fix(agent): resume truncated work after auto compact

commit: ddda7bddf2b96521888b29bdae546a7b6c76e87b
author: brice · 2mo ago · 2026-08-06 18:45 UTC
committer: brice · 2mo ago · 2026-08-06 19:02 UTC
parent: d7a9ef7
signature: GPG VerifiedCD88F9E564E0102646FB6077A1A07C7C7C3E349D · arch-gpg Verified for the committer email

7 files · +241 -23

M README.md +2 -2


@@ -405,7 +405,7 @@ Opens a dialog with every persistent setting. `up`/`down` to navigate, `enter` o
405	405		- fast mode — request the provider's fast tier where supported. It currently uses OpenAI's Fast service tier for OpenAI, OpenAI Responses, and OpenAI Codex models. Off by default; toggle it with `/fast` or `/settings`; changes apply on the next model call and persist as `fast_mode`. Other providers currently return an unsupported-provider error. Fast mode may cost more and depends on the selected model/account.
406	406		- lsp in main session — enable the built-in `lsp` tool and code diagnostics for the main agent. Enabled by default; changes persist as `lsp_enabled`.
407	407		- lsp in sub-agents — allow newly spawned background sub-agents to use the built-in `lsp` tool. Enabled by default; changes persist as `subagent_lsp_enabled` and apply when a child starts.
408		-	- auto-compact threshold — choose `off`, `70%`, `80%`, `85%` (default), or `90%` of the model's advertised context window. The selected percentage controls automatic compaction before and after interactive turns and persists as `auto_compact_threshold`. After a successful threshold compaction, zut automatically resumes an unfinished most-recent intent; a prompt queued during compaction takes priority. `off` disables percentage-based triggers but keeps manual `/compact` and automatic recovery from context-window and payload-too-large responses.
	408	+	- auto-compact threshold — choose `off`, `70%`, `80%`, `85%` (default), or `90%` of the model's advertised context window. The selected percentage controls automatic compaction before and after interactive turns and persists as `auto_compact_threshold`. After a successful threshold compaction, zut automatically resumes an unfinished most-recent intent, including truncated output; a prompt queued during compaction takes priority. `off` disables percentage-based triggers but keeps manual `/compact` and automatic recovery from context-window and payload-too-large responses.
409	409		- jail new sessions by default: start every new agent with tools confined to its working directory. Off by default. The setting applies to interactive, print, JSON, RPC, and background-agent runs, persists as `jail_by_default`, and immediately updates the current interactive session. `/jail` and `/unjail` remain session-scoped overrides and do not change this default.
410	410		- compact transcript rendering: reduce visual chrome in the chat transcript. Tool calls render as a quiet header plus indented output instead of a bordered box, and sent messages render without padded background bubbles. Off by default. Changes apply immediately and persist to `config.json` as `compact_mode`.
411	411		- show loaded resources at startup: list the active Zutfile agent (for `zut run`), loaded `AGENTS.md` paths, extensions, and user-installed skills in compact sections above the transcript. Built-in skills are omitted. Off by default. Changes apply immediately and persist to `config.json` as `show_instructions_at_startup`.

@@ -422,7 +422,7 @@ Opens a picker listing every discovered SKILL.md file, built-ins hidden. Each ro
422	422
423	423		Sends the current transcript through the model with a structured summarization prompt. The returned summary replaces the transcript as one synthetic user message, with the last few exchanges kept verbatim for continuity. The status bar's context meter resets. Use it when the context meter creeps past ~80%.
424	424
425		-	zut also auto-compacts in the background: after any turn that reaches the configured context threshold, the agent kicks off a condense pass on its own and, once it succeeds, resumes an unfinished most-recent intent automatically. Completed text answers settle normally, while a prompt queued during compaction runs instead of the automatic continuation. Choose `off`, `70%`, `80%`, `85%` (default), or `90%` under `/settings` → auto-compact threshold. You'll see `condensing history, esc to cancel` above the status bar and an `(auto)` tag next to the context percentage; `esc` aborts it without touching the transcript. Turning the percentage trigger off does not disable automatic compaction and retry after a context-window or payload-too-large response.
	425	+	zut also auto-compacts in the background: after any turn that reaches the configured context threshold, the agent kicks off a condense pass on its own and, once it succeeds, resumes an unfinished most-recent intent automatically, including output that stopped at the model's token limit. Completed text answers settle normally, while a prompt queued during compaction runs instead of the automatic continuation. Choose `off`, `70%`, `80%`, `85%` (default), or `90%` under `/settings` → auto-compact threshold. You'll see `condensing history, esc to cancel` above the status bar and an `(auto)` tag next to the context percentage; `esc` aborts it without touching the transcript. Turning the percentage trigger off does not disable automatic compaction and retry after a context-window or payload-too-large response.
426	426
427	427		### `/jail`
428	428


@@ -405,7 +405,7 @@ Opens a dialog with every persistent setting. `up`/`down` to navigate, `enter` o
405		- fast mode — request the provider's fast tier where supported. It currently uses OpenAI's Fast service tier for OpenAI, OpenAI Responses, and OpenAI Codex models. Off by default; toggle it with `/fast` or `/settings`; changes apply on the next model call and persist as `fast_mode`. Other providers currently return an unsupported-provider error. Fast mode may cost more and depends on the selected model/account.	405		- fast mode — request the provider's fast tier where supported. It currently uses OpenAI's Fast service tier for OpenAI, OpenAI Responses, and OpenAI Codex models. Off by default; toggle it with `/fast` or `/settings`; changes apply on the next model call and persist as `fast_mode`. Other providers currently return an unsupported-provider error. Fast mode may cost more and depends on the selected model/account.
406		- lsp in main session — enable the built-in `lsp` tool and code diagnostics for the main agent. Enabled by default; changes persist as `lsp_enabled`.	406		- lsp in main session — enable the built-in `lsp` tool and code diagnostics for the main agent. Enabled by default; changes persist as `lsp_enabled`.
407		- lsp in sub-agents — allow newly spawned background sub-agents to use the built-in `lsp` tool. Enabled by default; changes persist as `subagent_lsp_enabled` and apply when a child starts.	407		- lsp in sub-agents — allow newly spawned background sub-agents to use the built-in `lsp` tool. Enabled by default; changes persist as `subagent_lsp_enabled` and apply when a child starts.
408	-	- auto-compact threshold — choose `off`, `70%`, `80%`, `85%` (default), or `90%` of the model's advertised context window. The selected percentage controls automatic compaction before and after interactive turns and persists as `auto_compact_threshold`. After a successful threshold compaction, zut automatically resumes an unfinished most-recent intent; a prompt queued during compaction takes priority. `off` disables percentage-based triggers but keeps manual `/compact` and automatic recovery from context-window and payload-too-large responses.	408	+	- auto-compact threshold — choose `off`, `70%`, `80%`, `85%` (default), or `90%` of the model's advertised context window. The selected percentage controls automatic compaction before and after interactive turns and persists as `auto_compact_threshold`. After a successful threshold compaction, zut automatically resumes an unfinished most-recent intent, including truncated output; a prompt queued during compaction takes priority. `off` disables percentage-based triggers but keeps manual `/compact` and automatic recovery from context-window and payload-too-large responses.
409		- jail new sessions by default: start every new agent with tools confined to its working directory. Off by default. The setting applies to interactive, print, JSON, RPC, and background-agent runs, persists as `jail_by_default`, and immediately updates the current interactive session. `/jail` and `/unjail` remain session-scoped overrides and do not change this default.	409		- jail new sessions by default: start every new agent with tools confined to its working directory. Off by default. The setting applies to interactive, print, JSON, RPC, and background-agent runs, persists as `jail_by_default`, and immediately updates the current interactive session. `/jail` and `/unjail` remain session-scoped overrides and do not change this default.
410		- compact transcript rendering: reduce visual chrome in the chat transcript. Tool calls render as a quiet header plus indented output instead of a bordered box, and sent messages render without padded background bubbles. Off by default. Changes apply immediately and persist to `config.json` as `compact_mode`.	410		- compact transcript rendering: reduce visual chrome in the chat transcript. Tool calls render as a quiet header plus indented output instead of a bordered box, and sent messages render without padded background bubbles. Off by default. Changes apply immediately and persist to `config.json` as `compact_mode`.
411		- show loaded resources at startup: list the active Zutfile agent (for `zut run`), loaded `AGENTS.md` paths, extensions, and user-installed skills in compact sections above the transcript. Built-in skills are omitted. Off by default. Changes apply immediately and persist to `config.json` as `show_instructions_at_startup`.	411		- show loaded resources at startup: list the active Zutfile agent (for `zut run`), loaded `AGENTS.md` paths, extensions, and user-installed skills in compact sections above the transcript. Built-in skills are omitted. Off by default. Changes apply immediately and persist to `config.json` as `show_instructions_at_startup`.

