feat(agent): auto-compact long conversations in the runner

Runner.Compact folds the older portion of history into a single system-role summary when the previous turn's input tokens cross threshold_ratio × MaxContextWindow. When summarization fails, the runner falls back to truncateToBudget so a flaky summarize call never breaks the user's request. EstimatePromptTokens / totalPromptTokens give a conservative count (roughly 3 chars per token) used by both the compaction trigger and BuildMessages' new limitRAGContext / fitHistory helpers to cap the prompt inside the provider's reported window before the request goes out. Covers the first-turn case where no usage has been reported yet. Adds runner_compaction_test.go with table-driven coverage for the disabled, below-threshold, short-history, unknown-window, fallback and first-stream cases, plus a regression for the RAG-context trimmer.
2026-07-18 00:08:05 -07:00 · 2026-07-18 00:08:05 -07:00 · 7393812fba
commit 7393812fba
parent 4c5cad38f8
2 changed files with 804 additions and 2 deletions
--- a/internal/agent/runner.go
+++ b/internal/agent/runner.go
@ -5,12 +5,19 @@
 // straight Q&A flow: no tools, no multi-iteration reasoning. Calling the
 // underlying provider keeps the prompt under our full control so we can
 // inject RAG context into the system message exactly where we want it.
+//
+// Auto-compaction (see CompactionConfig / Compact) folds the older portion
+// of a long conversation into a single summary system message when the
+// previous turn's input tokens approach the model's reported context
+// window. It is opt-in via WithCompaction and short-circuits silently when
+// the window is unknown or the history is too short to bother.
 package agent

 import (
 	"context"
 	"fmt"
 	"iter"
+	"log/slog"
 	"strings"

 	"github.com/VictorVargas/rony-llm-agent/pkg/llm"
@ -40,6 +47,9 @@ type Runner struct {
 	store        *portfolio.Store
 	topK         int
 	usage        *Usage
+
+	compaction  CompactionConfig
+	lastCompact CompactionStats
 }

 type Usage struct {
@ -47,18 +57,64 @@ type Usage struct {
 	OutputTokens int
 }

+// CompactionConfig mirrors the YAML `compaction:` block. Zero value means
+// compaction is disabled.
+type CompactionConfig struct {
+	Enabled             bool
+	ThresholdRatio      float64 // 0–1, default 0.75 when Enabled
+	KeepRecentTurns     int     // default 4 when Enabled
+	SummarySystemPrompt string  // empty → built-in bilingual default
+}
+
+// CompactionStats reports the outcome of the most recent Compact call, so
+// the HTTP handler can surface a "context compacted" event to the client.
+type CompactionStats struct {
+	Happened      bool // true when older turns were folded into a summary
+	OlderTurns    int  // user turns that got summarized away
+	KeptTurns     int  // user turns kept verbatim after the summary
+	SummaryTokens int  // approx. token count of the summary text
+	WindowTokens  int  // model's reported context window at compaction time
+	UsedTokens    int  // input tokens reported by the previous turn
+}
+
 // New returns a Runner. The store may be nil (RAG disabled).
 // systemPrompt is the full hand-written prompt from configs/...yaml.
 func New(client llm.LLMClient, p llmpersona.Persona, systemPrompt string, store *portfolio.Store, topK int) *Runner {
 	return &Runner{client: client, persona: p, systemPrompt: systemPrompt, store: store, topK: topK, usage: &Usage{}}
 }

+// WithCompaction enables auto-compaction with the given config. Returns
+// the receiver for chaining. Pass a zero-value CompactionConfig to keep
+// compaction disabled.
+func (r *Runner) WithCompaction(cfg CompactionConfig) *Runner {
+	r.compaction = cfg
+	if cfg.Enabled {
+		if r.compaction.ThresholdRatio <= 0 || r.compaction.ThresholdRatio > 1 {
+			r.compaction.ThresholdRatio = 0.75
+		}
+		if r.compaction.KeepRecentTurns <= 0 {
+			r.compaction.KeepRecentTurns = 4
+		}
+	}
+	return r
+}
+
 // LastUsage returns the token usage recorded on the most recent call.
 func (r *Runner) LastUsage() Usage { return *r.usage }

+// LastCompaction returns the outcome of the most recent Compact call.
+// Useful for the HTTP handler to emit a "compaction" SSE event after a
+// stream finishes.
