proxy,proxy/config: add global TTL feature (#554)

Add a new configuration parameter globalTTL that all models will
inherit. The default value is 0 which matches the currently
functionality to never automatically unload a model.

The model.ttl's default has changed to -1, which means use the global
TTL value. Any model.ttl >=0 is now value with 0 meaning never unload.
This allows a model to override a globalTTL > 0 and be configured to
never unload.

Fixes #459
Closes #512

.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/containers.yml

@@ -17,12 +17,19 @@ on:
       - '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, rocm]
+        platform: [intel, cuda, cuda13, vulkan, cpu, musa, rocm]
       fail-fast: false
     steps:
       - name: Checkout code

AGENTS.md

@@ -0,0 +1,50 @@
+## 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.
+- 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.
+
+### 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

@@ -1,49 +1 @@
-## 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 (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 `gh` to create PRs and load issues
-- do include Co-Authored-By or created by when committing changes or creating PRs
-- keep PR descriptions short and focused on changes.
-  - never include a test plan
-
-## 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.
-
-### 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
+@AGENTS.md

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`
@@ -48,13 +48,27 @@ Built in Go for performance and simplicity, llama-swap has zero dependencies and
 
 ### 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="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" />
 
-<img width="1360" height="963" alt="image" src="https://github.com/user-attachments/assets/5f3edee6-d03a-4ae5-ae06-b20ac1f135bd" />
 
 ## Installation
 
@@ -69,6 +83,7 @@ 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

config-schema.json

@@ -48,6 +48,12 @@
             "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 +93,12 @@
             "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 +183,9 @@
                     },
                     "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",
@@ -194,11 +206,20 @@
                                 "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. Supports stripParams and setParams."
+                        "description": "Dictionary of filter settings. Supports stripParams, setParams, and setParamsByID."
                     },
                     "metadata": {
                         "type": "object",

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
@@ -70,6 +75,11 @@ sendLoadingState: true
 #   all fields except for Id so chat UIs can use the alias equivalent to the original.
 includeAliasesInList: false
 
+# 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
 # - macros are reusable snippets
@@ -121,7 +131,7 @@ apiKeys:
 # - 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
@@ -138,7 +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}
 
@@ -146,13 +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
@@ -167,14 +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
@@ -183,8 +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
 
@@ -192,7 +196,7 @@ 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
@@ -211,11 +215,38 @@ models:
       # - 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

docker/build-container.sh

@@ -27,7 +27,7 @@ ARCH=$1
 PUSH_IMAGES=${2:-false}
 
 # List of allowed architectures
-ALLOWED_ARCHS=("intel" "vulkan" "musa" "cuda" "cpu" "rocm")
+ALLOWED_ARCHS=("intel" "vulkan" "musa" "cuda" "cuda13" "cpu" "rocm")
 
 # Check if ARCH is in the allowed list
 if [[ ! " ${ALLOWED_ARCHS[@]} " =~ " ${ARCH} " ]]; then
@@ -142,7 +142,7 @@ for CONTAINER_TYPE in non-root root; do
   fi
 
   log_info "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} \
+  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} .
 
@@ -150,7 +150,7 @@ for CONTAINER_TYPE in non-root root; do
   case "$ARCH" in
     "musa" | "vulkan")
       log_info "Adding sd-server to $CONTAINER_TAG"
-      docker build -f llama-swap-sd.Containerfile \
+      docker build --provenance=false -f llama-swap-sd.Containerfile \
         --build-arg BASE=${CONTAINER_TAG} \
         --build-arg SD_IMAGE=${SD_IMAGE} --build-arg SD_TAG=${SD_TAG} \
         --build-arg UID=${USER_UID} --build-arg GID=${USER_GID} \

proxy/config/config.go

@@ -123,6 +123,8 @@ 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 */
@@ -201,6 +203,8 @@ func LoadConfigFromReader(r io.Reader) (Config, error) {
 		LogTimeFormat:      "",
 		LogToStdout:        LogToStdoutProxy,
 		MetricsMaxInMemory: 1000,
+		CaptureBuffer:      5,
+		GlobalTTL:          0,
 	}
 	if err = yaml.Unmarshal([]byte(yamlStr), &config); err != nil {
 		return Config{}, err
@@ -214,6 +218,10 @@ 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:
@@ -253,6 +261,15 @@ func LoadConfigFromReader(r io.Reader) (Config, error) {
 		modelConfig.Cmd = StripComments(modelConfig.Cmd)
 		modelConfig.CmdStop = StripComments(modelConfig.CmdStop)
 
+		// 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 {
@@ -292,6 +309,24 @@ func LoadConfigFromReader(r io.Reader) (Config, error) {
 			modelConfig.CheckEndpoint = strings.ReplaceAll(modelConfig.CheckEndpoint, macroSlug, macroStr)
 			modelConfig.Filters.StripParams = strings.ReplaceAll(modelConfig.Filters.StripParams, macroSlug, macroStr)
 
+			// 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 {
 				result, err := substituteMacroInValue(modelConfig.Metadata, entry.Name, entry.Value)
@@ -357,6 +392,34 @@ func LoadConfigFromReader(r io.Reader) (Config, error) {
 			}
 		}
 
+		// 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)
 		}

proxy/config/config_posix_test.go

@@ -215,6 +215,7 @@ groups:
 		},
 		HealthCheckTimeout: 15,
 		MetricsMaxInMemory: 1000,
+		CaptureBuffer:      5,
 		Profiles: map[string][]string{
 			"test": {"model1", "model2"},
 		},

proxy/config/config_test.go

@@ -848,6 +848,71 @@ func TestConfig_APIKeys_EnvMacros(t *testing.T) {
 	})
 }
 
+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")

proxy/config/config_windows_test.go

@@ -204,6 +204,7 @@ groups:
 		},
 		HealthCheckTimeout: 15,
 		MetricsMaxInMemory: 1000,
+		CaptureBuffer:      5,
 		Profiles: map[string][]string{
 			"test": {"model1", "model2"},
 		},

proxy/config/filters.go

@@ -20,6 +20,12 @@ type Filters struct {
 	// 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,
@@ -51,6 +57,33 @@ func (f Filters) SanitizedStripParams() []string {
 	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) {

proxy/config/filters_test.go

@@ -162,6 +162,123 @@ func TestFilters_SanitizedSetParams(t *testing.T) {
 	}
 }
 
+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/model_config.go

@@ -5,6 +5,10 @@ import (
 	"runtime"
 )
 
+const (
+	MODEL_CONFIG_DEFAULT_TTL = -1
+)
+
 type ModelConfig struct {
 	Cmd           string   `yaml:"cmd"`
 	CmdStop       string   `yaml:"cmdStop"`
@@ -47,7 +51,7 @@ 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,

proxy/config/model_config_test.go

@@ -73,6 +73,72 @@ models:
 	}
 }
 
+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:

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/metrics_monitor.go

@@ -28,6 +28,28 @@ 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"`
+}
+
+// Size returns the approximate memory usage of this capture in bytes
+func (c *ReqRespCapture) Size() int {
+	size := len(c.ReqPath) + len(c.ReqBody) + len(c.RespBody)
+	for k, v := range c.ReqHeaders {
+		size += len(k) + len(v)
+	}
+	for k, v := range c.RespHeaders {
+		size += len(k) + len(v)
+	}
+	return size
 }
 