@@ -422,7 +422,7 @@ Opens a picker listing every discovered SKILL.md file, built-ins hidden. Each ro
422			422
423		Sends the current transcript through the model with a structured summarization prompt. The returned summary replaces the transcript as one synthetic user message, with the last few exchanges kept verbatim for continuity. The status bar's context meter resets. Use it when the context meter creeps past ~80%.	423		Sends the current transcript through the model with a structured summarization prompt. The returned summary replaces the transcript as one synthetic user message, with the last few exchanges kept verbatim for continuity. The status bar's context meter resets. Use it when the context meter creeps past ~80%.
424			424
425	-	zut also auto-compacts in the background: after any turn that reaches the configured context threshold, the agent kicks off a condense pass on its own and, once it succeeds, resumes an unfinished most-recent intent automatically. Completed text answers settle normally, while a prompt queued during compaction runs instead of the automatic continuation. Choose `off`, `70%`, `80%`, `85%` (default), or `90%` under `/settings` → auto-compact threshold. You'll see `condensing history, esc to cancel` above the status bar and an `(auto)` tag next to the context percentage; `esc` aborts it without touching the transcript. Turning the percentage trigger off does not disable automatic compaction and retry after a context-window or payload-too-large response.	425	+	zut also auto-compacts in the background: after any turn that reaches the configured context threshold, the agent kicks off a condense pass on its own and, once it succeeds, resumes an unfinished most-recent intent automatically, including output that stopped at the model's token limit. Completed text answers settle normally, while a prompt queued during compaction runs instead of the automatic continuation. Choose `off`, `70%`, `80%`, `85%` (default), or `90%` under `/settings` → auto-compact threshold. You'll see `condensing history, esc to cancel` above the status bar and an `(auto)` tag next to the context percentage; `esc` aborts it without touching the transcript. Turning the percentage trigger off does not disable automatic compaction and retry after a context-window or payload-too-large response.
426			426
427		### `/jail`	427		### `/jail`
428			428

M packages/agent/modes/compact_queue_test.go +133 -5


@@ -153,7 +153,7 @@ func TestPromptSubmittedDuringCompactionStartsFollowUpTurn(t *testing.T) {
153	153		interactive := NewInteractive(InteractiveConfig{Agent: agent})
154	154		interactive.runCtx = context.Background()
155	155
156		-	interactive.runCompact(context.Background(), false)
	156	+	interactive.runCompact(context.Background(), false, false)
157	157		select {
158	158		case <-client.compactionStarted:
159	159		case <-time.After(2 * time.Second):

@@ -178,9 +178,12 @@ type thresholdAutoCompactClient struct {
178	178		compactionStarted chan struct{}
179	179		releaseCompaction chan struct{}
180	180		followUpRequest chan provider.Request
	181	+	firstStop provider.StopReason
	182	+	firstContent []provider.Content
181	183
182		-	mu sync.Mutex
183		-	calls int
	184	+	mu sync.Mutex
	185	+	calls int
	186	+	followUps int
184	187		}
185	188
186	189		func (c *thresholdAutoCompactClient) Name() string { return "threshold-auto-compact-test" }

@@ -196,12 +199,20 @@ func (c *thresholdAutoCompactClient) Stream(ctx context.Context, req provider.Re
196	199		defer close(out)
197	200		switch call {
198	201		case 1:
	202	+	stop := c.firstStop
	203	+	if stop == "" {
	204	+	stop = provider.StopEnd
	205	+	}
	206	+	content := c.firstContent
	207	+	if len(content) == 0 {
	208	+	content = []provider.Content{provider.ReasoningBlock{Summary: "still working"}}
	209	+	}
199	210		out <- provider.EventUsage{Usage: provider.Usage{InputTokens: 150000}}
200	211		out <- provider.EventDone{
201		-	Stop: provider.StopEnd,
	212	+	Stop: stop,
202	213		Message: provider.Message{
203	214		Role: provider.RoleAssistant,
204		-	Content: []provider.Content{provider.ReasoningBlock{Summary: "still working"}},
	215	+	Content: content,
205	216		},
206	217		}
207	218		case 2:

@@ -214,6 +225,9 @@ func (c *thresholdAutoCompactClient) Stream(ctx context.Context, req provider.Re
214	225		out <- provider.EventDone{Stop: provider.StopEnd}
215	226		}
216	227		default:
	228	+	c.mu.Lock()
	229	+	c.followUps++
	230	+	c.mu.Unlock()
217	231		c.followUpRequest <- req
218	232		out <- provider.EventDone{
219	233		Stop: provider.StopEnd,

