feat(subagents): bound resident children by context, not model steps (#202)

* feat(subagents): bound resident children by context, not model steps

A resident child inherits the parent's resolved step limit instead of a
hardcoded 50-step fallback, so an unlimited parent no longer caps a child
and an explicit --max-steps is still enforced.

The context reminder re-arms once per band at 85, 90, and 95 percent
instead of firing once per runner, and a provider context-overflow error
now recovers once per accepted turn: the child compacts its transcript,
keeps the two most recent messages, journals the replacement as a
child.compacted record, and continues the same turn. Recovery preserves an
explicit step allowance, and a compaction or persistence failure restores
the pre-compaction transcript and fails the turn.

Resume replay starts from the newest checkpoint; the history pager keeps
the append-only log. Reconciliation restarts tool bookkeeping from the
checkpoint so a call dropped by compaction cannot gain a synthetic result.

* fix(subagents): keep compaction checkpoints free of stale host context

Review of the context-bounded child change found four correctness gaps:

- The summarization request's usage became the per-turn gauge, so a
  successful recovery could inject a reminder claiming the compacted
  transcript sat at the top of the window. Recovery now clears that gauge.
- The checkpoint carried host-authored internal context, so a stale
  reminder was replayed verbatim on resume. core.IsInternalContextMessage
  is exported and the runner excludes those messages from the durable
  replacement transcript.
- A checkpoint could be attributed to a queued follow-up instead of the
  executing turn. RecordTurnStarted is now the only writer of the active
  turn ID.
- Reconciliation treated a checkpoint result without its call as
  repairable work and rewrote the lifecycle time on every pass. Dangling
  calls are now the exact repair trigger.

appendSync also applies the reader's newline-inclusive record bound, so the
checkpoint writer no longer duplicates that check, and the reminder bands
use a fixed array.

New regression tests cover the gauge reset, the checkpoint contents, turn
attribution, reconciliation idempotence, a partial step allowance, and an
overflow without enough history to compact.

* test(subagents): cover the checkpoint record bound and reminder recovery

Close the coverage gaps left after the compaction-checkpoint fixes:

- appendSync now has a boundary test for the largest record the readers
  accept and for the one-byte-larger record it must reject;
- a dangling call kept by a checkpoint is repaired once and stays a stable
  no-op on later passes;
- residentCheckpointMessages keeps non-internal developer messages;
- the gauge-reset test also proves a truthful reminder still fires on a
  later continuation step.

The reminder documentation now states what the code does: the per-turn
gauge resets, while the band ladder advances at most once per child runner.

commit: 0d89d165e1a01771405514d4fb1b3f703b89af80
author: bnema · 41d ago · 2026-09-11 05:45 UTC
committer: GitHub · 41d ago · 2026-09-11 05:45 UTC
parent: 21873e8
signature: GPG UnverifiedB5690EEEBB952194 The key is unknown

9 files · +1449 -146

M README.md +1 -1


@@ -442,7 +442,7 @@ Resident subagents run alongside the main session as independent in-process agen
442	442
443	443		The fair global scheduler defaults to six concurrent turns, one active turn per child. `subagents.max_concurrent` overrides that limit. Queueing has no default deadline; a positive `subagents.queue_timeout` durably fails a prompt that waits too long for a slot. Required child work remains an obligation until an explicit successful follow-up.
444	444
445		-	Each child runs without a cumulative token budget; execution stays bounded through a finite per-turn step default. A child whose last-turn prompt reaches 85% of its model context window receives one non-blocking reminder to finish up, injected into its next provider request. Retrieve a terminal result without a model call using `subagent_status` with `agent_id` and `include_result: true`. After checking the saved progress and side effects, an explicit resume continues the retained session. Required work stays unresolved until success. See [lifecycle and persistence](docs/subagents.md#lifecycle-and-persistence) for details.
	445	+	Each child runs without a cumulative token budget or step cap: it inherits the parent's resolved `--max-steps`, unlimited by default, and each turn is bounded by the model context window with one compaction recovery. A child whose last-turn prompt reaches 85%, 90%, or 95% of its model context window receives one non-blocking reminder per band to finish up, injected into its next provider request. Retrieve a terminal result without a model call using `subagent_status` with `agent_id` and `include_result: true`. After checking the saved progress and side effects, an explicit resume continues the retained session. Required work stays unresolved until success. See [lifecycle and persistence](docs/subagents.md#lifecycle-and-persistence) for details.
446	446
447	447		Worktree isolation captures a patch and changed-file list without applying it. Shared children use the host checkout. The available logical references are `subagent://<id>`, `subagent://<id>/history`, `subagent://<id>/result`, and `subagent://<id>/patch`.
448	448