+func (r *Runner) LastCompaction() CompactionStats { return r.lastCompact }
+
 // BuildMessages prepares the system prompt and turns the chat history into
 // the upstream message list. The system prompt includes RAG context for the
 // user's last message (if RAG is enabled).
+//
+// Compaction is NOT applied here — callers invoke Compact explicitly before
+// BuildMessages so the same Compact call doesn't run twice per request
+// (Stream calls BuildMessages internally, and handlers sometimes call it
+// first to extract the RAG context for the sources event).
 func (r *Runner) BuildMessages(ctx context.Context, history []Message) ([]Message, string, error) {
 	ragContext := ""
 	if r.store != nil && len(history) > 0 {
@ -69,7 +125,7 @@ func (r *Runner) BuildMessages(ctx context.Context, history []Message) ([]Messag
 				return nil, "", fmt.Errorf("rag search: %w", err)
 			}
 			if len(hits) > 0 {
-				ragContext = formatHits(hits)
+				ragContext = r.limitRAGContext(formatHits(hits), history)
 			}
 		}
 	}
@ -77,12 +133,205 @@ func (r *Runner) BuildMessages(ctx context.Context, history []Message) ([]Messag
 	// The system prompt comes from the YAML, not from the persona struct.
 	// Keep the persona around only for the UI greeting.
 	system := botpersona.BuildSystemPrompt(r.systemPrompt, ragContext)
+	history, err := r.fitHistory(system, history)
+	if err != nil {
+		return nil, "", err
+	}
 	msgs := make([]Message, 0, len(history)+1)
 	msgs = append(msgs, Message{Role: RoleSystem, Content: system})
 	msgs = append(msgs, history...)
 	return msgs, ragContext, nil
 }

+const maxContextSafetyMargin = 256
+
+func contextBudget(window int) int {
+	margin := int(float64(window) * 0.125)
+	if margin > maxContextSafetyMargin {
+		margin = maxContextSafetyMargin
+	}
+	if margin < 1 {
+		margin = 1
+	}
+	return window - margin
+}
+
+func (r *Runner) limitRAGContext(ragContext string, history []Message) string {
+	if strings.TrimSpace(ragContext) == "" {
+		return ""
+	}
+	budget := contextBudget(r.client.Capabilities().MaxContextWindow)
+	if budget <= 0 {
+		return ""
+	}
+
+	historyTokens := totalPromptTokens(history)
+	fitted := ""
+	for _, block := range strings.Split(strings.TrimSpace(ragContext), "\n\n") {
+		block = strings.TrimSpace(block)
+		if block == "" {
+			continue
+		}
+		candidate := block
+		if fitted != "" {
+			candidate = fitted + "\n\n" + block
+		}
+		if estimatePromptTokens(botpersona.BuildSystemPrompt(r.systemPrompt, candidate))+historyTokens > budget {
+			break
+		}
+		fitted = candidate
+	}
+	return fitted
+}
+
+func (r *Runner) fitHistory(system string, history []Message) ([]Message, error) {
+	window := r.client.Capabilities().MaxContextWindow
+	if window <= 0 {
+		return history, nil
+	}
+	budget := contextBudget(window)
+	if budget <= 0 {
+		return nil, fmt.Errorf("context window is too small")
+	}
+
+	history = dropLeadingAssistantMessages(history)
+	for {
+		if estimatePromptTokens(system)+totalPromptTokens(history) <= budget {
+			return history, nil
+		}
+		turnStarts := []int{}
+		for i, message := range history {
+			if message.Role == RoleUser {
+				turnStarts = append(turnStarts, i)
+			}
+		}
+		if len(turnStarts) < 2 {
+			return nil, fmt.Errorf("prompt exceeds context window (%d tokens)", window)
+		}
+		first, next := turnStarts[0], turnStarts[1]
+		kept := make([]Message, 0, len(history)-(next-first))
+		kept = append(kept, history[:first]...)
+		kept = append(kept, history[next:]...)
+		history = kept
+	}
+}
+
+func dropLeadingAssistantMessages(history []Message) []Message {
+	firstUser := -1
+	for i, message := range history {
+		if message.Role == RoleUser {
+			firstUser = i
+			break
+		}
+	}
+	if firstUser <= 0 {
+		return history
+	}
+	out := make([]Message, 0, len(history))
+	for _, message := range history[:firstUser] {
+		if message.Role != RoleAssistant {
+			out = append(out, message)
+		}
+	}
+	return append(out, history[firstUser:]...)
