refactor(coderd): optimize chatdebug (#27129)

Adds a bund of optimizations to chatdebug:

In `coderd/x/chatd/chatdebug`:
- Adds a benchmark (excluding LLM and database stuff)
- Replaces string concatenation with strings.Builder when accumulating
stream parts (~105,000ns -> ~50,00ns)
- Removes double JSON encode in RecordingTransport (114,000ns ->
64,000ns)

In `coderd/util/strings`:
- Adds a benchmark for Truncate
- Removes unnecessary allocations in Truncate (~110,000ns -> 1,550ns in
truncation case, 1 alloc -> 0 allocs in no truncation case)

> 🤖 Claude helped with this.

coderd/util/strings/strings.go

@@ -64,8 +64,19 @@ func Truncate(s string, n int, opts ...TruncateOption) string {
 	if n < 1 {
 		return ""
 	}
-	runes := []rune(s)
-	if len(runes) <= n {
+
+	// Find the byte offset of the (n+1)th rune, if any; early exit
+	// avoids decoding s past what's needed.
+	runeCount := 0
+	cutoff := -1
+	for i := range s {
+		runeCount++
+		if runeCount > n {
+			cutoff = i
+			break
+		}
+	}
+	if cutoff < 0 {
 		return s
 	}
 
@@ -73,6 +84,7 @@ func Truncate(s string, n int, opts ...TruncateOption) string {
 	if options&TruncateWithEllipsis != 0 {
 		maxLen--
 	}
+	runes := []rune(s[:cutoff])
 	var sb strings.Builder
 	if options&TruncateWithFullWords != 0 {
 		// Convert the rune-safe prefix to a string, then find

coderd/util/strings/strings_test.go

@@ -1,6 +1,7 @@
 package strings_test
 
 import (
+	"bytes"
 	"fmt"
 	"testing"
 
@@ -32,6 +33,7 @@ func TestTruncate(t *testing.T) {
 		{"foo", 1, "f", nil},
 		{"foo", 0, "", nil},
 		{"foo", -1, "", nil},
+		{"", 5, "", nil},
 		{"foo bar", 7, "foo bar", []strings.TruncateOption{strings.TruncateWithEllipsis}},
 		{"foo bar", 6, "foo b…", []strings.TruncateOption{strings.TruncateWithEllipsis}},
 		{"foo bar", 5, "foo …", []strings.TruncateOption{strings.TruncateWithEllipsis}},
@@ -81,6 +83,28 @@ func TestTruncate(t *testing.T) {
 	}
 }
 
+func BenchmarkTruncate(b *testing.B) {
+	b.Run("NoTruncationNeeded", func(b *testing.B) {
+		s := "a short string well under the limit"
+		b.ReportAllocs()
+		for b.Loop() {
+			strings.Truncate(s, 1000)
+		}
+	})
+
+	b.Run("ActualTruncation", func(b *testing.B) {
+		var buf bytes.Buffer
+		for range 2000 {
+			buf.WriteString("日本語テスト word ")
+		}
+		s := buf.String()
+		b.ReportAllocs()
+		for b.Loop() {
+			strings.Truncate(s, 100, strings.TruncateWithEllipsis, strings.TruncateWithFullWords)
+		}
+	})
+}
+
 func TestUISanitize(t *testing.T) {
 	t.Parallel()
 