@@ -442,7 +442,7 @@ Resident subagents run alongside the main session as independent in-process agen
442			442
443		The fair global scheduler defaults to six concurrent turns, one active turn per child. `subagents.max_concurrent` overrides that limit. Queueing has no default deadline; a positive `subagents.queue_timeout` durably fails a prompt that waits too long for a slot. Required child work remains an obligation until an explicit successful follow-up.	443		The fair global scheduler defaults to six concurrent turns, one active turn per child. `subagents.max_concurrent` overrides that limit. Queueing has no default deadline; a positive `subagents.queue_timeout` durably fails a prompt that waits too long for a slot. Required child work remains an obligation until an explicit successful follow-up.
444			444
445	-	Each child runs without a cumulative token budget; execution stays bounded through a finite per-turn step default. A child whose last-turn prompt reaches 85% of its model context window receives one non-blocking reminder to finish up, injected into its next provider request. Retrieve a terminal result without a model call using `subagent_status` with `agent_id` and `include_result: true`. After checking the saved progress and side effects, an explicit resume continues the retained session. Required work stays unresolved until success. See [lifecycle and persistence](docs/subagents.md#lifecycle-and-persistence) for details.	445	+	Each child runs without a cumulative token budget or step cap: it inherits the parent's resolved `--max-steps`, unlimited by default, and each turn is bounded by the model context window with one compaction recovery. A child whose last-turn prompt reaches 85%, 90%, or 95% of its model context window receives one non-blocking reminder per band to finish up, injected into its next provider request. Retrieve a terminal result without a model call using `subagent_status` with `agent_id` and `include_result: true`. After checking the saved progress and side effects, an explicit resume continues the retained session. Required work stays unresolved until success. See [lifecycle and persistence](docs/subagents.md#lifecycle-and-persistence) for details.
446			446
447		Worktree isolation captures a patch and changed-file list without applying it. Shared children use the host checkout. The available logical references are `subagent://<id>`, `subagent://<id>/history`, `subagent://<id>/result`, and `subagent://<id>/patch`.	447		Worktree isolation captures a patch and changed-file list without applying it. Shared children use the host checkout. The available logical references are `subagent://<id>`, `subagent://<id>/history`, `subagent://<id>/result`, and `subagent://<id>/patch`.
448			448

M docs/subagents.md +46 -14


@@ -52,7 +52,7 @@ The manager records each accepted child under its managed state root:
52	52
53	53		```text
54	54		subagents/<child-id>/
55		-	transcript.jsonl # authoritative accepted turns and finalized messages
	55	+	transcript.jsonl # accepted turns, finalized messages, and compaction checkpoints
56	56		owner.lock # host-ownership coordination; do not modify or remove
57	57		.transcript-backup-* # retained private pre-repair transcript, when repaired
58	58		metadata.json # rebuildable bounded state

@@ -63,7 +63,10 @@ subagents/<child-id>/
63	63		Acceptance is durable before `subagent_spawn` reports success. Finalized user,
64	64		assistant, tool-call, and tool-result messages are journaled; streamed deltas
65	65		and hidden reasoning are not exposed as ordinary history. A tool call is stored
66		-	before execution and reconciliation repairs incomplete call/result pairs.
	66	+	before execution and reconciliation repairs incomplete call/result pairs. A
	67	+	context-overflow recovery appends one `child.compacted` checkpoint carrying the
	68	+	replacement transcript; resume starts from the newest checkpoint, while the
	69	+	paged history view still shows everything appended before it.
67	70
68	71		All children stop when the host exits. On the next start, queued or running
69	72		turns are marked `interrupted`; zut never replays their task. Resume only with