+}
+
+// Compact reduces the older portion of history to a single system-role
+// summary message when the previous turn's input tokens exceed the
+// configured threshold. Returns the (possibly compacted) message slice —
+// unchanged when compaction is disabled, history is too short, the model
+// has no reported context window, or summarization fails.
+//
+// On summarization failure the runner falls back to dropping the oldest
+// turns until the remaining slice fits a conservative budget, so a flaky
+// summarize call never fails the user's request.
+func (r *Runner) Compact(ctx context.Context, history []Message) ([]Message, error) {
+	r.lastCompact = CompactionStats{}
+	if !r.compaction.Enabled || len(history) == 0 {
+		return history, nil
+	}
+
+	caps := r.client.Capabilities()
+	if caps.MaxContextWindow <= 0 {
+		return history, nil
+	}
+
+	used := r.usage.InputTokens
+	estimated := estimatePromptTokens(r.systemPrompt) + totalPromptTokens(history)
+	if estimated > used {
+		used = estimated
+	}
+	window := caps.MaxContextWindow
+	threshold := int(float64(window) * r.compaction.ThresholdRatio)
+	if used < threshold {
+		return history, nil
+	}
+
+	older, recent := splitByTurns(history, r.compaction.KeepRecentTurns)
+	if len(older) == 0 {
+		return history, nil
+	}
+
+	summary, err := r.summarize(ctx, older)
+	if err != nil {
+		slog.Warn("compaction: summarize failed, falling back to truncation",
+			"err", err, "older_turns", userTurnCount(older))
+		// Fall back: drop oldest turns until the slice is small enough to
+		// fit in (threshold) tokens, using a 4-chars-per-token heuristic.
+		dropped, kept := truncateToBudget(history, threshold)
+		r.lastCompact = CompactionStats{
+			Happened:     true,
+			OlderTurns:   userTurnCount(dropped),
+			KeptTurns:    userTurnCount(kept),
+			WindowTokens: window,
+			UsedTokens:   used,
+		}
+		return kept, nil
+	}
+
+	out := make([]Message, 0, 1+len(recent))
+	out = append(out, Message{
+		Role:    RoleSystem,
+		Content: "Earlier conversation summary:\n" + summary,
+	})
+	out = append(out, recent...)
+	r.lastCompact = CompactionStats{
+		Happened:      true,
+		OlderTurns:    userTurnCount(older),
+		KeptTurns:     userTurnCount(recent),
+		SummaryTokens: estimateTokens(summary),
+		WindowTokens:  window,
+		UsedTokens:    used,
+	}
+	return out, nil
+}
+
+// summarize calls the LLM (non-streaming) to condense older turns. We use
+// a small max_tokens cap so the compaction step itself stays cheap.
+func (r *Runner) summarize(ctx context.Context, older []Message) (string, error) {
+	transcript := renderTranscript(older)
+	sys := r.compaction.SummarySystemPrompt
+	if sys == "" {
+		sys = defaultSummaryPrompt
+	}
+	maxTokens := 512
+	resp, err := r.client.Generate(ctx, llm.CompletionRequest{
+		Messages: []Message{
+			{Role: RoleSystem, Content: sys},
+			{Role: RoleUser, Content: transcript},
+		},
+		MaxTokens: &maxTokens,
+	})
+	if err != nil {
+		return "", err
+	}
+	summary := strings.TrimSpace(resp.Content)
+	if summary == "" {
+		return "", fmt.Errorf("empty summary")
+	}
+	return summary, nil
+}
+
 // Stream runs the model and yields each streamed chunk. The caller
 // forwards chunk.Delta to the SSE stream; chunk.Usage on the last chunk
 // carries token counts.
@ -123,3 +372,131 @@ func formatHits(hits []portfolio.SearchResult) string {
 	}
 	return b.String()
 }
+
+// defaultSummaryPrompt is used when the YAML doesn't override it. The
+// instruction explicitly avoids headers / meta-commentary so the summary
+// drops in cleanly as a system message and the model treats it as facts
+// about the prior conversation rather than instructions from the user.
+const defaultSummaryPrompt = "Summarize the following conversation between a user and the assistant concisely but completely. " +
+	"Preserve decisions made, concrete facts (names, paths, IDs, preferences) and any pending task or context the assistant needs to keep helping. " +
+	"Respond only with the summary, in the same language as the conversation, with no headers or extra commentary."
+
+// splitByTurns divides messages into an older portion to summarize and a
+// recent tail to keep verbatim. A turn is a user message plus everything
+// up to (but not including) the next user message — so a kept turn's tool
+// calls, thoughts and response always stay together. Returns older == nil
+// when there are not enough turns to bother compacting.
