile, use just the model name
 			if key == "model" {
 				// # issue #69 allow custom model names to be sent to upstream
-				useModelName := pm.config.Models[modelID].UseModelName
-
 				if useModelName != "" {
 					fieldValue = useModelName
 				} else {
@@ -801,9 +969,50 @@ func (pm *ProxyManager) proxyOAIPostFormHandler(c *gin.Context) {
 	modifiedReq.ContentLength = int64(requestBuffer.Len())
 
 	// Use the modified request for proxying
-	if err := processGroup.ProxyRequest(modelID, c.Writer, modifiedReq); err != nil {
+	if err := nextHandler(modelID, c.Writer, modifiedReq); err != nil {
 		pm.sendErrorResponse(c, http.StatusInternalServerError, fmt.Sprintf("error proxying request: %s", err.Error()))
-		pm.proxyLogger.Errorf("Error Proxying Request for processGroup %s and model %s", processGroup.id, modelID)
+		pm.proxyLogger.Errorf("Error Proxying Request for model %s", modelID)
+		return
+	}
+}
+
+func (pm *ProxyManager) proxyGETModelHandler(c *gin.Context) {
+	requestedModel := c.Query("model")
+	if requestedModel == "" {
+		pm.sendErrorResponse(c, http.StatusBadRequest, "missing required 'model' query parameter")
+		return
+	}
+
+	var nextHandler func(modelID string, w http.ResponseWriter, r *http.Request) error
+	var modelID string
+
+	if realModelID, found := pm.config.RealModelName(requestedModel); found {
+		modelID = realModelID
+		if pm.matrix != nil {
+			nextHandler = pm.matrix.ProxyRequest
+		} else {
+			processGroup, err := pm.swapProcessGroup(realModelID)
+			if err != nil {
+				pm.sendErrorResponse(c, http.StatusInternalServerError, fmt.Sprintf("error swapping process group: %s", err.Error()))
+				return
+			}
+			nextHandler = processGroup.ProxyRequest
+		}
+		pm.proxyLogger.Debugf("ProxyManager using local Process for model: %s", requestedModel)
+	} else if pm.peerProxy != nil && pm.peerProxy.HasPeerModel(requestedModel) {
+		modelID = requestedModel
+		pm.proxyLogger.Debugf("ProxyManager using ProxyPeer for model: %s", requestedModel)
+		nextHandler = pm.peerProxy.ProxyRequest
+	}
+
+	if nextHandler == nil {
+		pm.sendErrorResponse(c, http.StatusBadRequest, fmt.Sprintf("could not find suitable handler for %s", requestedModel))
+		return
+	}
+
+	if err := nextHandler(modelID, c.Writer, c.Request); err != nil {
+		pm.sendErrorResponse(c, http.StatusInternalServerError, fmt.Sprintf("error proxying request: %s", err.Error()))
+		pm.proxyLogger.Errorf("Error Proxying GET Request for model %s", modelID)
 		return
 	}
 }
@@ -888,15 +1097,36 @@ func (pm *ProxyManager) listRunningProcessesHandler(context *gin.Context) {
 	context.Header("Content-Type", "application/json")
 	runningProcesses := make([]gin.H, 0) // Default to an empty response.
 
-	for _, processGroup := range pm.processGroups {
-		for _, process := range processGroup.processes {
-			if process.CurrentState() == StateReady {
+	if pm.matrix != nil {
+		for _, modelID := range pm.matrix.RunningModels() {
+			if process, ok := pm.matrix.GetProcess(modelID); ok {
 				runningProcesses = append(runningProcesses, gin.H{
-					"model": process.ID,
-					"state": process.state,
+					"model":       process.ID,
+					"state":       process.state,
+					"cmd":         process.config.Cmd,
+					"proxy":       process.config.Proxy,
+					"ttl":         process.config.UnloadAfter,
+					"name":        process.config.Name,
+					"description": process.config.Description,
 				})
 			}
 		}
+	} else {
+		for _, processGroup := range pm.processGroups {
+			for _, process := range processGroup.processes {
+				if process.CurrentState() == StateReady {
+					runningProcesses = append(runningProcesses, gin.H{
+						"model":       process.ID,
+						"state":       process.state,
+						"cmd":         process.config.Cmd,
+						"proxy":       process.config.Proxy,
+						"ttl":         process.config.UnloadAfter,
+						"name":        process.config.Name,
+						"description": process.config.Description,
+					})
+				}
+			}
+		}
 	}
 
 	// Put the results under the `running` key.

proxy/proxymanager_api.go

@@ -6,6 +6,7 @@ import (
 	"fmt"
 	"net/http"
 	"sort"
+	"strconv"
 	"strings"
 
 	"github.com/gin-gonic/gin"
@@ -13,12 +14,13 @@ import (
 )
 
 type Model struct {
-	Id          string `json:"id"`
-	Name        string `json:"name"`
-	Description string `json:"description"`
-	State       string `json:"state"`
-	Unlisted    bool   `json:"unlisted"`
-	PeerID      string `json:"peerID"`
+	Id          string   `json:"id"`
+	Name        string   `json:"name"`
+	Description string   `json:"description"`
+	State       string   `json:"state"`
+	Unlisted    bool     `json:"unlisted"`
+	PeerID      string   `json:"peerID"`
+	Aliases     []string `json:"aliases,omitempty"`
 }
 
 func addApiHandlers(pm *ProxyManager) {
@@ -31,6 +33,7 @@ func addApiHandlers(pm *ProxyManager) {
 		apiGroup.GET("/events", pm.apiSendEvents)
 		apiGroup.GET("/metrics", pm.apiGetMetrics)
 		apiGroup.GET("/version", pm.apiGetVersion)
+		apiGroup.GET("/captures/:id", pm.apiGetCapture)
 	}
 }
 
@@ -52,27 +55,28 @@ func (pm *ProxyManager) getModelStatus() []Model {
 	// Iterate over sorted keys
 	for _, modelID := range modelIDs {
 		// Get process state
-		processGroup := pm.findGroupByModelName(modelID)
 		state := "unknown"
-		if processGroup != nil {
-			process := processGroup.processes[modelID]
-			if process != nil {
-				var stateStr string
-				switch process.CurrentState() {
-				case StateReady:
-					stateStr = "ready"
-				case StateStarting:
-					stateStr = "starting"
-				case StateStopping:
-					stateStr = "stopping"
-				case StateShutdown:
-					stateStr = "shutdown"
-				case StateStopped:
-					stateStr = "stopped"
-				default:
-					stateStr = "unknown"
-				}
-				state = stateStr
+		var process *Process
+		if pm.matrix != nil {
+			process, _ = pm.matrix.GetProcess(modelID)
+		} else {
+			processGroup := pm.findGroupByModelName(modelID)
+			if processGroup != nil {
+				process = processGroup.processes[modelID]
+			}
+		}
+		if process != nil {
+			switch process.CurrentState() {
+			case StateReady:
+				state = "ready"
+			case StateStarting:
+				state = "starting"
+			case StateStopping:
+				state = "stopping"
+			case StateShutdown:
+				state = "shutdown"
+			case StateStopped:
+				state = "stopped"
 			}
 		}
 		models = append(models, Model{
@@ -81,6 +85,7 @@ func (pm *ProxyManager) getModelStatus() []Model {
 			Description: pm.config.Models[modelID].Description,
 			State:       state,
 			Unlisted:    pm.config.Models[modelID].Unlisted,
+			Aliases:     pm.config.Models[modelID].Aliases,
 		})
 	}
 
@@ -105,6 +110,7 @@ const (
 	msgTypeModelStatus messageType = "modelStatus"
 	msgTypeLogData     messageType = "logData"
 	msgTypeMetrics     messageType = "metrics"
+	msgTypeInFlight    messageType = "inflight"
 )
 
 type messageEnvelope struct {
@@ -164,6 +170,18 @@ func (pm *ProxyManager) apiSendEvents(c *gin.Context) {
 		}
 	}
 
+	sendInFlight := func(total int) {
+		jsonData, err := json.Marshal(gin.H{"total": total})
+		if err == nil {
+			select {
+			case sendBuffer <- messageEnvelope{Type: msgTypeInFlight, Data: string(jsonData)}:
+			case <-ctx.Done():
+				return
+			default:
+			}
+		}
+	}
+
 	/**
 	 * Send updated models list
 	 */
@@ -191,11 +209,19 @@ func (pm *ProxyManager) apiSendEvents(c *gin.Context) {
 		sendMetrics([]TokenMetrics{e.Metrics})
 	})()
 
+	/**
+	 * Send in-flight request stats related to token stats "Waiting: N" count.
+	 */
+	defer event.On(func(e InFlightRequestsEvent) {
+		sendInFlight(e.Total)
+	})()
+
 	// send initial batch of data
 	sendLogData("proxy", pm.proxyLogger.GetHistory())
 	sendLogData("upstream", pm.upstreamLogger.GetHistory())
 	sendModels()
 	sendMetrics(pm.metricsMonitor.getMetrics())
+	sendInFlight(pm.inFlightCounter.Current())
 
 	for {
 		select {
@@ -229,18 +255,23 @@ func (pm *ProxyManager) apiUnloadSingleModelHandler(c *gin.Context) {
 		return
 	}
 
-	processGroup := pm.findGroupByModelName(realModelName)
-	if processGroup == nil {
-		pm.sendErrorResponse(c, http.StatusInternalServerError, fmt.Sprintf("process group not found for model %s", requestedModel))
-		return
+	var stopErr error
+	if pm.matrix != nil {
+		stopErr = pm.matrix.StopProcess(realModelName, StopImmediately)
+	} else {
+		processGroup := pm.findGroupByModelName(realModelName)
+		if processGroup == nil {
+			pm.sendErrorResponse(c, http.StatusInternalServerError, fmt.Sprintf("process group not found for model %s", requestedModel))
+			return
+		}
+		stopErr = processGroup.StopProcess(realModelName, StopImmediately)
 	}
 
-	if err := processGroup.StopProcess(realModelName, StopImmediately); err != nil {
-		pm.sendErrorResponse(c, http.StatusInternalServerError, fmt.Sprintf("error stopping process: %s", err.Error()))
+	if stopErr != nil {
+		pm.sendErrorResponse(c, http.StatusInternalServerError, fmt.Sprintf("error stopping process: %s", stopErr.Error()))
 		return
-	} else {
-		c.String(http.StatusOK, "OK")
 	}
+	c.String(http.StatusOK, "OK")
 }
 
 func (pm *ProxyManager) apiGetVersion(c *gin.Context) {
@@ -250,3 +281,35 @@ func (pm *ProxyManager) apiGetVersion(c *gin.Context) {
 		"build_date": pm.buildDate,
 	})
 }
+
+func (pm *ProxyManager) apiGetCapture(c *gin.Context) {
+	idStr := c.Param("id")
+	id, err := strconv.Atoi(idStr)
+	if err != nil {
+		c.JSON(http.StatusBadRequest, gin.H{"error": "invalid capture ID"})
+		return
+	}
+
+	data, exists := pm.metricsMonitor.getCompressedBytes(id)
+	if !exists {
+		c.JSON(http.StatusNotFound, gin.H{"error": "capture not found"})
+		return
+	}
+
+	c.Header("Vary", "Accept-Encoding")
+
+	// ¯\_(ツ)_/¯ quality weights are too fancy for us anyway
+	hasZstd := strings.Contains(c.GetHeader("Accept-Encoding"), "zstd")
+
+	if hasZstd {
+		c.Header("Content-Encoding", "zstd")
+		c.Data(http.StatusOK, "application/json", data)
+	} else {
+		decompressed, err := decompressCapture(data)
+		if err != nil {
+			c.JSON(http.StatusInternalServerError, gin.H{"error": "failed to decompress capture"})
+			return
+		}
+		c.Data(http.StatusOK, "application/json", decompressed)
+	}
+}

proxy/ui_compress.go

@@ -0,0 +1,81 @@
+package proxy
+
+import (
+	"net/http"
+	"strings"
+)
+
+// selectEncoding chooses the best encoding based on Accept-Encoding header
+// Returns the encoding ("br", "gzip", or "") and the corresponding file extension
+func selectEncoding(acceptEncoding string) (encoding, ext string) {
+	if acceptEncoding == "" {
+		return "", ""
+	}
+
+	for _, part := range strings.Split(acceptEncoding, ",") {
+		enc := strings.TrimSpace(strings.SplitN(part, ";", 2)[0])
+		if enc == "br" {
+			return "br", ".br"
+		}
+	}
+
+	for _, part := range strings.Split(acceptEncoding, ",") {
+		enc := strings.TrimSpace(strings.SplitN(part, ";", 2)[0])
+		if enc == "gzip" {
+			return "gzip", ".gz"
+		}
+	}
+
+	return "", ""
+}
+
+// ServeCompressedFile serves a file with compression support.
+// It checks for pre-compressed versions and serves them with proper headers.
+func ServeCompressedFile(fs http.FileSystem, w http.ResponseWriter, r *http.Request, name string) {
+	encoding, ext := selectEncoding(r.Header.Get("Accept-Encoding"))
+
+	// Try to serve compressed version if client supports it
+	if encoding != "" {
+		if cf, err := fs.Open(name + ext); err == nil {
+			defer cf.Close()
+
+			// Verify it's a regular file (not a directory)
+			if stat, err := cf.Stat(); err == nil && !stat.IsDir() {
+				// Set the content encoding header
+				w.Header().Set("Content-Encoding", encoding)
+				w.Header().Add("Vary", "Accept-Encoding")
+
+				// Get original file info for content type detection
+				origFile, err := fs.Open(name)
+				if err == nil {
+					origFile.Close()
+				}
+
+				// Serve the compressed file
+				http.ServeContent(w, r, name, stat.ModTime(), cf)
+				return
+			}
+		}
+	}
+
+	// Fall back to serving the uncompressed file
+	file, err := fs.Open(name)
+	if err != nil {
+		http.Error(w, err.Error(), http.StatusNotFound)
+		return
+	}
+	defer file.Close()
+
+	stat, err := file.Stat()
+	if err != nil {
+		http.Error(w, err.Error(), http.StatusInternalServerError)
+		return
+	}
+
+	if stat.IsDir() {
+		http.Error(w, "is a directory", http.StatusForbidden)
+		return
+	}
+
+	http.ServeContent(w, r, name, stat.ModTime(), file)
+}

proxy/ui_compress_test.go

@@ -0,0 +1,283 @@
+package proxy
+
+import (
+	"bytes"
+	"compress/gzip"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+	"testing/fstest"
+	"time"
+)
+
+func TestServeCompressedFile_Brotli(t *testing.T) {
+	// Create test content
+	content := []byte("This is test content that should be compressed with brotli")
+	brContent := []byte("fake-brotli-compressed-data")
+
+	// Create a test filesystem
+	mapFS := fstest.MapFS{
+		"test.js":    {Data: content, ModTime: time.Now()},
+		"test.js.br": {Data: brContent, ModTime: time.Now()},
+		"test.js.gz": {Data: []byte("fake-gzip-data"), ModTime: time.Now()},
+	}
+	fs := http.FS(mapFS)
+
+	req := httptest.NewRequest(http.MethodGet, "/test.js", nil)
+	req.Header.Set("Accept-Encoding", "br, gzip")
+	w := httptest.NewRecorder()
+
+	ServeCompressedFile(fs, w, req, "test.js")
+
+	resp := w.Result()
+	body, _ := io.ReadAll(resp.Body)
+
+	if resp.StatusCode != http.StatusOK {
+		t.Errorf("Expected status 200, got %d", resp.StatusCode)
+	}
+
+	// Check that brotli is used (preferred over gzip)
+	if encoding := resp.Header.Get("Content-Encoding"); encoding != "br" {
+		t.Errorf("Expected Content-Encoding 'br', got '%s'", encoding)
+	}
+
+	if vary := resp.Header.Get("Vary"); vary != "Accept-Encoding" {
+		t.Errorf("Expected Vary 'Accept-Encoding', got '%s'", vary)
+	}
+
+	if !bytes.Equal(body, brContent) {
+		t.Errorf("Expected brotli content, got %s", string(body))
+	}
+}
+
+func TestServeCompressedFile_Gzip(t *testing.T) {
+	// Create test content
+	content := []byte("This is test content that should be compressed with gzip")
+	gzContent := []byte("fake-gzip-compressed-data")
+
+	// Create a test filesystem without brotli
+	mapFS := fstest.MapFS{
+		"test.js":    {Data: content, ModTime: time.Now()},
+		"test.js.gz": {Data: gzContent, ModTime: time.Now()},
+	}
+	fs := http.FS(mapFS)
+
+	req := httptest.NewRequest(http.MethodGet, "/test.js", nil)
+	req.Header.Set("Accept-Encoding", "gzip")
+	w := httptest.NewRecorder()
+
+	ServeCompressedFile(fs, w, req, "test.js")
+
+	resp := w.Result()
+	body, _ := io.ReadAll(resp.Body)
+
+	if resp.StatusCode != http.StatusOK {
+		t.Errorf("Expected status 200, got %d", resp.StatusCode)
+	}
+
+	if encoding := resp.Header.Get("Content-Encoding"); encoding != "gzip" {
+		t.Errorf("Expected Content-Encoding 'gzip', got '%s'", encoding)
+	}
+
+	if !bytes.Equal(body, gzContent) {
+		t.Errorf("Expected gzip content, got %s", string(body))
+	}
+}
+
+func TestServeCompressedFile_UncompressedFallback(t *testing.T) {
+	// Create test content
+	content := []byte("This is uncompressed test content")
+
+	// Create a test filesystem without compressed versions
+	mapFS := fstest.MapFS{
+		"test.js": {Data: content, ModTime: time.Now()},
+	}
+	fs := http.FS(mapFS)
+
+	req := httptest.NewRequest(http.MethodGet, "/test.js", nil)
+	req.Header.Set("Accept-Encoding", "br, gzip")
+	w := httptest.NewRecorder()
+
+	ServeCompressedFile(fs, w, req, "test.js")
+
+	resp := w.Result()
+	body, _ := io.ReadAll(resp.Body)
+
+	if resp.StatusCode != http.StatusOK {
+		t.Errorf("Expected status 200, got %d", resp.StatusCode)
+	}
+
+	// Should not have Content-Encoding header since we're serving uncompressed
+	if encoding := resp.Header.Get("Content-Encoding"); encoding != "" {
+		t.Errorf("Expected no Content-Encoding, got '%s'", encoding)
+	}
+
+	if !bytes.Equal(body, content) {
+		t.Errorf("Expected original content, got %s", string(body))
+	}
+}
+
+func TestServeCompressedFile_NoAcceptEncoding(t *testing.T) {
+	// Create test content
+	content := []byte("This is test content")
+
+	// Create a test filesystem with compressed versions
+	mapFS := fstest.MapFS{
+		"test.js":    {Data: content, ModTime: time.Now()},
+		"test.js.br": {Data: []byte("brotli"), ModTime: time.Now()},
+		"test.js.gz": {Data: []byte("gzip"), ModTime: time.Now()},
+	}
+	fs := http.FS(mapFS)
+
+	req := httptest.NewRequest(http.MethodGet, "/test.js", nil)
+	// No Accept-Encoding header
+	w := httptest.NewRecorder()
+
+	ServeCompressedFile(fs, w, req, "test.js")
+
+	resp := w.Result()
+	body, _ := io.ReadAll(resp.Body)
+
+	if resp.StatusCode != http.StatusOK {
+		t.Errorf("Expected status 200, got %d", resp.StatusCode)
+	}
+
+	// Should serve uncompressed content
+	if encoding := resp.Header.Get("Content-Encoding"); encoding != "" {
+		t.Errorf("Expected no Content-Encoding, got '%s'", encoding)
+	}
+
+	if !bytes.Equal(body, content) {
+		t.Errorf("Expected original content, got %s", string(body))
+	}
+}
+
+func TestServeCompressedFile_NotFound(t *testing.T) {
+	mapFS := fstest.MapFS{}
+	fs := http.FS(mapFS)
+
+	req := httptest.NewRequest(http.MethodGet, "/nonexistent.js", nil)
+	w := httptest.NewRecorder()
+
+	ServeCompressedFile(fs, w, req, "nonexistent.js")
+
+	resp := w.Result()
+
+	if resp.StatusCode != http.StatusNotFound {
+		t.Errorf("Expected status 404, got %d", resp.StatusCode)
+	}
+}
+
+func TestSelectEncoding(t *testing.T) {
+	tests := []struct {
+		acceptEncoding string
+		wantEncoding   string
+		wantExt        string
+	}{
+		{"br, gzip", "br", ".br"},
+		{"gzip, deflate", "gzip", ".gz"},
+		{"gzip", "gzip", ".gz"},
+		{"br", "br", ".br"},
+		{"", "", ""},
+		{"deflate", "", ""},
+		{"br;q=1.0, gzip;q=0.5", "br", ".br"},
+		{"gzip;q=1.0, br;q=0.5", "br", ".br"},
+		{"browser", "", ""},
+		{"compress, deflate", "", ""},
+	}
+
+	for _, tt := range tests {
+		gotEncoding, gotExt := selectEncoding(tt.acceptEncoding)
+		if gotEncoding != tt.wantEncoding || gotExt != tt.wantExt {
+			t.Errorf("selectEncoding(%q) = (%q, %q), want (%q, %q)",
+				tt.acceptEncoding, gotEncoding, gotExt, tt.wantEncoding, tt.wantExt)
+		}
+	}
+}
+
+// Test with actual pre-compressed files from ui_dist
+func TestServeCompressedFile_RealFiles(t *testing.T) {
+	// Check if ui_dist exists
+	if _, err := os.Stat("./ui_dist"); os.IsNotExist(err) {
+		t.Skip("ui_dist not found, skipping real file test")
+	}
+
+	// Find a .js or .css file that has compressed versions
+	entries, err := os.ReadDir("./ui_dist/assets")
+	if err != nil {
+		t.Skipf("Could not read ui_dist/assets: %v", err)
+	}
+
+	var testFile string
+	for _, entry := range entries {
+		name := entry.Name()
+		if strings.HasSuffix(name, ".js") && !strings.HasSuffix(name, ".js.gz") && !strings.HasSuffix(name, ".js.br") {
+			// Check if compressed versions exist
+			base := strings.TrimSuffix(name, ".js")
+			if _, err := os.Stat(filepath.Join("./ui_dist/assets", base+".js.gz")); err == nil {
+				testFile = "assets/" + name
+				break
+			}
+		}
+	}
+
+	if testFile == "" {
+		t.Skip("No suitable test file found with compressed versions")
+	}
+
+	fs := http.FS(os.DirFS("./ui_dist"))
+
+	// Test brotli
+	t.Run("brotli", func(t *testing.T) {
+		req := httptest.NewRequest(http.MethodGet, "/"+testFile, nil)
+		req.Header.Set("Accept-Encoding", "br")
+		w := httptest.NewRecorder()
+
+		ServeCompressedFile(fs, w, req, testFile)
+
+		resp := w.Result()
+		if resp.StatusCode != http.StatusOK {
+			t.Fatalf("Expected status 200, got %d", resp.StatusCode)
+		}
+
+		if encoding := resp.Header.Get("Content-Encoding"); encoding != "br" {
+			t.Errorf("Expected Content-Encoding 'br', got '%s'", encoding)
+		}
+	})
+
+	// Test gzip
+	t.Run("gzip", func(t *testing.T) {
+		req := httptest.NewRequest(http.MethodGet, "/"+testFile, nil)
+		req.Header.Set("Accept-Encoding", "gzip")
+		w := httptest.NewRecorder()
+
+		ServeCompressedFile(fs, w, req, testFile)
+
+		resp := w.Result()
+		if resp.StatusCode != http.StatusOK {
+			t.Fatalf("Expected status 200, got %d", resp.StatusCode)
+		}
+
+		if encoding := resp.Header.Get("Content-Encoding"); encoding != "gzip" {
+			t.Errorf("Expected Content-Encoding 'gzip', got '%s'", encoding)
+		}
+
+		// Verify it's valid gzip
+		reader, err := gzip.NewReader(resp.Body)
+		if err != nil {
+			t.Errorf("Expected valid gzip content: %v", err)
+			return
+		}
+		defer reader.Close()
+
+		// Just read to verify it's valid
+		_, err = io.Copy(io.Discard, reader)
+		if err != nil {
+			t.Errorf("Failed to decompress gzip: %v", err)
+		}
+	})
+}