coderd/x/chatd/chatdebug/bench_internal_test.go

@@ -0,0 +1,350 @@
+package chatdebug
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"fmt"
+	"io"
+	"net/http"
+	"slices"
+	"strings"
+	"testing"
+
+	"charm.land/fantasy"
+	"github.com/google/uuid"
+
+	"cdr.dev/slog/v3"
+	"github.com/coder/coder/v2/coderd/database"
+	"github.com/coder/coder/v2/coderd/database/pubsub"
+	"github.com/coder/coder/v2/coderd/x/chatd/chattest"
+)
+
+// noopStore implements database.Store by embedding it (nil) and
+// overriding only the methods chatdebug's hot paths call. Every
+// override returns immediately with a canned value and does no I/O,
+// so any time spent inside Service methods that use this store is
+// chatdebug's own CPU/allocation cost, not Postgres's.
+type noopStore struct {
+	database.Store
+}
+
+func (noopStore) GetChatDebugLoggingAllowUsers(context.Context) (bool, error) {
+	return true, nil
+}
+
+func (noopStore) GetUserChatDebugLoggingEnabled(context.Context, uuid.UUID) (bool, error) {
+	return true, nil
+}
+
+func (noopStore) InsertChatDebugStep(
+	_ context.Context,
+	arg database.InsertChatDebugStepParams,
+) (database.ChatDebugStep, error) {
+	return database.ChatDebugStep{
+		ID:         uuid.New(),
+		RunID:      arg.RunID,
+		ChatID:     arg.ChatID,
+		StepNumber: arg.StepNumber,
+		Operation:  arg.Operation,
+		Status:     arg.Status,
+	}, nil
+}
+
+func (noopStore) UpdateChatDebugStep(
+	_ context.Context,
+	arg database.UpdateChatDebugStepParams,
+) (database.ChatDebugStep, error) {
+	return database.ChatDebugStep{
+		ID:     arg.ID,
+		ChatID: arg.ChatID,
+	}, nil
+}
+
+// TouchChatDebugStepAndRun and GetChatDebugStepsByRunID aren't on the
+// hot path these benchmarks exercise (heartbeat interval and step
+// retry-collision handling respectively), but overriding them avoids
+// a nil-dereference panic if a future benchmark configuration change
+// reaches them through the embedded nil database.Store.
+func (noopStore) TouchChatDebugStepAndRun(context.Context, database.TouchChatDebugStepAndRunParams) error {
+	return nil
+}
+
+func (noopStore) GetChatDebugStepsByRunID(context.Context, uuid.UUID) ([]database.ChatDebugStep, error) {
+	return nil, nil
+}
+
+var _ database.Store = noopStore{}
+
+// benchService builds a Service backed by noopStore and an in-memory
+// pubsub so publishEvent still marshals and dispatches DebugEvent
+// payloads, mirroring a real deployment's wiring, without any network
+// or disk I/O.
+func benchService(b *testing.B) *Service {
+	b.Helper()
+	ps := pubsub.NewInMemory()
+	b.Cleanup(func() { _ = ps.Close() })
+	return NewService(noopStore{}, slog.Make(), ps)
+}
+
+// benchCall builds a realistic multi-turn prompt with tool
+// definitions so normalizeCall walks a representative message/tool
+// shape.
+func benchCall(nMessages, nTools int) fantasy.Call {
+	prompt := make(fantasy.Prompt, 0, nMessages)
+	for i := range nMessages {
+		if i%2 == 0 {
+			prompt = append(prompt, fantasy.NewUserMessage(
+				fmt.Sprintf("Please investigate issue #%d and summarize the relevant log lines around the failure.", i)))
+			continue
+		}
+		prompt = append(prompt, fantasy.Message{
+			Role: fantasy.MessageRoleAssistant,
+			Content: []fantasy.MessagePart{
+				fantasy.TextPart{Text: strings.Repeat("analysis details ", 40)},
+			},
+		})
+	}
+
+	tools := make([]fantasy.Tool, 0, nTools)
+	for i := range nTools {
+		tools = append(tools, fantasy.FunctionTool{
+			Name:        fmt.Sprintf("tool_%d", i),
+			Description: "Runs a workspace command and returns its output",
+			InputSchema: map[string]any{
+				"type": "object",
+				"properties": map[string]any{
+					"command": map[string]any{"type": "string"},
+					"timeout": map[string]any{"type": "integer"},
+				},
+				"required": []string{"command"},
+			},
+		})
+	}
+
+	return fantasy.Call{
+		Prompt: prompt,
+		Tools:  tools,
+	}
+}
+
+func benchStreamParts(nDeltas int) []fantasy.StreamPart {
+	parts := make([]fantasy.StreamPart, 0, nDeltas+2)
+	for i := range nDeltas {
+		parts = append(parts, fantasy.StreamPart{
+			Type:  fantasy.StreamPartTypeTextDelta,
+			ID:    "text-1",
+			Delta: fmt.Sprintf("token-%d ", i),
+		})
+	}
+	parts = append(parts,
+		fantasy.StreamPart{Type: fantasy.StreamPartTypeToolCall, ID: "tool-1", ToolCallName: "tool_0"},
+		fantasy.StreamPart{
+			Type: fantasy.StreamPartTypeFinish, FinishReason: fantasy.FinishReasonStop,
+			Usage: fantasy.Usage{InputTokens: 512, OutputTokens: int64(nDeltas), TotalTokens: 512 + int64(nDeltas)},
+		},
+	)
+	return parts
+}
+
+// runContext returns a fresh RunContext/context pair for one
+// benchmark iteration, mirroring a new chat turn.
+func runContext(chatID uuid.UUID) context.Context {
+	return ContextWithRun(context.Background(), &RunContext{RunID: uuid.New(), ChatID: chatID})
+}
+
+func BenchmarkWrapModel_Generate(b *testing.B) {
+	svc := benchService(b)
+	chatID, ownerID := uuid.New(), uuid.New()
+	call := benchCall(6, 4)
+	resp := &fantasy.Response{
+		Content: fantasy.ResponseContent{
+			fantasy.TextContent{Text: strings.Repeat("response text ", 100)},
+			fantasy.ToolCallContent{ToolCallID: "tool-1", ToolName: "tool_0", Input: `{"command":"ls"}`},
+		},
+		FinishReason: fantasy.FinishReasonStop,