+func splitByTurns(history []Message, keepRecentTurns int) (older, recent []Message) {
+	turnStarts := []int{}
+	for i, m := range history {
+		if m.Role == RoleUser {
+			turnStarts = append(turnStarts, i)
+		}
+	}
+	if len(turnStarts) <= keepRecentTurns {
+		return nil, history
+	}
+	cut := turnStarts[len(turnStarts)-keepRecentTurns]
+	return history[:cut], history[cut:]
+}
+
+// renderTranscript flattens chat messages into a plain User/Assistant
+// transcript for summarization. Tool messages and empty assistant
+// placeholders are skipped — neither contributes facts the model needs
+// to remember.
+func renderTranscript(messages []Message) string {
+	var b strings.Builder
+	for _, m := range messages {
+		switch m.Role {
+		case RoleUser:
+			b.WriteString("User: ")
+			b.WriteString(m.Content)
+			b.WriteString("\n")
+		case RoleAssistant:
+			if m.Content == "" {
+				continue
+			}
+			b.WriteString("Assistant: ")
+			b.WriteString(m.Content)
+			b.WriteString("\n")
+		}
+	}
+	return b.String()
+}
+
+// userTurnCount is the number of user-role messages in a slice — a
+// convenient proxy for "turns" in stats reporting (each turn starts
+// with a user message).
+func userTurnCount(messages []Message) int {
+	n := 0
+	for _, m := range messages {
+		if m.Role == RoleUser {
+			n++
+		}
+	}
+	return n
+}
+
+// estimateTokens gives a rough token count for s. ~4 chars per token is
+// the usual rule of thumb for English; close enough for Spanish / mixed
+// text that this only drives a trigger threshold.
+func estimateTokens(s string) int {
+	if s == "" {
+		return 0
+	}
+	return len(s)/4 + 1
+}
+
+func estimatePromptTokens(s string) int {
+	if s == "" {
+		return 0
+	}
+	return len(s)/3 + 1
+}
+
+func totalPromptTokens(messages []Message) int {
+	total := 0
+	for _, m := range messages {
+		total += estimatePromptTokens(m.Content)
+	}
+	return total
+}
+
+// truncateToBudget is the fallback when summarization fails. It drops
+// the oldest user-turns one at a time until the remaining slice fits in
+// `budget` tokens (conservative heuristic). At least the last user turn is
+// always preserved, so we never return an empty slice.
+//
+// Implementation note: as `cut` increases, history[turnStarts[cut]:]
+// shrinks, so `totalPromptTokens(...)` is monotonically non-increasing in
+// `cut`. The loop is therefore guaranteed to terminate by either
+// fitting the budget or hitting the "keep at least the last turn" floor.
+func truncateToBudget(history []Message, budget int) (dropped, kept []Message) {
+	if len(history) == 0 {
+		return nil, nil
+	}
+	turnStarts := []int{}
+	for i, m := range history {
+		if m.Role == RoleUser {
+			turnStarts = append(turnStarts, i)
+		}
+	}
+	if len(turnStarts) == 0 {
+		return nil, history
+	}
+	// cut = number of leading turns to drop. 0 = keep everything; we
+	// grow it until the remaining slice fits in budget (or we hit the
+	// floor of one turn kept).