@@ -261,6 +275,120 @@ func TestThresholdAutoCompactionContinuesMostRecentIntent(t *testing.T) {
261	275		}
262	276		}
263	277
	278	+	func TestThresholdAutoCompactionContinuesTruncatedText(t *testing.T) {
	279	+	client := &thresholdAutoCompactClient{
	280	+	compactionStarted: make(chan struct{}),
	281	+	releaseCompaction: make(chan struct{}),
	282	+	followUpRequest: make(chan provider.Request, 1),
	283	+	firstStop: provider.StopLength,
	284	+	firstContent: []provider.Content{provider.TextBlock{Text: "partial answer"}},
	285	+	}
	286	+	agent := core.NewAgent(client, "test-model", "", nil)
	287	+	threshold := 70
	288	+	interactive := NewInteractive(InteractiveConfig{
	289	+	Agent: agent,
	290	+	Provider: "anthropic",
	291	+	Model: "claude-sonnet-4-5-20250929",
	292	+	AutoCompactThreshold: &threshold,
	293	+	})
	294	+	interactive.runCtx = context.Background()
	295	+
	296	+	interactive.startTurn(context.Background(), "initial request")
	297	+	select {
	298	+	case <-client.compactionStarted:
	299	+	case <-time.After(2 * time.Second):
	300	+	t.Fatal("threshold auto-compaction did not start after truncated text")
	301	+	}
	302	+	close(client.releaseCompaction)
	303	+
	304	+	select {
	305	+	case req := <-client.followUpRequest:
	306	+	if got := requestUserTextCount(req, autoCompactContinuationPrompt); got != 1 {
	307	+	t.Fatalf("truncated continuation request contains auto-continue prompt %d times, want 1: %#v", got, req.Messages)
	308	+	}
	309	+	case <-time.After(2 * time.Second):
	310	+	t.Fatal("threshold auto-compaction did not continue truncated text")
	311	+	}
	312	+	}
	313	+
	314	+	func TestThresholdAutoCompactionForcedContinuationPrecedesQueuedPrompt(t *testing.T) {
	315	+	client := &thresholdAutoCompactClient{
	316	+	compactionStarted: make(chan struct{}),
	317	+	releaseCompaction: make(chan struct{}),
	318	+	followUpRequest: make(chan provider.Request, 2),
	319	+	firstStop: provider.StopLength,
	320	+	firstContent: []provider.Content{provider.TextBlock{Text: "partial answer"}},
	321	+	}
	322	+	agent := core.NewAgent(client, "test-model", "", nil)
	323	+	threshold := 70
	324	+	interactive := NewInteractive(InteractiveConfig{
	325	+	Agent: agent,
	326	+	Provider: "anthropic",
	327	+	Model: "claude-sonnet-4-5-20250929",
	328	+	AutoCompactThreshold: &threshold,
	329	+	})
	330	+	interactive.runCtx = context.Background()
	331	+
	332	+	interactive.startTurn(context.Background(), "initial request")
	333	+	select {
	334	+	case <-client.compactionStarted:
	335	+	case <-time.After(2 * time.Second):
	336	+	t.Fatal("threshold auto-compaction did not start after truncated text")
	337	+	}
	338	+	interactive.SubmitOrQueue("queued follow-up", nil)
	339	+	close(client.releaseCompaction)
	340	+
	341	+	select {
	342	+	case req := <-client.followUpRequest:
	343	+	if got := requestUserTextCount(req, autoCompactContinuationPrompt); got != 1 {
	344	+	t.Fatalf("first request contains auto-continue prompt %d times, want 1: %#v", got, req.Messages)
	345	+	}
	346	+	if got := requestUserTextCount(req, "queued follow-up"); got != 0 {
	347	+	t.Fatalf("forced continuation request included queued prompt %d times, want 0: %#v", got, req.Messages)
	348	+	}
	349	+	case <-time.After(2 * time.Second):
	350	+	t.Fatal("threshold auto-compaction did not continue truncated text")
	351	+	}
	352	+
	353	+	select {
	354	+	case req := <-client.followUpRequest:
	355	+	if got := requestUserTextCount(req, "queued follow-up"); got != 1 {
	356	+	t.Fatalf("second request contains queued prompt %d times, want 1: %#v", got, req.Messages)
	357	+	}
	358	+	case <-time.After(2 * time.Second):
	359	+	t.Fatal("queued prompt did not run after forced continuation")
	360	+	}
	361	+
	362	+	deadline := time.NewTimer(2 * time.Second)
	363	+	defer deadline.Stop()
	364	+	poll := time.NewTicker(time.Millisecond)
	365	+	defer poll.Stop()
	366	+	for {
	367	+	interactive.mu.Lock()
	368	+	idle := !interactive.busy
	369	+	interactive.mu.Unlock()
	370	+	if idle {
	371	+	break
	372	+	}
	373	+	select {
	374	+	case <-deadline.C:
	375	+	t.Fatal("interactive did not settle after forced continuation and queued prompt")
	376	+	case <-poll.C:
	377	+	}
	378	+	}
	379	+
	380	+	client.mu.Lock()
	381	+	calls := client.calls
	382	+	followUps := client.followUps
	383	+	client.mu.Unlock()
	384	+	if calls != 4 {
	385	+	t.Fatalf("provider called %d times, want initial request, compaction, forced continuation, and queued prompt", calls)
	386	+	}
	387	+	if followUps != 2 {
	388	+	t.Fatalf("follow-up requests = %d, want forced continuation and queued prompt", followUps)
	389	+	}
	390	+	}
	391	+
264	392		func TestPreTurnCompactionPreservesPromptImages(t *testing.T) {
265	393		client := &compactQueueClient{
266	394		compactionStarted: make(chan struct{}),


@@ -153,7 +153,7 @@ func TestPromptSubmittedDuringCompactionStartsFollowUpTurn(t *testing.T) {
153		interactive := NewInteractive(InteractiveConfig{Agent: agent})	153		interactive := NewInteractive(InteractiveConfig{Agent: agent})
154		interactive.runCtx = context.Background()	154		interactive.runCtx = context.Background()
155			155
156	-	interactive.runCompact(context.Background(), false)	156	+	interactive.runCompact(context.Background(), false, false)
157		select {	157		select {
158		case <-client.compactionStarted:	158		case <-client.compactionStarted:
159		case <-time.After(2 * time.Second):	159		case <-time.After(2 * time.Second):

@@ -178,9 +178,12 @@ type thresholdAutoCompactClient struct {
178		compactionStarted chan struct{}	178		compactionStarted chan struct{}
179		releaseCompaction chan struct{}	179		releaseCompaction chan struct{}
180		followUpRequest chan provider.Request	180		followUpRequest chan provider.Request
			181	+	firstStop provider.StopReason
			182	+	firstContent []provider.Content
181			183
182	-	mu sync.Mutex	184	+	mu sync.Mutex
183	-	calls int	185	+	calls int
			186	+	followUps int
184		}	187		}
185			188
186		func (c *thresholdAutoCompactClient) Name() string { return "threshold-auto-compact-test" }	189		func (c *thresholdAutoCompactClient) Name() string { return "threshold-auto-compact-test" }