ui-svelte/.gitignore

@@ -0,0 +1,2 @@
+node_modules
+.vite

ui-svelte/.npmrc

@@ -0,0 +1 @@
+legacy-peer-deps=true

ui-svelte/index.html

@@ -10,8 +10,8 @@
     <link rel="manifest" href="/site.webmanifest" />
     <title>llama-swap</title>
   </head>
-  <body >
-    <div id="root"></div>
-    <script type="module" src="/src/main.tsx"></script>
+  <body>
+    <div id="app"></div>
+    <script type="module" src="/src/main.ts"></script>
   </body>
 </html>

ui-svelte/package.json

@@ -0,0 +1,42 @@
+{
+  "name": "ui-svelte",
+  "private": true,
+  "version": "0.0.0",
+  "type": "module",
+  "scripts": {
+    "start": "vite",
+    "build": "vite build --emptyOutDir",
+    "preview": "vite preview",
+    "check": "svelte-check --tsconfig ./tsconfig.json",
+    "test": "vitest run",
+    "test:watch": "vitest"
+  },
+  "devDependencies": {
+    "@sveltejs/vite-plugin-svelte": "^7.0.0",
+    "@tailwindcss/vite": "^4.1.8",
+    "@tsconfig/svelte": "^5.0.4",
+    "@types/hast": "^3.0.4",
+    "@types/node": "^25.1.0",
+    "svelte": "^5.46.4",
+    "svelte-check": "^4.1.4",
+    "tailwindcss": "^4.1.8",
+    "typescript": "~5.8.3",
+    "vite": "^8.0.0",
+    "vite-plugin-compression2": "^2.5.1",
+    "vitest": "^4.1.0"
+  },
+  "dependencies": {
+    "highlight.js": "^11.11.1",
+    "katex": "^0.16.28",
+    "lucide-svelte": "^0.563.0",
+    "rehype-katex": "^7.0.1",
+    "rehype-stringify": "^10.0.1",
+    "remark-gfm": "^4.0.1",
+    "remark-math": "^6.0.0",
+    "remark-parse": "^11.0.0",
+    "remark-rehype": "^11.1.2",
+    "svelte-spa-router": "^4.0.1",
+    "unified": "^11.0.5",
+    "unist-util-visit": "^5.1.0"
+  }
+}

ui-svelte/src/App.svelte

@@ -0,0 +1,58 @@
+<script lang="ts">
+  import { onMount } from "svelte";
+  import Router from "svelte-spa-router";
+  import Header from "./components/Header.svelte";
+  import LogViewer from "./routes/LogViewer.svelte";
+  import Models from "./routes/Models.svelte";
+  import Activity from "./routes/Activity.svelte";
+  import Playground from "./routes/Playground.svelte";
+  import PlaygroundStub from "./routes/PlaygroundStub.svelte";
+  import { enableAPIEvents } from "./stores/api";
+  import { initScreenWidth, isDarkMode, appTitle, connectionState } from "./stores/theme";
+  import { currentRoute } from "./stores/route";
+
+  const routes = {
+    "/": PlaygroundStub,
+    "/models": Models,
+    "/logs": LogViewer,
+    "/activity": Activity,
+    "*": PlaygroundStub,
+  };
+
+  function handleRouteLoaded(event: { detail: { route: string | RegExp } }) {
+    const route = event.detail.route;
+    currentRoute.set(typeof route === "string" ? route : "/");
+  }
+
+  $effect(() => {
+    document.documentElement.setAttribute("data-theme", $isDarkMode ? "dark" : "light");
+  });
+
+  $effect(() => {
+    const icon = $connectionState === "connecting" ? "\u{1F7E1}" : $connectionState === "connected" ? "\u{1F7E2}" : "\u{1F534}";
+    document.title = `${icon} ${$appTitle}`;
+  });
+
+  onMount(() => {
+    const cleanupScreenWidth = initScreenWidth();
+    enableAPIEvents(true);
+
+    return () => {
+      cleanupScreenWidth();
+      enableAPIEvents(false);
+    };
+  });
+</script>
+
+<div class="flex flex-col h-screen">
+  <Header />
+
+  <main class="flex-1 overflow-auto p-4">
+    <div class="h-full" class:hidden={$currentRoute !== "/"}>
+      <Playground />
+    </div>
+    <div class="h-full" class:hidden={$currentRoute === "/"}>
+      <Router {routes} on:routeLoaded={handleRouteLoaded} />
+    </div>
+  </main>
+</div>

ui-svelte/src/components/ActivityStats.svelte

@@ -0,0 +1,72 @@
+<script lang="ts">
+  import { inFlightRequests, metrics } from "../stores/api";
+  import { persistentStore } from "../stores/persistent";
+  import { calculateHistogramData } from "../lib/histogram";
+  import TokenHistogram from "./TokenHistogram.svelte";
+
+  const nf = new Intl.NumberFormat();
+  const histogramCollapsed = persistentStore<boolean>("activity-histogram-collapsed", false);
+
+  let stats = $derived.by(() => {
+    const totalRequests = $metrics.length;
+    const totalInputTokens = $metrics.reduce((sum, m) => sum + m.input_tokens, 0);
+    const totalOutputTokens = $metrics.reduce((sum, m) => sum + m.output_tokens, 0);
+
+    const tokensPerSecond = $metrics
+      .filter((m) => m.tokens_per_second > 0)
+      .map((m) => m.tokens_per_second);
+
+    const histogramData = tokensPerSecond.length > 0
+      ? calculateHistogramData(tokensPerSecond, { minBins: 20, maxBins: 80, binScaling: 3 })
+      : null;
+
+    return {
+      totalRequests,
+      totalInputTokens,
+      totalOutputTokens,
+      inFlightRequests: $inFlightRequests,
+      histogramData,
+    };
+  });
+</script>
+
+<div class="card">
+  <button
+    class="flex items-center gap-1 px-4 pt-3 text-xs font-medium text-gray-500 dark:text-gray-400 hover:text-gray-700 dark:hover:text-gray-200 transition-colors"
+    onclick={() => $histogramCollapsed = !$histogramCollapsed}
+  >
+    <svg
+      class="w-3 h-3 transition-transform"
+      style="transform: rotate({$histogramCollapsed ? -90 : 0}deg)"
+      viewBox="0 0 16 16"
+      fill="currentColor"
+    >
+      <path d="M4.5 6l3.5 4 3.5-4H4.5z" />
+    </svg>
+    Tokens/sec Distribution
+  </button>
+  {#if !$histogramCollapsed}
+    {#if stats.histogramData}
+      <TokenHistogram data={stats.histogramData} />
+    {:else}
+      <div class="px-4 py-6 text-center text-sm text-gray-500 dark:text-gray-400">
+        No token speed data yet
+      </div>
+    {/if}
+  {/if}
+  <div class="grid grid-cols-3 gap-x-6 gap-y-1 px-4 pb-3 text-sm">
+    <div class="text-xs uppercase tracking-wider text-gray-500 dark:text-gray-400">Requests</div>
+    <div class="text-xs uppercase tracking-wider text-gray-500 dark:text-gray-400">Processed</div>
+    <div class="text-xs uppercase tracking-wider text-gray-500 dark:text-gray-400">Generated</div>
+    <div class="text-sm text-gray-700 dark:text-gray-300">
+      <span class="font-semibold">{nf.format(stats.totalRequests)}</span> completed,
+      <span class="font-semibold">{nf.format(stats.inFlightRequests)}</span> waiting
+    </div>
+    <div class="text-sm text-gray-700 dark:text-gray-300">
+      <span class="font-semibold">{nf.format(stats.totalInputTokens)}</span> tokens
+    </div>
+    <div class="text-sm text-gray-700 dark:text-gray-300">
+      <span class="font-semibold">{nf.format(stats.totalOutputTokens)}</span> tokens
+    </div>
+  </div>
+</div>

ui-svelte/src/components/CaptureDialog.svelte

@@ -0,0 +1,452 @@
+<script lang="ts">
+  import type { ReqRespCapture } from "../lib/types";
+
+  interface Props {
+    capture: ReqRespCapture | null;
+    open: boolean;
+    onclose: () => void;
+  }
+
+  let { capture, open, onclose }: Props = $props();
+
+  let dialogEl: HTMLDialogElement | undefined = $state();
+
+  type BodyTab = "raw" | "pretty" | "chat";
+  let reqBodyTab: BodyTab = $state("pretty");
+  let respBodyTab: BodyTab = $state("pretty");
+  let copiedReq = $state(false);
+  let copiedResp = $state(false);
+
+  $effect(() => {
+    if (open && dialogEl) {
+      dialogEl.showModal();
+    } else if (!open && dialogEl) {
+      dialogEl.close();
+    }
+  });
+
+  // Reset tabs when capture changes
+  $effect(() => {
+    if (capture) {
+      const reqCt = getContentType(capture.req_headers);
+      const respCt = getContentType(capture.resp_headers);
+      reqBodyTab = reqCt.includes("json") ? "pretty" : "raw";
+      respBodyTab = respCt.includes("text/event-stream")
+        ? "chat"
+        : respCt.includes("json")
+          ? "pretty"
+          : "raw";
+    }
+  });
+
+  function handleDialogClose() {
+    onclose();
+  }
+
+  function decodeBody(body: string | null | undefined): string {
+    if (!body) return "";
+    try {
+      const binary = atob(body);
+      const bytes = Uint8Array.from(binary, (c) => c.charCodeAt(0));
+      return new TextDecoder().decode(bytes);
+    } catch {
+      return body;
+    }
+  }
+
+  function formatJson(str: string): string {
+    try {
+      const parsed = JSON.parse(str);
+      return JSON.stringify(parsed, null, 2);
+    } catch {
+      return str;
+    }
+  }
+
+  function getContentType(
+    headers: Record<string, string> | null | undefined,
+  ): string {
+    if (!headers) return "";
+    const ct = headers["Content-Type"] || headers["content-type"] || "";
+    return ct.toLowerCase();
+  }
+
+  function isImageContentType(contentType: string): boolean {
+    return contentType.startsWith("image/");
+  }
+
+  function isTextContentType(contentType: string): boolean {
+    return (
+      contentType.startsWith("text/") ||
+      contentType.includes("application/json") ||
+      contentType.includes("application/xml") ||
+      contentType.includes("application/javascript")
+    );
+  }
+
+  function getImageDataUrl(body: string, contentType: string): string {
+    const mimeType = contentType.split(";")[0].trim();
+    return `data:${mimeType};base64,${body}`;
+  }
+
+  interface SSEChat {
+    reasoning: string;
+    content: string;
+  }
+
+  function parseSSEChat(text: string): SSEChat {
+    const result: SSEChat = { reasoning: "", content: "" };
+    for (const line of text.split("\n")) {
+      const trimmed = line.trim();
+      if (!trimmed || !trimmed.startsWith("data: ")) continue;
+      const data = trimmed.slice(6);
+      if (data === "[DONE]") continue;
+      try {
+        const parsed = JSON.parse(data);
+        const delta = parsed.choices?.[0]?.delta;
+        if (delta?.content) result.content += delta.content;
+        if (delta?.reasoning_content) result.reasoning += delta.reasoning_content;
+      } catch {
+        // skip unparseable lines
+      }
+    }
+    return result;
+  }
+
+  async function copyToClipboard(text: string, type: "req" | "resp") {
+    try {
+      await navigator.clipboard.writeText(text);
+      if (type === "req") {
+        copiedReq = true;
+        setTimeout(() => (copiedReq = false), 1500);
+      } else {
+        copiedResp = true;
+        setTimeout(() => (copiedResp = false), 1500);
+      }
+    } catch {
+      // ignore
+    }
+  }
+
+  function getCopyText(): string {
+    if (respBodyTab === "chat") {
+      let text = "";
+      if (sseChat.reasoning) text += sseChat.reasoning + "\n\n";
+      text += sseChat.content;
+      return text;
+    }
+    return displayedResponseBody;
+  }
+
+  // Request body derivations
+  let requestContentType = $derived(
+    capture ? getContentType(capture.req_headers) : "",
+  );
+  let isRequestJson = $derived(requestContentType.includes("json"));
+
+  let requestBodyRaw = $derived.by(() => {
+    if (!capture) return "";
+    return decodeBody(capture.req_body);
+  });
+
+  let requestBodyPretty = $derived.by(() => {
+    if (!isRequestJson) return requestBodyRaw;
+    return formatJson(requestBodyRaw);
+  });
+
+  let displayedRequestBody = $derived(
+    reqBodyTab === "pretty" ? requestBodyPretty : requestBodyRaw,
+  );
+
+  // Response body derivations
+  let responseContentType = $derived(
+    capture ? getContentType(capture.resp_headers) : "",
+  );
+  let isResponseImage = $derived(isImageContentType(responseContentType));
+  let isResponseText = $derived(isTextContentType(responseContentType));
+  let isResponseJson = $derived(responseContentType.includes("json"));
+  let isSSE = $derived(responseContentType.includes("text/event-stream"));
+
+  let responseBodyRaw = $derived.by(() => {
+    if (!capture) return "";
+    return decodeBody(capture.resp_body);
+  });
+
+  let responseBodyPretty = $derived.by(() => {
+    if (!isResponseJson) return responseBodyRaw;
+    return formatJson(responseBodyRaw);
+  });
+
+  let sseChat = $derived.by(() => {
+    if (!isSSE || !responseBodyRaw)
+      return { reasoning: "", content: "" } as SSEChat;
+    return parseSSEChat(responseBodyRaw);
+  });
+
+  let displayedResponseBody = $derived.by(() => {
+    if (respBodyTab === "pretty") return responseBodyPretty;
+    return responseBodyRaw;
+  });
+</script>
+
+<dialog
+  bind:this={dialogEl}
+  onclose={handleDialogClose}
+  class="bg-surface text-txtmain rounded-lg shadow-xl max-w-4xl w-full max-h-[90vh] p-0 backdrop:bg-black/50 m-auto"
+>
+  {#if capture}
+    <div class="flex flex-col max-h-[90vh]">
+      <div
+        class="flex justify-between items-center p-4 border-b border-card-border"
+      >
+        <h2 class="text-xl font-bold pb-0">Capture #{capture.id + 1}{#if capture.req_path} <span class="text-base font-mono font-normal text-txtsecondary">{capture.req_path}</span>{/if}</h2>
+        <button
+          onclick={() => dialogEl?.close()}
+          class="text-txtsecondary hover:text-txtmain text-2xl leading-none"
+        >
+          &times;
+        </button>
+      </div>
+
+      <div class="overflow-y-auto flex-1 p-4 space-y-4">
+        <!-- Request Headers -->
+        <details class="group" open>
+          <summary
+            class="cursor-pointer font-semibold text-sm uppercase tracking-wider text-txtsecondary hover:text-txtmain"
+          >
+            Request Headers
+          </summary>
+          <div
+            class="mt-2 bg-background rounded border border-card-border overflow-auto max-h-48"
+          >
+            <table class="w-full text-sm">
+              <tbody>
+                {#each Object.entries(capture.req_headers || {}) as [key, value]}
+                  <tr class="border-b border-card-border-inner last:border-0">
+                    <td class="px-3 py-1 font-mono text-primary whitespace-nowrap"
+                      >{key}</td
+                    >
+                    <td class="px-3 py-1 font-mono break-all">{value}</td>
+                  </tr>
+                {/each}
+              </tbody>
+            </table>
+          </div>
+        </details>
+
+        <!-- Request Body -->
+        <details class="group" open>
+          <summary
+            class="cursor-pointer font-semibold text-sm uppercase tracking-wider text-txtsecondary hover:text-txtmain"
+          >
+            Request Body
+          </summary>
+          {#if requestBodyRaw}
+            <div class="mt-2 flex items-center justify-between">
+              <div class="flex gap-1">
+                {#if isRequestJson}
+                  <button
+                    class="tab-btn"
+                    class:tab-btn-active={reqBodyTab === "pretty"}
+                    onclick={() => (reqBodyTab = "pretty")}>Pretty</button
+                  >
+                  <button
+                    class="tab-btn"
+                    class:tab-btn-active={reqBodyTab === "raw"}
+                    onclick={() => (reqBodyTab = "raw")}>Raw</button
+                  >
+                {/if}
+              </div>
+              <button
+                class="tab-btn"
+                onclick={() =>
+                  copyToClipboard(displayedRequestBody, "req")}
+              >
+                {#if copiedReq}
+                  Copied!
+                {:else}
+                  Copy
+                {/if}
+              </button>
+            </div>
+            <div
+              class="mt-1 bg-background rounded border border-card-border overflow-auto max-h-96"
+            >
+              <pre
+                class="p-3 text-sm font-mono whitespace-pre-wrap break-all">{displayedRequestBody}</pre>
+            </div>
+          {:else}
+            <div
+              class="mt-2 bg-background rounded border border-card-border overflow-auto max-h-96"
+            >
+              <pre class="p-3 text-sm font-mono whitespace-pre-wrap break-all"
+                >(empty)</pre
+              >
+            </div>
+          {/if}
+        </details>
+
+        <!-- Response Headers -->
+        <details class="group" open>
+          <summary
+            class="cursor-pointer font-semibold text-sm uppercase tracking-wider text-txtsecondary hover:text-txtmain"
+          >
+            Response Headers
+          </summary>
+          <div
+            class="mt-2 bg-background rounded border border-card-border overflow-auto max-h-48"
+          >
+            <table class="w-full text-sm">
+              <tbody>
+                {#each Object.entries(capture.resp_headers || {}) as [key, value]}
+                  <tr class="border-b border-card-border-inner last:border-0">
+                    <td class="px-3 py-1 font-mono text-primary whitespace-nowrap"
+                      >{key}</td
+                    >
+                    <td class="px-3 py-1 font-mono break-all">{value}</td>
+                  </tr>
+                {/each}
+              </tbody>
+            </table>
+          </div>
+        </details>
+
+        <!-- Response Body -->
+        <details class="group" open>
+          <summary
+            class="cursor-pointer font-semibold text-sm uppercase tracking-wider text-txtsecondary hover:text-txtmain"
+          >
+            Response Body
+          </summary>
+          {#if isResponseImage && capture.resp_body}
+            <div
+              class="mt-2 bg-background rounded border border-card-border overflow-auto max-h-96"
+            >
+              <div class="p-3 flex justify-center">
+                <img
+                  src={getImageDataUrl(capture.resp_body, responseContentType)}
+                  alt="Response"
+                  class="max-w-full h-auto"
+                />
+              </div>
+            </div>
+          {:else if isSSE || isResponseText}
+            <div class="mt-2 flex items-center justify-between">
+              <div class="flex gap-1">
+                {#if isSSE}
+                  <button
+                    class="tab-btn"
+                    class:tab-btn-active={respBodyTab === "chat"}
+                    onclick={() => (respBodyTab = "chat")}>Chat</button
+                  >
+                {/if}
+                {#if isResponseJson}
+                  <button
+                    class="tab-btn"
+                    class:tab-btn-active={respBodyTab === "pretty"}
+                    onclick={() => (respBodyTab = "pretty")}>Pretty</button
+                  >
+                {/if}
+                {#if isSSE || isResponseJson}
+                  <button
+                    class="tab-btn"
+                    class:tab-btn-active={respBodyTab === "raw"}
+                    onclick={() => (respBodyTab = "raw")}>Raw</button
+                  >
+                {/if}
+              </div>
+              <button
+                class="tab-btn"
+                onclick={() => copyToClipboard(getCopyText(), "resp")}
+              >
+                {#if copiedResp}
+                  Copied!
+                {:else}
+                  Copy
+                {/if}
+              </button>
+            </div>
+            <div
+              class="mt-1 bg-background rounded border border-card-border overflow-auto max-h-96"
+            >
+              {#if respBodyTab === "chat"}
+                <div class="p-3 text-sm space-y-3">
+                  {#if sseChat.reasoning}
+                    <div>
+                      <div
+                        class="text-xs font-semibold uppercase tracking-wider text-txtsecondary mb-1"
+                      >
+                        Reasoning
+                      </div>
+                      <pre
+                        class="font-mono whitespace-pre-wrap break-all text-txtsecondary">{sseChat.reasoning}</pre>
+                    </div>
+                  {/if}
+                  {#if sseChat.content}
+                    <div>
+                      {#if sseChat.reasoning}
+                        <div
+                          class="text-xs font-semibold uppercase tracking-wider text-txtsecondary mb-1"
+                        >
+                          Response
+                        </div>
+                      {/if}
+                      <pre
+                        class="font-mono whitespace-pre-wrap break-all">{sseChat.content}</pre>
+                    </div>
+                  {/if}
+                  {#if !sseChat.reasoning && !sseChat.content}
+                    <pre class="font-mono">(empty)</pre>
+                  {/if}
+                </div>
+              {:else}
+                <pre
+                  class="p-3 text-sm font-mono whitespace-pre-wrap break-all">{displayedResponseBody || "(empty)"}</pre>
+              {/if}
+            </div>
+          {:else if responseBodyRaw}
+            <div
+              class="mt-2 bg-background rounded border border-card-border overflow-auto max-h-96"
+            >
+              <div class="p-3 text-sm text-txtsecondary italic">
+                (binary data - {responseContentType || "unknown content type"})
+              </div>
+            </div>
+          {:else}
+            <div
+              class="mt-2 bg-background rounded border border-card-border overflow-auto max-h-96"
+            >
+              <pre class="p-3 text-sm font-mono">(empty)</pre>
+            </div>
+          {/if}
+        </details>
+      </div>
+
+      <div class="p-4 border-t border-card-border flex justify-end">
+        <button onclick={() => dialogEl?.close()} class="btn"> Close </button>
+      </div>
+    </div>
+  {/if}
+</dialog>
+
+<style>
+  .tab-btn {
+    padding: 2px 10px;
+    font-size: 0.75rem;
+    border-radius: 4px;
+    color: var(--color-txtsecondary);
+    cursor: pointer;
+    border: 1px solid transparent;
+    background: transparent;
+    transition: all 0.15s;
+  }
+  .tab-btn:hover {
+    color: var(--color-txtmain);
+    background: var(--color-secondary);
+  }
+  .tab-btn-active {
+    color: var(--color-primary);
+    background: color-mix(in srgb, var(--color-primary) 12%, transparent);
+    border-color: color-mix(in srgb, var(--color-primary) 25%, transparent);
+  }
+</style>