@@ -90,9 +93,10 @@ unlimited by default; a positive `subagents.queue_timeout` cancels an accepted
90	93		prompt that has not received a slot and records its terminal failure durably.
91	94		`subagents.allowed_tools` is an allowlist for child-visible tools and
92	95		`subagents.allowed_roots` limits eligible child workspaces. Resident children
93		-	have no cumulative token budget. Execution stays bounded through a finite
94		-	per-turn step default: a child whose parent resolution is unlimited runs with
95		-	`MaxSteps=50`; an explicit parent limit is preserved. Removed legacy
	96	+	have no cumulative token budget and no child-only step cap: a child inherits
	97	+	the parent's resolved limit, including the unlimited default. A turn is bounded
	98	+	by the provider context window, with one compaction recovery, and by an
	99	+	explicit parent `--max-steps`. Removed legacy
96	100		settings, including `tui_subagent_position` and `subagents.budget_ratio`, are
97	101		ignored. Resident journals written by older versions are read with explicit
98	102		translation (an old `budget_exhausted` outcome reconciles as failure) or, for

@@ -109,21 +113,49 @@ no longer exposes `budget` or `budget_source` fields.
109	113
110	114		### Context reminders
111	115
112		-	When a child's last-turn prompt reaches 85% of its model context window, the
113		-	runner injects one non-blocking reminder into the next provider request:
	116	+	When a child's last-turn prompt reaches 85%, 90%, or 95% of its model context
	117	+	window, the runner injects a non-blocking reminder into the next provider
	118	+	request:
114	119
115	120		```text
116	121		Context usage is at NN% of the model window. Finish the current task, report
117	122		results, limits, and remaining verifications, and do not start broad new work.
118	123		```
119	124
120		-	The reminder fires at most once per child runner and never blocks a turn.
121		-	It travels as developer context in the live request; the durable transcript
122		-	keeps only finalized history. Usage is the last-turn prompt size (input plus
123		-	cache read/write tokens) against the resolved context window; an unknown
124		-	window never triggers a reminder. The per-turn gauge resets when the host
125		-	restarts, so a restarted child is reminded again only once fresh usage past
126		-	the threshold arrives.
	125	+	Each band fires at most once per child runner and reminders never block a turn.
	126	+	A jump straight to a higher band delivers one reminder for that band, and usage
	127	+	that falls does not re-deliver a band. Reminders travel as developer context in
	128	+	the live request; the durable transcript keeps only finalized history. Usage is
	129	+	the last-turn prompt size (input plus cache read/write tokens) against the
	130	+	resolved context window; an unknown window never triggers a reminder. The
	131	+	per-turn gauge resets when the host restarts and when a recovery compacts the
	132	+	context, so neither event reports a stale percentage. The band ladder itself
	133	+	advances at most once per child runner: a restart starts a fresh ladder, while a
	134	+	compacted child is reminded again only when usage reaches a higher band.
	135	+
	136	+	### Context-overflow recovery
	137	+
	138	+	A provider request that exceeds the model context window ends the turn. When
	139	+	the transcript holds at least four messages, the child recovers once per turn:
	140	+	it compacts the transcript, keeps the two most recent messages verbatim,
	141	+	journals the replacement as a `child.compacted` record, and continues the same
	142	+	turn. A compaction or persistence failure restores the pre-compaction
	143	+	transcript and fails the turn, and a second overflow stays terminal. Recovery
	144	+	never replenishes an explicit parent `--max-steps`: when the allowance is
	145	+	already spent, the checkpoint is still written for a later resume, but the turn
	146	+	is not continued.
	147	+
	148	+	Explicit resume replays the newest checkpoint instead of the full append-only
	149	+	log. The history views (`/subagents logs`, `HistoryPage`, `subagent://` refs)