 // TokenMetricsEvent represents a token metrics event
@@ -46,19 +68,32 @@ type metricsMonitor struct {
 	maxMetrics int
 	nextID     int
 	logger     *LogMonitor
+
+	// capture fields
+	enableCaptures bool
+	captures       map[int]ReqRespCapture // map for O(1) lookup by ID
+	captureOrder   []int                  // track insertion order for FIFO eviction
+	captureSize    int                    // current total size in bytes
+	maxCaptureSize int                    // max bytes for captures
 }
 
-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]ReqRespCapture),
+		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 +104,49 @@ 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 capture exceeds maxCaptureSize.
+func (mp *metricsMonitor) addCapture(capture ReqRespCapture) {
+	if !mp.enableCaptures {
+		return
+	}
+
+	mp.mu.Lock()
+	defer mp.mu.Unlock()
+
+	captureSize := capture.Size()
+	if captureSize > mp.maxCaptureSize {
+		mp.logger.Warnf("capture size %d exceeds max %d, skipping", captureSize, mp.maxCaptureSize)
+		return
+	}
+
+	// Evict oldest (FIFO) until room available
+	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 {
+			mp.captureSize -= evicted.Size()
+			delete(mp.captures, oldestID)
+		}
+	}
+
+	mp.captures[capture.ID] = capture
+	mp.captureOrder = append(mp.captureOrder, capture.ID)
+	mp.captureSize += captureSize
+}
+
+// getCaptureByID returns a capture by its ID, or nil if not found.
+func (mp *metricsMonitor) getCaptureByID(id int) *ReqRespCapture {
+	mp.mu.RLock()
+	defer mp.mu.RUnlock()
+
+	if capture, exists := mp.captures[id]; exists {
+		return &capture
+	}
+	return nil
 }
 
 // getMetrics returns a copy of the current metrics
@@ -97,6 +175,28 @@ 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 +240,6 @@ 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 +252,14 @@ 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 +272,38 @@ 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,
+		}
+		// Only set HasCapture if the capture will actually be stored (not too large)
+		if capture.Size() <= mp.maxCaptureSize {
+			tm.HasCapture = true
+		}
+	}
+
+	metricID := mp.addMetrics(tm)
+
+	// Store capture if enabled
+	if capture != nil {
+		capture.ID = metricID
+		mp.addCapture(*capture)
+	}
+
 	return nil
 }
 
@@ -336,6 +474,24 @@ 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.

proxy/metrics_monitor_test.go

@@ -18,7 +18,7 @@ import (
 
 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 +37,7 @@ 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 +51,7 @@ 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 +71,7 @@ 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 +101,14 @@ 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 +128,7 @@ 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 +140,7 @@ 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 +168,7 @@ 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 +199,7 @@ 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 +239,7 @@ 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 +275,7 @@ 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 +295,7 @@ 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 +317,7 @@ 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 +341,7 @@ 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 +360,7 @@ 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 +384,75 @@ 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 +506,7 @@ 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 +533,7 @@ 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)
 
@@ -502,7 +571,7 @@ func TestMetricsMonitor_Concurrent(t *testing.T) {
 
 func TestMetricsMonitor_ParseMetrics(t *testing.T) {
 	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 +611,7 @@ 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 +646,7 @@ 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 +680,7 @@ 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
 
@@ -641,7 +710,7 @@ data: [DONE]
 	})
 
 	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 +738,7 @@ 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 +759,7 @@ 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 +780,7 @@ 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 +814,7 @@ 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 +848,7 @@ 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 +876,7 @@ 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 +901,228 @@ func TestMetricsMonitor_WrapHandler_Compression(t *testing.T) {
 		assert.Equal(t, 100, metrics[0].OutputTokens)
 	})
 }
+
+func TestReqRespCapture_Size(t *testing.T) {
+	t.Run("calculates size correctly", func(t *testing.T) {
+		capture := ReqRespCapture{
+			ID:      1,
+			ReqPath: "/v1/chat/completions", // 20 bytes
+			ReqHeaders: map[string]string{
+				"Content-Type": "application/json", // 12 + 16 = 28
+			},
+			ReqBody: []byte("request body"), // 12 bytes
+			RespHeaders: map[string]string{
+				"X-Test": "value", // 6 + 5 = 11
+			},
+			RespBody: []byte("response body"), // 13 bytes
+		}
+
+		// Expected: 20 + 12 + 13 + 28 + 11 = 84
+		assert.Equal(t, 84, capture.Size())
+	})
+
+	t.Run("handles empty capture", func(t *testing.T) {
+		capture := ReqRespCapture{}
+		assert.Equal(t, 0, capture.Size())
+	})
+}
+
+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))
+	})
+
+	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)
+		assert.NotNil(t, retrieved)
+		assert.Equal(t, 0, retrieved.ID)
+		assert.Equal(t, []byte("test request"), retrieved.ReqBody)
+		assert.Equal(t, []byte("test response"), retrieved.RespBody)
+	})
+
+	t.Run("evicts oldest when exceeding max size", func(t *testing.T) {
+		mm := newMetricsMonitor(testLogger, 10, 5)
+		mm.maxCaptureSize = 100 // Set small limit for test
+
+		// Add captures that will exceed the limit
+		capture1 := ReqRespCapture{ID: 0, ReqBody: make([]byte, 40)}
+		capture2 := ReqRespCapture{ID: 1, ReqBody: make([]byte, 40)}
+		capture3 := ReqRespCapture{ID: 2, ReqBody: make([]byte, 40)}
+
+		mm.addCapture(capture1)
+		mm.addCapture(capture2)
+		// Adding capture3 should evict capture1
+		mm.addCapture(capture3)
+
+		assert.Nil(t, mm.getCaptureByID(0), "capture 0 should be evicted")
+		assert.NotNil(t, mm.getCaptureByID(1), "capture 1 should exist")
+		assert.NotNil(t, mm.getCaptureByID(2), "capture 2 should exist")
+	})
+
+	t.Run("skips capture larger than max size", func(t *testing.T) {
+		mm := newMetricsMonitor(testLogger, 10, 5)
+		mm.maxCaptureSize = 100
+
+		// Add a capture larger than max
+		largeCapture := ReqRespCapture{ID: 0, ReqBody: make([]byte, 200)}
+		mm.addCapture(largeCapture)
+
+		assert.Nil(t, mm.getCaptureByID(0), "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))
+	})
+
+	t.Run("returns capture by ID", func(t *testing.T) {
+		mm := newMetricsMonitor(testLogger, 10, 5)
+
+		capture := ReqRespCapture{
+			ID:      42,
+			ReqBody: []byte("test"),
+		}
+		mm.addCapture(capture)
+
+		retrieved := mm.getCaptureByID(42)
+		assert.NotNil(t, retrieved)
+		assert.Equal(t, 42, retrieved.ID)
+	})
+}
+
+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
+		capture := mm.getCaptureByID(metricID)
+		assert.NotNil(t, capture)
+		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)
+		assert.Nil(t, capture)
+	})
+}

proxy/process_test.go

@@ -117,12 +117,12 @@ 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 +159,12 @@ 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++ {