ui-svelte/src/components/ConnectionStatus.svelte

@@ -0,0 +1,24 @@
+<script lang="ts">
+  import { connectionState } from "../stores/theme";
+  import { versionInfo } from "../stores/api";
+
+  let eventStatusColor = $derived.by(() => {
+    switch ($connectionState) {
+      case "connected":
+        return "bg-emerald-500";
+      case "connecting":
+        return "bg-amber-500";
+      case "disconnected":
+      default:
+        return "bg-red-500";
+    }
+  });
+
+  let tooltipText = $derived(
+    `Event Stream: ${$connectionState ?? "unknown"}\nAPI Version: ${$versionInfo?.version ?? "unknown"}\nCommit Hash: ${$versionInfo?.commit?.substring(0, 7) ?? "unknown"}\nBuild Date: ${$versionInfo?.build_date ?? "unknown"}`
+  );
+</script>
+
+<div class="flex items-center" title={tooltipText}>
+  <span class="inline-block w-3 h-3 rounded-full {eventStatusColor} mr-2"></span>
+</div>

ui-svelte/src/components/Header.svelte

@@ -0,0 +1,120 @@
+<script lang="ts">
+  import { link } from "svelte-spa-router";
+  import { screenWidth, toggleTheme, isDarkMode, appTitle, isNarrow } from "../stores/theme";
+  import { currentRoute } from "../stores/route";
+  import { playgroundActivity } from "../stores/playgroundActivity";
+  import ConnectionStatus from "./ConnectionStatus.svelte";
+
+  function handleTitleChange(newTitle: string): void {
+    const sanitized = newTitle.replace(/\n/g, "").trim().substring(0, 64) || "llama-swap";
+    appTitle.set(sanitized);
+  }
+
+  function handleKeyDown(e: KeyboardEvent): void {
+    if (e.key === "Enter") {
+      e.preventDefault();
+      const target = e.currentTarget as HTMLElement;
+      handleTitleChange(target.textContent || "(set title)");
+      target.blur();
+    }
+  }
+
+  function handleBlur(e: FocusEvent): void {
+    const target = e.currentTarget as HTMLElement;
+    handleTitleChange(target.textContent || "(set title)");
+  }
+
+  function isActive(path: string, current: string): boolean {
+    return path === "/" ? current === "/" : current.startsWith(path);
+  }
+
+</script>
+
+<header
+  class="flex items-center justify-between bg-surface border-b border-border px-4 {$isNarrow
+    ? 'py-1 h-[60px]'
+    : 'p-2 h-[75px]'}"
+>
+  {#if $screenWidth !== "xs" && $screenWidth !== "sm"}
+    <h1
+      contenteditable="true"
+      class="p-0 outline-none hover:bg-gray-100 dark:hover:bg-gray-700 rounded"
+      onblur={handleBlur}
+      onkeydown={handleKeyDown}
+    >
+      {$appTitle}
+    </h1>
+  {/if}
+
+  <menu class="flex items-center gap-4 overflow-x-auto">
+    <a
+      href="/"
+      use:link
+      class="p-1 whitespace-nowrap {isActive('/', $currentRoute) ? 'font-semibold' : ''} {$playgroundActivity ? 'activity-link' : 'text-gray-600 hover:text-black dark:text-gray-300 dark:hover:text-gray-100'}"
+    >
+      Playground
+    </a>
+    <a
+      href="/models"
+      use:link
+      class="text-gray-600 hover:text-black dark:text-gray-300 dark:hover:text-gray-100 p-1 whitespace-nowrap"
+      class:font-semibold={isActive("/models", $currentRoute)}
+    >
+      Models
+    </a>
+    <a
+      href="/activity"
+      use:link
+      class="text-gray-600 hover:text-black dark:text-gray-300 dark:hover:text-gray-100 p-1 whitespace-nowrap"
+      class:font-semibold={isActive("/activity", $currentRoute)}
+    >
+      Activity
+    </a>
+    <a
+      href="/logs"
+      use:link
+      class="text-gray-600 hover:text-black dark:text-gray-300 dark:hover:text-gray-100 p-1 whitespace-nowrap"
+      class:font-semibold={isActive("/logs", $currentRoute)}
+    >
+      Logs
+    </a>
+    <button onclick={toggleTheme} title="Toggle theme">
+      {#if $isDarkMode}
+        <svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 24 24" fill="currentColor" class="w-5 h-5">
+          <path
+            fill-rule="evenodd"
+            d="M9.528 1.718a.75.75 0 0 1 .162.819A8.97 8.97 0 0 0 9 6a9 9 0 0 0 9 9 8.97 8.97 0 0 0 3.463-.69.75.75 0 0 1 .981.98 10.503 10.503 0 0 1-9.694 6.46c-5.799 0-10.5-4.7-10.5-10.5 0-4.368 2.667-8.112 6.46-9.694a.75.75 0 0 1 .818.162Z"
+            clip-rule="evenodd"
+          />
+        </svg>
+      {:else}
+        <svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 24 24" fill="currentColor" class="w-5 h-5">
+          <path
+            d="M12 2.25a.75.75 0 0 1 .75.75v2.25a.75.75 0 0 1-1.5 0V3a.75.75 0 0 1 .75-.75ZM7.5 12a4.5 4.5 0 1 1 9 0 4.5 4.5 0 0 1-9 0ZM18.894 6.166a.75.75 0 0 0-1.06-1.06l-1.591 1.59a.75.75 0 1 0 1.06 1.061l1.591-1.59ZM21.75 12a.75.75 0 0 1-.75.75h-2.25a.75.75 0 0 1 0-1.5H21a.75.75 0 0 1 .75.75ZM17.834 18.894a.75.75 0 0 0 1.06-1.06l-1.59-1.591a.75.75 0 1 0-1.061 1.06l1.591 1.591ZM12 18a.75.75 0 0 1 .75.75V21a.75.75 0 0 1-1.5 0v-2.25A.75.75 0 0 1 12 18ZM7.758 17.303a.75.75 0 0 0-1.061-1.06l-1.591 1.59a.75.75 0 0 0 1.06 1.061l1.591-1.59ZM6 12a.75.75 0 0 1-.75.75H3a.75.75 0 0 1 0-1.5h2.25A.75.75 0 0 1 6 12ZM6.697 7.757a.75.75 0 0 0 1.06-1.06l-1.59-1.591a.75.75 0 0 0-1.061 1.06l1.59 1.591Z"
+          />
+        </svg>
+      {/if}
+    </button>
+    <ConnectionStatus />
+  </menu>
+</header>
+
+<style>
+  .activity-link {
+    background: linear-gradient(90deg, #6366f1, #8b5cf6, #a855f7, #8b5cf6, #6366f1);
+    background-size: 200% 100%;
+    -webkit-background-clip: text;
+    background-clip: text;
+    -webkit-text-fill-color: transparent;
+    animation: gradient-shift 2s linear infinite;
+  }
+
+  @keyframes gradient-shift {
+    0% {
+      background-position: 0% 50%;
+    }
+    100% {
+      background-position: 200% 50%;
+    }
+  }
+</style>

ui-svelte/src/components/LogPanel.svelte

@@ -0,0 +1,139 @@
+<script lang="ts">
+  import { persistentStore } from "../stores/persistent";
+
+  interface Props {
+    id: string;
+    title: string;
+    logData: string;
+  }
+
+  let { id, title, logData }: Props = $props();
+
+  let filterRegex = $state("");
+
+  // Create persistent stores for this panel (id is intentionally captured at init time)
+  // svelte-ignore state_referenced_locally
+  const fontSizeStore = persistentStore<"xxs" | "xs" | "small" | "normal">(`logPanel-${id}-fontSize`, "normal");
+  // svelte-ignore state_referenced_locally
+  const wrapTextStore = persistentStore<boolean>(`logPanel-${id}-wrapText`, false);
+  // svelte-ignore state_referenced_locally
+  const showFilterStore = persistentStore<boolean>(`logPanel-${id}-showFilter`, false);
+
+  let textWrapClass = $derived($wrapTextStore ? "whitespace-pre-wrap" : "whitespace-pre");
+
+  function toggleFontSize(): void {
+    fontSizeStore.update((prev) => {
+      switch (prev) {
+        case "xxs": return "xs";
+        case "xs": return "small";
+        case "small": return "normal";
+        case "normal": return "xxs";
+      }
+    });
+  }
+
+  function toggleWrapText(): void {
+    wrapTextStore.update((prev) => !prev);
+  }
+
+  function toggleFilter(): void {
+    if ($showFilterStore) {
+      showFilterStore.set(false);
+      filterRegex = "";
+    } else {
+      showFilterStore.set(true);
+    }
+  }
+
+  let fontSizeClass = $derived.by(() => {
+    switch ($fontSizeStore) {
+      case "xxs": return "text-[0.5rem]";
+      case "xs": return "text-[0.75rem]";
+      case "small": return "text-[0.875rem]";
+      case "normal": return "text-base";
+    }
+  });
+
+  let filteredLogs = $derived.by(() => {
+    if (!filterRegex) return logData;
+    try {
+      const regex = new RegExp(filterRegex, "i");
+      return logData.split("\n").filter((line) => regex.test(line)).join("\n");
+    } catch {
+      return logData;
+    }
+  });
+
+  let preElement: HTMLPreElement;
+  let userScrolledUp = $state(false);
+
+  function handleScroll() {
+    if (!preElement) return;
+    const { scrollTop, scrollHeight, clientHeight } = preElement;
+    userScrolledUp = scrollHeight - scrollTop - clientHeight > 40;
+  }
+
+  // Auto scroll to bottom when logs change, unless user has scrolled up
+  $effect(() => {
+    if (preElement && filteredLogs && !userScrolledUp) {
+      preElement.scrollTop = preElement.scrollHeight;
+    }
+  });
+</script>
+
+<div class="rounded-lg overflow-hidden flex flex-col bg-gray-950/5 dark:bg-white/10 h-full w-full p-1">
+  <div class="p-4">
+    <div class="flex items-center justify-between">
+      <h3 class="m-0 text-lg p-0">{title}</h3>
+
+      <div class="flex gap-2 items-center">
+        <button class="btn border-0" onclick={toggleFontSize} title="Change font size">
+          <svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 24 24" fill="currentColor" class="w-4 h-4">
+            <path d="M2 4v3h5v12h3V7h5V4H2zm19 5h-9v3h3v7h3v-7h3V9z"/>
+          </svg>
+        </button>
+        <button class="btn border-0" onclick={toggleWrapText} title="Toggle text wrap">
+          {#if $wrapTextStore}
+            <svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 24 24" fill="currentColor" class="w-4 h-4">
+              <path fill-rule="evenodd" d="M3 6.75A.75.75 0 0 1 3.75 6h16.5a.75.75 0 0 1 0 1.5H3.75A.75.75 0 0 1 3 6.75ZM3 12a.75.75 0 0 1 .75-.75h16.5a.75.75 0 0 1 0 1.5H3.75A.75.75 0 0 1 3 12Zm0 5.25a.75.75 0 0 1 .75-.75h16.5a.75.75 0 0 1 0 1.5H3.75a.75.75 0 0 1-.75-.75Z" clip-rule="evenodd" />
+            </svg>
+          {:else}
+            <svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 24 24" fill="currentColor" class="w-4 h-4">
+              <path fill-rule="evenodd" d="M3 6.75A.75.75 0 0 1 3.75 6h16.5a.75.75 0 0 1 0 1.5H3.75A.75.75 0 0 1 3 6.75ZM3 12a.75.75 0 0 1 .75-.75h10.5a.75.75 0 0 1 0 1.5H3.75A.75.75 0 0 1 3 12Zm0 5.25a.75.75 0 0 1 .75-.75h16.5a.75.75 0 0 1 0 1.5H3.75a.75.75 0 0 1-.75-.75Z" clip-rule="evenodd" />
+            </svg>
+          {/if}
+        </button>
+        <button class="btn border-0" onclick={toggleFilter} title="Toggle filter">
+          {#if $showFilterStore}
+            <svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 24 24" fill="currentColor" class="w-4 h-4">
+              <path fill-rule="evenodd" d="M10.5 3.75a6.75 6.75 0 1 0 0 13.5 6.75 6.75 0 0 0 0-13.5ZM2.25 10.5a8.25 8.25 0 1 1 14.59 5.28l4.69 4.69a.75.75 0 1 1-1.06 1.06l-4.69-4.69A8.25 8.25 0 0 1 2.25 10.5Z" clip-rule="evenodd" />
+            </svg>
+          {:else}
+            <svg xmlns="http://www.w3.org/2000/svg" fill="none" viewBox="0 0 24 24" stroke-width="1.5" stroke="currentColor" class="w-4 h-4">
+              <path stroke-linecap="round" stroke-linejoin="round" d="m21 21-5.197-5.197m0 0A7.5 7.5 0 1 0 5.196 5.196a7.5 7.5 0 0 0 10.607 10.607Z" />
+            </svg>
+          {/if}
+        </button>
+      </div>
+    </div>
+
+    {#if $showFilterStore}
+      <div class="mt-2 flex gap-2 items-center w-full">
+        <input
+          type="text"
+          class="w-full text-sm border border-gray-950/10 dark:border-white/5 p-2 rounded outline-none"
+          placeholder="Filter logs (regex)..."
+          bind:value={filterRegex}
+        />
+        <button class="pl-2" onclick={() => (filterRegex = "")} aria-label="Clear filter">
+          <svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 24 24" fill="currentColor" class="w-6 h-6">
+            <path fill-rule="evenodd" d="M12 2.25c-5.385 0-9.75 4.365-9.75 9.75s4.365 9.75 9.75 9.75 9.75-4.365 9.75-9.75S17.385 2.25 12 2.25Zm-1.72 6.97a.75.75 0 1 0-1.06 1.06L10.94 12l-1.72 1.72a.75.75 0 1 0 1.06 1.06L12 13.06l1.72 1.72a.75.75 0 1 0 1.06-1.06L13.06 12l1.72-1.72a.75.75 0 1 0-1.06-1.06L12 10.94l-1.72-1.72Z" clip-rule="evenodd" />
+          </svg>
+        </button>
+      </div>
+    {/if}
+  </div>
+  <div class="rounded-lg bg-background font-mono text-sm flex-1 overflow-hidden">
+    <pre bind:this={preElement} onscroll={handleScroll} class="{textWrapClass} {fontSizeClass} h-full overflow-auto p-4">{filteredLogs}</pre>
+  </div>
+</div>

ui-svelte/src/components/ModelsPanel.svelte

@@ -0,0 +1,211 @@
+<script lang="ts">
+  import { models, loadModel, unloadAllModels, unloadSingleModel } from "../stores/api";
+  import { isNarrow } from "../stores/theme";
+  import { persistentStore } from "../stores/persistent";
+  import type { Model } from "../lib/types";
+
+  let isUnloading = $state(false);
+  let menuOpen = $state(false);
+
+  const showUnlistedStore = persistentStore<boolean>("showUnlisted", true);
+  const showIdorNameStore = persistentStore<"id" | "name">("showIdorName", "id");
+
+  let filteredModels = $derived.by(() => {
+    const filtered = $models.filter((model) => $showUnlistedStore || !model.unlisted);
+    const peerModels = filtered.filter((m) => m.peerID);
+
+    // Group peer models by peerID
+    const grouped = peerModels.reduce(
+      (acc, model) => {
+        const peerId = model.peerID || "unknown";
+        if (!acc[peerId]) acc[peerId] = [];
+        acc[peerId].push(model);
+        return acc;
+      },
+      {} as Record<string, Model[]>
+    );
+
+    return {
+      regularModels: filtered.filter((m) => !m.peerID),
+      peerModelsByPeerId: grouped,
+    };
+  });
+
+  async function handleUnloadAllModels(): Promise<void> {
+    isUnloading = true;
+    try {
+      await unloadAllModels();
+    } catch (e) {
+      console.error(e);
+    } finally {
+      setTimeout(() => (isUnloading = false), 1000);
+    }
+  }
+
+  function toggleIdorName(): void {
+    showIdorNameStore.update((prev) => (prev === "name" ? "id" : "name"));
+  }
+
+  function toggleShowUnlisted(): void {
+    showUnlistedStore.update((prev) => !prev);
+  }
+
+  function getModelDisplay(model: Model): string {
+    return $showIdorNameStore === "id" ? model.id : (model.name || model.id);
+  }
+</script>
+
+<div class="card h-full flex flex-col">
+  <div class="shrink-0">
+    <div class="flex justify-between items-baseline">
+      <h2 class={$isNarrow ? "text-xl" : ""}>Models</h2>
+      {#if $isNarrow}
+        <div class="relative">
+          <button class="btn text-base flex items-center gap-2 py-1" onclick={() => (menuOpen = !menuOpen)} aria-label="Toggle menu">
+            <svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 24 24" fill="currentColor" class="w-5 h-5">
+              <path fill-rule="evenodd" d="M3 6.75A.75.75 0 0 1 3.75 6h16.5a.75.75 0 0 1 0 1.5H3.75A.75.75 0 0 1 3 6.75ZM3 12a.75.75 0 0 1 .75-.75h16.5a.75.75 0 0 1 0 1.5H3.75A.75.75 0 0 1 3 12Zm0 5.25a.75.75 0 0 1 .75-.75h16.5a.75.75 0 0 1 0 1.5H3.75a.75.75 0 0 1-.75-.75Z" clip-rule="evenodd" />
+            </svg>
+          </button>
+          {#if menuOpen}
+            <div class="absolute right-0 mt-2 w-48 bg-surface border border-gray-200 dark:border-white/10 rounded shadow-lg z-20">
+              <button
+                class="w-full text-left px-4 py-2 hover:bg-secondary-hover flex items-center gap-2"
+                onclick={() => { toggleIdorName(); menuOpen = false; }}
+              >
+                <svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 24 24" fill="currentColor" class="w-5 h-5">
+                  <path fill-rule="evenodd" d="M15.97 2.47a.75.75 0 0 1 1.06 0l4.5 4.5a.75.75 0 0 1 0 1.06l-4.5 4.5a.75.75 0 1 1-1.06-1.06l3.22-3.22H7.5a.75.75 0 0 1 0-1.5h11.69l-3.22-3.22a.75.75 0 0 1 0-1.06Zm-7.94 9a.75.75 0 0 1 0 1.06l-3.22 3.22H16.5a.75.75 0 0 1 0 1.5H4.81l3.22 3.22a.75.75 0 1 1-1.06 1.06l-4.5-4.5a.75.75 0 0 1 0-1.06l4.5-4.5a.75.75 0 0 1 1.06 0Z" clip-rule="evenodd" />
+                </svg>
+                {$showIdorNameStore === "id" ? "Show Name" : "Show ID"}
+              </button>
+              <button
+                class="w-full text-left px-4 py-2 hover:bg-secondary-hover flex items-center gap-2"
+                onclick={() => { toggleShowUnlisted(); menuOpen = false; }}
+              >
+                {#if $showUnlistedStore}
+                  <svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 24 24" fill="currentColor" class="w-5 h-5">
+                    <path d="M3.53 2.47a.75.75 0 0 0-1.06 1.06l18 18a.75.75 0 1 0 1.06-1.06l-18-18ZM22.676 12.553a11.249 11.249 0 0 1-2.631 4.31l-3.099-3.099a5.25 5.25 0 0 0-6.71-6.71L7.759 4.577a11.217 11.217 0 0 1 4.242-.827c4.97 0 9.185 3.223 10.675 7.69.12.362.12.752 0 1.113Z" />
+                    <path d="M15.75 12c0 .18-.013.357-.037.53l-4.244-4.243A3.75 3.75 0 0 1 15.75 12ZM12.53 15.713l-4.243-4.244a3.75 3.75 0 0 0 4.244 4.243Z" />
+                    <path d="M6.75 12c0-.619.107-1.213.304-1.764l-3.1-3.1a11.25 11.25 0 0 0-2.63 4.31c-.12.362-.12.752 0 1.114 1.489 4.467 5.704 7.69 10.675 7.69 1.5 0 2.933-.294 4.242-.827l-2.477-2.477A5.25 5.25 0 0 1 6.75 12Z" />
+                  </svg>
+                {:else}
+                  <svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 24 24" fill="currentColor" class="w-5 h-5">
+                    <path d="M12 15a3 3 0 1 0 0-6 3 3 0 0 0 0 6Z" />
+                    <path fill-rule="evenodd" d="M1.323 11.447C2.811 6.976 7.028 3.75 12.001 3.75c4.97 0 9.185 3.223 10.675 7.69.12.362.12.752 0 1.113-1.487 4.471-5.705 7.697-10.677 7.697-4.97 0-9.186-3.223-10.675-7.69a1.762 1.762 0 0 1 0-1.113ZM17.25 12a5.25 5.25 0 1 1-10.5 0 5.25 5.25 0 0 1 10.5 0Z" clip-rule="evenodd" />
+                  </svg>
+                {/if}
+                {$showUnlistedStore ? "Hide Unlisted" : "Show Unlisted"}
+              </button>
+              <button
+                class="w-full text-left px-4 py-2 hover:bg-secondary-hover flex items-center gap-2"
+                onclick={() => { handleUnloadAllModels(); menuOpen = false; }}
+                disabled={isUnloading}
+              >
+                <svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 24 24" fill="currentColor" class="w-6 h-6">
+                  <path fill-rule="evenodd" d="M12 2.25c-5.385 0-9.75 4.365-9.75 9.75s4.365 9.75 9.75 9.75 9.75-4.365 9.75-9.75S17.385 2.25 12 2.25Zm.53 5.47a.75.75 0 0 0-1.06 0l-3 3a.75.75 0 1 0 1.06 1.06l1.72-1.72v5.69a.75.75 0 0 0 1.5 0v-5.69l1.72 1.72a.75.75 0 1 0 1.06-1.06l-3-3Z" clip-rule="evenodd" />
+                </svg>
+                {isUnloading ? "Unloading..." : "Unload All"}
+              </button>
+            </div>
+          {/if}
+        </div>
+      {/if}
+    </div>
+    {#if !$isNarrow}
+      <div class="flex justify-between">
+        <div class="flex gap-2">
+          <button class="btn text-base flex items-center gap-2" onclick={toggleIdorName} style="line-height: 1.2">
+            <svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 24 24" fill="currentColor" class="w-5 h-5">
+              <path fill-rule="evenodd" d="M15.97 2.47a.75.75 0 0 1 1.06 0l4.5 4.5a.75.75 0 0 1 0 1.06l-4.5 4.5a.75.75 0 1 1-1.06-1.06l3.22-3.22H7.5a.75.75 0 0 1 0-1.5h11.69l-3.22-3.22a.75.75 0 0 1 0-1.06Zm-7.94 9a.75.75 0 0 1 0 1.06l-3.22 3.22H16.5a.75.75 0 0 1 0 1.5H4.81l3.22 3.22a.75.75 0 1 1-1.06 1.06l-4.5-4.5a.75.75 0 0 1 0-1.06l4.5-4.5a.75.75 0 0 1 1.06 0Z" clip-rule="evenodd" />
+            </svg>
+            {$showIdorNameStore === "id" ? "ID" : "Name"}
+          </button>
+
+          <button class="btn text-base flex items-center gap-2" onclick={toggleShowUnlisted} style="line-height: 1.2">
+            {#if $showUnlistedStore}
+              <svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 24 24" fill="currentColor" class="w-5 h-5">
+                <path d="M12 15a3 3 0 1 0 0-6 3 3 0 0 0 0 6Z" />
+                <path fill-rule="evenodd" d="M1.323 11.447C2.811 6.976 7.028 3.75 12.001 3.75c4.97 0 9.185 3.223 10.675 7.69.12.362.12.752 0 1.113-1.487 4.471-5.705 7.697-10.677 7.697-4.97 0-9.186-3.223-10.675-7.69a1.762 1.762 0 0 1 0-1.113ZM17.25 12a5.25 5.25 0 1 1-10.5 0 5.25 5.25 0 0 1 10.5 0Z" clip-rule="evenodd" />
+              </svg>
+            {:else}
+              <svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 24 24" fill="currentColor" class="w-5 h-5">
+                <path d="M3.53 2.47a.75.75 0 0 0-1.06 1.06l18 18a.75.75 0 1 0 1.06-1.06l-18-18ZM22.676 12.553a11.249 11.249 0 0 1-2.631 4.31l-3.099-3.099a5.25 5.25 0 0 0-6.71-6.71L7.759 4.577a11.217 11.217 0 0 1 4.242-.827c4.97 0 9.185 3.223 10.675 7.69.12.362.12.752 0 1.113Z" />
+                <path d="M15.75 12c0 .18-.013.357-.037.53l-4.244-4.243A3.75 3.75 0 0 1 15.75 12ZM12.53 15.713l-4.243-4.244a3.75 3.75 0 0 0 4.244 4.243Z" />
+                <path d="M6.75 12c0-.619.107-1.213.304-1.764l-3.1-3.1a11.25 11.25 0 0 0-2.63 4.31c-.12.362-.12.752 0 1.114 1.489 4.467 5.704 7.69 10.675 7.69 1.5 0 2.933-.294 4.242-.827l-2.477-2.477A5.25 5.25 0 0 1 6.75 12Z" />
+              </svg>
+            {/if}
+            unlisted
+          </button>
+        </div>
+        <button class="btn text-base flex items-center gap-2" onclick={handleUnloadAllModels} disabled={isUnloading}>
+          <svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 24 24" fill="currentColor" class="w-6 h-6">
+            <path fill-rule="evenodd" d="M12 2.25c-5.385 0-9.75 4.365-9.75 9.75s4.365 9.75 9.75 9.75 9.75-4.365 9.75-9.75S17.385 2.25 12 2.25Zm.53 5.47a.75.75 0 0 0-1.06 0l-3 3a.75.75 0 1 0 1.06 1.06l1.72-1.72v5.69a.75.75 0 0 0 1.5 0v-5.69l1.72 1.72a.75.75 0 1 0 1.06-1.06l-3-3Z" clip-rule="evenodd" />
+          </svg>
+          {isUnloading ? "Unloading..." : "Unload All"}
+        </button>
+      </div>
+    {/if}
+  </div>
+
+  <div class="flex-1 overflow-y-auto">
+    <table class="w-full">
+      <thead class="sticky top-0 bg-card z-10">
+        <tr class="text-left border-b border-gray-200 dark:border-white/10 bg-surface">
+          <th>{$showIdorNameStore === "id" ? "Model ID" : "Name"}</th>
+          <th></th>
+          <th>State</th>
+        </tr>
+      </thead>
+      <tbody>
+        {#each filteredModels.regularModels as model (model.id)}
+          <tr class="border-b hover:bg-secondary-hover border-gray-200">
+            <td class={model.unlisted ? "text-txtsecondary" : ""}>
+              <a href="/upstream/{model.id}/" class="font-semibold" target="_blank">
+                {getModelDisplay(model)}
+              </a>
+              {#if model.description}
+                <p class={model.unlisted ? "text-opacity-70" : ""}><em>{model.description}</em></p>
+              {/if}
+              {#if model.aliases && model.aliases.length > 0}
+                <p class="text-xs text-txtsecondary">Aliases: {model.aliases.join(", ")}</p>
+              {/if}
+            </td>
+            <td class="w-12">
+              {#if model.state === "stopped"}
+                <button class="btn btn--sm" onclick={() => loadModel(model.id)}>Load</button>
+              {:else}
+                <button class="btn btn--sm" onclick={() => unloadSingleModel(model.id)} disabled={model.state !== "ready"}>Unload</button>
+              {/if}
+            </td>
+            <td class="w-20">
+              <span class="w-16 text-center status status--{model.state}">{model.state}</span>
+            </td>
+          </tr>
+        {/each}
+      </tbody>
+    </table>
+
+    {#if Object.keys(filteredModels.peerModelsByPeerId).length > 0}
+      <h3 class="mt-8 mb-2">Peer Models</h3>
+      {#each Object.entries(filteredModels.peerModelsByPeerId).sort(([a], [b]) => a.localeCompare(b)) as [peerId, peerModels] (peerId)}
+        <div class="mb-4">
+          <table class="w-full">
+            <thead class="sticky top-0 bg-card z-10">
+              <tr class="text-left border-b border-gray-200 dark:border-white/10 bg-surface">
+                <th class="font-semibold">{peerId}</th>
+              </tr>
+            </thead>
+            <tbody>
+              {#each peerModels as model (model.id)}
+                <tr class="border-b hover:bg-secondary-hover border-gray-200">
+                  <td class="pl-8 {model.unlisted ? 'text-txtsecondary' : ''}">
+                    <span>{model.id}</span>
+                  </td>
+                </tr>
+              {/each}
+            </tbody>
+          </table>
+        </div>
+      {/each}
+    {/if}
+  </div>
+</div>