@@ -196,12 +199,20 @@ func (c *thresholdAutoCompactClient) Stream(ctx context.Context, req provider.Re
196		defer close(out)	199		defer close(out)
197		switch call {	200		switch call {
198		case 1:	201		case 1:
			202	+	stop := c.firstStop
			203	+	if stop == "" {
			204	+	stop = provider.StopEnd
			205	+	}
			206	+	content := c.firstContent
			207	+	if len(content) == 0 {
			208	+	content = []provider.Content{provider.ReasoningBlock{Summary: "still working"}}
			209	+	}
199		out <- provider.EventUsage{Usage: provider.Usage{InputTokens: 150000}}	210		out <- provider.EventUsage{Usage: provider.Usage{InputTokens: 150000}}
200		out <- provider.EventDone{	211		out <- provider.EventDone{
201	-	Stop: provider.StopEnd,	212	+	Stop: stop,
202		Message: provider.Message{	213		Message: provider.Message{
203		Role: provider.RoleAssistant,	214		Role: provider.RoleAssistant,
204	-	Content: []provider.Content{provider.ReasoningBlock{Summary: "still working"}},	215	+	Content: content,
205		},	216		},
206		}	217		}
207		case 2:	218		case 2:

@@ -214,6 +225,9 @@ func (c *thresholdAutoCompactClient) Stream(ctx context.Context, req provider.Re
214		out <- provider.EventDone{Stop: provider.StopEnd}	225		out <- provider.EventDone{Stop: provider.StopEnd}
215		}	226		}
216		default:	227		default:
			228	+	c.mu.Lock()
			229	+	c.followUps++
			230	+	c.mu.Unlock()
217		c.followUpRequest <- req	231		c.followUpRequest <- req
218		out <- provider.EventDone{	232		out <- provider.EventDone{
219		Stop: provider.StopEnd,	233		Stop: provider.StopEnd,

@@ -261,6 +275,120 @@ func TestThresholdAutoCompactionContinuesMostRecentIntent(t *testing.T) {
261		}	275		}
262		}	276		}
263			277
			278	+	func TestThresholdAutoCompactionContinuesTruncatedText(t *testing.T) {
			279	+	client := &thresholdAutoCompactClient{
			280	+	compactionStarted: make(chan struct{}),
			281	+	releaseCompaction: make(chan struct{}),
			282	+	followUpRequest: make(chan provider.Request, 1),
			283	+	firstStop: provider.StopLength,
			284	+	firstContent: []provider.Content{provider.TextBlock{Text: "partial answer"}},
			285	+	}
			286	+	agent := core.NewAgent(client, "test-model", "", nil)
			287	+	threshold := 70
			288	+	interactive := NewInteractive(InteractiveConfig{
			289	+	Agent: agent,
			290	+	Provider: "anthropic",
			291	+	Model: "claude-sonnet-4-5-20250929",
			292	+	AutoCompactThreshold: &threshold,
			293	+	})
			294	+	interactive.runCtx = context.Background()
			295	+
			296	+	interactive.startTurn(context.Background(), "initial request")
			297	+	select {
			298	+	case <-client.compactionStarted:
			299	+	case <-time.After(2 * time.Second):
			300	+	t.Fatal("threshold auto-compaction did not start after truncated text")
			301	+	}
			302	+	close(client.releaseCompaction)
			303	+
			304	+	select {
			305	+	case req := <-client.followUpRequest:
			306	+	if got := requestUserTextCount(req, autoCompactContinuationPrompt); got != 1 {
			307	+	t.Fatalf("truncated continuation request contains auto-continue prompt %d times, want 1: %#v", got, req.Messages)
			308	+	}
			309	+	case <-time.After(2 * time.Second):
			310	+	t.Fatal("threshold auto-compaction did not continue truncated text")
			311	+	}
			312	+	}
			313	+
			314	+	func TestThresholdAutoCompactionForcedContinuationPrecedesQueuedPrompt(t *testing.T) {
			315	+	client := &thresholdAutoCompactClient{
			316	+	compactionStarted: make(chan struct{}),
			317	+	releaseCompaction: make(chan struct{}),
			318	+	followUpRequest: make(chan provider.Request, 2),
			319	+	firstStop: provider.StopLength,
			320	+	firstContent: []provider.Content{provider.TextBlock{Text: "partial answer"}},
			321	+	}
			322	+	agent := core.NewAgent(client, "test-model", "", nil)
			323	+	threshold := 70
			324	+	interactive := NewInteractive(InteractiveConfig{
			325	+	Agent: agent,
			326	+	Provider: "anthropic",
			327	+	Model: "claude-sonnet-4-5-20250929",
			328	+	AutoCompactThreshold: &threshold,
			329	+	})
			330	+	interactive.runCtx = context.Background()
			331	+
			332	+	interactive.startTurn(context.Background(), "initial request")
			333	+	select {
			334	+	case <-client.compactionStarted:
			335	+	case <-time.After(2 * time.Second):
			336	+	t.Fatal("threshold auto-compaction did not start after truncated text")
			337	+	}
			338	+	interactive.SubmitOrQueue("queued follow-up", nil)
			339	+	close(client.releaseCompaction)
			340	+
			341	+	select {
			342	+	case req := <-client.followUpRequest:
			343	+	if got := requestUserTextCount(req, autoCompactContinuationPrompt); got != 1 {
			344	+	t.Fatalf("first request contains auto-continue prompt %d times, want 1: %#v", got, req.Messages)
			345	+	}
			346	+	if got := requestUserTextCount(req, "queued follow-up"); got != 0 {
			347	+	t.Fatalf("forced continuation request included queued prompt %d times, want 0: %#v", got, req.Messages)
			348	+	}
			349	+	case <-time.After(2 * time.Second):
			350	+	t.Fatal("threshold auto-compaction did not continue truncated text")
			351	+	}
			352	+
			353	+	select {
			354	+	case req := <-client.followUpRequest:
			355	+	if got := requestUserTextCount(req, "queued follow-up"); got != 1 {
			356	+	t.Fatalf("second request contains queued prompt %d times, want 1: %#v", got, req.Messages)
			357	+	}
			358	+	case <-time.After(2 * time.Second):
			359	+	t.Fatal("queued prompt did not run after forced continuation")
			360	+	}
			361	+
			362	+	deadline := time.NewTimer(2 * time.Second)
			363	+	defer deadline.Stop()
			364	+	poll := time.NewTicker(time.Millisecond)
			365	+	defer poll.Stop()
			366	+	for {
			367	+	interactive.mu.Lock()
			368	+	idle := !interactive.busy
			369	+	interactive.mu.Unlock()
			370	+	if idle {
			371	+	break
			372	+	}
			373	+	select {
			374	+	case <-deadline.C:
			375	+	t.Fatal("interactive did not settle after forced continuation and queued prompt")
			376	+	case <-poll.C:
			377	+	}
			378	+	}
			379	+
			380	+	client.mu.Lock()
			381	+	calls := client.calls
			382	+	followUps := client.followUps
			383	+	client.mu.Unlock()
			384	+	if calls != 4 {
			385	+	t.Fatalf("provider called %d times, want initial request, compaction, forced continuation, and queued prompt", calls)
			386	+	}
			387	+	if followUps != 2 {
			388	+	t.Fatalf("follow-up requests = %d, want forced continuation and queued prompt", followUps)
			389	+	}
			390	+	}
			391	+
264		func TestPreTurnCompactionPreservesPromptImages(t *testing.T) {	392		func TestPreTurnCompactionPreservesPromptImages(t *testing.T) {
265		client := &compactQueueClient{	393		client := &compactQueueClient{
266		compactionStarted: make(chan struct{}),	394		compactionStarted: make(chan struct{}),