	150	+	still show the append-only log, including pre-compaction records. A build that
	151	+	predates checkpoint support cannot resume such a child correctly; downgrading
	152	+	is unsupported.
	153	+
	154	+	The summarization request itself can be rejected when a single oversized
	155	+	message leaves nothing to summarize away. The turn then fails with the original
	156	+	provider error and no checkpoint is written, so the child keeps failing on
	157	+	resume until its transcript shrinks. Recovery is a best effort, not a
	158	+	guarantee.
127	159
128	160		Retrieve the saved result without executing a model:
129	161

 ```text
 subagents/<child-id>/
-  transcript.jsonl  # authoritative accepted turns and finalized messages
+  transcript.jsonl  # accepted turns, finalized messages, and compaction checkpoints
   owner.lock        # host-ownership coordination; do not modify or remove
   .transcript-backup-* # retained private pre-repair transcript, when repaired
   metadata.json     # rebuildable bounded state
 Acceptance is durable before `subagent_spawn` reports success. Finalized user,
 assistant, tool-call, and tool-result messages are journaled; streamed deltas
 and hidden reasoning are not exposed as ordinary history. A tool call is stored
-before execution and reconciliation repairs incomplete call/result pairs.
+before execution and reconciliation repairs incomplete call/result pairs. A
+context-overflow recovery appends one `child.compacted` checkpoint carrying the
+replacement transcript; resume starts from the newest checkpoint, while the
+paged history view still shows everything appended before it.
 All children stop when the host exits. On the next start, queued or running
 turns are marked `interrupted`; zut never replays their task. Resume only with
 prompt that has not received a slot and records its terminal failure durably.
 `subagents.allowed_tools` is an allowlist for child-visible tools and
 `subagents.allowed_roots` limits eligible child workspaces. Resident children
-have no cumulative token budget. Execution stays bounded through a finite
+have no cumulative token budget and no child-only step cap: a child inherits
-per-turn step default: a child whose parent resolution is unlimited runs with
+the parent's resolved limit, including the unlimited default. A turn is bounded
-`MaxSteps=50`; an explicit parent limit is preserved. Removed legacy
+by the provider context window, with one compaction recovery, and by an
+explicit parent `--max-steps`. Removed legacy
 settings, including `tui_subagent_position` and `subagents.budget_ratio`, are
 ignored. Resident journals written by older versions are read with explicit
 translation (an old `budget_exhausted` outcome reconciles as failure) or, for
 ### Context reminders
-When a child's last-turn prompt reaches 85% of its model context window, the
+When a child's last-turn prompt reaches 85%, 90%, or 95% of its model context
-runner injects one non-blocking reminder into the next provider request:
+window, the runner injects a non-blocking reminder into the next provider
+request:
 ```text
 Context usage is at NN% of the model window. Finish the current task, report
 results, limits, and remaining verifications, and do not start broad new work.
 ```
-The reminder fires at most once per child runner and never blocks a turn.
+Each band fires at most once per child runner and reminders never block a turn.
-It travels as developer context in the live request; the durable transcript
+A jump straight to a higher band delivers one reminder for that band, and usage
-keeps only finalized history. Usage is the last-turn prompt size (input plus
+that falls does not re-deliver a band. Reminders travel as developer context in
-cache read/write tokens) against the resolved context window; an unknown
+the live request; the durable transcript keeps only finalized history. Usage is
-window never triggers a reminder. The per-turn gauge resets when the host
+the last-turn prompt size (input plus cache read/write tokens) against the