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 +77,8 @@ type ProxyManager struct {
 
 	processGroups map[string]*ProcessGroup
 
+	inFlightCounter *InflightCounter
+
 	// shutdown signaling
 	shutdownCtx    context.Context
 	shutdownCancel context.CancelFunc
@@ -151,10 +187,12 @@ 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,
 
@@ -276,37 +314,37 @@ 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.proxyInferenceHandler)
+	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.GET("/v1/audio/voices", pm.apiKeyAuth(), pm.proxyGETModelHandler)
-	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)
 
 	pm.ginEngine.GET("/v1/models", pm.apiKeyAuth(), pm.listModelsHandler)
 
@@ -325,7 +363,7 @@ 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) {
@@ -389,6 +427,14 @@ 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)
@@ -674,6 +720,17 @@ func (pm *ProxyManager) proxyInferenceHandler(c *gin.Context) {
 			}
 		}
 
+		// 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
 	} else if pm.peerProxy != nil && pm.peerProxy.HasPeerModel(requestedModel) {

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)
 	}
 }
 
@@ -81,6 +84,7 @@ 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 +109,7 @@ const (
 	msgTypeModelStatus messageType = "modelStatus"
 	msgTypeLogData     messageType = "logData"
 	msgTypeMetrics     messageType = "metrics"
+	msgTypeInFlight    messageType = "inflight"
 )
 
 type messageEnvelope struct {
@@ -164,6 +169,18 @@ 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 +208,19 @@ 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 {
@@ -250,3 +275,20 @@ 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
+	}
+
+	capture := pm.metricsMonitor.getCaptureByID(id)
+	if capture == nil {
+		c.JSON(http.StatusNotFound, gin.H{"error": "capture not found"})
+		return
+	}
+
+	c.JSON(http.StatusOK, capture)
+}

proxy/proxymanager_test.go

@@ -730,7 +730,7 @@ func TestProxyManager_RunningEndpoint(t *testing.T) {
 		// Verify extended fields are present
 		assert.NotEmpty(t, response.Running[0].Cmd, "cmd should be populated")
 		assert.NotEmpty(t, response.Running[0].Proxy, "proxy should be populated")
-		assert.Equal(t, 0, response.Running[0].TTL, "ttl should default to 0")
+		assert.Equal(t, -1, response.Running[0].TTL, "ttl should default to -1 (use globalTTL)")
 	})
 }
 
@@ -1046,6 +1046,61 @@ func TestProxyManager_FiltersStripParams(t *testing.T) {
 	// t.Logf("%v", response)
 }
 
+func TestProxyManager_FiltersSetParamsByID(t *testing.T) {
+	// no explicit aliases — setParamsByID keys are auto-registered as aliases
+	configStr := strings.Replace(`
+logLevel: error
+models:
+  model1:
+    cmd: 'SRPATH --port ${PORT} --silent --respond model1'
+    proxy: "http://127.0.0.1:${PORT}"
+    filters:
+      setParams:
+        reasoning_effort: medium
+      setParamsByID:
+        "${MODEL_ID}:high":
+          reasoning_effort: high
+        "${MODEL_ID}:low":
+          reasoning_effort: low
+`, "SRPATH", simpleResponderPath, -1)
+
+	cfg, err := config.LoadConfigFromReader(strings.NewReader(configStr))
+	if !assert.NoError(t, err, "invalid test configuration") {
+		return
+	}
+
+	proxy := New(cfg)
+	defer proxy.StopProcesses(StopWaitForInflightRequest)
+
+	tests := []struct {
+		requestedModel string
+		wantEffort     string
+	}{
+		// setParams applies, no setParamsByID match
+		{requestedModel: "model1", wantEffort: "medium"},
+		// setParamsByID overrides setParams
+		{requestedModel: "model1:high", wantEffort: "high"},
+		{requestedModel: "model1:low", wantEffort: "low"},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.requestedModel, func(t *testing.T) {
+			reqBody := fmt.Sprintf(`{"model":%q}`, tt.requestedModel)
+			req := httptest.NewRequest("POST", "/v1/chat/completions", bytes.NewBufferString(reqBody))
+			w := CreateTestResponseRecorder()
+			proxy.ServeHTTP(w, req)
+			assert.Equal(t, http.StatusOK, w.Code)
+
+			var response map[string]interface{}
+			assert.NoError(t, json.Unmarshal(w.Body.Bytes(), &response))
+
+			requestBody, _ := response["request_body"].(string)
+			gotEffort := gjson.Get(requestBody, "reasoning_effort").String()
+			assert.Equal(t, tt.wantEffort, gotEffort, "reasoning_effort mismatch for model %s", tt.requestedModel)
+		})
+	}
+}
+
 func TestProxyManager_HealthEndpoint(t *testing.T) {
 	config := config.AddDefaultGroupToConfig(config.Config{
 		HealthCheckTimeout: 15,

ui-svelte/src/App.svelte

@@ -6,23 +6,28 @@
   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 = {
-    "/": Playground,
+    "/": PlaygroundStub,
     "/models": Models,
     "/logs": LogViewer,
     "/activity": Activity,
-    "*": Playground,
+    "*": PlaygroundStub,
   };
 
-  // Sync theme to document attribute
+  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");
   });
 
-  // Sync title to document
   $effect(() => {
     const icon = $connectionState === "connecting" ? "\u{1F7E1}" : $connectionState === "connected" ? "\u{1F7E2}" : "\u{1F534}";
     document.title = `${icon} ${$appTitle}`;
@@ -43,6 +48,11 @@
   <Header />
 
   <main class="flex-1 overflow-auto p-4">
-    <Router {routes} />
+    <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/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/Header.svelte

@@ -1,6 +1,8 @@
 <script lang="ts">
-  import { link, location } from "svelte-spa-router";
+  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 {
@@ -22,9 +24,10 @@
     handleTitleChange(target.textContent || "(set title)");
   }
 
-  function isActive(path: string, currentLocation: string): boolean {
-    return path === "/" ? currentLocation === "/" : currentLocation.startsWith(path);
+  function isActive(path: string, current: string): boolean {
+    return path === "/" ? current === "/" : current.startsWith(path);
   }
+
 </script>
 
 <header
@@ -47,8 +50,7 @@
     <a
       href="/"
       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("/", $location)}
+      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>
@@ -56,7 +58,7 @@
       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", $location)}
+      class:font-semibold={isActive("/models", $currentRoute)}
     >
       Models
     </a>
@@ -64,7 +66,7 @@
       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", $location)}
+      class:font-semibold={isActive("/activity", $currentRoute)}
     >
       Activity
     </a>
@@ -72,7 +74,7 @@
       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", $location)}
+      class:font-semibold={isActive("/logs", $currentRoute)}
     >
       Logs
     </a>
@@ -96,3 +98,23 @@
     <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

@@ -65,10 +65,17 @@
   });
 
   let preElement: HTMLPreElement;
+  let userScrolledUp = $state(false);
 