+	cut := 0
+	for cut < len(turnStarts)-1 {
+		start := turnStarts[cut+1]
+		if totalPromptTokens(history[start:]) <= budget {
+			break
+		}
+		cut++
+	}
+	if cut == 0 {
+		return nil, history
+	}
+	return history[:turnStarts[cut]], history[turnStarts[cut]:]
+}
--- a/internal/agent/runner_compaction_test.go
+++ b/internal/agent/runner_compaction_test.go
@ -0,0 +1,425 @@
+package agent
+
+import (
+	"context"
+	"errors"
+	"iter"
+	"strings"
+	"testing"
+
+	"github.com/VictorVargas/rony-llm-agent/pkg/llm"
+	llmpersona "github.com/VictorVargas/rony-llm-agent/pkg/persona"
+)
+
+// compactionStub is an llm.LLMClient that:
+//   - reports a configurable MaxContextWindow via Capabilities
+//   - records Generate calls (used for summarization)
+//   - returns a fixed Stream that also bumps the runner's r.usage via
+//     the usage chunk (so subsequent compactions see realistic usage)
+type compactionStub struct {
+	window        int
+	streamInput   int // input tokens reported on the streamed usage chunk
+	generateCalls int
+	generateErr   error
+	summary       string
+	lastGenerate  llm.CompletionRequest
+}
+
+func (s *compactionStub) Generate(_ context.Context, req llm.CompletionRequest) (llm.CompletionResponse, error) {
+	s.generateCalls++
+	s.lastGenerate = req
+	if s.generateErr != nil {
+		return llm.CompletionResponse{}, s.generateErr
+	}
+	return llm.CompletionResponse{
+		Content:    s.summary,
+		StopReason: llm.StopReasonEndTurn,
+	}, nil
+}
+
+func (s *compactionStub) Stream(_ context.Context, _ llm.CompletionRequest) iter.Seq2[llm.StreamChunk, error] {
+	return func(yield func(llm.StreamChunk, error) bool) {
+		yield(llm.StreamChunk{Delta: "ok", FinishReason: "stop", Usage: llm.TokenUsage{InputTokens: s.streamInput, OutputTokens: 3}}, nil)
+	}
+}
+
+func (s *compactionStub) Name() string { return "compaction-stub" }
+func (s *compactionStub) Capabilities() llm.ProviderCapabilities {
+	return llm.ProviderCapabilities{MaxContextWindow: s.window}
+}
+
+// pumpStream drives a single Stream call through the runner so the usage
+// chunk updates r.usage — needed because the compaction trigger reads
+// from r.usage, not from the request itself.
+func pumpStream(t *testing.T, r *Runner, history []Message) {
+	t.Helper()
+	for chunk, err := range r.Stream(context.Background(), history) {
+		if err != nil {
+			t.Fatal(err)
+		}
+		_ = chunk
+	}
+}
+
+// makeHistory builds a synthetic N-turn conversation (user + assistant
+// pairs), each "long" enough to be obviously past any small threshold.
+func makeHistory(turns int, contentLen int) []Message {
+	out := make([]Message, 0, turns*2)
+	payload := strings.Repeat("x", contentLen)
+	for i := 0; i < turns; i++ {
+		out = append(out,
+			Message{Role: RoleUser, Content: payload + " q"},
+			Message{Role: RoleAssistant, Content: payload + " a"},
+		)
+	}
+	return out
+}
+
+func personaMinimal() llmpersona.Persona {
+	return llmpersona.Persona{Name: "t", Tone: "concise", Language: "English"}
+}
+
+func TestCompactDisabledIsNoop(t *testing.T) {
+	cli := &compactionStub{window: 100, streamInput: 999, summary: "ignored"}
+	r := New(cli, personaMinimal(), "sys", nil, 5) // no WithCompaction
+	pumpStream(t, r, []Message{{Role: RoleUser, Content: "hi"}})
+	got, err := r.Compact(context.Background(), makeHistory(8, 200))
+	if err != nil {
+		t.Fatal(err)
+	}
+	if cli.generateCalls != 0 {
+		t.Errorf("Generate called %d times, want 0 (compaction disabled)", cli.generateCalls)
+	}
+	if len(got) != 16 {
+		t.Errorf("len(history) = %d, want 16 (unchanged)", len(got))
+	}
+	stats := r.LastCompaction()
+	if stats.Happened {
+		t.Errorf("stats.Happened = true, want false")
+	}
+}
+
+func TestCompactBelowThresholdIsNoop(t *testing.T) {
+	// window = 1000, threshold = 0.5 → 500. Previous turn reported only
+	// 50 input tokens — well below the trigger.
+	cli := &compactionStub{window: 1000, streamInput: 50, summary: "ignored"}
+	r := New(cli, personaMinimal(), "sys", nil, 5).
+		WithCompaction(CompactionConfig{Enabled: true, ThresholdRatio: 0.5, KeepRecentTurns: 2})
+	pumpStream(t, r, []Message{{Role: RoleUser, Content: "hi"}})
+	got, err := r.Compact(context.Background(), makeHistory(8, 40))
+	if err != nil {
+		t.Fatal(err)
+	}
+	if cli.generateCalls != 0 {
+		t.Errorf("Generate called %d times, want 0 (below threshold)", cli.generateCalls)
+	}
+	if len(got) != 16 {
+		t.Errorf("len(history) = %d, want 16 (unchanged)", len(got))
+	}
+}
+
+func TestCompactUsesEstimatedInputBeforeFirstStream(t *testing.T) {
+	cli := &compactionStub{window: 400, summary: "summary"}
+	r := New(cli, personaMinimal(), "system prompt", nil, 5).