ui-svelte/src/components/ResizablePanels.svelte

@@ -0,0 +1,152 @@
+<script lang="ts">
+  import type { Snippet } from "svelte";
+  import { onMount } from "svelte";
+
+  interface Props {
+    direction: "horizontal" | "vertical";
+    storageKey: string;
+    leftPanel: Snippet;
+    rightPanel: Snippet;
+    defaultSize?: number;
+    minSize?: number;
+  }
+
+  let { direction, storageKey, leftPanel, rightPanel, defaultSize = 50, minSize = 5 }: Props = $props();
+
+  let containerRef: HTMLDivElement;
+  let isDragging = $state(false);
+  // svelte-ignore state_referenced_locally
+  let leftSize = $state(defaultSize);
+
+  // Load saved size from localStorage
+  onMount(() => {
+    const saved = localStorage.getItem(`panel-size-${storageKey}`);
+    if (saved) {
+      const parsed = parseFloat(saved);
+      if (!isNaN(parsed) && parsed >= minSize && parsed <= 100 - minSize) {
+        leftSize = parsed;
+      }
+    }
+  });
+
+  function saveSize(): void {
+    localStorage.setItem(`panel-size-${storageKey}`, String(leftSize));
+  }
+
+  function handleMouseDown(e: MouseEvent): void {
+    e.preventDefault();
+    isDragging = true;
+    document.addEventListener("mousemove", handleMouseMove);
+    document.addEventListener("mouseup", handleMouseUp);
+  }
+
+  function handleTouchStart(_e: TouchEvent): void {
+    isDragging = true;
+    document.addEventListener("touchmove", handleTouchMove);
+    document.addEventListener("touchend", handleTouchEnd);
+  }
+
+  function handleMouseMove(e: MouseEvent): void {
+    if (!isDragging || !containerRef) return;
+    updateSize(e.clientX, e.clientY);
+  }
+
+  function handleTouchMove(e: TouchEvent): void {
+    if (!isDragging || !containerRef || e.touches.length === 0) return;
+    updateSize(e.touches[0].clientX, e.touches[0].clientY);
+  }
+
+  function updateSize(clientX: number, clientY: number): void {
+    const rect = containerRef.getBoundingClientRect();
+
+    let newSize: number;
+    if (direction === "horizontal") {
+      newSize = ((clientX - rect.left) / rect.width) * 100;
+    } else {
+      newSize = ((clientY - rect.top) / rect.height) * 100;
+    }
+
+    // Clamp size
+    newSize = Math.max(minSize, Math.min(100 - minSize, newSize));
+    leftSize = newSize;
+  }
+
+  function handleMouseUp(): void {
+    isDragging = false;
+    saveSize();
+    document.removeEventListener("mousemove", handleMouseMove);
+    document.removeEventListener("mouseup", handleMouseUp);
+  }
+
+  function handleTouchEnd(): void {
+    isDragging = false;
+    saveSize();
+    document.removeEventListener("touchmove", handleTouchMove);
+    document.removeEventListener("touchend", handleTouchEnd);
+  }
+
+  function handleKeyDown(e: KeyboardEvent): void {
+    const step = 2; // 2% increment for keyboard navigation
+    const key = e.key;
+
+    if (direction === "horizontal" && (key === "ArrowLeft" || key === "ArrowRight")) {
+      e.preventDefault();
+      const delta = key === "ArrowLeft" ? -step : step;
+      const newSize = Math.max(minSize, Math.min(100 - minSize, leftSize + delta));
+      leftSize = newSize;
+      saveSize();
+    } else if (direction === "vertical" && (key === "ArrowUp" || key === "ArrowDown")) {
+      e.preventDefault();
+      const delta = key === "ArrowUp" ? -step : step;
+      const newSize = Math.max(minSize, Math.min(100 - minSize, leftSize + delta));
+      leftSize = newSize;
+      saveSize();
+    }
+  }
+
+  let containerClass = $derived(direction === "horizontal" ? "flex-row" : "flex-col");
+
+  let handleClass = $derived(
+    direction === "horizontal"
+      ? "w-2 h-full cursor-col-resize"
+      : "w-full h-2 cursor-row-resize"
+  );
+
+  let leftStyle = $derived(
+    direction === "horizontal"
+      ? `width: ${leftSize}%; min-width: ${minSize}%`
+      : `height: ${leftSize}%; min-height: ${minSize}%`
+  );
+
+  let rightStyle = $derived(
+    direction === "horizontal"
+      ? `width: ${100 - leftSize}%; min-width: ${minSize}%`
+      : `height: ${100 - leftSize}%; min-height: ${minSize}%`
+  );
+</script>
+
+<div bind:this={containerRef} class="flex {containerClass} h-full w-full gap-2">
+  <div style={leftStyle} class="overflow-hidden">
+    {@render leftPanel()}
+  </div>
+
+  <!-- svelte-ignore a11y_no_noninteractive_tabindex -->
+  <!-- svelte-ignore a11y_no_noninteractive_element_interactions -->
+  <div
+    role="separator"
+    tabindex="0"
+    class="{handleClass} bg-primary hover:bg-success transition-colors rounded flex-shrink-0"
+    onmousedown={handleMouseDown}
+    ontouchstart={handleTouchStart}
+    onkeydown={handleKeyDown}
+    aria-label="Resize panels"
+    aria-orientation={direction}
+    aria-valuenow={Math.round(leftSize)}
+    aria-valuemin={minSize}
+    aria-valuemax={100 - minSize}
+  ></div>
+
+  <div style={rightStyle} class="overflow-hidden">
+    {@render rightPanel()}
+  </div>
+</div>

ui-svelte/src/components/TokenHistogram.svelte

@@ -0,0 +1,113 @@
+<script lang="ts">
+  import type { HistogramData } from "../lib/types";
+
+  let { data }: { data: HistogramData } = $props();
+
+  const height = 55;
+  const padding = { top: 5, right: 45, bottom: 15, left: 45 };
+  const viewBoxWidth = 1200;
+  const chartWidth = viewBoxWidth - padding.left - padding.right;
+  const chartHeight = height - padding.top - padding.bottom;
+
+  let maxCount = $derived(Math.max(...data.bins));
+  let barWidth = $derived(chartWidth / data.bins.length);
+  let range = $derived(data.max - data.min);
+
+  function getXPosition(value: number): number {
+    return padding.left + ((value - data.min) / range) * chartWidth;
+  }
+</script>
+
+<div class="mt-2 w-full">
+  <svg viewBox="0 0 {viewBoxWidth} {height}" class="w-full h-auto" preserveAspectRatio="xMidYMid meet">
+    <!-- Y-axis -->
+    <line
+      x1={padding.left}
+      y1={padding.top}
+      x2={padding.left}
+      y2={height - padding.bottom}
+      stroke="currentColor"
+      stroke-width="1"
+      opacity="0.3"
+    />
+
+    <!-- X-axis -->
+    <line
+      x1={padding.left}
+      y1={height - padding.bottom}
+      x2={viewBoxWidth - padding.right}
+      y2={height - padding.bottom}
+      stroke="currentColor"
+      stroke-width="1"
+      opacity="0.3"
+    />
+
+    <!-- Histogram bars -->
+    {#each data.bins as count, i}
+      {@const barHeight = maxCount > 0 ? (count / maxCount) * chartHeight : 0}
+      {@const x = padding.left + i * barWidth}
+      {@const y = height - padding.bottom - barHeight}
+      {@const binStart = data.min + i * data.binSize}
+      {@const binEnd = binStart + data.binSize}
+      <g>
+        <rect
+          {x}
+          {y}
+          width={Math.max(barWidth - 1, 1)}
+          height={barHeight}
+          fill="currentColor"
+          opacity="0.6"
+          class="text-blue-500 dark:text-blue-400 hover:opacity-90 transition-opacity cursor-pointer"
+        />
+        <title>{`${binStart.toFixed(1)} - ${binEnd.toFixed(1)} tokens/sec\nCount: ${count}`}</title>
+      </g>
+    {/each}
+
+    <!-- Percentile lines -->
+    <line
+      x1={getXPosition(data.p50)}
+      y1={padding.top}
+      x2={getXPosition(data.p50)}
+      y2={height - padding.bottom}
+      stroke="currentColor"
+      stroke-width="2"
+      stroke-dasharray="4 2"
+      opacity="0.7"
+      class="text-gray-600 dark:text-gray-400"
+    />
+
+    <line
+      x1={getXPosition(data.p95)}
+      y1={padding.top}
+      x2={getXPosition(data.p95)}
+      y2={height - padding.bottom}
+      stroke="currentColor"
+      stroke-width="2"
+      stroke-dasharray="4 2"
+      opacity="0.7"
+      class="text-orange-500 dark:text-orange-400"
+    />
+
+    <line
+      x1={getXPosition(data.p99)}
+      y1={padding.top}
+      x2={getXPosition(data.p99)}
+      y2={height - padding.bottom}
+      stroke="currentColor"
+      stroke-width="2"
+      stroke-dasharray="4 2"
+      opacity="0.7"
+      class="text-green-500 dark:text-green-400"
+    />
+
+    <!-- X-axis labels -->
+    <text x={padding.left} y={height - 5} font-size="10" fill="currentColor" opacity="0.6" text-anchor="start">
+      {data.min.toFixed(1)}
+    </text>
+
+    <text x={viewBoxWidth - padding.right} y={height - 5} font-size="10" fill="currentColor" opacity="0.6" text-anchor="end">
+      {data.max.toFixed(1)}
+    </text>
+
+  </svg>
+</div>

ui-svelte/src/components/Tooltip.svelte

@@ -0,0 +1,20 @@
+<script lang="ts">
+  interface Props {
+    content: string;
+  }
+
+  let { content }: Props = $props();
+</script>
+
+<div class="relative group inline-block">
+  <span class="cursor-help">&#9432;</span>
+  <div
+    class="absolute top-full left-1/2 transform -translate-x-1/2 mt-2
+           px-3 py-2 bg-gray-900 text-white text-sm rounded-md
+           opacity-0 group-hover:opacity-100 transition-opacity
+           duration-200 pointer-events-none whitespace-nowrap z-50 normal-case"
+  >
+    {content}
+    <div class="absolute bottom-full left-1/2 transform -translate-x-1/2 border-4 border-transparent border-b-gray-900"></div>
+  </div>
+</div>

ui-svelte/src/components/playground/AudioInterface.svelte

@@ -0,0 +1,256 @@
+<script lang="ts">
+  import { models } from "../../stores/api";
+  import { persistentStore } from "../../stores/persistent";
+  import { transcribeAudio } from "../../lib/audioApi";
+  import { playgroundStores } from "../../stores/playgroundActivity";
+  import ModelSelector from "./ModelSelector.svelte";
+
+  const selectedModelStore = persistentStore<string>("playground-audio-model", "");
+
+  let selectedFile = $state<File | null>(null);
+  let isTranscribing = $state(false);
+  let transcriptionResult = $state<string | null>(null);
+  let error = $state<string | null>(null);
+  let abortController = $state<AbortController | null>(null);
+  let isDragging = $state(false);
+  let fileInput = $state<HTMLInputElement | null>(null);
+  let copied = $state(false);
+
+  const ACCEPTED_FORMATS = ['.mp3', '.wav', '.ogg'];
+  const MAX_FILE_SIZE = 25 * 1024 * 1024; // 25MB
+
+  let hasModels = $derived($models.some((m) => !m.unlisted));
+
+  let canTranscribe = $derived(selectedFile !== null && $selectedModelStore !== "" && !isTranscribing);
+
+  $effect(() => {
+    playgroundStores.audioTranscribing.set(isTranscribing);
+  });
+
+  function validateFile(file: File): { valid: boolean; error?: string } {
+    const ext = '.' + file.name.split('.').pop()?.toLowerCase();
+
+    if (!ACCEPTED_FORMATS.includes(ext)) {
+      return { valid: false, error: 'Invalid file type. Accepted: MP3, WAV, OGG' };
+    }
+
+    if (file.size > MAX_FILE_SIZE) {
+      return { valid: false, error: 'File too large. Maximum: 25MB' };
+    }
+
+    return { valid: true };
+  }
+
+  function handleFileSelect(event: Event) {
+    const target = event.target as HTMLInputElement;
+    const file = target.files?.[0];
+    if (file) {
+      const validation = validateFile(file);
+      if (validation.valid) {
+        selectedFile = file;
+        error = null;
+        transcriptionResult = null;
+      } else {
+        error = validation.error || "Invalid file";
+        selectedFile = null;
+      }
+    }
+  }
+
+  function handleDragOver(event: DragEvent) {
+    event.preventDefault();
+    isDragging = true;
+  }
+
+  function handleDragLeave() {
+    isDragging = false;
+  }
+
+  function handleDrop(event: DragEvent) {
+    event.preventDefault();
+    isDragging = false;
+
+    const file = event.dataTransfer?.files[0];
+    if (file) {
+      const validation = validateFile(file);
+      if (validation.valid) {
+        selectedFile = file;
+        error = null;
+        transcriptionResult = null;
+      } else {
+        error = validation.error || "Invalid file";
+        selectedFile = null;
+      }
+    }
+  }
+
+  async function transcribe() {
+    if (!selectedFile || !$selectedModelStore || isTranscribing) return;
+
+    isTranscribing = true;
+    error = null;
+    transcriptionResult = null;
+    abortController = new AbortController();
+
+    try {
+      const response = await transcribeAudio(
+        $selectedModelStore,
+        selectedFile,
+        abortController.signal
+      );
+
+      transcriptionResult = response.text;
+    } catch (err) {
+      if (err instanceof Error && err.name === "AbortError") {
+        // User cancelled
+      } else {
+        error = err instanceof Error ? err.message : "An error occurred";
+      }
+    } finally {
+      isTranscribing = false;
+      abortController = null;
+    }
+  }
+
+  function cancelTranscription() {
+    abortController?.abort();
+  }
+
+  function clearAll() {
+    selectedFile = null;
+    transcriptionResult = null;
+    error = null;
+    if (fileInput) {
+      fileInput.value = '';
+    }
+  }
+
+  function copyToClipboard() {
+    if (transcriptionResult) {
+      navigator.clipboard.writeText(transcriptionResult);
+      copied = true;
+      setTimeout(() => {
+        copied = false;
+      }, 2000);
+    }
+  }
+
+  function formatFileSize(bytes: number): string {
+    if (bytes < 1024) return bytes + ' B';
+    if (bytes < 1024 * 1024) return (bytes / 1024).toFixed(1) + ' KB';
+    return (bytes / (1024 * 1024)).toFixed(1) + ' MB';
+  }
+</script>
+
+<div class="flex flex-col h-full">
+  <!-- Model selector -->
+  <div class="shrink-0 flex flex-wrap gap-2 mb-4">
+    <ModelSelector bind:value={$selectedModelStore} placeholder="Select an audio model..." disabled={isTranscribing} />
+  </div>
+
+  <!-- Empty state for no models configured -->
+  {#if !hasModels}
+    <div class="flex-1 flex items-center justify-center text-txtsecondary">
+      <p>No models configured. Add models to your configuration to transcribe audio.</p>
+    </div>
+  {:else}
+    <!-- File upload / Result display area -->
+    <div class="flex-1 overflow-auto mb-4 flex items-center justify-center bg-surface border border-gray-200 dark:border-white/10 rounded">
+      {#if isTranscribing}
+        <div class="text-center text-txtsecondary">
+          <div class="inline-block w-8 h-8 border-4 border-primary border-t-transparent rounded-full animate-spin mb-2"></div>
+          <p>Transcribing audio...</p>
+        </div>
+      {:else if error}
+        <div class="text-center text-red-500 p-4">
+          <p class="font-medium">Error</p>
+          <p class="text-sm mt-1">{error}</p>
+        </div>
+      {:else if transcriptionResult}
+        <div class="w-full h-full flex flex-col p-4">
+          <div class="flex justify-between items-center mb-2">
+            <h3 class="font-medium">Transcription Result</h3>
+            <button
+              class="btn btn-sm"
+              onclick={copyToClipboard}
+              title={copied ? 'Copied!' : 'Copy to clipboard'}
+            >
+              {#if copied}
+                <svg class="w-5 h-5 text-green-500" fill="none" stroke="currentColor" viewBox="0 0 24 24">
+                  <path stroke-linecap="round" stroke-linejoin="round" stroke-width="2" d="M5 13l4 4L19 7"></path>
+                </svg>
+              {:else}
+                <svg class="w-5 h-5" fill="none" stroke="currentColor" viewBox="0 0 24 24">
+                  <path stroke-linecap="round" stroke-linejoin="round" stroke-width="2" d="M8 16H6a2 2 0 01-2-2V6a2 2 0 012-2h8a2 2 0 012 2v2m-6 12h8a2 2 0 002-2v-8a2 2 0 00-2-2h-8a2 2 0 00-2 2v8a2 2 0 002 2z"></path>
+                </svg>
+              {/if}
+            </button>
+          </div>
+          <div class="flex-1 overflow-auto p-3 rounded border border-gray-200 dark:border-white/10 bg-background whitespace-pre-wrap">
+            {transcriptionResult}
+          </div>
+        </div>
+      {:else if selectedFile}
+        <div class="text-center text-txtsecondary p-4">
+          <p class="font-medium mb-2">File Selected</p>
+          <p class="text-sm">{selectedFile.name}</p>
+          <p class="text-xs mt-1">{formatFileSize(selectedFile.size)}</p>
+        </div>
+      {:else}
+        <div
+          role="region"
+          aria-label="Audio file drop zone"
+          class="w-full h-full flex items-center justify-center text-center text-txtsecondary p-8 {isDragging ? 'bg-primary/10' : ''}"
+          ondragover={handleDragOver}
+          ondragleave={handleDragLeave}
+          ondrop={handleDrop}
+        >
+          <div>
+            <p class="mb-2">Drag and drop an audio file here</p>
+            <p class="text-sm">or use the Browse button below</p>
+            <p class="text-xs mt-4">Accepted formats: MP3, WAV, OGG (max 25MB)</p>
+          </div>
+        </div>
+      {/if}
+    </div>
+
+    <!-- File input and transcribe button -->
+    <div class="shrink-0 flex gap-2">
+      <input
+        type="file"
+        accept=".mp3,.wav,.ogg"
+        class="hidden"
+        onchange={handleFileSelect}
+        bind:this={fileInput}
+      />
+      <button
+        class="btn"
+        onclick={() => fileInput?.click()}
+        disabled={isTranscribing}
+      >
+        Browse Files
+      </button>
+      <div class="flex-1"></div>
+      {#if isTranscribing}
+        <button class="btn bg-red-500 hover:bg-red-600 text-white" onclick={cancelTranscription}>
+          Cancel
+        </button>
+      {:else}
+        <button
+          class="btn bg-primary text-btn-primary-text hover:opacity-90"
+          onclick={transcribe}
+          disabled={!canTranscribe}
+        >
+          Transcribe
+        </button>
+        <button
+          class="btn"
+          onclick={clearAll}
+          disabled={!selectedFile && !transcriptionResult && !error}
+        >
+          Clear
+        </button>
+      {/if}
+    </div>
+  {/if}
+</div>

