proxy: support timings for /infill from llama-server

fixes: #463

.github/workflows/containers.yml

@@ -15,6 +15,7 @@ on:
     paths:
       - '.github/workflows/containers.yml'
       - 'docker/build-container.sh'
+      - 'docker/*.Containerfile'
 
 jobs:
   build-and-push:

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

@@ -3,9 +3,25 @@ name: Windows CI
 on:
   push:
     branches: [ "main" ]
+    # only run when backend source changes
+    # cmd/ is excluded because it contains utilities without tests
+    paths:
+      - '**/*.go'
+      - '!cmd/**'
+      - 'go.mod'
+      - 'go.sum'
+      - 'Makefile'
+      - '.github/workflows/go-ci-windows.yml'
 
   pull_request:
     branches: [ "main" ]
+    paths:
+      - '**/*.go'
+      - '!cmd/**'
+      - 'go.mod'
+      - 'go.sum'
+      - 'Makefile'
+      - '.github/workflows/go-ci-windows.yml'
 
   # Allows manual triggering of the workflow
   workflow_dispatch:

.github/workflows/go-ci.yml

@@ -3,9 +3,25 @@ name: Linux CI
 on:
   push:
     branches: [ "main" ]
+    # only run when backend source changes
+    # cmd/ is excluded because it contains utilities without tests
+    paths:
+      - '**/*.go'
+      - '!cmd/**'
+      - 'go.mod'
+      - 'go.sum'
+      - 'Makefile'
+      - '.github/workflows/go-ci.yml'
 
   pull_request:
     branches: [ "main" ]
+    paths:
+      - '**/*.go'
+      - '!cmd/**'
+      - 'go.mod'
+      - 'go.sum'
+      - 'Makefile'
+      - '.github/workflows/go-ci.yml'
 
   # Allows manual triggering of the workflow
   workflow_dispatch:

.github/workflows/release.yml

@@ -29,7 +29,7 @@ jobs:
       - name: Set up Node.js
         uses: actions/setup-node@v4
         with:
-          node-version: "23"
+          node-version: "24"
       - name: Install dependencies and build UI
         run: |
           cd ui-svelte

.github/workflows/ui-tests.yml

@@ -0,0 +1,42 @@
+name: UI Tests
+
+on:
+  push:
+    branches: [ "main" ]
+    paths:
+      - 'ui-svelte/**'
+      - '.github/workflows/ui-tests.yml'
+
+  pull_request:
+    branches: [ "main" ]
+    paths:
+      - 'ui-svelte/**'
+      - '.github/workflows/ui-tests.yml'
+
+  workflow_dispatch:
+
+jobs:
+
+  run-tests:
+    runs-on: ubuntu-latest
+    defaults:
+      run:
+        working-directory: ui-svelte
+    steps:
+    - uses: actions/checkout@v4
+
+    - name: Set up Node.js
+      uses: actions/setup-node@v4
+      with:
+        node-version: '24'
+        cache: 'npm'
+        cache-dependency-path: ui-svelte/package-lock.json
+
+    - name: Install dependencies
+      run: npm ci
+
+    - name: Type check
+      run: npm run check
+
+    - name: Run tests
+      run: npm test

README.md

@@ -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`
@@ -69,6 +69,7 @@ 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

@@ -87,6 +87,12 @@
             "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,

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

docker/build-container.sh

@@ -1,5 +1,7 @@
 #!/bin/bash
 
+set -euo pipefail
+
 cd $(dirname "$0")
 
 # use this to test locally, example:
@@ -8,6 +10,9 @@ cd $(dirname "$0")
 # the scopes: gist, read:org, repo, write:packages
 # then: gh auth login (and copy/paste the new token)
 
+LOG_DEBUG=${LOG_DEBUG:-0}
+DEBUG_ABORT_BUILD=${DEBUG_ABORT_BUILD:-}
+
 log_debug() {
     if [ "$LOG_DEBUG" = "1" ]; then
         echo "[DEBUG] $*"
@@ -31,7 +36,7 @@ if [[ ! " ${ALLOWED_ARCHS[@]} " =~ " ${ARCH} " ]]; then
 fi
 