M packages/agent/modes/interactive.go +14 -10


@@ -5562,7 +5562,7 @@ func (i *Interactive) runSlash(ctx context.Context, cmd string) (done bool) {
5562	5562		case "/skills":
5563	5563		i.openSkillsDialog()
5564	5564		case "/compact":
5565		-	i.runCompact(ctx, false)
	5565	+	i.runCompact(ctx, false, false)
5566	5566		case "/study":
5567	5567		// Canned prompt that tells the agent to read every file
5568	5568		// in some target so its later turns have the whole thing

@@ -6495,7 +6495,7 @@ func (i *Interactive) clearPendingCompactTurnLocked() {
6495	6495		i.pendingPostCompactNote = ""
6496	6496		}
6497	6497
6498		-	const autoCompactContinuationPrompt = `Resume work on the user's most recent intent. Re-read the kept recent messages after the summary to confirm what the user asked for last. If their latest request supersedes earlier plans recorded in the summary, follow the latest request. If there is nothing left to do, say so briefly instead of inventing further work.`
	6498	+	const autoCompactContinuationPrompt = `Context was compacted while work was in progress. Continue the user's most recent active request now; do not wait for them to type "continue". Re-read the context summary and kept recent messages, treating active constraints and preferences there as still in force, including delegation/subagent instructions. If a newer request supersedes earlier plans in the summary, follow the newer request. If there is truly nothing left to do, say so briefly instead of inventing work.`
6499	6499
6500	6500		// startAutoCompactContinuation adds an internal user turn that tells the
6501	6501		// model to resume the task after threshold compaction. Provider requests

@@ -6530,7 +6530,9 @@ func (i *Interactive) startAutoCompactContinuation(parent context.Context) {
6530	6530		// When auto is true the spinner message is pinned to "condensing
6531	6531		// history" and the status bar surfaces "(auto)" next to the context
6532	6532		// percentage so it's obvious the system triggered this, not the user.
6533		-	func (i *Interactive) runCompact(parent context.Context, auto bool) {
	6533	+	// forceAutoContinue tells the post-compact hand-off to resume even if the
	6534	+	// transcript tail has visible assistant text, for example after StopLength.
	6535	+	func (i *Interactive) runCompact(parent context.Context, auto bool, forceAutoContinue bool) {
6534	6536		if i.agent == nil {
6535	6537		i.mu.Lock()
6536	6538		i.statusErr = "not logged in. type /login first."

@@ -6564,7 +6566,7 @@ func (i *Interactive) runCompact(parent context.Context, auto bool) {
6564	6566		// Sink discards deltas — we don't stream the summary to the UI.
6565	6567		sink := func(delta string) {}
6566	6568		msgsBefore := i.agent.Messages()
6567		-	autoContinue := auto && shouldAutoContinueAfterCompaction(msgsBefore)
	6569	+	autoContinue := auto && (forceAutoContinue \|\| shouldAutoContinueAfterCompaction(msgsBefore))
6568	6570		// Keep the usual recent tail when possible, but never let it cover
6569	6571		// the whole transcript. A short session can still be at 70–90% of
6570	6572		// a model's window because one prompt or tool result is large; the

@@ -6650,14 +6652,16 @@ func (i *Interactive) runCompact(parent context.Context, auto bool) {
6650	6652		i.pendingCompactImages = nil
6651	6653		i.hasPendingCompactPrompt = false
6652	6654		hasNext = true
	6655	+	case forceAutoContinue && autoContinue:
	6656	+	continueAutomatically = true
6653	6657		case len(i.queued) > 0:
6654	6658		next, i.queued = i.queued[0], i.queued[1:]
6655	6659		hasNext = true
6656	6660		case autoContinue:
6657	6661		// A successful threshold compaction must hand control back to
6658		-	// an unfinished model turn. Completed text answers should settle
6659		-	// without an invented extra request; the user can still prompt
6660		-	// again normally.
	6662	+	// unfinished work. Completed text answers still settle normally
	6663	+	// unless the caller knows the text was truncated and forces this
	6664	+	// continuation.
6661	6665		continueAutomatically = true
6662	6666		}
6663	6667		}

@@ -6868,7 +6872,7 @@ func (i *Interactive) startTurnRequest(parent context.Context, prompt string, im
6868	6872		i.pendingPostCompactNote = "context auto-compacted; sending your last message"
6869	6873		i.mu.Unlock()
6870	6874		i.invalidate()
6871		-	i.runCompact(parent, true)
	6875	+	i.runCompact(parent, true, false)
6872	6876		return
6873	6877		}
6874	6878		i.mu.Unlock()

@@ -7063,9 +7067,9 @@ func (i *Interactive) startTurnRequest(parent context.Context, prompt string, im
7063	7067		case offer:
7064	7068		i.openRescueDialog(rescueProv, rescueFprov, rescueModel, rescueWhy, prompt, rescueImgs)
7065	7069		case recoverContextOverflow:
7066		-	i.runCompact(parent, true)
	7070	+	i.runCompact(parent, true, false)
7067	7071		case shouldAutoCompact:
7068		-	i.runCompact(parent, true)
	7072	+	i.runCompact(parent, true, lastStop == provider.StopLength)
7069	7073		}
7070	7074		}()
7071	7075		}