+		Usage:        fantasy.Usage{InputTokens: 512, OutputTokens: 128, TotalTokens: 640},
+	}
+	inner := &chattest.FakeModel{GenerateFn: func(context.Context, fantasy.Call) (*fantasy.Response, error) {
+		return resp, nil
+	}}
+	model := WrapModel(inner, svc, RecorderOptions{ChatID: chatID, OwnerID: ownerID})
+
+	b.ReportAllocs()
+	for b.Loop() {
+		ctx := runContext(chatID)
+		if _, err := model.Generate(ctx, call); err != nil {
+			b.Fatal(err)
+		}
+	}
+}
+
+func BenchmarkWrapModel_Stream(b *testing.B) {
+	svc := benchService(b)
+	chatID, ownerID := uuid.New(), uuid.New()
+	call := benchCall(6, 4)
+	parts := benchStreamParts(200)
+	inner := &chattest.FakeModel{StreamFn: func(context.Context, fantasy.Call) (fantasy.StreamResponse, error) {
+		return slices.Values(parts), nil
+	}}
+	model := WrapModel(inner, svc, RecorderOptions{ChatID: chatID, OwnerID: ownerID})
+
+	b.ReportAllocs()
+	for b.Loop() {
+		ctx := runContext(chatID)
+		seq, err := model.Stream(ctx, call)
+		if err != nil {
+			b.Fatal(err)
+		}
+		for range seq { //nolint:revive // draining the stream is the point of the benchmark.
+		}
+	}
+}
+
+// cannedRoundTripper replays a fixed response body for every request,
+// isolating RecordingTransport's own redaction/buffering cost from any
+// real network or provider latency.
+type cannedRoundTripper struct {
+	status    int
+	header    http.Header
+	body      []byte
+	chunkSize int // 0 means return the whole body in one Read.
+}
+
+func (rt *cannedRoundTripper) RoundTrip(req *http.Request) (*http.Response, error) {
+	var body io.ReadCloser
+	if rt.chunkSize > 0 {
+		body = io.NopCloser(&chunkedReader{data: rt.body, chunkSize: rt.chunkSize})
+	} else {
+		body = io.NopCloser(bytes.NewReader(rt.body))
+	}
+	header := rt.header.Clone()
+	return &http.Response{
+		StatusCode:    rt.status,
+		Header:        header,
+		Body:          body,
+		ContentLength: -1,
+		Request:       req,
+	}, nil
+}
+
+// chunkedReader splits data into fixed-size Read() calls so the
+// benchmark exercises recordingBody's incremental accumulation path
+// the way a real SSE stream would, instead of handing back the whole
+// body in a single Read.
+type chunkedReader struct {
+	data      []byte
+	chunkSize int
+}
+
+func (c *chunkedReader) Read(p []byte) (int, error) {
+	if len(c.data) == 0 {
+		return 0, io.EOF
+	}
+	n := min(c.chunkSize, len(c.data), len(p))
+	copy(p, c.data[:n])
+	c.data = c.data[n:]
+	return n, nil
+}
+
+func benchJSONResponseBody(n int) []byte {
+	type choice struct {
+		Index int    `json:"index"`
+		Text  string `json:"text"`
+	}
+	payload := struct {
+		ID      string   `json:"id"`
+		Object  string   `json:"object"`
+		Model   string   `json:"model"`
+		Choices []choice `json:"choices"`
+		Usage   struct {
+			PromptTokens     int `json:"prompt_tokens"`
+			CompletionTokens int `json:"completion_tokens"`
+		} `json:"usage"`
+	}{
+		ID:     "resp-bench",
+		Object: "chat.completion",
+		Model:  "bench-model",
+	}
+	for i := range n {
+		payload.Choices = append(payload.Choices, choice{
+			Index: i,
+			Text:  strings.Repeat("word ", 50),
+		})
+	}
+	data, err := json.Marshal(payload)
+	if err != nil {
+		panic(err)
+	}
+	return data
+}
+
+func benchSSEResponseBody(nEvents int) []byte {
+	var buf bytes.Buffer
+	for i := range nEvents {
+		// bytes.Buffer.Write* never returns an error.
+		_, _ = fmt.Fprintf(&buf, "data: {\"delta\":\"token-%d \",\"index\":%d}\n\n", i, i)
+	}
+	_, _ = buf.WriteString("data: [DONE]\n\n")
+	return buf.Bytes()
+}
+
+func newBenchRequest(b *testing.B, sink *attemptSink) *http.Request {
+	b.Helper()
+	req, err := http.NewRequestWithContext(
+		withAttemptSink(context.Background(), sink),
+		http.MethodPost,
+		"https://api.example.com/v1/chat/completions",
+		bytes.NewReader([]byte(`{"model":"bench-model","messages":[{"role":"user","content":"hi"}]}`)),
+	)
+	if err != nil {
+		b.Fatal(err)
+	}
+	req.Header.Set("Authorization", "Bearer sk-should-be-redacted")
+	req.Header.Set("Content-Type", "application/json")
+	return req
+}
+
+func BenchmarkRecordingTransport_RoundTrip_JSON(b *testing.B) {
+	header := http.Header{"Content-Type": []string{"application/json"}}
+	body := benchJSONResponseBody(20)
+	transport := &RecordingTransport{Base: &cannedRoundTripper{status: 200, header: header, body: body}}
+
+	b.ReportAllocs()
+	for b.Loop() {
+		sink := &attemptSink{}
+		req := newBenchRequest(b, sink)
+		resp, err := transport.RoundTrip(req)
+		if err != nil {
+			b.Fatal(err)
+		}
+		if _, err := io.Copy(io.Discard, resp.Body); err != nil {
+			b.Fatal(err)
+		}
+		if err := resp.Body.Close(); err != nil {
+			b.Fatal(err)
+		}
+	}
+}
+
+func BenchmarkRecordingTransport_RoundTrip_SSE(b *testing.B) {
+	header := http.Header{"Content-Type": []string{"text/event-stream"}}
+	body := benchSSEResponseBody(200)
+	transport := &RecordingTransport{Base: &cannedRoundTripper{status: 200, header: header, body: body, chunkSize: 256}}
+
+	b.ReportAllocs()
+	for b.Loop() {
+		sink := &attemptSink{}
+		req := newBenchRequest(b, sink)
+		resp, err := transport.RoundTrip(req)
+		if err != nil {
+			b.Fatal(err)
+		}
+		if _, err := io.Copy(io.Discard, resp.Body); err != nil {
+			b.Fatal(err)
+		}
+		if err := resp.Body.Close(); err != nil {
+			b.Fatal(err)
+		}
+	}
+}