+		WithCompaction(CompactionConfig{Enabled: true, ThresholdRatio: 0.5, KeepRecentTurns: 2})
+
+	got, err := r.Compact(context.Background(), makeHistory(8, 200))
+	if err != nil {
+		t.Fatal(err)
+	}
+	if cli.generateCalls != 1 {
+		t.Fatalf("Generate called %d times, want 1", cli.generateCalls)
+	}
+	if len(got) != 5 {
+		t.Fatalf("len(compacted) = %d, want 5", len(got))
+	}
+}
+
+func TestCompactShortHistoryIsNoop(t *testing.T) {
+	// Above threshold but only 3 turns total (3 user, 3 assistant = 6
+	// messages). KeepRecentTurns = 4, so splitByTurns has nothing older
+	// to summarize.
+	cli := &compactionStub{window: 100, streamInput: 999, summary: "ignored"}
+	r := New(cli, personaMinimal(), "sys", nil, 5).
+		WithCompaction(CompactionConfig{Enabled: true, ThresholdRatio: 0.5, KeepRecentTurns: 4})
+	pumpStream(t, r, []Message{{Role: RoleUser, Content: "hi"}})
+	history := makeHistory(3, 200) // 6 messages, 3 user turns
+	got, err := r.Compact(context.Background(), history)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if cli.generateCalls != 0 {
+		t.Errorf("Generate called %d times, want 0 (not enough turns)", cli.generateCalls)
+	}
+	if len(got) != len(history) {
+		t.Errorf("len(history) = %d, want %d (unchanged)", len(got), len(history))
+	}
+}
+
+func TestCompactUnknownWindowIsNoop(t *testing.T) {
+	// Provider doesn't report a window (Capabilities returns 0).
+	cli := &compactionStub{window: 0, streamInput: 999, summary: "ignored"}
+	r := New(cli, personaMinimal(), "sys", nil, 5).
+		WithCompaction(CompactionConfig{Enabled: true, ThresholdRatio: 0.5, KeepRecentTurns: 2})
+	pumpStream(t, r, []Message{{Role: RoleUser, Content: "hi"}})
+	got, err := r.Compact(context.Background(), makeHistory(8, 200))
+	if err != nil {
+		t.Fatal(err)
+	}
+	if cli.generateCalls != 0 {
+		t.Errorf("Generate called %d times, want 0 (window unknown)", cli.generateCalls)
+	}
+	if len(got) != 16 {
+		t.Errorf("len(history) = %d, want 16 (unchanged)", len(got))
+	}
+}
+
+func TestCompactTriggersSummarize(t *testing.T) {
+	// 8 user turns × ~200 chars each = ~1600 chars ≈ 400 tokens. With
+	// window=400, threshold 0.5 = 200, and the stub reports 999 input
+	// tokens on the previous turn, compaction fires.
+	cli := &compactionStub{window: 400, streamInput: 999, summary: "user asked 8 things, here's the gist"}
+	r := New(cli, personaMinimal(), "sys", nil, 5).
+		WithCompaction(CompactionConfig{
+			Enabled:         true,
+			ThresholdRatio:  0.5,
+			KeepRecentTurns: 2,
+		})
+	pumpStream(t, r, []Message{{Role: RoleUser, Content: "hi"}})
+	history := makeHistory(8, 200)
+	got, err := r.Compact(context.Background(), history)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if cli.generateCalls != 1 {
+		t.Fatalf("Generate called %d times, want 1 (compaction fired)", cli.generateCalls)
+	}
+	// The summary call's messages should be [system, user(transcript)].
+	if len(cli.lastGenerate.Messages) != 2 {
+		t.Fatalf("summarize call has %d messages, want 2", len(cli.lastGenerate.Messages))
+	}
+	if cli.lastGenerate.Messages[0].Role != RoleSystem {
+		t.Errorf("summarize call[0].Role = %q, want system", cli.lastGenerate.Messages[0].Role)
+	}
+	if cli.lastGenerate.Messages[1].Role != RoleUser {
+		t.Errorf("summarize call[1].Role = %q, want user", cli.lastGenerate.Messages[1].Role)
+	}
+	if !strings.Contains(cli.lastGenerate.Messages[1].Content, "User:") {
+		t.Errorf("summarize transcript missing 'User:' lines")
+	}
+	if got[0].Role != RoleSystem {
+		t.Errorf("compacted[0].Role = %q, want system (summary)", got[0].Role)
+	}
+	if !strings.Contains(got[0].Content, "user asked 8 things") {
+		t.Errorf("compacted[0] missing summary text: %q", got[0].Content)
+	}
+	// Recent tail: 2 turns = 4 messages, exactly the last 4 of the
+	// original history.