 	case "/skills":
 		i.openSkillsDialog()
 	case "/compact":
-		i.runCompact(ctx, false)
+		i.runCompact(ctx, false, false)
 	case "/study":
 		// Canned prompt that tells the agent to read every file
 		// in some target so its later turns have the whole thing
 	i.pendingPostCompactNote = ""
 }
-const autoCompactContinuationPrompt = `Resume work on the user's most recent intent. Re-read the kept recent messages after the summary to confirm what the user asked for last. If their latest request supersedes earlier plans recorded in the summary, follow the latest request. If there is nothing left to do, say so briefly instead of inventing further work.`
+const autoCompactContinuationPrompt = `Context was compacted while work was in progress. Continue the user's most recent active request now; do not wait for them to type "continue". Re-read the context summary and kept recent messages, treating active constraints and preferences there as still in force, including delegation/subagent instructions. If a newer request supersedes earlier plans in the summary, follow the newer request. If there is truly nothing left to do, say so briefly instead of inventing work.`
 // startAutoCompactContinuation adds an internal user turn that tells the
 // model to resume the task after threshold compaction. Provider requests
 // When auto is true the spinner message is pinned to "condensing
 // history" and the status bar surfaces "(auto)" next to the context
 // percentage so it's obvious the system triggered this, not the user.
-func (i *Interactive) runCompact(parent context.Context, auto bool) {
+// forceAutoContinue tells the post-compact hand-off to resume even if the
+// transcript tail has visible assistant text, for example after StopLength.
+func (i *Interactive) runCompact(parent context.Context, auto bool, forceAutoContinue bool) {
 	if i.agent == nil {
 		i.mu.Lock()
 		i.statusErr = "not logged in. type /login first."
 		// Sink discards deltas — we don't stream the summary to the UI.
 		sink := func(delta string) {}
 		msgsBefore := i.agent.Messages()
-		autoContinue := auto && shouldAutoContinueAfterCompaction(msgsBefore)
+		autoContinue := auto && (forceAutoContinue || shouldAutoContinueAfterCompaction(msgsBefore))
 		// Keep the usual recent tail when possible, but never let it cover
 		// the whole transcript. A short session can still be at 70–90% of
 		// a model's window because one prompt or tool result is large; the
 				i.pendingCompactImages = nil
 				i.hasPendingCompactPrompt = false
 				hasNext = true
+			case forceAutoContinue && autoContinue:
+				continueAutomatically = true
 			case len(i.queued) > 0:
 				next, i.queued = i.queued[0], i.queued[1:]
 				hasNext = true
 			case autoContinue:
 				// A successful threshold compaction must hand control back to
-				// an unfinished model turn. Completed text answers should settle
+				// unfinished work. Completed text answers still settle normally
-				// without an invented extra request; the user can still prompt
+				// unless the caller knows the text was truncated and forces this
-				// again normally.
+				// continuation.
 				continueAutomatically = true
 			}
 		}
 		i.pendingPostCompactNote = "context auto-compacted; sending your last message"
 		i.mu.Unlock()
 		i.invalidate()
-		i.runCompact(parent, true)
+		i.runCompact(parent, true, false)
 		return
 	}
 	i.mu.Unlock()
 		case offer:
 			i.openRescueDialog(rescueProv, rescueFprov, rescueModel, rescueWhy, prompt, rescueImgs)
 		case recoverContextOverflow:
-			i.runCompact(parent, true)
+			i.runCompact(parent, true, false)
 		case shouldAutoCompact:
-			i.runCompact(parent, true)
+			i.runCompact(parent, true, lastStop == provider.StopLength)
 		}
 	}()
 }

M packages/agent/systemprompt.go +5 -0


@@ -64,6 +64,9 @@ func BuildSystemPrompt(o SystemPromptOpts) string {
64	64		sb.WriteString(defaultIdentity)
65	65		}
66	66
	67	+	sb.WriteString("\n\n")
	68	+	sb.WriteString(compactedSummaryHandoffInstruction)
	69	+
67	70		if strings.TrimSpace(o.ZutDocsDir) != "" {
68	71		sb.WriteString("\n\nZut's own docs are installed under ")
69	72		sb.WriteString(o.ZutDocsDir)

@@ -87,3 +90,5 @@ const defaultIdentity = `You are an expert coding assistant operating inside zut
87	90		Your output renders in a TUI that understands markdown for prose and plain text for tool output. Use markdown freely, keep answers concise, and let tool calls speak for themselves rather than narrating them in prose before you invoke them. Act first, then summarise what you did.
88	91
89	92		For focused changes to an existing file, inspect its current contents and use edit with verbatim oldText taken from that same file. Include only enough context to make each match unambiguous. Use write when creating a file or replacing it wholesale. Do not mutate files through bash redirections or commands such as cat, echo, sed, or tee, because those changes appear as opaque shell output instead of a readable edit diff.`
	93	+
	94	+	const compactedSummaryHandoffInstruction = `When you see a "## Context Summary (compacted)" message, treat it as a handoff from earlier work. Keep its active constraints and preferences in force, continue the most recent unresolved user request without waiting for the user to type "continue", and follow a newer user request when it supersedes the summary.`


@@ -64,6 +64,9 @@ func BuildSystemPrompt(o SystemPromptOpts) string {
64	sb.WriteString(defaultIdentity)	64		sb.WriteString(defaultIdentity)
65	}	65		}
66		66
		67	+	sb.WriteString("\n\n")
		68	+	sb.WriteString(compactedSummaryHandoffInstruction)
		69	+
67	if strings.TrimSpace(o.ZutDocsDir) != "" {	70		if strings.TrimSpace(o.ZutDocsDir) != "" {
68	sb.WriteString("\n\nZut's own docs are installed under ")	71		sb.WriteString("\n\nZut's own docs are installed under ")
69	sb.WriteString(o.ZutDocsDir)	72		sb.WriteString(o.ZutDocsDir)

@@ -87,3 +90,5 @@ const defaultIdentity = `You are an expert coding assistant operating inside zut
87	Your output renders in a TUI that understands markdown for prose and plain text for tool output. Use markdown freely, keep answers concise, and let tool calls speak for themselves rather than narrating them in prose before you invoke them. Act first, then summarise what you did.	90		Your output renders in a TUI that understands markdown for prose and plain text for tool output. Use markdown freely, keep answers concise, and let tool calls speak for themselves rather than narrating them in prose before you invoke them. Act first, then summarise what you did.
88		91
89	For focused changes to an existing file, inspect its current contents and use edit with verbatim oldText taken from that same file. Include only enough context to make each match unambiguous. Use write when creating a file or replacing it wholesale. Do not mutate files through bash redirections or commands such as cat, echo, sed, or tee, because those changes appear as opaque shell output instead of a readable edit diff.`	92		For focused changes to an existing file, inspect its current contents and use edit with verbatim oldText taken from that same file. Include only enough context to make each match unambiguous. Use write when creating a file or replacing it wholesale. Do not mutate files through bash redirections or commands such as cat, echo, sed, or tee, because those changes appear as opaque shell output instead of a readable edit diff.`
		93	+
		94	+	const compactedSummaryHandoffInstruction = `When you see a "## Context Summary (compacted)" message, treat it as a handoff from earlier work. Keep its active constraints and preferences in force, continue the most recent unresolved user request without waiting for the user to type "continue", and follow a newer user request when it supersedes the summary.`

A packages/agent/systemprompt_test.go +19 -0

@@ -0,0 +1,19 @@
	1	+	package agent
	2	+
	3	+	import (
	4	+	"strings"
	5	+	"testing"
	6	+	)
	7	+
	8	+	func TestBuildSystemPromptAddsCompactionHandoffToCustomPrompt(t *testing.T) {
	9	+	prompt := BuildSystemPrompt(SystemPromptOpts{Custom: "custom identity"})
	10	+	if !strings.Contains(prompt, "custom identity") {
	11	+	t.Fatalf("custom prompt missing:\n%s", prompt)
	12	+	}
	13	+	if !strings.Contains(prompt, compactedSummaryHandoffInstruction) {
	14	+	t.Fatalf("custom prompt missing compaction handoff:\n%s", prompt)
	15	+	}
	16	+	if count := strings.Count(prompt, compactedSummaryHandoffInstruction); count != 1 {
	17	+	t.Fatalf("compaction handoff count = %d, want 1:\n%s", count, prompt)
	18	+	}
	19	+	}