coderd/x/chatd/chatdebug/model.go

@@ -7,6 +7,7 @@ import (
 	"fmt"
 	"iter"
 	"reflect"
+	"strings"
 	"sync"
 	"sync/atomic"
 	"unicode/utf8"
@@ -460,6 +461,9 @@ func wrapStreamSeq(
 			finishSeen       bool
 			finishReason     fantasy.FinishReason
 			content          []normalizedContentPart
+			currentText      *strings.Builder
+			currentTextIdx   int
+			argBuilders      = make(map[string]*strings.Builder)
 			warnings         []normalizedWarning
 			streamDebugBytes int
 			streamError      any
@@ -480,6 +484,7 @@ func wrapStreamSeq(
 			close(heartbeatDone)
 
 			summary.FinishReason = string(finishReason)
+			materializeStreamContent(content, currentText, currentTextIdx, argBuilders)
 
 			resp := normalizedResponsePayload{
 				Content:      content,
@@ -528,7 +533,8 @@ func wrapStreamSeq(
 					streamComplete.Store(true)
 				}
 
-				content = appendNormalizedStreamContent(content, part, &streamDebugBytes)
+				content, currentText, currentTextIdx = appendNormalizedStreamContent(
+					content, currentText, currentTextIdx, argBuilders, part, &streamDebugBytes)
 
 				if part.Type == fantasy.StreamPartTypeError || part.Error != nil {
 					summary.ErrorCount++
@@ -758,14 +764,18 @@ func safeMarshalJSON(label string, value any) json.RawMessage {
 	return append(json.RawMessage(nil), data...)
 }
 
+// appendStreamContentText accumulates a delta into the current
+// text/reasoning run.
 func appendStreamContentText(
 	content []normalizedContentPart,
+	currentText *strings.Builder,
+	currentTextIdx int,
 	partType string,
 	delta string,
 	streamDebugBytes *int,
-) []normalizedContentPart {
+) ([]normalizedContentPart, *strings.Builder, int) {
 	if delta == "" {
-		return content
+		return content, currentText, currentTextIdx
 	}
 
 	remaining := maxStreamDebugTextBytes
@@ -773,7 +783,7 @@ func appendStreamContentText(
 		remaining -= *streamDebugBytes
 	}
 	if remaining <= 0 {
-		return content
+		return content, currentText, currentTextIdx
 	}
 	if len(delta) > remaining {
 		cut := 0
@@ -790,18 +800,26 @@ func appendStreamContentText(
 		delta = delta[:cut]
 	}
 	if delta == "" {
-		return content
+		return content, currentText, currentTextIdx
 	}
 
 	if len(content) == 0 || content[len(content)-1].Type != partType {
+		// A different-typed part is interrupting (or this is the
+		// first part): flush the outgoing run into content now, since
+		// currentText is about to be replaced and materializeStreamContent
+		// only ever sees whichever run is still open at the end.
+		if currentText != nil && currentTextIdx < len(content) {
+			content[currentTextIdx].Text = currentText.String()
+		}
 		content = append(content, normalizedContentPart{Type: partType})
+		currentText = &strings.Builder{}
+		currentTextIdx = len(content) - 1
 	}
-	last := &content[len(content)-1]
-	last.Text += delta
+	_, _ = currentText.WriteString(delta)
 	if streamDebugBytes != nil {
 		*streamDebugBytes += len(delta)
 	}
-	return content
+	return content, currentText, currentTextIdx
 }
 
 // appendStreamToolInput accumulates incremental tool-input deltas
@@ -809,6 +827,7 @@ func appendStreamContentText(
 // remain distinguishable in interrupted stream debug payloads.
 func appendStreamToolInput(
 	content []normalizedContentPart,
+	argBuilders map[string]*strings.Builder,
 	part fantasy.StreamPart,
 	streamDebugBytes *int,
 ) []normalizedContentPart {
@@ -842,26 +861,19 @@ func appendStreamToolInput(
 		return content
 	}
 
-	// Find the existing tool_input part for this specific tool call ID.
-	// Scan backwards through all content; tool_input deltas for the
-	// same call may be separated by text, reasoning, or source parts
-	// when streams interleave multiple tool invocations.
-	for i := len(content) - 1; i >= 0; i-- {
-		if content[i].Type == "tool_input" && content[i].ToolCallID == part.ID {
-			content[i].Arguments += delta
-			if streamDebugBytes != nil {
-				*streamDebugBytes += len(delta)
-			}
-			return content
-		}
+	// argBuilders indexes by ToolCallID directly, so resuming an
+	// interleaved call needs no scan through content.
+	b := argBuilders[part.ID]
+	if b == nil {
+		b = &strings.Builder{}
+		argBuilders[part.ID] = b
+		content = append(content, normalizedContentPart{
+			Type:       "tool_input",
+			ToolCallID: part.ID,
+			ToolName:   part.ToolCallName,
+		})
 	}
-
-	content = append(content, normalizedContentPart{
-		Type:       "tool_input",
-		ToolCallID: part.ID,
-		ToolName:   part.ToolCallName,
-		Arguments:  delta,
-	})
+	_, _ = b.WriteString(delta)
 	if streamDebugBytes != nil {
 		*streamDebugBytes += len(delta)
 	}
@@ -879,16 +891,42 @@ func canonicalContentType(partType string) string {
 	}
 }
 