-restarts, so a restarted child is reminded again only once fresh usage past
+resolved context window; an unknown window never triggers a reminder. The
-the threshold arrives.
+per-turn gauge resets when the host restarts and when a recovery compacts the
+context, so neither event reports a stale percentage. The band ladder itself
+advances at most once per child runner: a restart starts a fresh ladder, while a
+compacted child is reminded again only when usage reaches a higher band.
+### Context-overflow recovery
+A provider request that exceeds the model context window ends the turn. When
+the transcript holds at least four messages, the child recovers once per turn:
+it compacts the transcript, keeps the two most recent messages verbatim,
+journals the replacement as a `child.compacted` record, and continues the same
+turn. A compaction or persistence failure restores the pre-compaction
+transcript and fails the turn, and a second overflow stays terminal. Recovery
+never replenishes an explicit parent `--max-steps`: when the allowance is
+already spent, the checkpoint is still written for a later resume, but the turn
+is not continued.
+Explicit resume replays the newest checkpoint instead of the full append-only
+log. The history views (`/subagents logs`, `HistoryPage`, `subagent://` refs)
+still show the append-only log, including pre-compaction records. A build that
+predates checkpoint support cannot resume such a child correctly; downgrading
+is unsupported.
+The summarization request itself can be rejected when a single oversized
+message leaves nothing to summarize away. The turn then fails with the original
+provider error and no checkpoint is written, so the child keeps failing on
+resume until its transcript shrinks. Recovery is a best effort, not a
+guarantee.
 Retrieve the saved result without executing a model:

M packages/agent/subagent_factory.go +132 -26


@@ -2,6 +2,7 @@ package agent
2	2
3	3		import (
4	4		"context"
	5	+	"errors"
5	6		"fmt"
6	7		"strings"
7	8		"sync"

@@ -9,6 +10,7 @@ import (
9	10		"github.com/bnema/zut/packages/agent/subagents"
10	11		"github.com/bnema/zut/packages/agent/tools"
11	12		"github.com/bnema/zut/packages/core"
	13	+	"github.com/bnema/zut/packages/provider"
12	14		)
13	15
14	16		// newResidentChildRunner is the host-owned construction boundary for one

@@ -37,13 +39,11 @@ func newResidentChildRunner(args Args, spec subagents.ResidentChildSpec, journal
37	39		}
38	40		}
39	41		agent := core.NewAgent(resolved.NewClient(), resolved.Model, system, registry)
	42	+	// A resident child inherits the parent's resolved step limit, including the
	43	+	// unlimited default. Its ceiling is the provider context window, with one
	44	+	// compaction recovery per accepted turn, plus any explicit parent
	45	+	// --max-steps. There is no child-only step budget.
40	46		agent.MaxSteps = resolved.MaxSteps
41		-	if agent.MaxSteps <= 0 {
42		-	// Resident execution must stay bounded once the cumulative rollout
43		-	// budget is gone. An unlimited parent default must not propagate to
44		-	// children; explicit parent limits are preserved above.
45		-	agent.MaxSteps = 50
46		-	}
47	47		agent.ContextWindow = resolved.ContextWindow
48	48		agent.MaxTokens = resolved.MaxOutput
49	49		agent.Reasoning = resolved.Reasoning

@@ -54,8 +54,8 @@ func newResidentChildRunner(args Args, spec subagents.ResidentChildSpec, journal
54	54		baseline = journal.ConfigureUsage(resolved.ContextWindow, resolved.AuthMethod == "oauth")
55	55		agent.SeedCost(baseline.Usage)
56	56		}
57		-	// The nudge reads the agent's own per-turn usage gauge. No cumulative
58		-	// budget remains: execution stays bounded through the MaxSteps default.
	57	+	// The nudge reads the agent's own per-turn usage gauge. It stays advisory:
	58	+	// the provider context window is the hard ceiling on a child turn.
59	59		configureResidentContextNudge(agent, resolved.ContextWindow)
60	60		rootCacheID := strings.TrimSpace(spec.RootCacheID)
61	61		if rootCacheID == "" {

@@ -78,6 +78,7 @@ func newResidentChildRunner(args Args, spec subagents.ResidentChildSpec, journal