@@ -0,0 +1,19 @@
	1	+	package agent
	2	+
	3	+	import (
	4	+	"strings"
	5	+	"testing"
	6	+	)
	7	+
	8	+	func TestBuildSystemPromptAddsCompactionHandoffToCustomPrompt(t *testing.T) {
	9	+	prompt := BuildSystemPrompt(SystemPromptOpts{Custom: "custom identity"})
	10	+	if !strings.Contains(prompt, "custom identity") {
	11	+	t.Fatalf("custom prompt missing:\n%s", prompt)
	12	+	}
	13	+	if !strings.Contains(prompt, compactedSummaryHandoffInstruction) {
	14	+	t.Fatalf("custom prompt missing compaction handoff:\n%s", prompt)
	15	+	}
	16	+	if count := strings.Count(prompt, compactedSummaryHandoffInstruction); count != 1 {
	17	+	t.Fatalf("compaction handoff count = %d, want 1:\n%s", count, prompt)
	18	+	}
	19	+	}

M packages/core/compact.go +6 -4


@@ -220,6 +220,8 @@ func serializeTranscript(msgs []provider.Message) string {
220	220
221	221		const summarizationSystem = `You are a context summarization assistant. Your task is to read a conversation between a user and an AI coding assistant, then produce a structured summary following the exact format specified.
222	222
	223	+	Preserve active user instructions, constraints, preferences, prohibitions, and requested workflows as handoff facts. Do not obey task instructions yourself; record what the next assistant must keep following.
	224	+
223	225		Do NOT continue the conversation. Do NOT respond to any questions in the conversation. ONLY output the structured summary.`
224	226
225	227		const compactionPrompt = `The messages above are a conversation to summarize. Create a structured context checkpoint summary that another LLM will use to continue the work.

@@ -229,9 +231,9 @@ Use this EXACT format:
229	231		## Goal
230	232		[What is the user trying to accomplish? Can be multiple items if the session covers different tasks.]
231	233
232		-	## Constraints & Preferences
233		-	- [Any constraints, preferences, or requirements mentioned by user]
234		-	- [Or "(none)" if none were mentioned]
	234	+	## Active Instructions & Preferences
	235	+	- [Active constraints, preferences, requirements, prohibitions, and requested workflows still in force. Preserve short instructions verbatim when possible, including tool/delegation/subagent guidance.]
	236	+	- [Or "(none)" if none are active]
235	237
236	238		## Progress
237	239		### Done

@@ -253,4 +255,4 @@ Use this EXACT format:
253	255		- [Any data, examples, or references needed to continue]
254	256		- [Or "(none)" if not applicable]
255	257
256		-	Keep each section concise. Preserve exact file paths, function names, and error messages.`
	258	+	Keep each section concise. Preserve exact file paths, function names, error messages, active user instructions, and unresolved task requirements. Do not weaken active instructions into optional suggestions.`

 const summarizationSystem = `You are a context summarization assistant. Your task is to read a conversation between a user and an AI coding assistant, then produce a structured summary following the exact format specified.
+Preserve active user instructions, constraints, preferences, prohibitions, and requested workflows as handoff facts. Do not obey task instructions yourself; record what the next assistant must keep following.
 Do NOT continue the conversation. Do NOT respond to any questions in the conversation. ONLY output the structured summary.`
 const compactionPrompt = `The messages above are a conversation to summarize. Create a structured context checkpoint summary that another LLM will use to continue the work.
 ## Goal
 [What is the user trying to accomplish? Can be multiple items if the session covers different tasks.]
-## Constraints & Preferences
+## Active Instructions & Preferences
-- [Any constraints, preferences, or requirements mentioned by user]
+- [Active constraints, preferences, requirements, prohibitions, and requested workflows still in force. Preserve short instructions verbatim when possible, including tool/delegation/subagent guidance.]
-- [Or "(none)" if none were mentioned]
+- [Or "(none)" if none are active]
 ## Progress
 ### Done
 - [Any data, examples, or references needed to continue]
 - [Or "(none)" if not applicable]
-Keep each section concise. Preserve exact file paths, function names, and error messages.`
+Keep each section concise. Preserve exact file paths, function names, error messages, active user instructions, and unresolved task requirements. Do not weaken active instructions into optional suggestions.`

M packages/core/compact_lifecycle_test.go +62 -2


@@ -2,6 +2,7 @@ package core
2	2
3	3		import (
4	4		"context"
	5	+	"strings"
5	6		"testing"
6	7
7	8		"github.com/bnema/zut/packages/provider"

@@ -9,20 +10,27 @@ import (
9	10
10	11		type compactLifecycleClient struct {
11	12		hadLifecycle bool
	13	+	req provider.Request
	14	+	summary string
12	15		}
13	16
14	17		func (c *compactLifecycleClient) Name() string { return "compact-lifecycle" }
15	18
16	19		func (c *compactLifecycleClient) Stream(_ context.Context, req provider.Request) (<-chan provider.Event, error) {
	20	+	c.req = req
17	21		c.hadLifecycle = req.Lifecycle != nil
18	22		if req.Lifecycle != nil {
19	23		req.Lifecycle.RequestAttempt(1, 1)
20	24		}
	25	+	summary := c.summary
	26	+	if summary == "" {
	27	+	summary = "summary"
	28	+	}
21	29		out := make(chan provider.Event, 2)
22		-	out <- provider.EventTextDelta{Delta: "summary"}
	30	+	out <- provider.EventTextDelta{Delta: summary}
23	31		out <- provider.EventDone{Stop: provider.StopEnd, Message: provider.Message{
24	32		Role: provider.RoleAssistant,
25		-	Content: []provider.Content{provider.TextBlock{Text: "summary"}},
	33	+	Content: []provider.Content{provider.TextBlock{Text: summary}},
26	34		}}
27	35		close(out)
28	36		return out, nil