+// materializeStreamContent copies currentText (if any) and each
+// argBuilders entry into the corresponding content part's
+// Text/Arguments. It should be called once, after streaming completes
+// and before content is read. It is safe to call more than once.
+func materializeStreamContent(
+	content []normalizedContentPart,
+	currentText *strings.Builder,
+	currentTextIdx int,
+	argBuilders map[string]*strings.Builder,
+) {
+	if currentText != nil && currentTextIdx < len(content) {
+		content[currentTextIdx].Text = currentText.String()
+	}
+	for i := range content {
+		if content[i].Type != "tool_input" {
+			continue
+		}
+		if b, ok := argBuilders[content[i].ToolCallID]; ok {
+			content[i].Arguments = b.String()
+		}
+	}
+}
+
 func appendNormalizedStreamContent(
 	content []normalizedContentPart,
+	currentText *strings.Builder,
+	currentTextIdx int,
+	argBuilders map[string]*strings.Builder,
 	part fantasy.StreamPart,
 	streamDebugBytes *int,
-) []normalizedContentPart {
+) ([]normalizedContentPart, *strings.Builder, int) {
 	switch part.Type {
 	case fantasy.StreamPartTypeTextDelta:
-		return appendStreamContentText(content, "text", part.Delta, streamDebugBytes)
+		return appendStreamContentText(content, currentText, currentTextIdx, "text", part.Delta, streamDebugBytes)
 	case fantasy.StreamPartTypeReasoningStart, fantasy.StreamPartTypeReasoningDelta:
-		return appendStreamContentText(content, "reasoning", part.Delta, streamDebugBytes)
+		return appendStreamContentText(content, currentText, currentTextIdx, "reasoning", part.Delta, streamDebugBytes)
 	case fantasy.StreamPartTypeToolInputStart,
 		fantasy.StreamPartTypeToolInputDelta,
 		fantasy.StreamPartTypeToolInputEnd:
@@ -896,31 +934,35 @@ func appendNormalizedStreamContent(
 		// tool_call summary. Attribute each chunk to its tool call
 		// so interrupted streams can reconstruct which partial input
 		// belonged to which invocation.
-		return appendStreamToolInput(content, part, streamDebugBytes)
+		content = appendStreamToolInput(content, argBuilders, part, streamDebugBytes)
+		return content, currentText, currentTextIdx
 	case fantasy.StreamPartTypeToolCall:
-		return append(content, normalizedContentPart{
+		content = append(content, normalizedContentPart{
 			Type:        canonicalContentType(string(part.Type)),
 			ToolCallID:  part.ID,
 			ToolName:    part.ToolCallName,
 			Arguments:   boundText(part.ToolCallInput),
 			InputLength: utf8.RuneCountInString(part.ToolCallInput),
 		})
+		return content, currentText, currentTextIdx
 	case fantasy.StreamPartTypeToolResult:
-		return append(content, normalizedContentPart{
+		content = append(content, normalizedContentPart{
 			Type:       canonicalContentType(string(part.Type)),
 			ToolCallID: part.ID,
 			ToolName:   part.ToolCallName,
 			Result:     boundText(part.ToolCallInput),
 		})
+		return content, currentText, currentTextIdx
 	case fantasy.StreamPartTypeSource:
-		return append(content, normalizedContentPart{
+		content = append(content, normalizedContentPart{
 			Type:       string(part.Type),
 			SourceType: string(part.SourceType),
 			Title:      part.Title,
 			URL:        part.URL,
 		})
+		return content, currentText, currentTextIdx
 	default:
-		return content
+		return content, currentText, currentTextIdx
 	}
 }
 

coderd/x/chatd/chatdebug/model_normalization_internal_test.go

@@ -132,16 +132,30 @@ func TestNormalizers_SkipTypedNilInterfaceValues(t *testing.T) {
 func TestAppendNormalizedStreamContent_PreservesOrderAndCanonicalTypes(t *testing.T) {
 	t.Parallel()
 
-	var content []normalizedContentPart
+	var (
+		content        []normalizedContentPart
+		currentText    *strings.Builder
+		currentTextIdx int
+	)
+	argBuilders := make(map[string]*strings.Builder)
 	streamDebugBytes := 0
 	for _, part := range []fantasy.StreamPart{
-		{Type: fantasy.StreamPartTypeTextDelta, Delta: "before "},
+		// "before " and "after" each arrive as two consecutive
+		// same-type deltas (rather than one) so this also exercises
+		// accumulating into an already-existing text part across
+		// multiple appendNormalizedStreamContent calls, not just
+		// starting a fresh one.
+		{Type: fantasy.StreamPartTypeTextDelta, Delta: "bef"},
+		{Type: fantasy.StreamPartTypeTextDelta, Delta: "ore "},
 		{Type: fantasy.StreamPartTypeToolCall, ID: "call-1", ToolCallName: "search_docs", ToolCallInput: `{"query":"debug"}`},
 		{Type: fantasy.StreamPartTypeToolResult, ID: "call-1", ToolCallName: "search_docs", ToolCallInput: `{"matches":1}`},
-		{Type: fantasy.StreamPartTypeTextDelta, Delta: "after"},
+		{Type: fantasy.StreamPartTypeTextDelta, Delta: "aft"},
+		{Type: fantasy.StreamPartTypeTextDelta, Delta: "er"},
 	} {
-		content = appendNormalizedStreamContent(content, part, &streamDebugBytes)
+		content, currentText, currentTextIdx = appendNormalizedStreamContent(
+			content, currentText, currentTextIdx, argBuilders, part, &streamDebugBytes)
 	}
+	materializeStreamContent(content, currentText, currentTextIdx, argBuilders)
 
 	require.Equal(t, []normalizedContentPart{
 		{Type: "text", Text: "before "},
@@ -154,7 +168,12 @@ func TestAppendNormalizedStreamContent_PreservesOrderAndCanonicalTypes(t *testin
 func TestAppendNormalizedStreamContent_ToolInputAttributionPerCall(t *testing.T) {
 	t.Parallel()
 
-	var content []normalizedContentPart
+	var (
+		content        []normalizedContentPart
+		currentText    *strings.Builder
+		currentTextIdx int
+	)
+	argBuilders := make(map[string]*strings.Builder)
 	streamDebugBytes := 0
 	for _, part := range []fantasy.StreamPart{
 		{Type: fantasy.StreamPartTypeToolInputStart, ID: "call-a", ToolCallName: "search", Delta: `{"q`},
@@ -164,8 +183,10 @@ func TestAppendNormalizedStreamContent_ToolInputAttributionPerCall(t *testing.T)
 		{Type: fantasy.StreamPartTypeToolInputDelta, ID: "call-a", ToolCallName: "search", Delta: `":"x"}`},
 		{Type: fantasy.StreamPartTypeToolInputEnd, ID: "call-b", ToolCallName: "calc", Delta: `":"add"}`},
 	} {
-		content = appendNormalizedStreamContent(content, part, &streamDebugBytes)
+		content, currentText, currentTextIdx = appendNormalizedStreamContent(
+			content, currentText, currentTextIdx, argBuilders, part, &streamDebugBytes)
 	}
+	materializeStreamContent(content, currentText, currentTextIdx, argBuilders)
 
 	require.Equal(t, []normalizedContentPart{
 		{Type: "tool_input", ToolCallID: "call-a", ToolName: "search", Arguments: `{"query":"x"}`},
@@ -176,7 +197,12 @@ func TestAppendNormalizedStreamContent_ToolInputAttributionPerCall(t *testing.T)
 func TestAppendNormalizedStreamContent_ToolInputAcrossInterleavedText(t *testing.T) {
 	t.Parallel()
 