78	78		agent.SetMessages(messages)
79	79		}
80	80		}
	81	+	limit := agent.MaxSteps
81	82		return func(ctx context.Context, prompt string) error {
82	83		if ctx == nil {
83	84		ctx = context.Background()

@@ -86,7 +87,23 @@ func newResidentChildRunner(args Args, spec subagents.ResidentChildSpec, journal
86	87		defer cancel()
87	88		var journalErr error
88	89		var journalMu sync.Mutex
89		-	err := agent.Prompt(turnCtx, prompt, nil, func(event core.AgentEvent) {
	90	+	checkJournal := func() error {
	91	+	journalMu.Lock()
	92	+	defer journalMu.Unlock()
	93	+	if journalErr == nil {
	94	+	return nil
	95	+	}
	96	+	return fmt.Errorf("persist resident child transcript: %w", journalErr)
	97	+	}
	98	+	// usedSteps is the step high-water mark of this accepted turn. The sink
	99	+	// runs on the runner goroutine, so it needs no lock; the mark bounds the
	100	+	// continuation after a compaction recovery so an explicit parent limit
	101	+	// is never replenished.
	102	+	usedSteps := 0
	103	+	sink := func(event core.AgentEvent) {
	104	+	if turnStart, ok := event.(core.EvTurnStart); ok && turnStart.Step > usedSteps {
	105	+	usedSteps = turnStart.Step
	106	+	}
90	107		if err := journal.RecordAgentEvent(event); err != nil {
91	108		journalMu.Lock()
92	109		if journalErr == nil {

@@ -95,44 +112,133 @@ func newResidentChildRunner(args Args, spec subagents.ResidentChildSpec, journal
95	112		}
96	113		journalMu.Unlock()
97	114		}
98		-	})
99		-	journalMu.Lock()
100		-	defer journalMu.Unlock()
101		-	if journalErr != nil {
102		-	return fmt.Errorf("persist resident child transcript: %w", journalErr)
	115	+	}
	116	+	err := agent.Prompt(turnCtx, prompt, nil, sink)
	117	+	if journalFailure := checkJournal(); journalFailure != nil {
	118	+	return journalFailure
	119	+	}
	120	+	if err != nil && turnCtx.Err() == nil && provider.IsContextOverflowError(err) {
	121	+	err = recoverResidentContextOverflow(turnCtx, agent, journal, sink, limit, usedSteps, err)
	122	+	if journalFailure := checkJournal(); journalFailure != nil {
	123	+	return journalFailure
	124	+	}
103	125		}
104	126		return err
105	127		}, nil
106	128		}
107	129
108		-	// residentContextNudgePercent is the single context-usage threshold for the
109		-	// resident reminder. It is a completion aid, not a cost guard: execution
110		-	// stays bounded through the finite MaxSteps default.
111		-	const residentContextNudgePercent = 85
	130	+	// A compaction recovery never rewrites a transcript with nothing worth
	131	+	// summarizing: the child task and the current exchange must survive verbatim.
	132	+	const (
	133	+	residentContextRecoveryMinimumMessages = 4
	134	+	residentContextRecoveryKeepTail = 2
	135	+	)
	136	+
	137	+	// recoverResidentContextOverflow gives one accepted resident turn a second
	138	+	// chance after the provider rejected it for exceeding the model context
	139	+	// window. It compacts the transcript, journals the replacement transcript, and
	140	+	// continues the same turn. Recovery happens at most once per call, and any
	141	+	// compaction or persistence failure restores the exact pre-compaction
	142	+	// transcript before returning the original error.
	143	+	func recoverResidentContextOverflow(ctx context.Context, agent core.Agent, journal subagents.ResidentJournal, sink func(core.AgentEvent), limit, usedSteps int, overflowErr error) error {
	144	+	if agent == nil \|\| journal == nil {
	145	+	return overflowErr
	146	+	}
	147	+	before := agent.Messages()
	148	+	if len(before) < residentContextRecoveryMinimumMessages {
	149	+	// Summarizing a transcript this short would replace the instruction
	150	+	// rather than shrink prior work.
	151	+	return overflowErr
	152	+	}
	153	+	if _, err := agent.CompactWithEvents(ctx, residentContextRecoveryKeepTail, sink); err != nil {
	154	+	agent.SetMessages(before)
	155	+	return errors.Join(overflowErr, fmt.Errorf("compact resident child transcript: %w", err))
	156	+	}
	157	+	if err := journal.RecordCompacted(residentCheckpointMessages(agent.Messages())); err != nil {