@@ -68,3 +76,55 @@ func TestCompactWithEventsEmitsRequestLifecycle(t *testing.T) {
68	76		t.Fatalf("fourth event = %#v, want summary text delta", events[3])
69	77		}
70	78		}
	79	+
	80	+	func TestCompactPromptPreservesActiveInstructions(t *testing.T) {
	81	+	const activeInstruction = "delegate independent work to subagents"
	82	+	client := &compactLifecycleClient{summary: "Active instruction still in force: " + activeInstruction}
	83	+	agent := NewAgent(client, "compact-model", "system", Registry{})
	84	+	agent.SetMessages([]provider.Message{
	85	+	{
	86	+	Role: provider.RoleUser,
	87	+	Content: []provider.Content{provider.TextBlock{Text: "When coding, delegate independent work to subagents."}},
	88	+	},
	89	+	{
	90	+	Role: provider.RoleAssistant,
	91	+	Content: []provider.Content{provider.TextBlock{Text: "I will do that."}},
	92	+	},
	93	+	})
	94	+
	95	+	if _, err := agent.Compact(context.Background(), 0, nil); err != nil {
	96	+	t.Fatalf("Compact: %v", err)
	97	+	}
	98	+	messages := agent.Messages()
	99	+	if len(messages) != 1 {
	100	+	t.Fatalf("compacted messages = %d, want 1", len(messages))
	101	+	}
	102	+	if messages[0].Role != provider.RoleUser {
	103	+	t.Fatalf("compacted message role = %q, want user", messages[0].Role)
	104	+	}
	105	+	if messages[0].Meta["compaction"] != "true" {
	106	+	t.Fatalf("compacted message meta = %#v, want compaction=true", messages[0].Meta)
	107	+	}
	108	+	compacted, ok := messages[0].Content[0].(provider.TextBlock)
	109	+	if !ok {
	110	+	t.Fatalf("compacted message content = %#v, want TextBlock", messages[0].Content[0])
	111	+	}
	112	+	if !strings.Contains(compacted.Text, "## Context Summary (compacted)") \|\| !strings.Contains(compacted.Text, activeInstruction) {
	113	+	t.Fatalf("compacted transcript did not preserve active instruction:\n%s", compacted.Text)
	114	+	}
	115	+	if !strings.Contains(client.req.System, "Preserve active user instructions") {
	116	+	t.Fatalf("compaction system prompt does not preserve active instructions: %q", client.req.System)
	117	+	}
	118	+	if len(client.req.Messages) != 1 {
	119	+	t.Fatalf("compaction request messages = %d, want 1", len(client.req.Messages))
	120	+	}
	121	+	text, ok := client.req.Messages[0].Content[0].(provider.TextBlock)
	122	+	if !ok {
	123	+	t.Fatalf("compaction request content = %#v, want TextBlock", client.req.Messages[0].Content[0])
	124	+	}
	125	+	for _, want := range []string{"## Active Instructions & Preferences", "delegation/subagent", activeInstruction} {
	126	+	if !strings.Contains(text.Text, want) {
	127	+	t.Fatalf("compaction prompt missing %q:\n%s", want, text.Text)
	128	+	}
	129	+	}
	130	+	}


@@ -2,6 +2,7 @@ package core
2			2
3		import (	3		import (
4		"context"	4		"context"
			5	+	"strings"
5		"testing"	6		"testing"
6			7
7		"github.com/bnema/zut/packages/provider"	8		"github.com/bnema/zut/packages/provider"

@@ -9,20 +10,27 @@ import (
9			10
10		type compactLifecycleClient struct {	11		type compactLifecycleClient struct {
11		hadLifecycle bool	12		hadLifecycle bool
			13	+	req provider.Request
			14	+	summary string
12		}	15		}
13			16
14		func (c *compactLifecycleClient) Name() string { return "compact-lifecycle" }	17		func (c *compactLifecycleClient) Name() string { return "compact-lifecycle" }
15			18
16		func (c *compactLifecycleClient) Stream(_ context.Context, req provider.Request) (<-chan provider.Event, error) {	19		func (c *compactLifecycleClient) Stream(_ context.Context, req provider.Request) (<-chan provider.Event, error) {
			20	+	c.req = req
17		c.hadLifecycle = req.Lifecycle != nil	21		c.hadLifecycle = req.Lifecycle != nil
18		if req.Lifecycle != nil {	22		if req.Lifecycle != nil {
19		req.Lifecycle.RequestAttempt(1, 1)	23		req.Lifecycle.RequestAttempt(1, 1)
20		}	24		}
			25	+	summary := c.summary
			26	+	if summary == "" {
			27	+	summary = "summary"
			28	+	}
21		out := make(chan provider.Event, 2)	29		out := make(chan provider.Event, 2)
22	-	out <- provider.EventTextDelta{Delta: "summary"}	30	+	out <- provider.EventTextDelta{Delta: summary}
23		out <- provider.EventDone{Stop: provider.StopEnd, Message: provider.Message{	31		out <- provider.EventDone{Stop: provider.StopEnd, Message: provider.Message{
24		Role: provider.RoleAssistant,	32		Role: provider.RoleAssistant,
25	-	Content: []provider.Content{provider.TextBlock{Text: "summary"}},	33	+	Content: []provider.Content{provider.TextBlock{Text: summary}},
26		}}	34		}}
27		close(out)	35		close(out)
28		return out, nil	36		return out, nil

@@ -68,3 +76,55 @@ func TestCompactWithEventsEmitsRequestLifecycle(t *testing.T) {
68		t.Fatalf("fourth event = %#v, want summary text delta", events[3])	76		t.Fatalf("fourth event = %#v, want summary text delta", events[3])
69		}	77		}
70		}	78		}
			79	+
			80	+	func TestCompactPromptPreservesActiveInstructions(t *testing.T) {
			81	+	const activeInstruction = "delegate independent work to subagents"
			82	+	client := &compactLifecycleClient{summary: "Active instruction still in force: " + activeInstruction}
			83	+	agent := NewAgent(client, "compact-model", "system", Registry{})
			84	+	agent.SetMessages([]provider.Message{
			85	+	{
			86	+	Role: provider.RoleUser,
			87	+	Content: []provider.Content{provider.TextBlock{Text: "When coding, delegate independent work to subagents."}},
			88	+	},
			89	+	{
			90	+	Role: provider.RoleAssistant,
			91	+	Content: []provider.Content{provider.TextBlock{Text: "I will do that."}},
			92	+	},
			93	+	})
			94	+
			95	+	if _, err := agent.Compact(context.Background(), 0, nil); err != nil {
			96	+	t.Fatalf("Compact: %v", err)
			97	+	}
			98	+	messages := agent.Messages()
			99	+	if len(messages) != 1 {
			100	+	t.Fatalf("compacted messages = %d, want 1", len(messages))
			101	+	}
			102	+	if messages[0].Role != provider.RoleUser {
			103	+	t.Fatalf("compacted message role = %q, want user", messages[0].Role)
			104	+	}
			105	+	if messages[0].Meta["compaction"] != "true" {
			106	+	t.Fatalf("compacted message meta = %#v, want compaction=true", messages[0].Meta)
			107	+	}
			108	+	compacted, ok := messages[0].Content[0].(provider.TextBlock)
			109	+	if !ok {
			110	+	t.Fatalf("compacted message content = %#v, want TextBlock", messages[0].Content[0])
			111	+	}
			112	+	if !strings.Contains(compacted.Text, "## Context Summary (compacted)") \|\| !strings.Contains(compacted.Text, activeInstruction) {
			113	+	t.Fatalf("compacted transcript did not preserve active instruction:\n%s", compacted.Text)
			114	+	}
			115	+	if !strings.Contains(client.req.System, "Preserve active user instructions") {
			116	+	t.Fatalf("compaction system prompt does not preserve active instructions: %q", client.req.System)
			117	+	}
			118	+	if len(client.req.Messages) != 1 {
			119	+	t.Fatalf("compaction request messages = %d, want 1", len(client.req.Messages))
			120	+	}
			121	+	text, ok := client.req.Messages[0].Content[0].(provider.TextBlock)
			122	+	if !ok {
			123	+	t.Fatalf("compaction request content = %#v, want TextBlock", client.req.Messages[0].Content[0])
			124	+	}
			125	+	for _, want := range []string{"## Active Instructions & Preferences", "delegation/subagent", activeInstruction} {
			126	+	if !strings.Contains(text.Text, want) {
			127	+	t.Fatalf("compaction prompt missing %q:\n%s", want, text.Text)
			128	+	}
			129	+	}
			130	+	}