-	var content []normalizedContentPart
+	var (
+		content        []normalizedContentPart
+		currentText    *strings.Builder
+		currentTextIdx int
+	)
+	argBuilders := make(map[string]*strings.Builder)
 	streamDebugBytes := 0
 	for _, part := range []fantasy.StreamPart{
 		{Type: fantasy.StreamPartTypeToolInputStart, ID: "call-a", ToolCallName: "search", Delta: `{"q`},
@@ -184,8 +210,10 @@ func TestAppendNormalizedStreamContent_ToolInputAcrossInterleavedText(t *testing
 		{Type: fantasy.StreamPartTypeTextDelta, Delta: "thinking..."},
 		{Type: fantasy.StreamPartTypeToolInputDelta, ID: "call-a", ToolCallName: "search", Delta: `uery":"x"}`},
 	} {
-		content = appendNormalizedStreamContent(content, part, &streamDebugBytes)
+		content, currentText, currentTextIdx = appendNormalizedStreamContent(
+			content, currentText, currentTextIdx, argBuilders, part, &streamDebugBytes)
 	}
+	materializeStreamContent(content, currentText, currentTextIdx, argBuilders)
 
 	require.Equal(t, []normalizedContentPart{
 		{Type: "tool_input", ToolCallID: "call-a", ToolName: "search", Arguments: `{"query":"x"}`},
@@ -198,14 +226,21 @@ func TestAppendNormalizedStreamContent_GlobalTextCap(t *testing.T) {
 
 	streamDebugBytes := 0
 	long := strings.Repeat("a", maxStreamDebugTextBytes)
-	var content []normalizedContentPart
+	var (
+		content        []normalizedContentPart
+		currentText    *strings.Builder
+		currentTextIdx int
+	)
+	argBuilders := make(map[string]*strings.Builder)
 	for _, part := range []fantasy.StreamPart{
 		{Type: fantasy.StreamPartTypeTextDelta, Delta: long},
 		{Type: fantasy.StreamPartTypeToolCall, ID: "call-1", ToolCallName: "search_docs", ToolCallInput: `{}`},
 		{Type: fantasy.StreamPartTypeTextDelta, Delta: "tail"},
 	} {
-		content = appendNormalizedStreamContent(content, part, &streamDebugBytes)
+		content, currentText, currentTextIdx = appendNormalizedStreamContent(
+			content, currentText, currentTextIdx, argBuilders, part, &streamDebugBytes)
 	}
+	materializeStreamContent(content, currentText, currentTextIdx, argBuilders)
 
 	require.Len(t, content, 2)
 	require.Equal(t, strings.Repeat("a", maxStreamDebugTextBytes), content[0].Text)

coderd/x/chatd/chatdebug/redaction.go

@@ -88,6 +88,44 @@ func RedactHeaders(h http.Header) map[string]string {
 	return redacted
 }
 
+var (
+	errNotValidJSON  = xerrors.New("chatdebug: body is not valid JSON")
+	errExtraJSONData = xerrors.New("chatdebug: body contains extra JSON values")
+)
+
+func decodeCompleteJSON(data []byte) (any, error) {
+	decoder := json.NewDecoder(bytes.NewReader(data))
+	// Preserve precision: callers may re-marshal the result.
+	decoder.UseNumber()
+
+	var value any
+	if err := decoder.Decode(&value); err != nil {
+		return nil, errNotValidJSON
+	}
+	var extra any
+	if err := decoder.Decode(&extra); !errors.Is(err, io.EOF) {
+		return nil, errExtraJSONData
+	}
+	return value, nil
+}
+
+// redactDecodedJSON redacts an already-decoded JSON value. ok is
+// false when nothing needed redacting or re-marshaling the redacted
+// value failed; either way the caller should fall back to the
+// original bytes.
+func redactDecodedJSON(value any) (encoded []byte, ok bool) {
+	redacted, changed := redactJSONValue(value)
+	if !changed {
+		return nil, false
+	}
+
+	data, err := json.Marshal(redacted)
+	if err != nil {
+		return nil, false
+	}
+	return data, true
+}
+
 // RedactJSONSecrets redacts sensitive JSON values by key name. When
 // the input is not valid JSON (truncated body, HTML error page, etc.)
 // the raw bytes are replaced entirely with a diagnostic placeholder
@@ -97,29 +135,20 @@ func RedactJSONSecrets(data []byte) []byte {
 		return data
 	}
 
-	decoder := json.NewDecoder(bytes.NewReader(data))
-	decoder.UseNumber()
-
-	var value any
-	if err := decoder.Decode(&value); err != nil {
+	value, err := decodeCompleteJSON(data)
+	if err != nil {
+		if errors.Is(err, errExtraJSONData) {
+			return []byte(`{"error":"chatdebug: body contains extra JSON values, redacted for safety"}`)
+		}
 		// Cannot parse: replace entirely to prevent credential leaks
 		// from non-JSON error responses (HTML pages, partial bodies).
 		return []byte(`{"error":"chatdebug: body is not valid JSON, redacted for safety"}`)
 	}
-	if err := consumeJSONEOF(decoder); err != nil {
-		return []byte(`{"error":"chatdebug: body contains extra JSON values, redacted for safety"}`)
-	}
 