ui-svelte/src/components/playground/ChatInterface.svelte

@@ -0,0 +1,466 @@
+<script lang="ts">
+  import { models } from "../../stores/api";
+  import { persistentStore } from "../../stores/persistent";
+  import { streamChatCompletion } from "../../lib/chatApi";
+  import { playgroundStores } from "../../stores/playgroundActivity";
+  import type { ChatMessage, ContentPart } from "../../lib/types";
+  import ChatMessageComponent from "./ChatMessage.svelte";
+  import ModelSelector from "./ModelSelector.svelte";
+  import ExpandableTextarea from "./ExpandableTextarea.svelte";
+
+  const selectedModelStore = persistentStore<string>("playground-selected-model", "");
+  const systemPromptStore = persistentStore<string>("playground-system-prompt", "");
+  const temperatureStore = persistentStore<number>("playground-temperature", 0.7);
+
+  function loadMessages(): ChatMessage[] {
+    try {
+      const saved = localStorage.getItem("playground-messages");
+      return saved ? JSON.parse(saved) : [];
+    } catch {
+      return [];
+    }
+  }
+
+  let messages = $state<ChatMessage[]>(loadMessages());
+  let userInput = $state("");
+  let isStreaming = $state(false);
+  let isReasoning = $state(false);
+  let reasoningStartTime = $state<number>(0);
+  let abortController = $state<AbortController | null>(null);
+  let messagesContainer: HTMLDivElement | undefined = $state();
+  let showSettings = $state(false);
+  let attachedImages = $state<string[]>([]);
+  let fileInput = $state<HTMLInputElement | null>(null);
+  let imageError = $state<string | null>(null);
+
+  let hasModels = $derived($models.some((m) => !m.unlisted));
+  let userScrolledUp = $state(false);
+
+  $effect(() => {
+    playgroundStores.chatStreaming.set(isStreaming);
+  });
+
+  function handleMessagesScroll() {
+    if (!messagesContainer) return;
+    const { scrollTop, scrollHeight, clientHeight } = messagesContainer;
+    // Consider "at bottom" if within 40px of the bottom
+    userScrolledUp = scrollHeight - scrollTop - clientHeight > 40;
+  }
+
+  // Auto-scroll when messages change — skip if user scrolled up
+  $effect(() => {
+    if (messages.length > 0 && messagesContainer && !userScrolledUp) {
+      messagesContainer.scrollTo({
+        top: messagesContainer.scrollHeight,
+        behavior: isStreaming ? "instant" : "smooth",
+      });
+    }
+  });
+
+  // Persist messages to localStorage (throttled to once per 2s)
+  let lastSaveTime = 0;
+  $effect(() => {
+    const json = JSON.stringify(messages);
+    const elapsed = Date.now() - lastSaveTime;
+    const save = () => {
+      try { localStorage.setItem("playground-messages", json); } catch {}
+      lastSaveTime = Date.now();
+    };
+    if (elapsed >= 2000) {
+      save();
+      return;
+    }
+    const timer = setTimeout(save, 2000 - elapsed);
+    return () => clearTimeout(timer);
+  });
+
+  async function sendMessage() {
+    const trimmedInput = userInput.trim();
+    if ((!trimmedInput && attachedImages.length === 0) || !$selectedModelStore || isStreaming) return;
+
+    userScrolledUp = false;
+
+    // Build message content (multimodal if images attached)
+    let content: string | ContentPart[];
+    if (attachedImages.length > 0) {
+      const parts: ContentPart[] = [];
+      if (trimmedInput) {
+        parts.push({ type: "text", text: trimmedInput });
+      }
+      for (const url of attachedImages) {
+        parts.push({ type: "image_url", image_url: { url } });
+      }
+      content = parts;
+    } else {
+      content = trimmedInput;
+    }
+
+    // Add user message
+    messages = [...messages, { role: "user", content }];
+    userInput = "";
+    attachedImages = [];
+    imageError = null;
+
+    // Generate response from the new user message
+    await regenerateFromIndex(messages.length - 1);
+  }
+
+  function cancelStreaming() {
+    abortController?.abort();
+  }
+
+  function newChat() {
+    if (isStreaming) {
+      cancelStreaming();
+    }
+    messages = [];
+    isReasoning = false;
+    reasoningStartTime = 0;
+  }
+
+  async function regenerateFromIndex(idx: number) {
+    // Remove all messages after the edited user message
+    messages = messages.slice(0, idx + 1);
+
+    // Add empty assistant message for the new response
+    messages = [...messages, { role: "assistant", content: "" }];
+
+    isStreaming = true;
+    isReasoning = false;
+    reasoningStartTime = 0;
+    abortController = new AbortController();
+
+    try {
+      // Build messages array with optional system prompt
+      const apiMessages: ChatMessage[] = [];
+      if ($systemPromptStore.trim()) {
+        apiMessages.push({ role: "system", content: $systemPromptStore.trim() });
+      }
+      apiMessages.push(...messages.slice(0, -1)); // Add all messages except the empty assistant one
+
+      const stream = streamChatCompletion(
+        $selectedModelStore,
+        apiMessages,
+        abortController.signal,
+        { temperature: $temperatureStore }
+      );
+
+      for await (const chunk of stream) {
+        if (chunk.done) break;
+
+        // Handle reasoning content
+        if (chunk.reasoning_content) {
+          // Start timing on first reasoning content
+          if (!isReasoning) {
+            isReasoning = true;
+            reasoningStartTime = Date.now();
+          }
+
+          // Update the last message with reasoning content
+          messages = messages.map((msg, i) =>
+            i === messages.length - 1
+              ? { ...msg, reasoning_content: (msg.reasoning_content || "") + chunk.reasoning_content }
+              : msg
+          );
+        }
+
+        // Handle regular content - end reasoning phase when we get content
+        if (chunk.content) {
+          if (isReasoning) {
+            // Calculate reasoning time
+            const reasoningTimeMs = Date.now() - reasoningStartTime;
+            isReasoning = false;
+
+            // Update message with reasoning time
+            messages = messages.map((msg, i) =>
+              i === messages.length - 1
+                ? { ...msg, reasoningTimeMs }
+                : msg
+            );
+          }
+
+          // Update the last message (assistant) with new content
+          messages = messages.map((msg, i) =>
+            i === messages.length - 1
+              ? { ...msg, content: msg.content + chunk.content }
+              : msg
+          );
+        }
+      }
+    } catch (error) {
+      if (error instanceof Error && error.name === "AbortError") {
+        // User cancelled, keep partial response
+        // If we were still reasoning, record the time
+        if (isReasoning && reasoningStartTime > 0) {
+          const reasoningTimeMs = Date.now() - reasoningStartTime;
+          messages = messages.map((msg, i) =>
+            i === messages.length - 1
+              ? { ...msg, reasoningTimeMs }
+              : msg
+          );
+        }
+      } else {
+        // Show error in the assistant message
+        const errorMessage = error instanceof Error ? error.message : "An error occurred";
+        messages = messages.map((msg, i) =>
+          i === messages.length - 1
+            ? { ...msg, content: msg.content + `\n\n**Error:** ${errorMessage}` }
+            : msg
+        );
+      }
+    } finally {
+      isStreaming = false;
+      isReasoning = false;
+      abortController = null;
+    }
+  }
+
+  async function editMessage(idx: number, newContent: string) {
+    if (isStreaming || !$selectedModelStore) return;
+
+    // Update the user message at the specified index
+    messages = messages.map((msg, i) =>
+      i === idx ? { ...msg, content: newContent } : msg
+    );
+
+    // Trigger a new chat request with the updated messages
+    await regenerateFromIndex(idx);
+  }
+
+  function handleKeyDown(event: KeyboardEvent) {
+    if (event.key === "Enter" && !event.shiftKey) {
+      event.preventDefault();
+      sendMessage();
+    }
+  }
+
+  const ACCEPTED_IMAGE_FORMATS = ["image/jpeg", "image/png", "image/gif", "image/webp"];
+  const MAX_IMAGE_SIZE = 20 * 1024 * 1024; // 20MB
+  const MAX_IMAGES_PER_MESSAGE = 5;
+
+  function validateImageFile(file: File): string | null {
+    if (!ACCEPTED_IMAGE_FORMATS.includes(file.type)) {
+      return `Invalid file type: ${file.type}. Accepted formats: JPG, PNG, GIF, WEBP`;
+    }
+    if (file.size > MAX_IMAGE_SIZE) {
+      return `File too large: ${(file.size / 1024 / 1024).toFixed(1)}MB. Maximum size: 20MB`;
+    }
+    return null;
+  }
+
+  function fileToDataUrl(file: File): Promise<string> {
+    return new Promise((resolve, reject) => {
+      const reader = new FileReader();
+      reader.onload = () => resolve(reader.result as string);
+      reader.onerror = () => reject(new Error("Failed to read file"));
+      reader.readAsDataURL(file);
+    });
+  }
+
+  async function processImageFiles(files: File[]): Promise<void> {
+    imageError = null;
+
+    if (attachedImages.length + files.length > MAX_IMAGES_PER_MESSAGE) {
+      imageError = `Maximum ${MAX_IMAGES_PER_MESSAGE} images per message`;
+      return;
+    }
+
+    for (const file of files) {
+      const error = validateImageFile(file);
+      if (error) {
+        imageError = error;
+        return;
+      }
+    }
+
+    try {
+      const dataUrls = await Promise.all(files.map(fileToDataUrl));
+      attachedImages = [...attachedImages, ...dataUrls];
+    } catch (error) {
+      imageError = error instanceof Error ? error.message : "Failed to process images";
+    }
+  }
+
+  function handleImageSelect(event: Event) {
+    const input = event.target as HTMLInputElement;
+    if (input.files && input.files.length > 0) {
+      processImageFiles(Array.from(input.files));
+    }
+    // Reset the input so the same file can be selected again
+    input.value = "";
+  }
+
+  function removeImage(idx: number) {
+    attachedImages = attachedImages.filter((_, i) => i !== idx);
+    imageError = null;
+  }
+</script>
+
+<div class="flex flex-col h-full">
+  <!-- Model selector and controls -->
+  <div class="shrink-0 flex flex-wrap gap-2 mb-4">
+    <ModelSelector bind:value={$selectedModelStore} placeholder="Select a model..." disabled={isStreaming} />
+    <div class="flex gap-2">
+      <button
+        class="btn"
+        onclick={() => (showSettings = !showSettings)}
+        title="Settings"
+      >
+        <svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 20 20" fill="currentColor" class="w-5 h-5">
+          <path fill-rule="evenodd" d="M8.34 1.804A1 1 0 0 1 9.32 1h1.36a1 1 0 0 1 .98.804l.295 1.473c.497.144.971.342 1.416.587l1.25-.834a1 1 0 0 1 1.262.125l.962.962a1 1 0 0 1 .125 1.262l-.834 1.25c.245.445.443.919.587 1.416l1.473.295a1 1 0 0 1 .804.98v1.36a1 1 0 0 1-.804.98l-1.473.295a6.95 6.95 0 0 1-.587 1.416l.834 1.25a1 1 0 0 1-.125 1.262l-.962.962a1 1 0 0 1-1.262.125l-1.25-.834a6.953 6.953 0 0 1-1.416.587l-.295 1.473a1 1 0 0 1-.98.804H9.32a1 1 0 0 1-.98-.804l-.295-1.473a6.957 6.957 0 0 1-1.416-.587l-1.25.834a1 1 0 0 1-1.262-.125l-.962-.962a1 1 0 0 1-.125-1.262l.834-1.25a6.957 6.957 0 0 1-.587-1.416l-1.473-.295A1 1 0 0 1 1 10.68V9.32a1 1 0 0 1 .804-.98l1.473-.295c.144-.497.342-.971.587-1.416l-.834-1.25a1 1 0 0 1 .125-1.262l.962-.962A1 1 0 0 1 5.38 3.03l1.25.834a6.957 6.957 0 0 1 1.416-.587l.294-1.473ZM13 10a3 3 0 1 1-6 0 3 3 0 0 1 6 0Z" clip-rule="evenodd" />
+        </svg>
+      </button>
+      <button class="btn" onclick={newChat} disabled={messages.length === 0 && !isStreaming}>
+        New Chat
+      </button>
+    </div>
+  </div>
+
+  <!-- Settings panel -->
+  {#if showSettings}
+    <div class="shrink-0 mb-4 p-4 bg-surface border border-gray-200 dark:border-white/10 rounded">
+      <div class="mb-4">
+        <label class="block text-sm font-medium mb-1" for="system-prompt">System Prompt</label>
+        <textarea
+          id="system-prompt"
+          class="w-full px-3 py-2 rounded border border-gray-200 dark:border-white/10 bg-card focus:outline-none focus:ring-2 focus:ring-primary resize-none"
+          placeholder="You are a helpful assistant..."
+          rows="3"
+          bind:value={$systemPromptStore}
+          disabled={isStreaming}
+        ></textarea>
+      </div>
+      <div>
+        <label class="block text-sm font-medium mb-1" for="temperature">
+          Temperature: {$temperatureStore.toFixed(2)}
+        </label>
+        <input
+          id="temperature"
+          type="range"
+          min="0"
+          max="2"
+          step="0.05"
+          class="w-full"
+          bind:value={$temperatureStore}
+          disabled={isStreaming}
+        />
+        <div class="flex justify-between text-xs text-txtsecondary mt-1">
+          <span>Precise (0)</span>
+          <span>Creative (2)</span>
+        </div>
+      </div>
+    </div>
+  {/if}
+
+  <!-- Empty state for no models configured -->
+  {#if !hasModels}
+    <div class="flex-1 flex items-center justify-center text-txtsecondary">
+      <p>No models configured. Add models to your configuration to start chatting.</p>
+    </div>
+  {:else}
+    <!-- Messages area -->
+    <div
+      class="flex-1 overflow-y-auto mb-4 px-2"
+      bind:this={messagesContainer}
+      onscroll={handleMessagesScroll}
+    >
+      {#if messages.length === 0}
+        <div class="h-full flex items-center justify-center text-txtsecondary">
+          <p>Start a conversation by typing a message below.</p>
+        </div>
+      {:else}
+        {#each messages as message, idx (idx)}
+          <ChatMessageComponent
+            role={message.role}
+            content={message.content}
+            reasoning_content={message.reasoning_content}
+            reasoningTimeMs={message.reasoningTimeMs}
+            isStreaming={isStreaming && idx === messages.length - 1 && message.role === "assistant"}
+            isReasoning={isReasoning && idx === messages.length - 1 && message.role === "assistant"}
+            onEdit={message.role === "user" ? (newContent) => editMessage(idx, newContent) : undefined}
+            onRegenerate={message.role === "assistant" && idx > 0 && messages[idx - 1].role === "user"
+              ? () => regenerateFromIndex(idx - 1)
+              : undefined}
+          />
+        {/each}
+      {/if}
+    </div>
+
+    <!-- Input area -->
+    <div class="shrink-0">
+      <!-- Image preview strip -->
+      {#if attachedImages.length > 0}
+        <div class="mb-2 flex flex-wrap gap-2">
+          {#each attachedImages as imageUrl, idx (idx)}
+            <div class="relative group">
+              <img
+                src={imageUrl}
+                alt="Attached image {idx + 1}"
+                class="w-20 h-20 object-cover rounded border border-gray-200 dark:border-white/10"
+              />
+              <button
+                class="absolute -top-2 -right-2 bg-red-500 text-white rounded-full w-6 h-6 flex items-center justify-center opacity-0 group-hover:opacity-100 transition-opacity"
+                onclick={() => removeImage(idx)}
+                title="Remove image"
+              >
+                ×
+              </button>
+            </div>
+          {/each}
+        </div>
+      {/if}
+
+      <!-- Error message -->
+      {#if imageError}
+        <div class="mb-2 p-2 bg-red-100 dark:bg-red-900/20 text-red-700 dark:text-red-400 rounded text-sm">
+          {imageError}
+        </div>
+      {/if}
+
+      <div class="flex gap-2">
+        <!-- Hidden file input -->
+        <input
+          type="file"
+          accept=".jpg,.jpeg,.png,.gif,.webp"
+          multiple
+          class="hidden"
+          bind:this={fileInput}
+          onchange={handleImageSelect}
+        />
+
+        <ExpandableTextarea
+          bind:value={userInput}
+          placeholder="Type a message..."
+          rows={3}
+          onkeydown={handleKeyDown}
+          disabled={isStreaming || !$selectedModelStore}
+        />
+        <div class="flex flex-col gap-2">
+          {#if isStreaming}
+            <button class="btn bg-red-500 hover:bg-red-600 text-white" onclick={cancelStreaming}>
+              Cancel
+            </button>
+          {:else}
+            <button
+              class="btn"
+              onclick={() => fileInput?.click()}
+              disabled={isStreaming || !$selectedModelStore}
+              title="Attach image"
+            >
+              <svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 20 20" fill="currentColor" class="w-5 h-5">
+                <path fill-rule="evenodd" d="M1 5.25A2.25 2.25 0 0 1 3.25 3h13.5A2.25 2.25 0 0 1 19 5.25v9.5A2.25 2.25 0 0 1 16.75 17H3.25A2.25 2.25 0 0 1 1 14.75v-9.5Zm1.5 5.81v3.69c0 .414.336.75.75.75h13.5a.75.75 0 0 0 .75-.75v-2.69l-2.22-2.219a.75.75 0 0 0-1.06 0l-1.91 1.909.47.47a.75.75 0 1 1-1.06 1.06L6.53 8.091a.75.75 0 0 0-1.06 0l-2.97 2.97ZM12 7a1 1 0 1 1-2 0 1 1 0 0 1 2 0Z" clip-rule="evenodd" />
+              </svg>
+            </button>
+            <button
+              class="btn bg-primary text-btn-primary-text hover:opacity-90"
+              onclick={sendMessage}
+              disabled={(!userInput.trim() && attachedImages.length === 0) || !$selectedModelStore}
+            >
+              Send
+            </button>
+          {/if}
+        </div>
+      </div>
+    </div>
+  {/if}
+</div>