	158	+	// Memory must not run ahead of durable history: a later resume would
	159	+	// otherwise replay the pre-compaction transcript.
	160	+	agent.SetMessages(before)
	161	+	return errors.Join(overflowErr, fmt.Errorf("persist resident child compaction: %w", err))
	162	+	}
	163	+	// The summarization request's usage describes the pre-compaction
	164	+	// transcript. It must not arm the next reminder, which would report a
	165	+	// false percentage for the compacted transcript the continuation sends.
	166	+	agent.SeedLastTurnUsage(provider.Usage{})
	167	+	if limit > 0 {
	168	+	remaining := limit - usedSteps
	169	+	if remaining <= 0 {
	170	+	// An explicit parent limit is already spent. The checkpoint stays
	171	+	// durable so an explicit resume starts from the compacted
	172	+	// transcript, but this turn is not continued.
	173	+	return overflowErr
	174	+	}
	175	+	agent.MaxSteps = remaining
	176	+	defer func() { agent.MaxSteps = limit }()
	177	+	}
	178	+	return agent.Continue(ctx, sink)
	179	+	}
	180	+
	181	+	// residentCheckpointMessages builds the durable replacement transcript of a
	182	+	// compaction. Host-authored internal context is dropped: it belongs to the
	183	+	// live request, and a replayed reminder would survive as a stale instruction
	184	+	// that the resumed child would still obey.
	185	+	func residentCheckpointMessages(messages []provider.Message) []provider.Message {
	186	+	checkpoint := make([]provider.Message, 0, len(messages))
	187	+	for _, message := range messages {
	188	+	if core.IsInternalContextMessage(message) {
	189	+	continue
	190	+	}
	191	+	checkpoint = append(checkpoint, message)
	192	+	}
	193	+	return checkpoint
	194	+	}
	195	+
	196	+	// residentContextReminderBands are the context-usage percentages that arm the
	197	+	// resident reminder. Each band fires at most once per child runner, so a child
	198	+	// whose prompt keeps growing toward the window is told to wrap up again
	199	+	// without turning the reminder into a budget.
	200	+	var residentContextReminderBands = [...]int{85, 90, 95}
112	201
113	202		func configureResidentContextNudge(agent *core.Agent, contextMax int) {
114	203		if agent == nil {
115	204		return
116	205		}
117		-	// notified is owned by the runner closure and never reset from Resume
118		-	// or journal paths. The injected developer-context message persists via
119		-	// appendDynamicContext and is not replayed verbatim from the child
120		-	// journal after restart; re-nudging after restart is acceptable.
121		-	var notified bool
	206	+	// deliveredBand is owned by the runner closure and never reset from
	207	+	// Resume or journal paths. The injected developer-context message persists
	208	+	// via appendDynamicContext and is not replayed verbatim from the child
	209	+	// journal after restart; restarting a child rests the ladder.
	210	+	deliveredBand := -1
122	211		agent.BeforeTurnContext = func(_ context.Context, _ int) (bool, string, string) {
123		-	if notified \|\| contextMax <= 0 {
	212	+	if contextMax <= 0 {
124	213		return true, "", ""
125	214		}
126	215		usage := agent.LastTurnUsage()
127	216		used := usage.InputTokens + usage.CacheReadTokens + usage.CacheWriteTokens
128		-	if used100 < contextMaxresidentContextNudgePercent {
	217	+	band := residentContextReminderBand(used, contextMax)
	218	+	if band <= deliveredBand {
129	219		return true, "", ""
130	220		}
131		-	notified = true
	221	+	deliveredBand = band
132	222		return true, "", fmt.Sprintf("Context usage is at %d%% of the model window. Finish the current task, report results, limits, and remaining verifications, and do not start broad new work.", used*100/contextMax)
133	223		}
134	224		}
135	225
	226	+	// residentContextReminderBand returns the highest armed band crossed by the
	227	+	// current prompt size, or -1 while usage stays below the first band. A direct
	228	+	// jump to a higher band delivers one reminder for that band.
	229	+	func residentContextReminderBand(used, contextMax int) int {
	230	+	if contextMax <= 0 {
	231	+	return -1
	232	+	}
	233	+	band := -1
	234	+	for _, candidate := range residentContextReminderBands {