-	redacted, changed := redactJSONValue(value)
-	if !changed {
-		return data
+	if encoded, changed := redactDecodedJSON(value); changed {
+		return encoded
 	}
-
-	encoded, err := json.Marshal(redacted)
-	if err != nil {
-		return data
-	}
-	return encoded
+	return data
 }
 
 // RedactNDJSONSecrets redacts sensitive values in newline-delimited
@@ -150,18 +179,6 @@ func RedactNDJSONSecrets(data []byte) []byte {
 	return bytes.Join(lines, []byte("\n"))
 }
 
-func consumeJSONEOF(decoder *json.Decoder) error {
-	var extra any
-	err := decoder.Decode(&extra)
-	if errors.Is(err, io.EOF) {
-		return nil
-	}
-	if err == nil {
-		return xerrors.New("chatdebug: extra JSON values")
-	}
-	return err
-}
-
 // safeRateLimitHeaderNames lists rate-limit headers that contain
 // "token" in the name but carry numeric usage counters, not
 // credentials. They are checked in isSensitiveName before the

coderd/x/chatd/chatdebug/transport.go

@@ -2,7 +2,6 @@ package chatdebug
 
 import (
 	"bytes"
-	"encoding/json"
 	"errors"
 	"io"
 	"mime"
@@ -177,6 +176,7 @@ func captureRequestBody(req *http.Request) ([]byte, error) {
 			if len(limited) > maxRecordedRequestBodyBytes {
 				return []byte("[TRUNCATED]"), nil
 			}
+			// Request bodies have no completeness pre-check to reuse; always decode fresh.
 			return RedactJSONSecrets(limited), nil
 		}
 	}
@@ -316,12 +316,12 @@ func (r *recordingBody) Close() error {
 			// The SSE EOF path already appended a completed attempt.
 			// inner.Close() surfaced a transport error, so upgrade
 			// that entry to failed instead of losing the close error.
-			upgraded := r.buildAttemptLocked(closeErr)
+			upgraded := r.buildAttemptLocked(closeErr, nil)
 			r.sink.replaceByNumber(upgraded.Number, upgraded)
 			r.recordedProvisional = false
 		} else {
 			r.recordOnce.Do(func() {
-				r.sink.record(r.buildAttemptLocked(closeErr))
+				r.sink.record(r.buildAttemptLocked(closeErr, nil))
 			})
 		}
 		r.mu.Unlock()
@@ -344,20 +344,35 @@ func (r *recordingBody) Close() error {
 	responseBody = append([]byte(nil), r.buf.Bytes()...)
 	r.mu.Unlock()
 
+	// decodedJSON is nil unless the completeness check below decoded
+	// these bytes; non-nil even for a stored nil value (JSON "null").
+	// A truncated buffer is an incomplete prefix, so skip the check.
+	var decodedJSON *any
+	if contentLength < 0 && !truncated {
+		if value, ok := decodeUnknownLengthJSONBody(contentType, responseBody); ok {
+			decodedJSON = &value
+		}
+	}
+
+	recordReusingDecodedJSON := func(err error) {
+		if decodedJSON != nil {
+			r.recordJSON(err, *decodedJSON)
+			return
+		}
+		r.record(err)
+	}
+
 	switch {
-	// Only check JSON completeness when the recording buffer is
-	// not truncated. A truncated buffer is an incomplete prefix
-	// of the body, so the completeness check would false-positive.
-	case sawEOF && !truncated && contentLength < 0 && isJSONLikeContentType(contentType) && !isCompleteUnknownLengthJSONBody(contentType, responseBody):
+	case sawEOF && !truncated && contentLength < 0 && isJSONLikeContentType(contentType) && decodedJSON == nil:
 		r.record(io.ErrUnexpectedEOF)
 	case sawEOF:
-		r.record(io.EOF)
+		recordReusingDecodedJSON(io.EOF)
 	case responseHasNoBody(r.base.Method, r.base.ResponseStatus):
 		r.record(nil)
 	case contentLength >= 0 && bytesRead >= contentLength:
 		r.record(nil)
-	case contentLength < 0 && !truncated && isCompleteUnknownLengthJSONBody(contentType, responseBody):
-		r.record(nil)
+	case contentLength < 0 && !truncated && decodedJSON != nil:
+		recordReusingDecodedJSON(nil)
 	// Truncated unknown-length bodies: the caller consumed the
 	// response successfully but the recording buffer exceeded
 	// maxRecordedResponseBodyBytes. This is not a transport
@@ -437,30 +452,39 @@ func (r *recordingBody) drainToEOF() error {
 }
 
 func isCompleteUnknownLengthJSONBody(contentType string, body []byte) bool {
+	_, ok := decodeUnknownLengthJSONBody(contentType, body)
+	return ok
+}
+
+// decodeUnknownLengthJSONBody reports whether body is a single,
+// complete JSON document for the given content type. It returns the
+// decoded value too, to avoid re-decoding during redaction.
+func decodeUnknownLengthJSONBody(contentType string, body []byte) (any, bool) {
 	if !isJSONLikeContentType(contentType) {
-		return false
+		return nil, false
 	}
 
 	trimmed := bytes.TrimSpace(body)
 	if len(trimmed) == 0 {
-		return false
+		return nil, false
 	}
 
-	decoder := json.NewDecoder(bytes.NewReader(trimmed))
-	var value any
-	if err := decoder.Decode(&value); err != nil {
-		return false
+	value, err := decodeCompleteJSON(trimmed)
+	if err != nil {
+		return nil, false
 	}
-	var extra any
-	return errors.Is(decoder.Decode(&extra), io.EOF)
+	return value, true
 }
 