-  // Auto scroll to bottom when logs change
+  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) {
+    if (preElement && filteredLogs && !userScrolledUp) {
       preElement.scrollTop = preElement.scrollHeight;
     }
   });
@@ -82,7 +89,7 @@
       <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 fill-rule="evenodd" d="M10.5 3.75a6 6 0 0 0-5.98 6.496A5.25 5.25 0 0 0 6.75 20.25H18a4.5 4.5 0 0 0 2.206-8.423 3.75 3.75 0 0 0-4.133-4.303A6.001 6.001 0 0 0 10.5 3.75Zm2.25 6a.75.75 0 0 0-1.5 0v4.94l-1.72-1.72a.75.75 0 0 0-1.06 1.06l3 3a.75.75 0 0 0 1.06 0l3-3a.75.75 0 1 0-1.06-1.06l-1.72 1.72V9.75Z" clip-rule="evenodd" />
+            <path d="M2 4v3h5v12h3V7h5V4H2zm19 5h-9v3h3v7h3v-7h3V9z"/>
           </svg>
         </button>
         <button class="btn border-0" onclick={toggleWrapText} title="Toggle text wrap">
@@ -127,6 +134,6 @@
     {/if}
   </div>
   <div class="rounded-lg bg-background font-mono text-sm flex-1 overflow-hidden">
-    <pre bind:this={preElement} class="{textWrapClass} {fontSizeClass} h-full overflow-auto p-4">{filteredLogs}</pre>
+    <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

@@ -165,6 +165,9 @@
               {#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"}

ui-svelte/src/components/StatsPanel.svelte

@@ -1,5 +1,5 @@
 <script lang="ts">
-  import { metrics } from "../stores/api";
+  import { inFlightRequests, metrics } from "../stores/api";
   import TokenHistogram from "./TokenHistogram.svelte";
 
   interface HistogramData {
@@ -15,7 +15,14 @@
   let stats = $derived.by(() => {
     const totalRequests = $metrics.length;
     if (totalRequests === 0) {
-      return { totalRequests: 0, totalInputTokens: 0, totalOutputTokens: 0, tokenStats: { p99: "0", p95: "0", p50: "0" }, histogramData: null };
+      return {
+        totalRequests: 0,
+        totalInputTokens: 0,
+        totalOutputTokens: 0,
+        inFlightRequests: $inFlightRequests,
+        tokenStats: { p99: "0", p95: "0", p50: "0" },
+        histogramData: null,
+      };
     }
 
     const totalInputTokens = $metrics.reduce((sum, m) => sum + m.input_tokens, 0);
@@ -24,7 +31,14 @@
     // Calculate token statistics using output_tokens and duration_ms
     const validMetrics = $metrics.filter((m) => m.duration_ms > 0 && m.output_tokens > 0);
     if (validMetrics.length === 0) {
-      return { totalRequests, totalInputTokens, totalOutputTokens, tokenStats: { p99: "0", p95: "0", p50: "0" }, histogramData: null };
+      return {
+        totalRequests,
+        totalInputTokens,
+        totalOutputTokens,
+        inFlightRequests: $inFlightRequests,
+        tokenStats: { p99: "0", p95: "0", p50: "0" },
+        histogramData: null,
+      };
     }
 
     // Calculate tokens/second for each valid metric
@@ -63,6 +77,7 @@
       totalRequests,
       totalInputTokens,
       totalOutputTokens,
+      inFlightRequests: $inFlightRequests,
       tokenStats: {
         p99: p99.toFixed(2),
         p95: p95.toFixed(2),
@@ -95,7 +110,12 @@
 
       <tbody class="bg-surface divide-y divide-card-border-inner">
         <tr class="hover:bg-secondary">
-          <td class="px-4 py-4 text-sm font-semibold text-gray-900 dark:text-white">{stats.totalRequests}</td>
+          <td class="px-4 py-4 text-sm font-semibold text-gray-900 dark:text-white">
+            <div class="flex flex-col gap-1">
+              <span class="text-xs font-medium text-gray-500 dark:text-gray-400">Completed: {nf.format(stats.totalRequests)}</span>
+              <span class="text-xs font-medium text-gray-500 dark:text-gray-400">Waiting: {nf.format(stats.inFlightRequests)}</span>
+            </div>
+          </td>
 
           <td class="px-4 py-4 text-sm text-gray-700 dark:text-gray-300 border-l border-gray-200 dark:border-white/10">
             <div class="flex items-center gap-2">

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

@@ -2,6 +2,7 @@
   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", "");
@@ -15,18 +16,22 @@
   let fileInput = $state<HTMLInputElement | null>(null);
   let copied = $state(false);
 
-  const ACCEPTED_FORMATS = ['.mp3', '.wav'];
+  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' };
+      return { valid: false, error: 'Invalid file type. Accepted: MP3, WAV, OGG' };
     }
 
     if (file.size > MAX_FILE_SIZE) {
@@ -203,7 +208,7 @@
           <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 (max 25MB)</p>
+            <p class="text-xs mt-4">Accepted formats: MP3, WAV, OGG (max 25MB)</p>
           </div>
         </div>
       {/if}
@@ -213,7 +218,7 @@
     <div class="shrink-0 flex gap-2">
       <input
         type="file"
-        accept=".mp3,.wav"
+        accept=".mp3,.wav,.ogg"
         class="hidden"
         onchange={handleFileSelect}
         bind:this={fileInput}

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

@@ -2,6 +2,7 @@
   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";
@@ -11,7 +12,16 @@
   const systemPromptStore = persistentStore<string>("playground-system-prompt", "");
   const temperatureStore = persistentStore<number>("playground-temperature", 0.7);
 
-  let messages = $state<ChatMessage[]>([]);
+  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);
@@ -24,21 +34,52 @@
   let imageError = $state<string | null>(null);
 
   let hasModels = $derived($models.some((m) => !m.unlisted));
+  let userScrolledUp = $state(false);
 
-  // Auto-scroll when messages change
   $effect(() => {
-    if (messages.length > 0 && messagesContainer) {
+    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: "smooth",
+        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) {
@@ -321,6 +362,7 @@
     <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">

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

@@ -1,5 +1,6 @@
 <script lang="ts">
-  import { renderMarkdown, escapeHtml } from "../../lib/markdown";
+  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";
@@ -22,11 +23,17 @@
   let hasImages = $derived(imageUrls.length > 0);
   let canEdit = $derived(onEdit !== undefined && !hasImages);
 
-  let renderedContent = $derived(
-    role === "assistant" && !isStreaming
-      ? renderMarkdown(textContent)
-      : escapeHtml(textContent).replace(/\n/g, '<br>')
-  );
+  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);
@@ -109,13 +116,54 @@
       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 max-w-[85%] rounded-lg px-4 py-2 {role === 'user'
-      ? 'bg-primary text-btn-primary-text'
-      : 'bg-surface border border-gray-200 dark:border-white/10'}"
+    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}
@@ -167,8 +215,11 @@
       {#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">
-          {@html renderedContent}
+        <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}
@@ -289,14 +340,42 @@
 
 <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;