ui-svelte/src/components/playground/ChatMessage.svelte

@@ -0,0 +1,467 @@
+<script lang="ts">
+  import { renderMarkdown, escapeHtml, renderStreamingMarkdown, createStreamingCache } from "../../lib/markdown";
+  import type { RenderedBlock } from "../../lib/markdown";
+  import { Copy, Check, Pencil, X, Save, RefreshCw, ChevronDown, ChevronRight, Brain, Code } from "lucide-svelte";
+  import { getTextContent, getImageUrls } from "../../lib/types";
+  import type { ContentPart } from "../../lib/types";
+
+  interface Props {
+    role: "user" | "assistant" | "system";
+    content: string | ContentPart[];
+    reasoning_content?: string;
+    reasoningTimeMs?: number;
+    isStreaming?: boolean;
+    isReasoning?: boolean;
+    onEdit?: (newContent: string) => void;
+    onRegenerate?: () => void;
+  }
+
+  let { role, content, reasoning_content = "", reasoningTimeMs = 0, isStreaming = false, isReasoning = false, onEdit, onRegenerate }: Props = $props();
+
+  let textContent = $derived(getTextContent(content));
+  let imageUrls = $derived(getImageUrls(content));
+  let hasImages = $derived(imageUrls.length > 0);
+  let canEdit = $derived(onEdit !== undefined && !hasImages);
+
+  let streamingCache = createStreamingCache();
+  let renderedParts = $derived.by(() => {
+    if (role !== "assistant") {
+      return { blocks: [{ id: -1, html: escapeHtml(textContent).replace(/\n/g, '<br>') }] as RenderedBlock[], pendingHtml: "" };
+    }
+    if (!isStreaming) {
+      streamingCache = createStreamingCache();
+      return { blocks: [{ id: -1, html: renderMarkdown(textContent) }] as RenderedBlock[], pendingHtml: "" };
+    }
+    return renderStreamingMarkdown(textContent, streamingCache);
+  });
+  let copied = $state(false);
+  let showRaw = $state(false);
+  let isEditing = $state(false);
+  let editContent = $state("");
+  let showReasoning = $state(false);
+  let modalImageUrl = $state<string | null>(null);
+
+  function formatDuration(ms: number): string {
+    if (ms < 1000) {
+      return `${ms.toFixed(0)}ms`;
+    }
+    return `${(ms / 1000).toFixed(1)}s`;
+  }
+
+  async function copyToClipboard() {
+    try {
+      if (navigator.clipboard && window.isSecureContext) {
+        await navigator.clipboard.writeText(textContent);
+      } else {
+        // Fallback for non-secure contexts (HTTP)
+        const textarea = document.createElement("textarea");
+        textarea.value = textContent;
+        textarea.style.position = "fixed";
+        textarea.style.left = "-9999px";
+        document.body.appendChild(textarea);
+        textarea.select();
+        document.execCommand("copy");
+        document.body.removeChild(textarea);
+      }
+      copied = true;
+      setTimeout(() => (copied = false), 2000);
+    } catch (err) {
+      console.error("Failed to copy:", err);
+    }
+  }
+
+  function startEdit() {
+    editContent = textContent;
+    isEditing = true;
+  }
+
+  function cancelEdit() {
+    isEditing = false;
+    editContent = "";
+  }
+
+  function saveEdit() {
+    if (onEdit && editContent.trim() !== textContent) {
+      onEdit(editContent.trim());
+    }
+    isEditing = false;
+    editContent = "";
+  }
+
+  function openModal(imageUrl: string) {
+    modalImageUrl = imageUrl;
+    document.body.style.overflow = "hidden";
+  }
+
+  function closeModal(event?: MouseEvent) {
+    // Only close if clicking the background, not the image
+    if (event && event.target !== event.currentTarget) {
+      return;
+    }
+    modalImageUrl = null;
+    document.body.style.overflow = "";
+  }
+
+  function handleModalKeyDown(event: KeyboardEvent) {
+    if (event.key === "Escape") {
+      closeModal();
+    }
+  }
+
+  function handleKeyDown(event: KeyboardEvent) {
+    if (event.key === "Enter" && !event.shiftKey) {
+      event.preventDefault();
+      saveEdit();
+    } else if (event.key === "Escape") {
+      cancelEdit();
+    }
+  }
+
+  const COPY_SVG = `<svg xmlns="http://www.w3.org/2000/svg" width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><rect width="14" height="14" x="8" y="8" rx="2" ry="2"/><path d="M4 16c-1.1 0-2-.9-2-2V4c0-1.1.9-2 2-2h10c1.1 0 2 .9 2 2"/></svg>`;
+  const CHECK_SVG = `<svg xmlns="http://www.w3.org/2000/svg" width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><path d="M20 6 9 17l-5-5"/></svg>`;
+
+  function codeBlockCopy(node: HTMLElement) {
+    function attachButtons() {
+      node.querySelectorAll<HTMLPreElement>('pre:not([data-copy-btn])').forEach(pre => {
+        pre.setAttribute('data-copy-btn', 'true');
+        const btn = document.createElement('button');
+        btn.className = 'code-copy-btn';
+        btn.title = 'Copy code';
+        btn.innerHTML = COPY_SVG;
+        btn.addEventListener('click', async () => {
+          const text = pre.querySelector('code')?.textContent ?? pre.textContent ?? '';
+          try {
+            if (navigator.clipboard && window.isSecureContext) {
+              await navigator.clipboard.writeText(text);
+            } else {
+              const ta = document.createElement('textarea');
+              ta.value = text;
+              ta.style.cssText = 'position:fixed;left:-9999px';
+              document.body.appendChild(ta);
+              ta.select();
+              document.execCommand('copy');
+              document.body.removeChild(ta);
+            }
+            btn.innerHTML = CHECK_SVG;
+            btn.classList.add('copied');
+            setTimeout(() => { btn.innerHTML = COPY_SVG; btn.classList.remove('copied'); }, 2000);
+          } catch (e) {
+            console.error('copy failed', e);
+          }
+        });
+        pre.appendChild(btn);
+      });
+    }
+    attachButtons();
+    const mo = new MutationObserver(attachButtons);
+    mo.observe(node, { childList: true, subtree: true });
+    return { destroy: () => mo.disconnect() };
+  }
+</script>
+
+<div class="flex {role === 'user' ? 'justify-end' : 'justify-start'} mb-4">
+  <div
+    class="relative group rounded-lg px-4 py-2 {role === 'user'
+      ? 'max-w-[85%] bg-primary text-btn-primary-text'
+      : 'w-full sm:w-4/5 bg-surface border border-gray-200 dark:border-white/10'}"
+  >
+    {#if role === "assistant"}
+      {#if reasoning_content || isReasoning}
+        <div class="mb-3 border border-gray-200 dark:border-white/10 rounded overflow-hidden">
+          <button
+            class="w-full flex items-center gap-2 px-3 py-2 bg-gray-50 dark:bg-white/5 hover:bg-gray-100 dark:hover:bg-white/10 transition-colors text-sm"
+            onclick={() => showReasoning = !showReasoning}
+          >
+            {#if showReasoning}
+              <ChevronDown class="w-4 h-4" />
+            {:else}
+              <ChevronRight class="w-4 h-4" />
+            {/if}
+            <Brain class="w-4 h-4" />
+            <span class="font-medium">Reasoning</span>
+            <span class="text-txtsecondary ml-2">
+              ({reasoning_content.length} chars{#if !isReasoning && reasoningTimeMs > 0}, {formatDuration(reasoningTimeMs)}{/if})
+            </span>
+            {#if isReasoning}
+              <span class="ml-auto flex items-center gap-1 text-txtsecondary">
+                <span class="w-1.5 h-1.5 bg-primary rounded-full animate-pulse"></span>
+                reasoning...
+              </span>
+            {/if}
+          </button>
+          {#if showReasoning}
+            <div class="px-3 py-2 bg-gray-50/50 dark:bg-white/[0.02] text-sm text-txtsecondary whitespace-pre-wrap font-mono">
+              {reasoning_content}{#if isReasoning}<span class="inline-block w-1.5 h-4 bg-current animate-pulse ml-0.5"></span>{/if}
+            </div>
+          {/if}
+        </div>
+      {/if}
+      {#if hasImages}
+        <div class="mb-3 flex flex-wrap gap-2">
+          {#each imageUrls as imageUrl, idx (idx)}
+            <button
+              onclick={() => openModal(imageUrl)}
+              class="cursor-pointer rounded border border-gray-200 dark:border-white/10 hover:opacity-80 transition-opacity"
+            >
+              <img
+                src={imageUrl}
+                alt="Image {idx + 1}"
+                class="max-h-64 rounded"
+              />
+            </button>
+          {/each}
+        </div>
+      {/if}
+      {#if showRaw}
+        <div class="whitespace-pre-wrap font-mono text-sm">{textContent}</div>
+      {:else}
+        <div class="prose prose-sm dark:prose-invert max-w-none" use:codeBlockCopy>
+          {#each renderedParts.blocks as block (block.id)}
+            {@html block.html}
+          {/each}
+          {@html renderedParts.pendingHtml}
+          {#if isStreaming && !isReasoning}
+            <span class="inline-block w-2 h-4 bg-current animate-pulse ml-0.5"></span>
+          {/if}
+        </div>
+      {/if}
+      {#if !isStreaming}
+        <div class="flex gap-1 mt-2 pt-1 border-t border-gray-200 dark:border-white/10">
+          {#if onRegenerate}
+            <button
+              class="p-1 rounded hover:bg-black/10 dark:hover:bg-white/10 text-txtsecondary"
+              onclick={onRegenerate}
+              title="Regenerate response"
+            >
+              <RefreshCw class="w-4 h-4" />
+            </button>
+          {/if}
+          <button
+            class="p-1 rounded hover:bg-black/10 dark:hover:bg-white/10 text-txtsecondary"
+            onclick={copyToClipboard}
+            title={copied ? "Copied!" : "Copy to clipboard"}
+          >
+            {#if copied}
+              <Check class="w-4 h-4 text-green-500" />
+            {:else}
+              <Copy class="w-4 h-4" />
+            {/if}
+          </button>
+          <button
+            class="p-1 rounded hover:bg-black/10 dark:hover:bg-white/10 {showRaw ? 'text-primary' : 'text-txtsecondary'}"
+            onclick={() => showRaw = !showRaw}
+            title={showRaw ? "Show rendered" : "Show raw"}
+          >
+            <Code class="w-4 h-4" />
+          </button>
+        </div>
+      {/if}
+    {:else}
+      {#if isEditing}
+        <div class="flex flex-col gap-2 min-w-[300px]">
+          <textarea
+            class="w-full px-3 py-2 rounded border border-gray-200 dark:border-white/10 bg-surface text-txtmain focus:outline-none focus:ring-2 focus:ring-primary resize-none"
+            rows="3"
+            bind:value={editContent}
+            onkeydown={handleKeyDown}
+          ></textarea>
+          <div class="flex justify-end gap-2">
+            <button
+              class="p-1.5 rounded hover:bg-white/20"
+              onclick={cancelEdit}
+              title="Cancel"
+            >
+              <X class="w-4 h-4" />
+            </button>
+            <button
+              class="p-1.5 rounded hover:bg-white/20"
+              onclick={saveEdit}
+              title="Save"
+            >
+              <Save class="w-4 h-4" />
+            </button>
+          </div>
+        </div>
+      {:else}
+        {#if hasImages}
+          <div class="mb-2 flex flex-wrap gap-2">
+            {#each imageUrls as imageUrl, idx (idx)}
+              <button
+                onclick={() => openModal(imageUrl)}
+                class="cursor-pointer rounded border border-white/20 hover:opacity-80 transition-opacity"
+              >
+                <img
+                  src={imageUrl}
+                  alt="Image {idx + 1}"
+                  class="max-w-[200px] rounded"
+                />
+              </button>
+            {/each}
+          </div>
+        {/if}
+        <div class="whitespace-pre-wrap pr-8">{textContent}</div>
+        {#if canEdit}
+          <button
+            class="absolute top-2 right-2 p-1.5 rounded-lg opacity-0 group-hover:opacity-100 transition-opacity bg-white/20 hover:bg-white/30 shadow-sm"
+            onclick={startEdit}
+            title="Edit message"
+          >
+            <Pencil class="w-4 h-4" />
+          </button>
+        {/if}
+      {/if}
+    {/if}
+  </div>
+</div>
+
+<!-- Full-size image modal -->
+{#if modalImageUrl}
+  <div
+    class="fixed inset-0 z-50 flex items-center justify-center bg-black/80 p-4"
+    onclick={(e) => closeModal(e)}
+    onkeydown={handleModalKeyDown}
+    role="button"
+    tabindex="-1"
+  >
+    <button
+      class="absolute top-4 right-4 p-2 rounded-lg bg-white/10 hover:bg-white/20 text-white transition-colors"
+      onclick={() => closeModal()}
+      title="Close"
+    >
+      <X class="w-6 h-6" />
+    </button>
+    <img
+      src={modalImageUrl}
+      alt=""
+      class="max-w-full max-h-full rounded pointer-events-none"
+    />
+  </div>
+{/if}
+
+<style>
+  .prose :global(pre) {
+    position: relative;
+    background-color: var(--color-surface);
+    border: 1px solid var(--color-border, rgba(128, 128, 128, 0.2));
+    border-radius: 0.375rem;
+    padding: 0.75rem;
+    padding-right: 2.5rem;
+    overflow-x: auto;
+    margin: 0.5rem 0;
+  }
+
+  .prose :global(.code-copy-btn) {
+    position: absolute;
+    top: 0.375rem;
+    right: 0.375rem;
+    display: flex;
+    align-items: center;
+    justify-content: center;
+    padding: 0.25rem;
+    border-radius: 0.25rem;
+    border: 1px solid var(--color-border);
+    background: var(--color-surface);
+    color: var(--color-txtsecondary);
+    cursor: pointer;
+    transition: background-color 0.15s;
+    line-height: 0;
+  }
+
+  .prose :global(.code-copy-btn:hover) {
+    background: var(--color-secondary);
+  }
+
+  .prose :global(.code-copy-btn.copied) {
+    color: var(--color-success);
+    opacity: 1;
+  }
+
+  .prose :global(code) {
+    font-family: ui-monospace, SFMono-Regular, Menlo, Monaco, Consolas, monospace;
+    font-size: 0.875em;
+  }
+
+  .prose :global(pre code) {
+    background: none;
+    padding: 0;
+  }
+
+  .prose :global(code:not(pre code)) {
+    background-color: var(--color-surface);
+    padding: 0.125rem 0.25rem;
+    border-radius: 0.25rem;
+    border: 1px solid var(--color-border, rgba(128, 128, 128, 0.2));
+  }
+
+  .prose :global(p) {
+    margin: 0.5rem 0;
+  }
+
+  .prose :global(p:first-child) {
+    margin-top: 0;
+  }
+
+  .prose :global(p:last-child) {
+    margin-bottom: 0;
+  }
+
+  .prose :global(ul),
+  .prose :global(ol) {
+    margin: 0.5rem 0;
+    padding-left: 1.5rem;
+  }
+
+  .prose :global(li) {
+    margin: 0.25rem 0;
+  }
+
+  .prose :global(h1),
+  .prose :global(h2),
+  .prose :global(h3),
+  .prose :global(h4) {
+    margin: 1rem 0 0.5rem 0;
+    font-weight: 600;
+  }
+
+  .prose :global(h1:first-child),
+  .prose :global(h2:first-child),
+  .prose :global(h3:first-child),
+  .prose :global(h4:first-child) {
+    margin-top: 0;
+  }
+
+  .prose :global(blockquote) {
+    border-left: 3px solid var(--color-primary);
+    padding-left: 1rem;
+    margin: 0.5rem 0;
+    font-style: italic;
+  }
+
+  .prose :global(a) {
+    color: var(--color-primary);
+    text-decoration: underline;
+  }
+
+  .prose :global(table) {
+    width: 100%;
+    border-collapse: collapse;
+    margin: 0.5rem 0;
+  }
+
+  .prose :global(th),
+  .prose :global(td) {
+    border: 1px solid var(--color-border, rgba(128, 128, 128, 0.2));
+    padding: 0.5rem;
+    text-align: left;
+  }
+
+  .prose :global(th) {
+    background-color: var(--color-surface);
+    font-weight: 600;
+  }
+
+  /* Highlight.js theme overrides for dark mode */
+  :global(.dark) .prose :global(.hljs) {
+    background: transparent;
+  }
+</style>

ui-svelte/src/components/playground/ExpandableTextarea.svelte

@@ -0,0 +1,121 @@
+<script lang="ts">
+  import { untrack } from "svelte";
+  import { Maximize2, X } from "lucide-svelte";
+
+  interface Props {
+    value: string;
+    placeholder?: string;
+    rows?: number;
+    disabled?: boolean;
+    onkeydown?: (event: KeyboardEvent) => void;
+  }
+
+  let {
+    value = $bindable(),
+    placeholder = "",
+    rows = 3,
+    disabled = false,
+    onkeydown,
+  }: Props = $props();
+
+  let isExpanded = $state(false);
+  let expandedValue = $state("");
+  let expandedTextarea: HTMLTextAreaElement | undefined = $state();
+
+  function openExpanded() {
+    expandedValue = value;
+    isExpanded = true;
+  }
+
+  function closeExpanded() {
+    isExpanded = false;
+  }
+
+  function saveExpanded() {
+    value = expandedValue;
+    isExpanded = false;
+  }
+
+  function handleKeyDown(event: KeyboardEvent) {
+    if (event.key === "Escape") {
+      closeExpanded();
+    }
+  }
+
+  // Focus the textarea when expanded view opens
+  $effect(() => {
+    if (isExpanded && expandedTextarea) {
+      expandedTextarea.focus();
+      const len = untrack(() => expandedValue.length);
+      expandedTextarea.setSelectionRange(len, len);
+    }
+  });
+</script>
+
+<div class="flex-1 relative group flex items-stretch min-h-0">
+  <textarea
+    class="w-full px-3 py-2 pr-10 rounded border border-gray-200 dark:border-white/10 bg-surface focus:outline-none focus:ring-2 focus:ring-inset focus:ring-primary resize-none"
+    {placeholder}
+    {rows}
+    bind:value
+    {onkeydown}
+    {disabled}
+  ></textarea>
+  <button
+    class="absolute top-2 right-2 p-1.5 rounded-lg opacity-60 md:opacity-0 group-hover:opacity-100 transition-opacity bg-surface/90 hover:bg-surface border border-gray-200 dark:border-white/10 shadow-sm"
+    onclick={openExpanded}
+    title="Expand to edit"
+    type="button"
+    {disabled}
+  >
+    <Maximize2 class="w-4 h-4" />
+  </button>
+</div>
+
+{#if isExpanded}
+  <div class="fixed inset-0 z-50 flex items-center justify-center bg-black/50 p-4">
+    <div class="w-full max-w-4xl h-[80vh] flex flex-col bg-surface rounded-lg shadow-xl border border-gray-200 dark:border-white/10">
+      <!-- Header -->
+      <div class="flex justify-between items-center p-4 border-b border-gray-200 dark:border-white/10">
+        <h3 class="font-medium">Edit Text</h3>
+        <button
+          class="p-1.5 rounded-lg hover:bg-gray-100 dark:hover:bg-white/10"
+          onclick={closeExpanded}
+          title="Close"
+          type="button"
+        >
+          <X class="w-5 h-5" />
+        </button>
+      </div>
+
+      <!-- Textarea -->
+      <div class="flex-1 p-4">
+        <textarea
+          bind:this={expandedTextarea}
+          class="w-full h-full px-4 py-3 rounded border border-gray-200 dark:border-white/10 bg-card focus:outline-none focus:ring-2 focus:ring-primary resize-none"
+          placeholder={placeholder}
+          bind:value={expandedValue}
+          onkeydown={handleKeyDown}
+        ></textarea>
+      </div>
+
+      <!-- Footer -->
+      <div class="flex justify-end gap-2 p-4 border-t border-gray-200 dark:border-white/10">
+        <button
+          class="btn"
+          onclick={closeExpanded}
+          type="button"
+        >
+          Cancel
+        </button>
+        <button
+          class="btn bg-primary text-btn-primary-text hover:opacity-90"
+          onclick={saveExpanded}
+          type="button"
+        >
+          Done
+        </button>
+      </div>
+    </div>
+  </div>
+{/if}