 // buildAttemptLocked materializes the final Attempt from the current
-// buffered response data plus err. Callers use this from both the
-// record-once append path and the provisional-upgrade replace path so
-// both sites apply the same redaction and status rules. The caller
-// must hold r.mu for the duration of the call.
-func (r *recordingBody) buildAttemptLocked(err error) Attempt {
+// buffered response data plus err. The caller must hold r.mu for the
+// duration of the call.
+//
+// decodedJSON is nil when nothing was decoded ahead of time. A
+// non-nil decodedJSON - even one pointing at a stored nil interface,
+// e.g. a JSON "null" body - is redacted directly instead of decoding
+// the raw bytes again.
+func (r *recordingBody) buildAttemptLocked(err error, decodedJSON *any) Attempt {
 	finishedAt := time.Now()
 
 	truncated := r.truncated
@@ -472,6 +496,12 @@ func (r *recordingBody) buildAttemptLocked(err error) Attempt {
 	switch {
 	case truncated:
 		base.ResponseBody = []byte("[TRUNCATED]")
+	case decodedJSON != nil:
+		if encoded, changed := redactDecodedJSON(*decodedJSON); changed {
+			base.ResponseBody = encoded
+		} else {
+			base.ResponseBody = responseBody
+		}
 	case isNDJSONContentType(contentType):
 		base.ResponseBody = RedactNDJSONSecrets(responseBody)
 	case contentType == "" || isJSONLikeContentType(contentType):
@@ -508,7 +538,17 @@ func (r *recordingBody) record(err error) {
 	r.mu.Lock()
 	defer r.mu.Unlock()
 	r.recordOnce.Do(func() {
-		r.sink.record(r.buildAttemptLocked(err))
+		r.sink.record(r.buildAttemptLocked(err, nil))
+	})
+}
+
+// recordJSON behaves like record but reuses an already-decoded JSON
+// value, avoiding a second full decode of the same bytes.
+func (r *recordingBody) recordJSON(err error, value any) {
+	r.mu.Lock()
+	defer r.mu.Unlock()
+	r.recordOnce.Do(func() {
+		r.sink.record(r.buildAttemptLocked(err, &value))
 	})
 }
 
@@ -523,7 +563,7 @@ func (r *recordingBody) recordProvisional(err error) {
 	r.mu.Lock()
 	defer r.mu.Unlock()
 	r.recordOnce.Do(func() {
-		r.sink.record(r.buildAttemptLocked(err))
+		r.sink.record(r.buildAttemptLocked(err, nil))
 		r.recordedProvisional = true
 	})
 }

coderd/x/chatd/chatdebug/transport_internal_test.go

@@ -615,6 +615,46 @@ func TestRecordingTransport_CloseAfterDecoderConsumesUnknownLengthJSONSucceeds(t
 	require.Len(t, attempts, 1)
 	require.Equal(t, attemptStatusCompleted, attempts[0].Status)
 	require.Empty(t, attempts[0].Error)
+	require.JSONEq(t, `{"token":"[REDACTED]","safe":"ok"}`, string(attempts[0].ResponseBody))
+}
+
+// TestRecordingTransport_CloseAfterDecoderConsumesUnknownLengthJSONWithNoSensitiveKeysSucceeds
+// exercises buildAttemptLocked's decodedJSON!=nil-but-changed==false path:
+// the completeness check precomputes a decoded value, but nothing in it
+// needs redacting, so the original bytes must be recorded unchanged.
+func TestRecordingTransport_CloseAfterDecoderConsumesUnknownLengthJSONWithNoSensitiveKeysSucceeds(t *testing.T) {
+	t.Parallel()
+
+	ctx, sink := newTestSinkContext(t)
+	client := &http.Client{
+		Transport: &RecordingTransport{
+			Base: roundTripFunc(func(req *http.Request) (*http.Response, error) {
+				return &http.Response{ //nolint:exhaustruct // Test response exercises unknown-length close semantics.
+					StatusCode:    http.StatusOK,
+					Header:        http.Header{"Content-Type": []string{"application/json"}},
+					Body:          &scriptedReadCloser{chunks: [][]byte{[]byte(`{"safe":"ok","count":1}`)}},
+					ContentLength: -1,
+				}, nil
+			}),
+		},
+	}
+
+	req, err := http.NewRequestWithContext(ctx, http.MethodGet, "http://example.invalid", nil)
+	require.NoError(t, err)
+
+	resp, err := client.Do(req)
+	require.NoError(t, err)
+
+	var decoded map[string]any
+	require.NoError(t, json.NewDecoder(resp.Body).Decode(&decoded))
+	require.Equal(t, "ok", decoded["safe"])
+	require.NoError(t, resp.Body.Close())
+
+	attempts := sink.snapshot()
+	require.Len(t, attempts, 1)
+	require.Equal(t, attemptStatusCompleted, attempts[0].Status)
+	require.Empty(t, attempts[0].Error)
+	require.JSONEq(t, `{"safe":"ok","count":1}`, string(attempts[0].ResponseBody))
 }
 
 func TestRecordingTransport_CloseAfterDecoderConsumesUnknownLengthJSONWithTrailingDocumentMarksFailed(t *testing.T) {
@@ -649,6 +689,9 @@ func TestRecordingTransport_CloseAfterDecoderConsumesUnknownLengthJSONWithTraili
 	require.Len(t, attempts, 1)
 	require.Equal(t, attemptStatusFailed, attempts[0].Status)
 	require.Equal(t, io.ErrUnexpectedEOF.Error(), attempts[0].Error)
+	require.JSONEq(t,
+		`{"error":"chatdebug: body contains extra JSON values, redacted for safety"}`,
+		string(attempts[0].ResponseBody))
 }
 
 func TestRecordingTransport_CloseAfterDecoderConsumesUnknownLengthNDJSONMarksFailed(t *testing.T) {