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

@@ -2,6 +2,7 @@
   import { models } from "../../stores/api";
   import { persistentStore } from "../../stores/persistent";
   import { generateImage } from "../../lib/imageApi";
+  import { playgroundStores } from "../../stores/playgroundActivity";
   import ModelSelector from "./ModelSelector.svelte";
   import ExpandableTextarea from "./ExpandableTextarea.svelte";
 
@@ -17,6 +18,10 @@
 
   let hasModels = $derived($models.some((m) => !m.unlisted));
 
+  $effect(() => {
+    playgroundStores.imageGenerating.set(isGenerating);
+  });
+
   async function generate() {
     const trimmedPrompt = prompt.trim();
     if (!trimmedPrompt || !$selectedModelStore || isGenerating) return;

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

@@ -25,6 +25,11 @@
       <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}

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

@@ -2,6 +2,7 @@
   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";
 
@@ -20,11 +21,9 @@
   let availableVoices = $state<string[]>(["coral", "alloy", "echo", "fable", "onyx", "nova", "shimmer"]);
   let isLoadingVoices = $state(false);
 
-  // Default voices to fall back to if API call fails
   const defaultVoices = ["coral", "alloy", "echo", "fable", "onyx", "nova", "shimmer"];
   const CACHE_KEY = "playground-speech-voices-cache";
 
-  // Load voices cache from localStorage
   function getVoicesCache(): Record<string, string[]> {
     if (typeof window === "undefined") return {};
     try {
@@ -35,7 +34,6 @@
     }
   }
 
-  // Save voices cache to localStorage
   function saveVoicesCache(cache: Record<string, string[]>) {
     if (typeof window === "undefined") return;
     try {
@@ -47,9 +45,12 @@
 
   let hasModels = $derived($models.some((m) => !m.unlisted));
 
-  // Track if this is the initial page load to avoid fetching on refresh
   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;

ui-svelte/src/lib/markdown.test.ts

@@ -1,5 +1,5 @@
 import { describe, it, expect } from "vitest";
-import { renderMarkdown, escapeHtml } from "./markdown";
+import { renderMarkdown, escapeHtml, splitCompleteBlocks, closePendingBlock, normalizeLatexDelimiters, renderStreamingMarkdown, createStreamingCache } from "./markdown";
 
 describe("renderMarkdown", () => {
   describe("basic markdown", () => {
@@ -130,6 +130,35 @@ More text here.
       expect(result).toContain("katex");
       expect(result).toContain("sqrt");
     });
+
+    it("renders \\[...\\] display math", () => {
+      const result = renderMarkdown("\\[\nx^2 + y^2 = z^2\n\\]");
+      expect(result).toContain("katex");
+    });
+
+    it("renders \\(...\\) inline math", () => {
+      const result = renderMarkdown("The equation \\(E = mc^2\\) is famous.");
+      expect(result).toContain("katex");
+    });
+  });
+
+  describe("normalizeLatexDelimiters", () => {
+    it("converts \\[...\\] to $$...$$", () => {
+      expect(normalizeLatexDelimiters("\\[\nx^2\n\\]")).toBe("$$\nx^2\n$$");
+    });
+
+    it("converts \\(...\\) to $...$", () => {
+      expect(normalizeLatexDelimiters("\\(x^2\\)")).toBe("$x^2$");
+    });
+
+    it("leaves $$ and $ delimiters unchanged", () => {
+      const text = "$$x^2$$ and $y$";
+      expect(normalizeLatexDelimiters(text)).toBe(text);
+    });
+
+    it("handles multiple occurrences", () => {
+      expect(normalizeLatexDelimiters("\\(a\\) and \\(b\\)")).toBe("$a$ and $b$");
+    });
   });
 