 # Check if GITHUB_TOKEN is set and not empty
-if [[ -z "$GITHUB_TOKEN" ]]; then
+if [[ -z "${GITHUB_TOKEN:-}" ]]; then
   log_info "Error: GITHUB_TOKEN is not set or is empty."
   exit 1
 fi
@@ -39,6 +44,7 @@ fi
 # Set llama.cpp base image, customizable using the BASE_LLAMACPP_IMAGE environment
 # variable, this permits testing with forked llama.cpp repositories
 BASE_IMAGE=${BASE_LLAMACPP_IMAGE:-ghcr.io/ggml-org/llama.cpp}
+SD_IMAGE=${BASE_SDCPP_IMAGE:-ghcr.io/leejet/stable-diffusion.cpp}
 
 # Set llama-swap repository, automatically uses GITHUB_REPOSITORY variable
 # to enable easy container builds on forked repos
@@ -105,6 +111,8 @@ else
     BASE_TAG=server-${ARCH}-${LCPP_TAG}
 fi
 
+SD_TAG=master-${ARCH}
+
 # Abort if LCPP_TAG is empty.
 if [[ -z "$LCPP_TAG" ]]; then
     log_info "Abort: Could not find llama-server container for arch: $ARCH"
@@ -137,6 +145,18 @@ for CONTAINER_TYPE in non-root root; do
   docker build -f llama-swap.Containerfile --build-arg BASE_TAG=${BASE_TAG} --build-arg LS_VER=${LS_VER} --build-arg UID=${USER_UID} \
     --build-arg LS_REPO=${LS_REPO} --build-arg GID=${USER_GID} --build-arg USER_HOME=${USER_HOME} -t ${CONTAINER_TAG} -t ${CONTAINER_LATEST} \
     --build-arg BASE_IMAGE=${BASE_IMAGE} .
+
+  # For architectures with stable-diffusion.cpp support, layer sd-server on top
+  case "$ARCH" in
+    "musa" | "vulkan")
+      log_info "Adding sd-server to $CONTAINER_TAG"
+      docker build -f llama-swap-sd.Containerfile \
+        --build-arg BASE=${CONTAINER_TAG} \
+        --build-arg SD_IMAGE=${SD_IMAGE} --build-arg SD_TAG=${SD_TAG} \
+        --build-arg UID=${USER_UID} --build-arg GID=${USER_GID} \
+        -t ${CONTAINER_TAG} -t ${CONTAINER_LATEST} . ;;
+  esac
+
   if [ "$PUSH_IMAGES" == "true" ]; then
     docker push ${CONTAINER_TAG}
     docker push ${CONTAINER_LATEST}

docker/config.example.yaml

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

docker/llama-swap-sd.Containerfile

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

proxy/config/config.go

@@ -123,6 +123,7 @@ type Config struct {
 	LogTimeFormat      string                 `yaml:"logTimeFormat"`
 	LogToStdout        string                 `yaml:"logToStdout"`
 	MetricsMaxInMemory int                    `yaml:"metricsMaxInMemory"`
+	CaptureBuffer      int                    `yaml:"captureBuffer"`
 	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 +202,7 @@ func LoadConfigFromReader(r io.Reader) (Config, error) {
 		LogTimeFormat:      "",
 		LogToStdout:        LogToStdoutProxy,
 		MetricsMaxInMemory: 1000,
+		CaptureBuffer:      5,
 	}
 	if err = yaml.Unmarshal([]byte(yamlStr), &config); err != nil {
 		return Config{}, 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_windows_test.go

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

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/proxymanager.go

@@ -151,7 +151,7 @@ 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),
 

proxy/proxymanager_api.go

@@ -6,6 +6,7 @@ import (
 	"fmt"
 	"net/http"
 	"sort"
+	"strconv"
 	"strings"
 
 	"github.com/gin-gonic/gin"
@@ -31,6 +32,7 @@ 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)
 	}
 }
 
@@ -250,3 +252,20 @@ 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)
+}

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/playground/AudioInterface.svelte