ui-svelte/src/components/playground/ImageInterface.svelte

@@ -0,0 +1,521 @@
+<script lang="ts">
+  import { models } from "../../stores/api";
+  import { persistentStore } from "../../stores/persistent";
+  import { generateImage } from "../../lib/imageApi";
+  import { generateSdImage, fetchSdLoras } from "../../lib/sdApi";
+  import { playgroundStores } from "../../stores/playgroundActivity";
+  import ModelSelector from "./ModelSelector.svelte";
+  import ExpandableTextarea from "./ExpandableTextarea.svelte";
+  import type { ImageApiMode, SdApiLora, SdApiLoraRef } from "../../lib/types";
+
+  const selectedModelStore = persistentStore<string>("playground-image-model", "");
+  const selectedSizeStore = persistentStore<string>("playground-image-size", "1024x1024");
+  const apiModeStore = persistentStore<ImageApiMode>("playground-image-api-mode", "openai");
+
+  // SDAPI persistent settings
+  const sdNegativePromptStore = persistentStore<string>("playground-sdapi-negative-prompt", "");
+  const sdStepsStore = persistentStore<number>("playground-sdapi-steps", 20);
+  const sdCfgScaleStore = persistentStore<number>("playground-sdapi-cfg-scale", 7);
+  const sdSeedStore = persistentStore<number>("playground-sdapi-seed", -1);
+  const sdSamplerStore = persistentStore<string>("playground-sdapi-sampler", "");
+  const sdSchedulerStore = persistentStore<string>("playground-sdapi-scheduler", "");
+  const sdBatchSizeStore = persistentStore<number>("playground-sdapi-batch-size", 1);
+
+  let prompt = $state("");
+  let isGenerating = $state(false);
+  let generatedImages = $state<string[]>([]);
+  let error = $state<string | null>(null);
+  let abortController = $state<AbortController | null>(null);
+  let showFullscreen = $state(false);
+  let fullscreenIndex = $state(0);
+  let showSettings = $state(false);
+
+  // SDAPI lora state
+  let availableLoras = $state<SdApiLora[]>([]);
+  let selectedLoras = $state<SdApiLoraRef[]>([]);
+  let isLoadingLoras = $state(false);
+  let lorasLoaded = $state(false);
+  let loraError = $state<string | null>(null);
+
+  let hasModels = $derived($models.some((m) => !m.unlisted));
+  let isSdapi = $derived($apiModeStore === "sdapi");
+
+  $effect(() => {
+    playgroundStores.imageGenerating.set(isGenerating);
+  });
+
+  async function loadLoras() {
+    if (!$selectedModelStore || isLoadingLoras) return;
+    isLoadingLoras = true;
+    loraError = null;
+    try {
+      const loras = await fetchSdLoras($selectedModelStore);
+      availableLoras = loras;
+      lorasLoaded = true;
+    } catch (err) {
+      availableLoras = [];
+      loraError = err instanceof Error ? err.message : "Failed to load LoRAs";
+      lorasLoaded = false;
+    } finally {
+      isLoadingLoras = false;
+    }
+  }
+
+  function addLora(event: Event) {
+    const select = event.target as HTMLSelectElement;
+    const path = select.value;
+    if (!path) return;
+
+    const lora = availableLoras.find((l) => l.path === path);
+    if (lora && !selectedLoras.some((l) => l.path === path)) {
+      selectedLoras = [...selectedLoras, { path: lora.path, multiplier: 1.0 }];
+    }
+    select.value = "";
+  }
+
+  function removeLora(path: string) {
+    selectedLoras = selectedLoras.filter((l) => l.path !== path);
+  }
+
+  function updateLoraMultiplier(path: string, multiplier: number) {
+    selectedLoras = selectedLoras.map((l) =>
+      l.path === path ? { ...l, multiplier } : l
+    );
+  }
+
+  function getLoraName(path: string): string {
+    return availableLoras.find((l) => l.path === path)?.name ?? path;
+  }
+
+  async function generate() {
+    const trimmedPrompt = prompt.trim();
+    if (!trimmedPrompt || !$selectedModelStore || isGenerating) return;
+
+    isGenerating = true;
+    error = null;
+    abortController = new AbortController();
+
+    try {
+      if (isSdapi) {
+        const [w, h] = $selectedSizeStore.split("x").map(Number);
+        const request = {
+          model: $selectedModelStore,
+          prompt: trimmedPrompt,
+          negative_prompt: $sdNegativePromptStore || undefined,
+          width: w,
+          height: h,
+          steps: $sdStepsStore,
+          cfg_scale: $sdCfgScaleStore,
+          seed: $sdSeedStore,
+          batch_size: $sdBatchSizeStore,
+          sampler_name: $sdSamplerStore || undefined,
+          scheduler: $sdSchedulerStore || undefined,
+          lora: selectedLoras.length > 0 ? selectedLoras : undefined,
+        };
+
+        const response = await generateSdImage(request, abortController.signal);
+        if (response.images && response.images.length > 0) {
+          generatedImages = response.images.map(
+            (img) => `data:image/png;base64,${img}`
+          );
+        }
+      } else {
+        const response = await generateImage(
+          $selectedModelStore,
+          trimmedPrompt,
+          $selectedSizeStore,
+          abortController.signal
+        );
+
+        if (response.data && response.data.length > 0) {
+          const imageData = response.data[0];
+          if (imageData.b64_json) {
+            generatedImages = [`data:image/png;base64,${imageData.b64_json}`];
+          } else if (imageData.url) {
+            generatedImages = [imageData.url];
+          }
+        }
+      }
+    } catch (err) {
+      if (err instanceof Error && err.name === "AbortError") {
+        // User cancelled
+      } else {
+        error = err instanceof Error ? err.message : "An error occurred";
+      }
+    } finally {
+      isGenerating = false;
+      abortController = null;
+    }
+  }
+
+  function cancelGeneration() {
+    abortController?.abort();
+  }
+
+  function clearImage() {
+    generatedImages = [];
+    error = null;
+    prompt = "";
+  }
+
+  function downloadImage(index: number = 0) {
+    const img = generatedImages[index];
+    if (!img) return;
+
+    const link = document.createElement("a");
+    link.href = img;
+    link.download = `generated-image-${Date.now()}-${index}.png`;
+    document.body.appendChild(link);
+    link.click();
+    document.body.removeChild(link);
+  }
+
+  function openFullscreen(index: number = 0) {
+    fullscreenIndex = index;
+    showFullscreen = true;
+  }
+
+  function closeFullscreen(event?: MouseEvent) {
+    if (event && event.target !== event.currentTarget) {
+      return;
+    }
+    showFullscreen = false;
+  }
+
+  function handleKeyDown(event: KeyboardEvent) {
+    if (event.key === "Enter" && !event.shiftKey) {
+      event.preventDefault();
+      generate();
+    }
+  }
+</script>
+
+<div class="flex flex-col h-full">
+  <!-- Model selector and mode toggle -->
+  <div class="shrink-0 flex flex-wrap gap-2 mb-4">
+    <ModelSelector bind:value={$selectedModelStore} placeholder="Select an image model..." disabled={isGenerating} />
+
+    <select
+      class="px-3 py-2 rounded border border-gray-200 dark:border-white/10 bg-surface focus:outline-none focus:ring-2 focus:ring-primary"
+      bind:value={$apiModeStore}
+      disabled={isGenerating}
+    >
+      <option value="openai">OpenAI</option>
+      <option value="sdapi">SDAPI</option>
+    </select>
+
+    <select
+      class="px-3 py-2 rounded border border-gray-200 dark:border-white/10 bg-surface focus:outline-none focus:ring-2 focus:ring-primary"
+      bind:value={$selectedSizeStore}
+      disabled={isGenerating}
+    >
+      <optgroup label="Square">
+        <option value="512x512">512x512</option>
+        <option value="1024x1024">1024x1024</option>
+      </optgroup>
+      <optgroup label="Landscape">
+        <option value="1024x768">1024x768 (4:3)</option>
+        <option value="1280x720">1280x720 (16:9)</option>
+        <option value="1792x1024">1792x1024 (SDXL)</option>
+      </optgroup>
+      <optgroup label="Portrait">
+        <option value="768x1024">768x1024 (3:4)</option>
+        <option value="720x1280">720x1280 (9:16)</option>
+        <option value="1024x1792">1024x1792 (SDXL)</option>
+      </optgroup>
+    </select>
+
+    {#if isSdapi}
+      <button
+        class="px-3 py-2 rounded border border-gray-200 dark:border-white/10 bg-surface hover:bg-secondary-hover transition-colors"
+        onclick={() => showSettings = !showSettings}
+      >
+        {showSettings ? "Hide Settings" : "Settings"}
+      </button>
+    {/if}
+  </div>
+
+  <!-- SDAPI Settings Panel -->
+  {#if isSdapi && showSettings}
+    <div class="shrink-0 mb-4 p-4 rounded border border-gray-200 dark:border-white/10 bg-surface">
+      <div class="grid grid-cols-2 md:grid-cols-4 gap-3 mb-3">
+        <label class="flex flex-col gap-1">
+          <span class="text-xs text-txtsecondary">Steps</span>
+          <input
+            type="number"
+            class="px-2 py-1 rounded border border-gray-200 dark:border-white/10 bg-surface focus:outline-none focus:ring-2 focus:ring-primary"
+            bind:value={$sdStepsStore}
+            min="1"
+            max="150"
+          />
+        </label>
+        <label class="flex flex-col gap-1">
+          <span class="text-xs text-txtsecondary">CFG Scale</span>
+          <input
+            type="number"
+            class="px-2 py-1 rounded border border-gray-200 dark:border-white/10 bg-surface focus:outline-none focus:ring-2 focus:ring-primary"
+            bind:value={$sdCfgScaleStore}
+            min="1"
+            max="30"
+            step="0.5"
+          />
+        </label>
+        <label class="flex flex-col gap-1">
+          <span class="text-xs text-txtsecondary">Seed (-1 = random)</span>
+          <input
+            type="number"
+            class="px-2 py-1 rounded border border-gray-200 dark:border-white/10 bg-surface focus:outline-none focus:ring-2 focus:ring-primary"
+            bind:value={$sdSeedStore}
+            min="-1"
+          />
+        </label>
+        <label class="flex flex-col gap-1">
+          <span class="text-xs text-txtsecondary">Batch Size</span>
+          <input
+            type="number"
+            class="px-2 py-1 rounded border border-gray-200 dark:border-white/10 bg-surface focus:outline-none focus:ring-2 focus:ring-primary"
+            bind:value={$sdBatchSizeStore}
+            min="1"
+            max="8"
+          />
+        </label>
+        <label class="flex flex-col gap-1">
+          <span class="text-xs text-txtsecondary">Sampler</span>
+          <select
+            class="px-2 py-1 rounded border border-gray-200 dark:border-white/10 bg-surface focus:outline-none focus:ring-2 focus:ring-primary"
+            bind:value={$sdSamplerStore}
+          >
+            <option value="">Default</option>
+            <option value="euler_a">euler_a</option>
+            <option value="euler">euler</option>
+            <option value="heun">heun</option>
+            <option value="dpm2">dpm2</option>
+            <option value="dpmpp2s_a">dpmpp2s_a</option>
+            <option value="dpmpp2m">dpmpp2m</option>
+            <option value="dpmpp2mv2">dpmpp2mv2</option>
+            <option value="ipndm">ipndm</option>
+            <option value="ipndm_v">ipndm_v</option>
+            <option value="lcm">lcm</option>
+            <option value="ddim_trailing">ddim_trailing</option>
+            <option value="tcd">tcd</option>
+          </select>
+        </label>
+        <label class="flex flex-col gap-1">
+          <span class="text-xs text-txtsecondary">Scheduler</span>
+          <select
+            class="px-2 py-1 rounded border border-gray-200 dark:border-white/10 bg-surface focus:outline-none focus:ring-2 focus:ring-primary"
+            bind:value={$sdSchedulerStore}
+          >
+            <option value="">Auto for model</option>
+            <option value="discrete">discrete</option>
+            <option value="karras">karras</option>
+            <option value="exponential">exponential</option>
+            <option value="ays">ays</option>
+            <option value="gits">gits</option>
+          </select>
+        </label>
+      </div>
+
+      <label class="flex flex-col gap-1 mb-3">
+        <span class="text-xs text-txtsecondary">Negative Prompt</span>
+        <textarea
+          class="px-2 py-1 rounded border border-gray-200 dark:border-white/10 bg-surface focus:outline-none focus:ring-2 focus:ring-primary resize-y text-sm"
+          bind:value={$sdNegativePromptStore}
+          rows="2"
+          placeholder="Elements to avoid..."
+        ></textarea>
+      </label>
+
+      <!-- LoRA Selection -->
+      <div>
+        <span class="text-xs text-txtsecondary block mb-1">LoRAs</span>
+        <div class="flex items-center gap-2 mb-2">
+          <button
+            class="px-3 py-1.5 text-sm rounded border border-gray-200 dark:border-white/10 bg-surface hover:bg-secondary-hover transition-colors disabled:opacity-50"
+            onclick={loadLoras}
+            disabled={!$selectedModelStore || isLoadingLoras}
+          >
+            {isLoadingLoras ? "Loading..." : lorasLoaded ? "Reload LoRAs" : "Load LoRAs"}
+          </button>
+          {#if lorasLoaded && availableLoras.length > 0}
+            <select
+              class="flex-1 px-2 py-1.5 text-sm rounded border border-gray-200 dark:border-white/10 bg-surface focus:outline-none focus:ring-2 focus:ring-primary"
+              onchange={addLora}
+            >
+              <option value="">Add a LoRA...</option>
+              {#each availableLoras.filter((l) => !selectedLoras.some((s) => s.path === l.path)) as lora}
+                <option value={lora.path}>{lora.name}</option>
+              {/each}
+            </select>
+          {/if}
+        </div>
+        {#if loraError}
+          <p class="text-xs text-red-500 mb-1">{loraError}</p>
+        {/if}
+        {#if lorasLoaded && availableLoras.length === 0}
+          <p class="text-xs text-txtsecondary">No LoRAs available</p>
+        {/if}
+        {#if selectedLoras.length > 0}
+          <div class="flex flex-col gap-1.5">
+            {#each selectedLoras as lora}
+              <div class="flex items-center gap-2 text-sm">
+                <span class="flex-1 truncate">{getLoraName(lora.path)}</span>
+                <input
+                  type="number"
+                  class="w-20 px-1.5 py-1 text-xs rounded border border-gray-200 dark:border-white/10 bg-surface focus:outline-none focus:ring-1 focus:ring-primary"
+                  value={lora.multiplier}
+                  oninput={(e) => updateLoraMultiplier(lora.path, parseFloat((e.target as HTMLInputElement).value) || 1)}
+                  min="0"
+                  max="2"
+                  step="0.1"
+                />
+                <button
+                  class="px-1.5 py-0.5 text-xs rounded border border-gray-200 dark:border-white/10 hover:bg-red-500 hover:text-white hover:border-red-500 transition-colors"
+                  onclick={() => removeLora(lora.path)}
+                  aria-label="Remove LoRA"
+                >
+                  x
+                </button>
+              </div>
+            {/each}
+          </div>
+        {/if}
+      </div>
+    </div>
+  {/if}
+
+  <!-- Empty state for no models configured -->
+  {#if !hasModels}
+    <div class="flex-1 flex items-center justify-center text-txtsecondary">
+      <p>No models configured. Add models to your configuration to generate images.</p>
+    </div>
+  {:else}
+    <!-- Image display area -->
+    <div class="flex-1 overflow-auto mb-4 flex items-center justify-center bg-surface border border-gray-200 dark:border-white/10 rounded">
+      {#if isGenerating}
+        <div class="text-center text-txtsecondary">
+          <div class="inline-block w-8 h-8 border-4 border-primary border-t-transparent rounded-full animate-spin mb-2"></div>
+          <p>Generating image...</p>
+        </div>
+      {:else if error}
+        <div class="text-center text-red-500 p-4">
+          <p class="font-medium">Error</p>
+          <p class="text-sm mt-1">{error}</p>
+        </div>
+      {:else if generatedImages.length > 1}
+        <!-- Grid for multiple images (batch) -->
+        <div class="grid grid-cols-2 gap-2 p-2 w-full h-full overflow-auto">
+          {#each generatedImages as img, i}
+            <div class="relative flex items-center justify-center">
+              <button
+                class="p-0 border-0 bg-transparent cursor-pointer"
+                onclick={() => openFullscreen(i)}
+                aria-label="View fullscreen"
+              >
+                <img
+                  src={img}
+                  alt="AI generated content {i + 1}"
+                  class="max-w-full max-h-full object-contain hover:opacity-90 transition-opacity"
+                />
+              </button>
+              <button
+                class="absolute bottom-2 right-2 p-1.5 bg-black/60 hover:bg-black/80 text-white rounded-full transition-colors"
+                onclick={(e) => { e.stopPropagation(); downloadImage(i); }}
+                aria-label="Download image"
+              >
+                <svg class="w-4 h-4" fill="none" stroke="currentColor" viewBox="0 0 24 24">
+                  <path stroke-linecap="round" stroke-linejoin="round" stroke-width="2" d="M4 16v1a3 3 0 003 3h10a3 3 0 003-3v-1m-4-4l-4 4m0 0l-4-4m4 4V4"></path>
+                </svg>
+              </button>
+            </div>
+          {/each}
+        </div>
+      {:else if generatedImages.length === 1}
+        <div class="relative max-w-full max-h-full flex items-center justify-center">
+          <button
+            class="p-0 border-0 bg-transparent cursor-pointer"
+            onclick={() => openFullscreen(0)}
+            aria-label="View fullscreen"
+          >
+            <img
+              src={generatedImages[0]}
+              alt="AI generated content"
+              class="max-w-full max-h-full object-contain hover:opacity-90 transition-opacity"
+            />
+          </button>
+          <button
+            class="absolute bottom-2 right-2 p-2 bg-black/60 hover:bg-black/80 text-white rounded-full transition-colors"
+            onclick={(e) => { e.stopPropagation(); downloadImage(0); }}
+            aria-label="Download image"
+          >
+            <svg class="w-5 h-5" fill="none" stroke="currentColor" viewBox="0 0 24 24">
+              <path stroke-linecap="round" stroke-linejoin="round" stroke-width="2" d="M4 16v1a3 3 0 003 3h10a3 3 0 003-3v-1m-4-4l-4 4m0 0l-4-4m4 4V4"></path>
+            </svg>
+          </button>
+        </div>
+      {:else}
+        <div class="text-center text-txtsecondary">
+          <p>Enter a prompt below to generate an image</p>
+        </div>
+      {/if}
+    </div>
+
+    <!-- Prompt input area -->
+    <div class="shrink-0 flex flex-col md:flex-row gap-2">
+      <ExpandableTextarea
+        bind:value={prompt}
+        placeholder="Describe the image you want to generate..."
+        rows={3}
+        onkeydown={handleKeyDown}
+        disabled={isGenerating || !$selectedModelStore}
+      />
+      <div class="flex flex-row md:flex-col gap-2">
+        {#if isGenerating}
+          <button class="btn bg-red-500 hover:bg-red-600 text-white flex-1 md:flex-none" onclick={cancelGeneration}>
+            Cancel
+          </button>
+        {:else}
+          <button
+            class="btn bg-primary text-btn-primary-text hover:opacity-90 flex-1 md:flex-none"
+            onclick={generate}
+            disabled={!prompt.trim() || !$selectedModelStore}
+          >
+            Generate
+          </button>
+          <button
+            class="btn flex-1 md:flex-none"
+            onclick={clearImage}
+            disabled={generatedImages.length === 0 && !error && !prompt.trim()}
+          >
+            Clear
+          </button>
+        {/if}
+      </div>
+    </div>
+  {/if}
+</div>
+
+<!-- Fullscreen dialog -->
+{#if showFullscreen && generatedImages[fullscreenIndex]}
+  <div
+    class="fixed inset-0 bg-black/90 z-50 flex items-center justify-center p-4"
+    onclick={(e) => closeFullscreen(e)}
+    onkeydown={(e) => e.key === 'Escape' && closeFullscreen()}
+    role="dialog"
+    aria-modal="true"
+    tabindex="-1"
+  >
+    <button
+      class="absolute top-4 right-4 text-white hover:text-gray-300 text-2xl w-10 h-10 flex items-center justify-center rounded-full hover:bg-white/10 transition-colors"
+      onclick={() => closeFullscreen()}
+      aria-label="Close fullscreen"
+    >
+      ×
+    </button>
+    <img
+      src={generatedImages[fullscreenIndex]}
+      alt="AI generated content"
+      class="max-w-full max-h-full object-contain pointer-events-none"
+    />
+  </div>
+{/if}

ui-svelte/src/components/playground/ModelSelector.svelte

@@ -0,0 +1,44 @@
+<script lang="ts">
+  import { models } from "../../stores/api";
+  import { groupModels } from "../../lib/modelUtils";
+
+  interface Props {
+    value: string;
+    placeholder?: string;
+    disabled?: boolean;
+  }
+
+  let { value = $bindable(), placeholder = "Select a model...", disabled = false }: Props = $props();
+
+  let grouped = $derived(groupModels($models));
+  let hasModels = $derived(grouped.local.length > 0 || Object.keys(grouped.peersByProvider).length > 0);
+</script>
+
+{#if hasModels}
+  <select
+    class="min-w-0 flex-1 basis-48 px-3 py-2 rounded border border-gray-200 dark:border-white/10 bg-surface focus:outline-none focus:ring-2 focus:ring-primary"
+    bind:value
+    {disabled}
+  >
+    <option value="">{placeholder}</option>
+    {#if grouped.local.length > 0}
+      <optgroup label="Local">
+        {#each grouped.local as model (model.id)}
+          <option value={model.id}>{model.id}</option>
+          {#if model.aliases}
+            {#each model.aliases as alias (alias)}
+              <option value={alias}>  ↳ {alias}</option>
+            {/each}
+          {/if}
+        {/each}
+      </optgroup>
+    {/if}
+    {#each Object.entries(grouped.peersByProvider).sort(([a], [b]) => a.localeCompare(b)) as [peerId, peerModels] (peerId)}
+      <optgroup label="Peer: {peerId}">
+        {#each peerModels as model (model.id)}
+          <option value={model.id}>{model.id}</option>
+        {/each}
+      </optgroup>
+    {/each}
+  </select>
+{/if}

ui-svelte/src/components/playground/PlaceholderTab.svelte

@@ -0,0 +1,14 @@
+<script lang="ts">
+  interface Props {
+    featureName: string;
+  }
+
+  let { featureName }: Props = $props();
+</script>
+
+<div class="flex items-center justify-center h-full">
+  <div class="text-center text-txtsecondary">
+    <p class="text-lg">{featureName}</p>
+    <p class="text-sm mt-2">To be implemented</p>
+  </div>
+</div>