   describe("escapeHtml", () => {
@@ -158,3 +187,237 @@ More text here.
     });
   });
 });
+
+describe("splitCompleteBlocks", () => {
+  it("returns everything as pending when no blank line", () => {
+    const result = splitCompleteBlocks("Hello world");
+    expect(result.complete).toBe("");
+    expect(result.pending).toBe("Hello world");
+  });
+
+  it("returns empty for empty input", () => {
+    const result = splitCompleteBlocks("");
+    expect(result.complete).toBe("");
+    expect(result.pending).toBe("");
+  });
+
+  it("splits on blank line between paragraphs", () => {
+    const result = splitCompleteBlocks("First paragraph.\n\nSecond paragraph");
+    expect(result.complete).toBe("First paragraph.\n");
+    expect(result.pending).toBe("Second paragraph");
+  });
+
+  it("splits multiple paragraphs at last blank line", () => {
+    const result = splitCompleteBlocks("Para 1.\n\nPara 2.\n\nPara 3");
+    expect(result.complete).toBe("Para 1.\n\nPara 2.\n");
+    expect(result.pending).toBe("Para 3");
+  });
+
+  it("treats closed code fence as complete boundary", () => {
+    const text = "```js\nconst x = 1;\n```\nMore text";
+    const result = splitCompleteBlocks(text);
+    expect(result.complete).toBe("```js\nconst x = 1;\n```");
+    expect(result.pending).toBe("More text");
+  });
+
+  it("treats unclosed code fence as pending", () => {
+    const text = "Done paragraph.\n\n```js\nconst x = 1;";
+    const result = splitCompleteBlocks(text);
+    expect(result.complete).toBe("Done paragraph.\n");
+    expect(result.pending).toBe("```js\nconst x = 1;");
+  });
+
+  it("does not split on blank lines inside code fences", () => {
+    const text = "```\nline1\n\nline2\n```";
+    const result = splitCompleteBlocks(text);
+    expect(result.complete).toBe("```\nline1\n\nline2\n```");
+    expect(result.pending).toBe("");
+  });
+
+  it("handles tilde fences", () => {
+    const text = "~~~py\nprint('hi')\n~~~\nAfter";
+    const result = splitCompleteBlocks(text);
+    expect(result.complete).toBe("~~~py\nprint('hi')\n~~~");
+    expect(result.pending).toBe("After");
+  });
+
+  it("does not close backtick fence with tilde fence", () => {
+    const text = "```\ncode\n~~~\nstill code";
+    const result = splitCompleteBlocks(text);
+    // The ~~~ should not close a backtick fence, so everything from ``` onward is pending
+    expect(result.complete).toBe("");
+    expect(result.pending).toBe("```\ncode\n~~~\nstill code");
+  });
+
+  it("treats closed math block as complete boundary", () => {
+    const text = "$$\nx^2\n$$\nAfter";
+    const result = splitCompleteBlocks(text);
+    expect(result.complete).toBe("$$\nx^2\n$$");
+    expect(result.pending).toBe("After");
+  });
+
+  it("treats unclosed math block as pending", () => {
+    const text = "Before.\n\n$$\nx^2";
+    const result = splitCompleteBlocks(text);
+    expect(result.complete).toBe("Before.\n");
+    expect(result.pending).toBe("$$\nx^2");
+  });
+
+  it("treats closed \\[...\\] math block as complete boundary", () => {
+    const text = "\\[\nx^2\n\\]\nAfter";
+    const result = splitCompleteBlocks(text);
+    expect(result.complete).toBe("\\[\nx^2\n\\]");
+    expect(result.pending).toBe("After");
+  });
+
+  it("treats unclosed \\[ math block as pending", () => {
+    const text = "Before.\n\n\\[\nx^2";
+    const result = splitCompleteBlocks(text);
+    expect(result.complete).toBe("Before.\n");
+    expect(result.pending).toBe("\\[\nx^2");
+  });
+
+  it("handles trailing blank line making everything complete", () => {
+    const text = "Hello world.\n";
+    const result = splitCompleteBlocks(text);
+    // Last line is empty string after split, which is a blank line
+    expect(result.complete).toBe("Hello world.\n");
+    expect(result.pending).toBe("");
+  });
+});
+
+describe("closePendingBlock", () => {
+  it("returns empty string for empty input", () => {
+    expect(closePendingBlock("")).toBe("");
+  });
+
+  it("returns plain text unchanged", () => {
+    expect(closePendingBlock("Hello world")).toBe("Hello world");
+  });
+
+  it("closes an open backtick code fence", () => {
+    const result = closePendingBlock("```python\nprint('hi')");
+    expect(result).toBe("```python\nprint('hi')\n```");
+  });
+
+  it("closes an open tilde code fence", () => {
+    const result = closePendingBlock("~~~js\nconst x = 1;");
+    expect(result).toBe("~~~js\nconst x = 1;\n~~~");
+  });
+
+  it("does not modify already-closed code fence", () => {
+    const text = "```py\ncode\n```";
+    expect(closePendingBlock(text)).toBe(text);
+  });
+
+  it("closes an open math block", () => {
+    const result = closePendingBlock("$$\nx^2 + y^2");
+    expect(result).toBe("$$\nx^2 + y^2\n$$");
+  });
+
+  it("does not modify already-closed math block", () => {
+    const text = "$$\nx^2\n$$";
+    expect(closePendingBlock(text)).toBe(text);
+  });
+
+  it("closes an open \\[ math block with \\]", () => {
+    const result = closePendingBlock("\\[\nx^2 + y^2");
+    expect(result).toBe("\\[\nx^2 + y^2\n\\]");
+  });
+
+  it("does not modify already-closed \\[...\\] math block", () => {
+    const text = "\\[\nx^2\n\\]";
+    expect(closePendingBlock(text)).toBe(text);
+  });
+
+  it("closes code fence when preceded by regular text", () => {
+    const result = closePendingBlock("Some text\n```\ncode");
+    expect(result).toBe("Some text\n```\ncode\n```");
+  });
+
+  it("leaves headers unchanged", () => {
+    expect(closePendingBlock("## Hello")).toBe("## Hello");
+  });
+
+  it("leaves tables unchanged", () => {
+    const table = "| a | b |\n| --- | --- |\n| 1 | 2 |";
+    expect(closePendingBlock(table)).toBe(table);
+  });
+
+  it("leaves lists unchanged", () => {
+    expect(closePendingBlock("- item 1\n- item 2")).toBe("- item 1\n- item 2");
+  });
+});
+
+describe("renderStreamingMarkdown", () => {
+  it("renders complete blocks and pending as markdown", () => {
+    const cache = createStreamingCache();
+    const text = "# Hello\n\nWorld";
+    const { blocks, pendingHtml } = renderStreamingMarkdown(text, cache);
+    expect(blocks).toHaveLength(1);
+    expect(blocks[0].html).toContain("<h1>Hello</h1>");
+    expect(pendingHtml).toContain("World");
+    expect(pendingHtml).toContain("<p>");
+  });
+
+  it("preserves existing blocks when complete portion is unchanged", () => {
+    const cache = createStreamingCache();
+    renderStreamingMarkdown("# Hello\n\nWor", cache);
+    const firstBlocks = cache.blocks;
+
+    const { blocks } = renderStreamingMarkdown("# Hello\n\nWorld", cache);
+    // Same block array reference — nothing changed in the complete section
+    expect(blocks).toBe(firstBlocks);
+    expect(cache.completeKey).toBe("# Hello\n");
+  });
+
+  it("appends a new block when a new section completes", () => {
+    const cache = createStreamingCache();
+    renderStreamingMarkdown("# Hello\n\nParagraph", cache);
+    expect(cache.blocks).toHaveLength(1);
+    const firstBlock = cache.blocks[0];
+
+    renderStreamingMarkdown("# Hello\n\nParagraph.\n\nMore", cache);
+    expect(cache.blocks).toHaveLength(2);
+    // First block is preserved with the same id and html
+    expect(cache.blocks[0].id).toBe(firstBlock.id);
+    expect(cache.blocks[0].html).toBe(firstBlock.html);
+    // Second block contains the new paragraph
+    expect(cache.blocks[1].html).toContain("Paragraph.");
+  });
+
+  it("assigns unique stable ids to each block", () => {
+    const cache = createStreamingCache();
+    renderStreamingMarkdown("A.\n\nB.\n\nC", cache);
+    expect(cache.blocks).toHaveLength(1);
+    const id0 = cache.blocks[0].id;
+
+    renderStreamingMarkdown("A.\n\nB.\n\nC.\n\nD", cache);
+    expect(cache.blocks).toHaveLength(2);
+    expect(cache.blocks[0].id).toBe(id0);
+    expect(cache.blocks[1].id).toBe(id0 + 1);
+  });
+
+  it("renders pending code block with syntax highlighting", () => {
+    const cache = createStreamingCache();
+    const text = "Done.\n\n```python\nprint('hello')";
+    const { pendingHtml } = renderStreamingMarkdown(text, cache);
+    expect(pendingHtml).toContain("<code");
+    expect(pendingHtml).toContain("hljs");
+  });
+
+  it("renders pending table as markdown", () => {
+    const cache = createStreamingCache();
+    const text = "Done.\n\n| a | b |\n| --- | --- |\n| 1 | 2 |";
+    const { pendingHtml } = renderStreamingMarkdown(text, cache);
+    expect(pendingHtml).toContain("<table>");
+    expect(pendingHtml).toContain("<td>");
+  });
+
+  it("renders pending portion through markdown pipeline", () => {
+    const cache = createStreamingCache();
+    const text = "Done.\n\nSome **bold** text";
+    const { pendingHtml } = renderStreamingMarkdown(text, cache);
+    expect(pendingHtml).toContain("<strong>bold</strong>");
+  });
+});

ui-svelte/src/lib/markdown.ts

@@ -69,13 +69,189 @@ const processor = unified()
   .use(rehypeHighlight)
   .use(rehypeStringify, { allowDangerousHtml: true });
 