+	if len(got) != 1+4 {
+		t.Fatalf("len(compacted) = %d, want 5 (summary + 4)", len(got))
+	}
+	for i := 0; i < 4; i++ {
+		if got[1+i].Content != history[12+i].Content {
+			t.Errorf("kept[%d] content differs from history[%d]", i, 12+i)
+		}
+	}
+	stats := r.LastCompaction()
+	if !stats.Happened {
+		t.Fatal("stats.Happened = false, want true")
+	}
+	if stats.OlderTurns != 6 {
+		t.Errorf("OlderTurns = %d, want 6", stats.OlderTurns)
+	}
+	if stats.KeptTurns != 2 {
+		t.Errorf("KeptTurns = %d, want 2", stats.KeptTurns)
+	}
+	if stats.WindowTokens != 400 {
+		t.Errorf("WindowTokens = %d, want 400", stats.WindowTokens)
+	}
+	wantUsed := estimatePromptTokens("sys") + totalPromptTokens(history)
+	if stats.UsedTokens != wantUsed {
+		t.Errorf("UsedTokens = %d, want %d", stats.UsedTokens, wantUsed)
+	}
+	if stats.SummaryTokens == 0 {
+		t.Error("SummaryTokens = 0, want > 0")
+	}
+}
+
+func TestCompactSummarizeFailsFallsBackToTruncation(t *testing.T) {
+	cli := &compactionStub{
+		window:      400,
+		streamInput: 999,
+		generateErr: errors.New("summarize boom"),
+	}
+	r := New(cli, personaMinimal(), "sys", nil, 5).
+		WithCompaction(CompactionConfig{
+			Enabled:         true,
+			ThresholdRatio:  0.5,
+			KeepRecentTurns: 2,
+		})
+	pumpStream(t, r, []Message{{Role: RoleUser, Content: "hi"}})
+	// 7 completed turns + 1 trailing user message (the request's new
+	// question). Real production requests always end on a user message.
+	history := append(makeHistory(7, 200), Message{Role: RoleUser, Content: "new question"})
+	got, err := r.Compact(context.Background(), history)
+	if err != nil {
+		t.Fatalf("Compact should swallow summarize failure, got err: %v", err)
+	}
+	if cli.generateCalls != 1 {
+		t.Errorf("Generate called %d times, want 1 (tried once before fallback)", cli.generateCalls)
+	}
+	// Fallback should have dropped the oldest turns. The trailing user
+	// message (the current question) must always be preserved.
+	last := got[len(got)-1]
+	if last.Role != RoleUser {
+		t.Errorf("last message role = %q, want user", last.Role)
+	}
+	if last.Content != "new question" {
+		t.Errorf("last message content = %q, want %q", last.Content, "new question")
+	}
+	stats := r.LastCompaction()
+	if !stats.Happened {
+		t.Error("stats.Happened = false, want true (fallback counts as compaction)")
+	}
+	if stats.OlderTurns == 0 {
+		t.Error("OlderTurns = 0, want > 0 (fallback dropped turns)")
+	}
+}
+
+func TestCompactEmptyHistoryIsNoop(t *testing.T) {
+	cli := &compactionStub{window: 100, streamInput: 999}
+	r := New(cli, personaMinimal(), "sys", nil, 5).