@@ -141,9 +141,6 @@
   <!-- Model selector -->
   <div class="shrink-0 flex flex-wrap gap-2 mb-4">
     <ModelSelector bind:value={$selectedModelStore} placeholder="Select an audio model..." disabled={isTranscribing} />
-    <button class="btn" onclick={clearAll} disabled={!selectedFile && !transcriptionResult && !error}>
-      Clear
-    </button>
   </div>
 
   <!-- Empty state for no models configured -->
@@ -241,6 +238,13 @@
         >
           Transcribe
         </button>
+        <button
+          class="btn"
+          onclick={clearAll}
+          disabled={!selectedFile && !transcriptionResult && !error}
+        >
+          Clear
+        </button>
       {/if}
     </div>
   {/if}

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

@@ -1,4 +1,5 @@
 <script lang="ts">
+  import { untrack } from "svelte";
   import { Maximize2, X } from "lucide-svelte";
 
   interface Props {
@@ -45,7 +46,8 @@
   $effect(() => {
     if (isExpanded && expandedTextarea) {
       expandedTextarea.focus();
-      expandedTextarea.setSelectionRange(expandedValue.length, expandedValue.length);
+      const len = untrack(() => expandedValue.length);
+      expandedTextarea.setSelectionRange(len, len);
     }
   });
 </script>
@@ -60,7 +62,7 @@
     {disabled}
   ></textarea>
   <button
-    class="absolute top-2 right-2 p-1.5 rounded-lg opacity-0 group-hover:opacity-100 transition-opacity bg-surface/90 hover:bg-surface border border-gray-200 dark:border-white/10 shadow-sm"
+    class="absolute top-2 right-2 p-1.5 rounded-lg opacity-60 md:opacity-0 group-hover:opacity-100 transition-opacity bg-surface/90 hover:bg-surface border border-gray-200 dark:border-white/10 shadow-sm"
     onclick={openExpanded}
     title="Expand to edit"
     type="button"

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

@@ -60,6 +60,7 @@
   function clearImage() {
     generatedImage = null;
     error = null;
+    prompt = "";
   }
 
   function downloadImage() {
@@ -117,9 +118,6 @@
         <option value="1024x1792">1024x1792 (SDXL)</option>
       </optgroup>
     </select>
-    <button class="btn" onclick={clearImage} disabled={!generatedImage && !error}>
-      Clear
-    </button>
   </div>
 
   <!-- Empty state for no models configured -->
@@ -192,6 +190,13 @@
           >
             Generate
           </button>
+          <button
+            class="btn flex-1 md:flex-none"
+            onclick={clearImage}
+            disabled={!generatedImage && !error && !prompt.trim()}
+          >
+            Clear
+          </button>
         {/if}
       </div>
     </div>

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

@@ -12,7 +12,6 @@
   let inputText = $state("");
   let isGenerating = $state(false);
   let generatedAudioUrl = $state<string | null>(null);
-  let generatedText = $state<string | null>(null);
   let generatedVoice = $state<string | null>(null);
   let generatedTimestamp = $state<Date | null>(null);
   let error = $state<string | null>(null);
@@ -148,7 +147,6 @@
 
       // Create object URL for the audio blob and store metadata
       generatedAudioUrl = URL.createObjectURL(audioBlob);
-      generatedText = trimmedText;
       generatedVoice = $selectedVoiceStore;
       generatedTimestamp = new Date();
     } catch (err) {
@@ -167,18 +165,6 @@
     abortController?.abort();
   }
 
-  function clearAudio() {
-    if (generatedAudioUrl) {
-      URL.revokeObjectURL(generatedAudioUrl);
-    }
-    generatedAudioUrl = null;
-    generatedText = null;
-    generatedVoice = null;
-    generatedTimestamp = null;
-    error = null;
-    inputText = "";
-  }
-
   function clearInput() {
     inputText = "";
   }

ui-svelte/src/lib/types.ts

@@ -21,6 +21,16 @@ 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 {

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/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 } from "../lib/types";
 import { connectionState } from "./theme";
 
 const LOG_LENGTH_LIMIT = 1024 * 100; /* 100KB of log data */
@@ -172,3 +172,19 @@ 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;
+  }
+}