+export function splitCompleteBlocks(text: string): { complete: string; pending: string } {
+  if (!text) {
+    return { complete: "", pending: "" };
+  }
+
+  const lines = text.split("\n");
+  let lastCompleteBoundary = -1; // index of last line that ends a complete block
+  let inFence = false;
+  let fenceChar = "";
+  let inMathBlock = false;
+
+  for (let i = 0; i < lines.length; i++) {
+    const trimmed = lines[i].trimEnd();
+
+    if (inFence) {
+      // Check for closing fence: same character, at least 3, no other content
+      if (new RegExp(`^\\s*${fenceChar.replace(/~/g, "\\~")}{3,}\\s*$`).test(trimmed)) {
+        inFence = false;
+        fenceChar = "";
+        lastCompleteBoundary = i;
+      }
+      continue;
+    }
+
+    if (inMathBlock) {
+      if (trimmed === "$$" || trimmed === "\\]") {
+        inMathBlock = false;
+        lastCompleteBoundary = i;
+      }
+      continue;
+    }
+
+    // Check for opening fence
+    const fenceMatch = trimmed.match(/^(\s*)(```|~~~)/);
+    if (fenceMatch) {
+      // Check if it's an opening fence (may have language info after)
+      // A line with just ``` or ~~~ could be opening or closing, but since we're not in a fence it's opening
+      fenceChar = fenceMatch[2][0]; // '`' or '~'
+      inFence = true;
+      continue;
+    }
+
+    // Check for opening math block
+    if (trimmed === "$$" || trimmed === "\\[") {
+      inMathBlock = true;
+      continue;
+    }
+
+    // Outside fences/math: blank line marks a complete boundary
+    if (trimmed === "") {
+      lastCompleteBoundary = i;
+    }
+  }
+
+  if (lastCompleteBoundary < 0) {
+    return { complete: "", pending: text };
+  }
+
+  const completeLines = lines.slice(0, lastCompleteBoundary + 1);
+  const pendingLines = lines.slice(lastCompleteBoundary + 1);
+
+  return {
+    complete: completeLines.join("\n"),
+    pending: pendingLines.join("\n"),
+  };
+}
+
+export function closePendingBlock(pending: string): string {
+  if (!pending) return "";
+
+  const lines = pending.split("\n");
+  let inFence = false;
+  let fenceStr = "";
+  let inMathBlock = false;
+  let mathClose = "";
+
+  for (const line of lines) {
+    const trimmed = line.trimEnd();
+
+    if (inFence) {
+      if (new RegExp(`^\\s*${fenceStr[0] === "~" ? "~~~" : "\\`\\`\\`"}\\s*$`).test(trimmed)) {
+        inFence = false;
+        fenceStr = "";
+      }
+      continue;
+    }
+
+    if (inMathBlock) {
+      if (trimmed === "$$" || trimmed === "\\]") {
+        inMathBlock = false;
+        mathClose = "";
+      }
+      continue;
+    }
+
+    const fenceMatch = trimmed.match(/^(\s*)(```|~~~)/);
+    if (fenceMatch) {
+      fenceStr = fenceMatch[2];
+      inFence = true;
+      continue;
+    }
+
+    if (trimmed === "$$") {
+      inMathBlock = true;
+      mathClose = "$$";
+      continue;
+    }
+
+    if (trimmed === "\\[") {
+      inMathBlock = true;
+      mathClose = "\\]";
+      continue;
+    }
+  }
+
+  if (inFence) return pending + "\n" + fenceStr;
+  if (inMathBlock) return pending + "\n" + mathClose;
+  return pending;
+}
+
+export interface RenderedBlock {
+  id: number;
+  html: string;
+}
+
+export interface StreamingCache {
+  blocks: RenderedBlock[];
+  nextId: number;
+  completeKey: string;
+}
+
+export function createStreamingCache(): StreamingCache {
+  return { blocks: [], nextId: 0, completeKey: "" };
+}
+
+export function renderStreamingMarkdown(
+  text: string,
+  cache: StreamingCache,
+): { blocks: RenderedBlock[]; pendingHtml: string } {
+  const { complete, pending } = splitCompleteBlocks(text);
+
+  if (complete) {
+    if (cache.completeKey !== complete) {
+      if (complete.startsWith(cache.completeKey) && cache.completeKey.length > 0) {
+        // Complete section grew — render only the new part as a new block
+        const newPart = complete.slice(cache.completeKey.length);
+        cache.blocks = [...cache.blocks, { id: cache.nextId++, html: renderMarkdown(newPart) }];
+      } else {
+        // Complete section changed unexpectedly — re-render as single block
+        cache.blocks = [{ id: cache.nextId++, html: renderMarkdown(complete) }];
+      }
+      cache.completeKey = complete;
+    }
+  } else if (cache.blocks.length > 0) {
+    cache.blocks = [];
+    cache.completeKey = "";
+  }
+
+  let pendingHtml = "";
+  if (pending) {
+    const closed = closePendingBlock(pending);
+    pendingHtml = renderMarkdown(closed);
+  }
+
+  return { blocks: cache.blocks, pendingHtml };
+}
+
+// Convert \[...\] to $$...$$ and \(...\) to $...$
+export function normalizeLatexDelimiters(text: string): string {
+  // Display math: \[...\] → $$...$$  (may span multiple lines)
+  text = text.replace(/\\\[([\s\S]*?)\\\]/g, (_match, inner) => `$$${inner}$$`);
+  // Inline math: \(...\) → $...$
+  text = text.replace(/\\\(([\s\S]*?)\\\)/g, (_match, inner) => `$${inner}$`);
+  return text;
+}
+
 export function renderMarkdown(content: string): string {
   if (!content) {
     return "";
   }
 
   try {
-    const result = processor.processSync(content);
+    const result = processor.processSync(normalizeLatexDelimiters(content));
     return String(result);
   } catch {
     // Fallback to escaped plain text if markdown parsing fails

ui-svelte/src/lib/rerankApi.ts

@@ -0,0 +1,27 @@
+export interface RerankResult {
+  index: number;
+  relevance_score: number;
+}
+
+export interface RerankResponse {
+  model: string;
+  object: string;
+  usage: { prompt_tokens: number; total_tokens: number };
+  results: RerankResult[];
+}
+
+export async function rerank(
+  model: string,
+  query: string,
+  documents: string[],
+  signal: AbortSignal
+): Promise<RerankResponse> {
+  const response = await fetch("/v1/rerank", {
+    method: "POST",
+    headers: { "Content-Type": "application/json" },
+    body: JSON.stringify({ model, query, documents }),
+    signal,
+  });
+  if (!response.ok) throw new Error(`${response.status} ${response.statusText}`);
+  return response.json();
+}

ui-svelte/src/lib/types.ts

@@ -9,6 +9,7 @@ export interface Model {
   description: string;
   unlisted: boolean;
   peerID: string;
+  aliases?: string[];
 }
 
 export interface Metrics {
@@ -21,6 +22,16 @@ export interface Metrics {
   prompt_per_second: number;
   tokens_per_second: number;
   duration_ms: number;
+  has_capture: boolean;
+}
+
+export interface ReqRespCapture {
+  id: number;
+  req_path: string;
+  req_headers: Record<string, string>;
+  req_body: string; // base64 encoded bytes
+  resp_headers: Record<string, string>;
+  resp_body: string; // base64 encoded bytes
 }
 
 export interface LogData {
@@ -28,8 +39,12 @@ export interface LogData {
   data: string;
 }
 