+		WithCompaction(CompactionConfig{Enabled: true, ThresholdRatio: 0.1, KeepRecentTurns: 2})
+	got, err := r.Compact(context.Background(), nil)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if got != nil {
+		t.Errorf("got = %v, want nil", got)
+	}
+	if cli.generateCalls != 0 {
+		t.Errorf("Generate called %d times, want 0", cli.generateCalls)
+	}
+}
+
+func TestFitHistoryDropsOldTurnsToFit(t *testing.T) {
+	cli := &compactionStub{window: 400}
+	r := New(cli, personaMinimal(), "sys", nil, 5)
+	history := makeHistory(4, 400)
+
+	got, err := r.fitHistory("sys", history)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(got) != 2 {
+		t.Fatalf("len(history) = %d, want 2", len(got))
+	}
+	if got[0].Role != RoleUser || got[1].Role != RoleAssistant {
+		t.Errorf("remaining turn roles = %q, %q", got[0].Role, got[1].Role)
+	}
+}
+
+func TestSplitByTurns(t *testing.T) {
+	history := []Message{
+		{Role: RoleUser, Content: "u1"},
+		{Role: RoleAssistant, Content: "a1"},
+		{Role: RoleUser, Content: "u2"},
+		{Role: RoleAssistant, Content: "a2"},
+		{Role: RoleUser, Content: "u3"},
+		{Role: RoleAssistant, Content: "a3"},
+		{Role: RoleUser, Content: "u4"},
+		{Role: RoleAssistant, Content: "a4"},
+	}
+	older, recent := splitByTurns(history, 2)
+	if len(older) != 4 {
+		t.Errorf("len(older) = %d, want 4", len(older))
+	}
+	if len(recent) != 4 {
+		t.Errorf("len(recent) = %d, want 4", len(recent))
+	}
+	if older[0].Content != "u1" || older[3].Content != "a2" {
+		t.Errorf("older = %v, want [u1,a1,u2,a2]", older)
+	}
+	if recent[0].Content != "u3" || recent[3].Content != "a4" {
+		t.Errorf("recent = %v, want [u3,a3,u4,a4]", recent)
+	}
+}
+
+func TestSplitByTurnsNothingToCompact(t *testing.T) {
+	history := []Message{
+		{Role: RoleUser, Content: "u1"},
+		{Role: RoleAssistant, Content: "a1"},
+		{Role: RoleUser, Content: "u2"},
+	}
+	older, recent := splitByTurns(history, 4)
+	if older != nil {
+		t.Errorf("older = %v, want nil", older)
+	}
+	if len(recent) != 3 {
+		t.Errorf("len(recent) = %d, want 3 (everything)", len(recent))
+	}
+}
+
+func TestRenderTranscriptSkipsToolAndEmpty(t *testing.T) {
+	msgs := []Message{
+		{Role: RoleUser, Content: "hi"},
+		{Role: RoleAssistant, Content: "hello!"},
+		{Role: RoleAssistant, Content: ""}, // streaming placeholder
+		{Role: RoleSystem, Content: "internal"},
+	}
+	got := renderTranscript(msgs)
+	want := "User: hi\nAssistant: hello!\n"
+	if got != want {
+		t.Errorf("renderTranscript =\n%q\nwant\n%q", got, want)
+	}
+}
+
+func TestEstimateTokens(t *testing.T) {
+	cases := []struct {
+		in   string
+		want int
+	}{
+		{"", 0},
+		{"a", 1},
+		{"abcd", 2},
+		{strings.Repeat("x", 400), 101},
+	}
+	for _, c := range cases {
+		if got := estimateTokens(c.in); got != c.want {
+			t.Errorf("estimateTokens(%q) = %d, want %d", c.in, got, c.want)
+		}
+	}
+}
+
+func TestEstimatePromptTokensIsConservative(t *testing.T) {
+	text := strings.Repeat("x", 300)
+	if got := estimatePromptTokens(text); got <= estimateTokens(text) {
+		t.Errorf("estimatePromptTokens = %d, want greater than estimateTokens = %d", got, estimateTokens(text))
+	}
+}
+
+func TestTruncateToBudgetAlwaysKeepsLastUserTurn(t *testing.T) {
+	// 3 long completed turns + a short trailing user message (the new
+	// question). Budget so small only the last turn can fit.
+	long := strings.Repeat("x", 400)
+	history := []Message{
+		{Role: RoleUser, Content: long + " q1"},
+		{Role: RoleAssistant, Content: long + " a1"},
+		{Role: RoleUser, Content: long + " q2"},
+		{Role: RoleAssistant, Content: long + " a2"},
+		{Role: RoleUser, Content: long + " q3"},
+		{Role: RoleAssistant, Content: long + " a3"},
+		{Role: RoleUser, Content: "current question"},
+	}
+	dropped, kept := truncateToBudget(history, 10)
+	if len(kept) == 0 {
+		t.Fatal("kept is empty")
+	}
+	if kept[len(kept)-1].Content != "current question" {
+		t.Errorf("last kept message content = %q, want %q",
+			kept[len(kept)-1].Content, "current question")
+	}
+	if len(dropped)+len(kept) != len(history) {
+		t.Errorf("dropped+kept = %d, want %d", len(dropped)+len(kept), len(history))
+	}
+}