ui-svelte/src/components/playground/RerankInterface.svelte

@@ -0,0 +1,406 @@
+<script lang="ts">
+  import { models } from "../../stores/api";
+  import { persistentStore } from "../../stores/persistent";
+  import { rerank } from "../../lib/rerankApi";
+  import { playgroundStores } from "../../stores/playgroundActivity";
+  import ModelSelector from "./ModelSelector.svelte";
+
+  type RerankRow = { doc: string; score: number | null };
+  type SortOrder = "none" | "asc" | "desc";
+  type EditorMode = "table" | "json";
+
+  const selectedModelStore = persistentStore<string>("playground-rerank-model", "");
+
+  const defaultQuery = "How do LLM's work?";
+  const defaultDocs = [
+    "Large language models (LLMs) use transformer architectures to predict the next token in a sequence based on massive amounts of text data.",
+    "LLMs are trained on diverse internet text, learning statistical patterns of language that allow them to generate coherent responses.",
+    "During training, LLMs minimize a loss function that measures the difference between predicted and actual tokens across billions of examples.",
+    "Attention mechanisms in transformers enable LLMs to weigh the importance of different words when generating output.",
+    "Fine\u2011tuning allows a pre\u2011trained LLM to adapt to a specific downstream task with a smaller dataset.",
+    "Neural networks consist of layers of interconnected neurons that adjust their weights during back\u2011propagation.",
+    "The history of the Roman Empire spanned over a thousand years.",
+    "Soccer is the most popular sport in many countries around the world.",
+    "Quantum computing uses qubits to perform calculations that are intractable for classical computers.",
+  ];
+
+  let query = $state(defaultQuery);
+  let rows = $state<RerankRow[]>([
+    ...defaultDocs.map((doc) => ({ doc, score: null })),
+    { doc: "", score: null },
+  ]);
+  let isLoading = $state(false);
+  let error = $state<string | null>(null);
+  let usage = $state<{ prompt_tokens: number; total_tokens: number } | null>(null);
+  let abortController: AbortController | null = null;
+  let sortOrder = $state<SortOrder>("desc");
+  let editorMode = $state<EditorMode>("table");
+  let jsonText = $state("");
+  let jsonError = $state<string | null>(null);
+
+  let hasModels = $derived($models.some((m) => !m.unlisted));
+
+  let canSubmit = $derived((() => {
+    if (!$selectedModelStore || isLoading) return false;
+    if (editorMode === "json") {
+      try {
+        const parsed = JSON.parse(jsonText) as Record<string, unknown>;
+        return (
+          typeof parsed.query === "string" &&
+          parsed.query.trim() !== "" &&
+          Array.isArray(parsed.documents) &&
+          (parsed.documents as unknown[]).some(
+            (d) => typeof d === "string" && (d as string).trim() !== ""
+          )
+        );
+      } catch {
+        return false;
+      }
+    }
+    return query.trim() !== "" && rows.some((r) => r.doc.trim() !== "");
+  })());
+
+  // Display rows with sort applied (display-only transform, rows[] is never mutated by sorting)
+  let displayRows = $derived((() => {
+    const indexed = rows.map((row, i) => ({ row, i }));
+    if (sortOrder === "none") return indexed;
+    return [...indexed].sort((a, b) => {
+      if (a.row.score === null && b.row.score === null) return 0;
+      if (a.row.score === null) return 1;
+      if (b.row.score === null) return -1;
+      return sortOrder === "desc"
+        ? b.row.score - a.row.score
+        : a.row.score - b.row.score;
+    });
+  })());
+
+  // Auto-add a new empty row when the last row gets content (table mode only)
+  $effect(() => {
+    if (editorMode === "table" && rows[rows.length - 1]?.doc.trim() !== "") {
+      rows = [...rows, { doc: "", score: null }];
+    }
+  });
+
+  // Sync loading state to activity store
+  $effect(() => {
+    playgroundStores.rerankLoading.set(isLoading);
+  });
+
+  function switchToJson() {
+    if (editorMode === "json") return;
+    const docs = rows.filter((r) => r.doc.trim() !== "").map((r) => r.doc);
+    jsonText = JSON.stringify({ query, documents: docs }, null, 2);
+    jsonError = null;
+    editorMode = "json";
+  }
+
+  function switchToTable() {
+    if (editorMode === "table") return;
+    if (jsonText.trim() === "") {
+      query = "";
+      rows = [{ doc: "", score: null }];
+      jsonError = null;
+      editorMode = "table";
+      return;
+    }
+    try {
+      const parsed = JSON.parse(jsonText) as unknown;
+      if (typeof parsed !== "object" || parsed === null || Array.isArray(parsed)) {
+        throw new Error("Expected a JSON object");
+      }
+      const obj = parsed as Record<string, unknown>;
+      if (typeof obj.query !== "string") throw new Error('"query" must be a string');
+      if (!Array.isArray(obj.documents)) throw new Error('"documents" must be an array');
+      query = obj.query;
+      const newRows: RerankRow[] = (obj.documents as unknown[]).map((d) => ({
+        doc: typeof d === "string" ? d : String(d),
+        score: null,
+      }));
+      if (newRows.length === 0 || newRows[newRows.length - 1].doc.trim() !== "") {
+        newRows.push({ doc: "", score: null });
+      }
+      rows = newRows;
+      jsonError = null;
+      editorMode = "table";
+    } catch (err) {
+      jsonError = err instanceof Error ? err.message : "Invalid JSON";
+    }
+  }
+
+  function cycleSortOrder() {
+    sortOrder = sortOrder === "none" ? "desc" : sortOrder === "desc" ? "asc" : "none";
+  }
+
+  function sortIndicator(): string {
+    if (sortOrder === "desc") return " ↓";
+    if (sortOrder === "asc") return " ↑";
+    return "";
+  }
+
+  async function submit() {
+    if (!canSubmit) return;
+
+    let submitQuery: string;
+    let nonEmptyEntries: { originalIndex: number; doc: string }[];
+
+    if (editorMode === "json") {
+      // Parse JSON, sync state to table, then submit
+      try {
+        const parsed = JSON.parse(jsonText) as Record<string, unknown>;
+        submitQuery = parsed.query as string;
+        const docs = (parsed.documents as string[]).filter((d) => d.trim() !== "");
+        const newRows: RerankRow[] = docs.map((d) => ({ doc: d, score: null }));
+        newRows.push({ doc: "", score: null });
+        rows = newRows;
+        query = submitQuery;
+        editorMode = "table";
+      } catch {
+        error = "Invalid JSON — fix before submitting";
+        return;
+      }
+      nonEmptyEntries = rows
+        .map((r, i) => ({ originalIndex: i, doc: r.doc }))
+        .filter((e) => e.doc.trim() !== "");
+    } else {
+      submitQuery = query;
+      nonEmptyEntries = rows
+        .map((r, i) => ({ originalIndex: i, doc: r.doc }))
+        .filter((e) => e.doc.trim() !== "");
+    }
+
+    isLoading = true;
+    error = null;
+    usage = null;
+
+    // Clear previous scores
+    rows = rows.map((r) => ({ ...r, score: null }));
+
+    abortController = new AbortController();
+
+    try {
+      const response = await rerank(
+        $selectedModelStore,
+        submitQuery,
+        nonEmptyEntries.map((e) => e.doc),
+        abortController.signal
+      );
+
+      usage = response.usage;
+
+      // Map result.index (position in submitted docs array) back to original rows[] index
+      const updated = rows.map((r) => ({ ...r }));
+      for (const result of response.results) {
+        const entry = nonEmptyEntries[result.index];
+        if (entry !== undefined) {
+          updated[entry.originalIndex].score = result.relevance_score;
+        }
+      }
+      rows = updated;
+    } catch (err) {
+      if (err instanceof Error && err.name === "AbortError") {
+        // User cancelled
+      } else {
+        error = err instanceof Error ? err.message : "An error occurred";
+      }
+    } finally {
+      isLoading = false;
+      abortController = null;
+    }
+  }
+
+  function cancel() {
+    abortController?.abort();
+  }
+
+  function clear() {
+    query = defaultQuery;
+    rows = [...defaultDocs.map((doc) => ({ doc, score: null })), { doc: "", score: null }];
+    error = null;
+    usage = null;
+    sortOrder = "desc";
+    jsonText = "";
+    jsonError = null;
+  }
+
+  function deleteRow(originalIndex: number) {
+    if (rows.length <= 1) return;
+    rows = rows.filter((_, i) => i !== originalIndex);
+  }
+
+  function updateDoc(originalIndex: number, value: string) {
+    const updated = rows.map((r) => ({ ...r }));
+    updated[originalIndex].doc = value;
+    rows = updated;
+  }
+
+  function scoreColor(score: number | null): string {
+    if (score === null) return "text-txtsecondary";
+    if (score > 0) return "text-green-600 dark:text-green-400";
+    return "text-red-500 dark:text-red-400";
+  }
+
+  function formatScore(score: number | null): string {
+    if (score === null) return "—";
+    return score.toFixed(3);
+  }
+
+  function handleKeyDown(e: KeyboardEvent) {
+    if (e.key === "Enter" && !e.shiftKey) {
+      e.preventDefault();
+      submit();
+    }
+  }
+
+  let isCleared = $derived(
+    query === defaultQuery &&
+    rows.every((r, i) => r.score === null && r.doc === (defaultDocs[i] ?? "")) &&
+    rows.length === defaultDocs.length + 1 &&
+    !jsonText.trim() &&
+    !error &&
+    !usage
+  );
+</script>
+
+<div class="flex flex-col h-full">
+  <!-- Top bar: model selector + query input (table mode) + mode toggle -->
+  <div class="shrink-0 flex flex-wrap gap-2 mb-4">
+    <ModelSelector bind:value={$selectedModelStore} placeholder="Select a rerank model..." disabled={isLoading} />
+    {#if editorMode === "table"}
+      <input
+        type="text"
+        class="min-w-0 flex-1 basis-48 px-3 py-2 rounded border border-gray-200 dark:border-white/10 bg-surface focus:outline-none focus:ring-2 focus:ring-primary"
+        placeholder="Query..."
+        bind:value={query}
+        disabled={isLoading}
+        onkeydown={handleKeyDown}
+      />
+    {/if}
+    <!-- Table / JSON toggle -->
+    <div class="flex rounded border border-gray-200 dark:border-white/10 overflow-hidden shrink-0">
+      <button
+        class="px-3 py-1.5 text-sm transition-colors {editorMode === 'table'
+          ? 'bg-primary text-btn-primary-text'
+          : 'bg-surface hover:bg-secondary-hover'}"
+        onclick={switchToTable}
+        disabled={isLoading}
+      >
+        Table
+      </button>
+      <button
+        class="px-3 py-1.5 text-sm border-l border-gray-200 dark:border-white/10 transition-colors {editorMode === 'json'
+          ? 'bg-primary text-btn-primary-text'
+          : 'bg-surface hover:bg-secondary-hover'}"
+        onclick={switchToJson}
+        disabled={isLoading}
+      >
+        JSON
+      </button>
+    </div>
+  </div>
+
+  {#if !hasModels}
+    <div class="flex-1 flex items-center justify-center text-txtsecondary">
+      <p>No models configured. Add models to your configuration to use reranking.</p>
+    </div>
+  {:else if editorMode === "json"}
+    <!-- JSON editor -->
+    <div class="flex-1 flex flex-col min-h-0 mb-4">
+      <textarea
+        class="flex-1 w-full font-mono text-sm px-3 py-2 rounded border border-gray-200 dark:border-white/10 bg-surface focus:outline-none focus:ring-2 focus:ring-primary resize-none"
+        bind:value={jsonText}
+        disabled={isLoading}
+        placeholder={'{\n  "query": "your search query",\n  "documents": [\n    "document one",\n    "document two"\n  ]\n}'}
+        spellcheck={false}
+      ></textarea>
+      {#if jsonError}
+        <p class="mt-1 text-sm text-red-500">{jsonError}</p>
+      {/if}
+    </div>
+  {:else}
+    <!-- Document table -->
+    <div class="flex-1 overflow-y-auto mb-4 border border-gray-200 dark:border-white/10 rounded">
+      <table class="w-full border-collapse table-fixed">
+        <colgroup>
+          <col class="w-auto" />
+          <col style="width: 120px" />
+          <col style="width: 40px" />
+        </colgroup>
+        <thead class="sticky top-0 bg-surface border-b border-gray-200 dark:border-white/10">
+          <tr>
+            <th class="px-3 py-2 text-left text-sm font-medium text-txtsecondary">Document</th>
+            <th
+              class="px-3 py-2 text-right text-sm font-medium text-txtsecondary cursor-pointer select-none hover:text-txtprimary transition-colors"
+              onclick={cycleSortOrder}
+            >
+              Score{sortIndicator()}
+            </th>
+            <th class="px-2 py-2"></th>
+          </tr>
+        </thead>
+        <tbody>
+          {#each displayRows as { row, i } (i)}
+            <tr class="border-b border-gray-100 dark:border-white/5 last:border-0">
+              <td class="px-3 py-1.5">
+                <input
+                  type="text"
+                  class="w-full bg-transparent focus:outline-none focus:ring-1 focus:ring-primary rounded px-1 py-0.5"
+                  placeholder={i === rows.length - 1 ? "Add document..." : "Document text..."}
+                  value={row.doc}
+                  oninput={(e) => updateDoc(i, (e.target as HTMLInputElement).value)}
+                  disabled={isLoading}
+                  onkeydown={handleKeyDown}
+                />
+              </td>
+              <td class="px-3 py-1.5 text-right font-mono text-sm {scoreColor(row.score)}">
+                {#if isLoading && row.score === null && row.doc.trim() !== ""}
+                  <span class="inline-block w-4 h-4 border-2 border-current border-t-transparent rounded-full animate-spin align-middle"></span>
+                {:else}
+                  {formatScore(row.score)}
+                {/if}
+              </td>
+              <td class="px-2 py-1.5 text-center">
+                <button
+                  class="w-7 h-7 flex items-center justify-center text-txtsecondary hover:text-red-500 transition-colors rounded disabled:opacity-30 disabled:cursor-not-allowed"
+                  onclick={() => deleteRow(i)}
+                  disabled={rows.length <= 1}
+                  tabindex="-1"
+                  aria-label="Remove row"
+                >
+                  ×
+                </button>
+              </td>
+            </tr>
+          {/each}
+        </tbody>
+      </table>
+    </div>
+  {/if}
+
+  <!-- Bottom toolbar -->
+  {#if hasModels}
+    <div class="shrink-0 flex flex-wrap items-center gap-2">
+      {#if isLoading}
+        <button class="btn bg-red-500 hover:bg-red-600 text-white" onclick={cancel}>
+          Cancel
+        </button>
+      {:else}
+        <button
+          class="btn bg-primary text-btn-primary-text hover:opacity-90"
+          onclick={submit}
+          disabled={!canSubmit}
+        >
+          Rerank
+        </button>
+        <button class="btn" onclick={clear} disabled={isCleared}>
+          Clear
+        </button>
+      {/if}
+
+      {#if error}
+        <span class="text-sm text-red-500 ml-2">{error}</span>
+      {:else if usage}
+        <span class="text-sm text-txtsecondary ml-2">{usage.total_tokens} tokens</span>
+      {/if}
+    </div>
+  {/if}
+</div>

ui-svelte/src/components/playground/SpeechInterface.svelte

@@ -0,0 +1,360 @@
+<script lang="ts">
+  import { models } from "../../stores/api";
+  import { persistentStore } from "../../stores/persistent";
+  import { generateSpeech } from "../../lib/speechApi";
+  import { playgroundStores } from "../../stores/playgroundActivity";
+  import ModelSelector from "./ModelSelector.svelte";
+  import ExpandableTextarea from "./ExpandableTextarea.svelte";
+
+  const selectedModelStore = persistentStore<string>("playground-speech-model", "");
+  const selectedVoiceStore = persistentStore<string>("playground-speech-voice", "coral");
+  const autoPlayStore = persistentStore<boolean>("playground-speech-autoplay", false);
+
+  let inputText = $state("");
+  let isGenerating = $state(false);
+  let generatedAudioUrl = $state<string | null>(null);
+  let generatedVoice = $state<string | null>(null);
+  let generatedTimestamp = $state<Date | null>(null);
+  let error = $state<string | null>(null);
+  let abortController = $state<AbortController | null>(null);
+  let audioElement = $state<HTMLAudioElement | null>(null);
+  let availableVoices = $state<string[]>(["coral", "alloy", "echo", "fable", "onyx", "nova", "shimmer"]);
+  let isLoadingVoices = $state(false);
+
+  const defaultVoices = ["coral", "alloy", "echo", "fable", "onyx", "nova", "shimmer"];
+  const CACHE_KEY = "playground-speech-voices-cache";
+
+  function getVoicesCache(): Record<string, string[]> {
+    if (typeof window === "undefined") return {};
+    try {
+      const saved = localStorage.getItem(CACHE_KEY);
+      return saved ? JSON.parse(saved) : {};
+    } catch {
+      return {};
+    }
+  }
+
+  function saveVoicesCache(cache: Record<string, string[]>) {
+    if (typeof window === "undefined") return;
+    try {
+      localStorage.setItem(CACHE_KEY, JSON.stringify(cache));
+    } catch (e) {
+      console.error("Error saving voices cache", e);
+    }
+  }
+
+  let hasModels = $derived($models.some((m) => !m.unlisted));
+
+  let isInitialLoad = $state(true);
+
+  $effect(() => {
+    playgroundStores.speechGenerating.set(isGenerating);
+  });
+
+  // On page load, restore cached voices for the selected model if available
+  $effect(() => {
+    const model = $selectedModelStore;
+
+    if (isInitialLoad) {
+      isInitialLoad = false;
+      // If we have cached voices for this model, use them
+      const cache = getVoicesCache();
+      if (model && cache[model]) {
+        availableVoices = cache[model];
+      }
+    }
+  });
+
+  async function refreshVoices() {
+    const model = $selectedModelStore;
+    if (!model || isLoadingVoices) return;
+
+    isLoadingVoices = true;
+
+    try {
+      const response = await fetch(`/v1/audio/voices?model=${encodeURIComponent(model)}`);
+      if (!response.ok) {
+        // Fall back to default voices if API call fails
+        availableVoices = defaultVoices;
+        const cache = getVoicesCache();
+        cache[model] = defaultVoices;
+        saveVoicesCache(cache);
+        selectedVoiceStore.set(defaultVoices[0]);
+        return;
+      }
+      const data = await response.json();
+      // Expect response to be an array of voice strings or an object with a voices array
+      const voices = Array.isArray(data) ? data : (data.voices || defaultVoices);
+      const newVoices = voices.length > 0 ? voices : defaultVoices;
+
+      availableVoices = newVoices;
+      const cache = getVoicesCache();
+      cache[model] = newVoices;
+      saveVoicesCache(cache);
+
+      // Reset to first available voice
+      selectedVoiceStore.set(newVoices[0]);
+    } catch {
+      // Fall back to default voices on error
+      availableVoices = defaultVoices;
+      const cache = getVoicesCache();
+      cache[model] = defaultVoices;
+      saveVoicesCache(cache);
+      selectedVoiceStore.set(defaultVoices[0]);
+    } finally {
+      isLoadingVoices = false;
+    }
+  }
+
+  function handleVoiceChange(event: Event) {
+    const value = (event.target as HTMLSelectElement).value;
+    if (value === "(refresh)") {
+      refreshVoices();
+    } else {
+      selectedVoiceStore.set(value);
+    }
+  }
+
+  // Auto-play effect when new audio is generated
+  $effect(() => {
+    if (generatedAudioUrl && $autoPlayStore && audioElement) {
+      audioElement.load();
+      audioElement.play().catch(() => {
+        // Ignore auto-play errors (e.g., browser policy blocks)
+      });
+    }
+  });
+
+  async function generate() {
+    const trimmedText = inputText.trim();
+    if (!trimmedText || !$selectedModelStore || isGenerating) return;
+
+    isGenerating = true;
+    error = null;
+    abortController = new AbortController();
+
+    try {
+      const audioBlob = await generateSpeech(
+        $selectedModelStore,
+        trimmedText,
+        $selectedVoiceStore,
+        abortController.signal
+      );
+
+      // Revoke previous URL to prevent memory leaks
+      if (generatedAudioUrl) {
+        URL.revokeObjectURL(generatedAudioUrl);
+      }
+
+      // Create object URL for the audio blob and store metadata
+      generatedAudioUrl = URL.createObjectURL(audioBlob);
+      generatedVoice = $selectedVoiceStore;
+      generatedTimestamp = new Date();
+    } catch (err) {
+      if (err instanceof Error && err.name === "AbortError") {
+        // User cancelled
+      } else {
+        error = err instanceof Error ? err.message : "An error occurred";
+      }
+    } finally {
+      isGenerating = false;
+      abortController = null;
+    }
+  }
+
+  function cancelGeneration() {
+    abortController?.abort();
+  }
+
+  function clearInput() {
+    inputText = "";
+  }
+
+  function downloadAudio() {
+    if (!generatedAudioUrl) return;
+
+    const timestamp = (generatedTimestamp || new Date()).toISOString().replace(/[:.]/g, '-').slice(0, -5);
+    const voice = generatedVoice || 'speech';
+    const filename = `${voice}-${timestamp}.mp3`;
+
+    const a = document.createElement('a');
+    a.href = generatedAudioUrl;
+    a.download = filename;
+    document.body.appendChild(a);
+    a.click();
+    document.body.removeChild(a);
+  }
+
+  function formatTimestamp(date: Date): string {
+    return date.toLocaleString(undefined, {
+      month: 'short',
+      day: 'numeric',
+      hour: 'numeric',
+      minute: '2-digit',
+      hour12: true
+    });
+  }
+
+  function handleKeyDown(event: KeyboardEvent) {
+    if (event.key === "Enter" && !event.shiftKey) {
+      event.preventDefault();
+      generate();
+    }
+  }
+</script>
+
+<div class="flex flex-col h-full">
+  <!-- Model and voice selectors -->
+  <div class="shrink-0 flex gap-2 mb-4">
+    <ModelSelector bind:value={$selectedModelStore} placeholder="Select a speech model..." disabled={isGenerating} />
+    <div class="flex gap-2">
+      <select
+        class="shrink-0 px-3 py-2 rounded border border-gray-200 dark:border-white/10 bg-surface focus:outline-none focus:ring-2 focus:ring-primary"
+        value={$selectedVoiceStore}
+        onchange={handleVoiceChange}
+        disabled={isGenerating || isLoadingVoices || !$selectedModelStore}
+      >
+        {#each availableVoices as voice (voice)}
+          <option value={voice}>{voice}</option>
+        {/each}
+        <option value="(refresh)">(refresh)</option>
+      </select>
+      {#if $selectedModelStore && !getVoicesCache()[$selectedModelStore]}
+        <button
+          class="btn shrink-0"
+          onclick={refreshVoices}
+          disabled={isLoadingVoices}
+          title={isLoadingVoices ? "Loading voices..." : "Load voices for this model"}
+        >
+          {#if isLoadingVoices}
+            <svg class="w-5 h-5 animate-spin" fill="none" viewBox="0 0 24 24">
+              <circle class="opacity-25" cx="12" cy="12" r="10" stroke="currentColor" stroke-width="4"></circle>
+              <path class="opacity-75" fill="currentColor" d="M4 12a8 8 0 018-8V0C5.373 0 0 5.373 0 12h4zm2 5.291A7.962 7.962 0 014 12H0c0 3.042 1.135 5.824 3 7.938l3-2.647z"></path>
+            </svg>
+          {:else}
+            <svg class="w-5 h-5" fill="none" stroke="currentColor" viewBox="0 0 24 24">
+              <path stroke-linecap="round" stroke-linejoin="round" stroke-width="2" d="M4 4v5h.582m15.356 2A8.001 8.001 0 004.582 9m0 0H9m11 11v-5h-.581m0 0a8.003 8.003 0 01-15.357-2m15.357 2H15"></path>
+            </svg>
+          {/if}
+        </button>
+      {/if}
+    </div>
+  </div>
+
+  <!-- Empty state for no models configured -->
+  {#if !hasModels}
+    <div class="flex-1 flex items-center justify-center text-txtsecondary">
+      <p>No models configured. Add models to your configuration to generate speech.</p>
+    </div>
+  {:else}
+    <!-- Audio display area -->
+    <div class="shrink-0 mb-4 bg-surface border border-gray-200 dark:border-white/10 rounded p-4 md:p-6">
+      {#if isGenerating}
+        <div class="flex items-center justify-center text-txtsecondary py-8">
+          <div class="text-center">
+            <div class="inline-block w-8 h-8 border-4 border-primary border-t-transparent rounded-full animate-spin mb-2"></div>
+            <p>Generating speech...</p>
+          </div>
+        </div>
+      {:else if error}
+        <div class="flex items-center justify-center py-8">
+          <div class="text-center text-red-500">
+            <p class="font-medium">Error</p>
+            <p class="text-sm mt-1">{error}</p>
+          </div>
+        </div>
+      {:else if generatedAudioUrl}
+        <div class="flex flex-col gap-4">
+          <!-- Header with metadata and download -->
+          <div class="flex items-center justify-between gap-4">
+            <div class="flex flex-wrap gap-3 text-sm text-txtsecondary">
+              {#if generatedVoice}
+                <span class="flex items-center gap-1">
+                  <svg class="w-4 h-4" fill="none" stroke="currentColor" viewBox="0 0 24 24">
+                    <path stroke-linecap="round" stroke-linejoin="round" stroke-width="2" d="M19 11a7 7 0 01-7 7m0 0a7 7 0 01-7-7m7 7v4m0 0H8m4 0h4m-4-8a3 3 0 01-3-3V5a3 3 0 116 0v6a3 3 0 01-3 3z"></path>
+                  </svg>
+                  {generatedVoice}
+                </span>
+              {/if}
+              {#if generatedTimestamp}
+                <span class="flex items-center gap-1">
+                  <svg class="w-4 h-4" fill="none" stroke="currentColor" viewBox="0 0 24 24">
+                    <path stroke-linecap="round" stroke-linejoin="round" stroke-width="2" d="M12 8v4l3 3m6-3a9 9 0 11-18 0 9 9 0 0118 0z"></path>
+                  </svg>
+                  {formatTimestamp(generatedTimestamp)}
+                </span>
+              {/if}
+            </div>
+            <button
+              class="btn shrink-0"
+              onclick={downloadAudio}
+              title="Download audio file"
+            >
+              <svg class="w-5 h-5" fill="none" stroke="currentColor" viewBox="0 0 24 24">
+                <path stroke-linecap="round" stroke-linejoin="round" stroke-width="2" d="M4 16v1a3 3 0 003 3h10a3 3 0 003-3v-1m-4-4l-4 4m0 0l-4-4m4 4V4"></path>
+              </svg>
+            </button>
+          </div>
+
+          <!-- Audio player with larger controls -->
+          <div class="w-full">
+            <audio bind:this={audioElement} controls class="w-full h-12 md:h-16">
+              <source src={generatedAudioUrl} type="audio/mpeg" />
+              Your browser does not support the audio element.
+            </audio>
+          </div>
+        </div>
+      {:else}
+        <div class="flex items-center justify-center text-txtsecondary py-8">
+          <div class="text-center">
+            <svg class="w-12 h-12 md:w-16 md:h-16 mx-auto mb-2 opacity-40" fill="none" stroke="currentColor" viewBox="0 0 24 24">
+              <path stroke-linecap="round" stroke-linejoin="round" stroke-width="2" d="M19 11a7 7 0 01-7 7m0 0a7 7 0 01-7-7m7 7v4m0 0H8m4 0h4m-4-8a3 3 0 01-3-3V5a3 3 0 116 0v6a3 3 0 01-3 3z"></path>
+            </svg>
+            <p>Enter text below to convert to speech</p>
+          </div>
+        </div>
+      {/if}
+    </div>
+
+    <!-- Text input area -->
+    <div class="flex-1 flex flex-col md:flex-row gap-2 min-h-0">
+      <ExpandableTextarea
+        bind:value={inputText}
+        placeholder="Enter text to convert to speech..."
+        rows={8}
+        onkeydown={handleKeyDown}
+        disabled={isGenerating || !$selectedModelStore}
+      />
+      <div class="shrink-0 flex md:flex-col gap-2">
+        {#if isGenerating}
+          <button class="btn bg-red-500 hover:bg-red-600 text-white flex-1 md:flex-none" onclick={cancelGeneration}>
+            Cancel
+          </button>
+        {:else}
+          <button
+            class="btn bg-primary text-btn-primary-text hover:opacity-90 flex-1 md:flex-none"
+            onclick={generate}
+            disabled={!inputText.trim() || !$selectedModelStore}
+          >
+            Generate
+          </button>
+          <button
+            class="btn flex-1 md:flex-none"
+            onclick={clearInput}
+            disabled={!inputText.trim()}
+          >
+            Clear
+          </button>
+          <label class="flex items-center justify-center gap-2 text-sm cursor-pointer">
+            <input
+              type="checkbox"
+              bind:checked={$autoPlayStore}
+              class="cursor-pointer"
+            />
+            Auto-play
+          </label>
+        {/if}
+      </div>
+    </div>
+  {/if}
+</div>