+export interface InFlightStats {
+  total: number;
+}
+
 export interface APIEventEnvelope {
-  type: "modelStatus" | "logData" | "metrics";
+  type: "modelStatus" | "logData" | "metrics" | "inflight";
   data: string;
 }
 

ui-svelte/src/routes/Activity.svelte

@@ -1,6 +1,8 @@
 <script lang="ts">
-  import { metrics } from "../stores/api";
+  import { metrics, getCapture } from "../stores/api";
   import Tooltip from "../components/Tooltip.svelte";
+  import CaptureDialog from "../components/CaptureDialog.svelte";
+  import type { ReqRespCapture } from "../lib/types";
 
   function formatSpeed(speed: number): string {
     return speed < 0 ? "unknown" : speed.toFixed(2) + " t/s";
@@ -38,6 +40,25 @@
   }
 
   let sortedMetrics = $derived([...$metrics].sort((a, b) => b.id - a.id));
+
+  let selectedCapture = $state<ReqRespCapture | null>(null);
+  let dialogOpen = $state(false);
+  let loadingCaptureId = $state<number | null>(null);
+
+  async function viewCapture(id: number) {
+    loadingCaptureId = id;
+    const capture = await getCapture(id);
+    loadingCaptureId = null;
+    if (capture) {
+      selectedCapture = capture;
+      dialogOpen = true;
+    }
+  }
+
+  function closeDialog() {
+    dialogOpen = false;
+    selectedCapture = null;
+  }
 </script>
 
 <div class="p-2">
@@ -65,6 +86,7 @@
             <th class="px-6 py-3">Prompt Processing</th>
             <th class="px-6 py-3">Generation Speed</th>
             <th class="px-6 py-3">Duration</th>
+            <th class="px-6 py-3">Capture</th>
           </tr>
         </thead>
         <tbody class="divide-y">
@@ -79,6 +101,19 @@
               <td class="px-6 py-4">{formatSpeed(metric.prompt_per_second)}</td>
               <td class="px-6 py-4">{formatSpeed(metric.tokens_per_second)}</td>
               <td class="px-6 py-4">{formatDuration(metric.duration_ms)}</td>
+              <td class="px-6 py-4">
+                {#if metric.has_capture}
+                  <button
+                    onclick={() => viewCapture(metric.id)}
+                    disabled={loadingCaptureId === metric.id}
+                    class="btn btn--sm"
+                  >
+                    {loadingCaptureId === metric.id ? "..." : "View"}
+                  </button>
+                {:else}
+                  <span class="text-txtsecondary">-</span>
+                {/if}
+              </td>
             </tr>
           {/each}
         </tbody>
@@ -86,3 +121,5 @@
     </div>
   {/if}
 </div>
+
+<CaptureDialog capture={selectedCapture} open={dialogOpen} onclose={closeDialog} />

ui-svelte/src/routes/Playground.svelte

@@ -4,8 +4,9 @@
   import ImageInterface from "../components/playground/ImageInterface.svelte";
   import AudioInterface from "../components/playground/AudioInterface.svelte";
   import SpeechInterface from "../components/playground/SpeechInterface.svelte";
+  import RerankInterface from "../components/playground/RerankInterface.svelte";
 
-  type Tab = "chat" | "images" | "speech" | "audio";
+  type Tab = "chat" | "images" | "speech" | "audio" | "rerank";
 
   const selectedTabStore = persistentStore<Tab>("playground-selected-tab", "chat");
   let mobileMenuOpen = $state(false);
@@ -15,6 +16,7 @@
     { id: "images", label: "Images" },
     { id: "speech", label: "Speech" },
     { id: "audio", label: "Transcription" },
+    { id: "rerank", label: "Rerank" },
   ];
 
   function selectTab(tab: Tab) {
@@ -89,6 +91,9 @@
     <div class="h-full" class:tab-hidden={$selectedTabStore !== "audio"}>
       <AudioInterface />
     </div>
+    <div class="h-full" class:tab-hidden={$selectedTabStore !== "rerank"}>
+      <RerankInterface />
+    </div>
   </div>
 </div>
 

ui-svelte/src/routes/PlaygroundStub.svelte

@@ -0,0 +1 @@
+<!-- empty: real Playground is always mounted in App.svelte -->

ui-svelte/src/stores/api.ts

@@ -1,5 +1,5 @@
 import { writable } from "svelte/store";
-import type { Model, Metrics, VersionInfo, LogData, APIEventEnvelope } from "../lib/types";
+import type { Model, Metrics, VersionInfo, LogData, APIEventEnvelope, ReqRespCapture, InFlightStats } from "../lib/types";
 import { connectionState } from "./theme";
 
 const LOG_LENGTH_LIMIT = 1024 * 100; /* 100KB of log data */
@@ -9,6 +9,7 @@ export const models = writable<Model[]>([]);
 export const proxyLogs = writable<string>("");
 export const upstreamLogs = writable<string>("");
 export const metrics = writable<Metrics[]>([]);
+export const inFlightRequests = writable<number>(0);
 export const versionInfo = writable<VersionInfo>({
   build_date: "unknown",
   commit: "unknown",
@@ -29,6 +30,7 @@ export function enableAPIEvents(enabled: boolean): void {
     apiEventSource?.close();
     apiEventSource = null;
     metrics.set([]);
+    inFlightRequests.set(0);
     return;
   }
 
@@ -46,6 +48,7 @@ export function enableAPIEvents(enabled: boolean): void {
       proxyLogs.set("");
       upstreamLogs.set("");
       metrics.set([]);
+      inFlightRequests.set(0);
       models.set([]);
       retryCount = 0;
       connectionState.set("connected");
@@ -83,6 +86,11 @@ export function enableAPIEvents(enabled: boolean): void {
             metrics.update((prevMetrics) => [...newMetrics, ...prevMetrics]);
             break;
           }
+          case "inflight": {
+            const stats = JSON.parse(message.data) as InFlightStats;
+            inFlightRequests.set(stats.total ?? 0);
+            break;
+          }
         }
       } catch (err) {
         console.error(e.data, err);
@@ -172,3 +180,19 @@ export async function loadModel(model: string): Promise<void> {
     throw error;
   }
 }
+
+export async function getCapture(id: number): Promise<ReqRespCapture | null> {
+  try {
+    const response = await fetch(`/api/captures/${id}`);
+    if (response.status === 404) {
+      return null;
+    }
+    if (!response.ok) {
+      throw new Error(`Failed to fetch capture: ${response.status}`);
+    }
+    return await response.json();
+  } catch (error) {
+    console.error("Failed to fetch capture:", error);
+    return null;
+  }
+}

ui-svelte/src/stores/playgroundActivity.ts

@@ -0,0 +1,20 @@
+import { writable, derived } from "svelte/store";
+
+const chatStreaming = writable(false);
+const imageGenerating = writable(false);
+const speechGenerating = writable(false);
+const audioTranscribing = writable(false);
+const rerankLoading = writable(false);
+
+export const playgroundActivity = derived(
+  [chatStreaming, imageGenerating, speechGenerating, audioTranscribing, rerankLoading],
+  ([$chat, $image, $speech, $audio, $rerank]) => $chat || $image || $speech || $audio || $rerank
+);
+
+export const playgroundStores = {
+  chatStreaming,
+  imageGenerating,
+  speechGenerating,
+  audioTranscribing,
+  rerankLoading,
+};

ui-svelte/src/stores/route.ts

@@ -0,0 +1,3 @@
+import { writable } from "svelte/store";
+
+export const currentRoute = writable("/");