	235	+	if used100 >= contextMaxcandidate {
	236	+	band = candidate
	237	+	}
	238	+	}
	239	+	return band
	240	+	}
	241	+
136	242		// residentChildArgs applies the complete durable child specification without
137	243		// rediscovering a mutable profile file. That preserves the profile inheritance
138	244		// decision accepted at spawn time across explicit resume and restart.


@@ -2,6 +2,7 @@ package agent
2			2
3		import (	3		import (
4		"context"	4		"context"
			5	+	"errors"
5		"fmt"	6		"fmt"
6		"strings"	7		"strings"
7		"sync"	8		"sync"

@@ -9,6 +10,7 @@ import (
9		"github.com/bnema/zut/packages/agent/subagents"	10		"github.com/bnema/zut/packages/agent/subagents"
10		"github.com/bnema/zut/packages/agent/tools"	11		"github.com/bnema/zut/packages/agent/tools"
11		"github.com/bnema/zut/packages/core"	12		"github.com/bnema/zut/packages/core"
			13	+	"github.com/bnema/zut/packages/provider"
12		)	14		)
13			15
14		// newResidentChildRunner is the host-owned construction boundary for one	16		// newResidentChildRunner is the host-owned construction boundary for one

@@ -37,13 +39,11 @@ func newResidentChildRunner(args Args, spec subagents.ResidentChildSpec, journal
37		}	39		}
38		}	40		}
39		agent := core.NewAgent(resolved.NewClient(), resolved.Model, system, registry)	41		agent := core.NewAgent(resolved.NewClient(), resolved.Model, system, registry)
			42	+	// A resident child inherits the parent's resolved step limit, including the
			43	+	// unlimited default. Its ceiling is the provider context window, with one
			44	+	// compaction recovery per accepted turn, plus any explicit parent
			45	+	// --max-steps. There is no child-only step budget.
40		agent.MaxSteps = resolved.MaxSteps	46		agent.MaxSteps = resolved.MaxSteps
41	-	if agent.MaxSteps <= 0 {
42	-	// Resident execution must stay bounded once the cumulative rollout
43	-	// budget is gone. An unlimited parent default must not propagate to
44	-	// children; explicit parent limits are preserved above.
45	-	agent.MaxSteps = 50
46	-	}
47		agent.ContextWindow = resolved.ContextWindow	47		agent.ContextWindow = resolved.ContextWindow
48		agent.MaxTokens = resolved.MaxOutput	48		agent.MaxTokens = resolved.MaxOutput
49		agent.Reasoning = resolved.Reasoning	49		agent.Reasoning = resolved.Reasoning

@@ -54,8 +54,8 @@ func newResidentChildRunner(args Args, spec subagents.ResidentChildSpec, journal
54		baseline = journal.ConfigureUsage(resolved.ContextWindow, resolved.AuthMethod == "oauth")	54		baseline = journal.ConfigureUsage(resolved.ContextWindow, resolved.AuthMethod == "oauth")
55		agent.SeedCost(baseline.Usage)	55		agent.SeedCost(baseline.Usage)
56		}	56		}
57	-	// The nudge reads the agent's own per-turn usage gauge. No cumulative	57	+	// The nudge reads the agent's own per-turn usage gauge. It stays advisory:
58	-	// budget remains: execution stays bounded through the MaxSteps default.	58	+	// the provider context window is the hard ceiling on a child turn.
59		configureResidentContextNudge(agent, resolved.ContextWindow)	59		configureResidentContextNudge(agent, resolved.ContextWindow)
60		rootCacheID := strings.TrimSpace(spec.RootCacheID)	60		rootCacheID := strings.TrimSpace(spec.RootCacheID)
61		if rootCacheID == "" {	61		if rootCacheID == "" {

@@ -78,6 +78,7 @@ func newResidentChildRunner(args Args, spec subagents.ResidentChildSpec, journal
78		agent.SetMessages(messages)	78		agent.SetMessages(messages)
79		}	79		}
80		}	80		}
			81	+	limit := agent.MaxSteps
81		return func(ctx context.Context, prompt string) error {	82		return func(ctx context.Context, prompt string) error {
82		if ctx == nil {	83		if ctx == nil {
83		ctx = context.Background()	84		ctx = context.Background()

@@ -86,7 +87,23 @@ func newResidentChildRunner(args Args, spec subagents.ResidentChildSpec, journal
86		defer cancel()	87		defer cancel()
87		var journalErr error	88		var journalErr error
88		var journalMu sync.Mutex	89		var journalMu sync.Mutex
89	-	err := agent.Prompt(turnCtx, prompt, nil, func(event core.AgentEvent) {	90	+	checkJournal := func() error {
			91	+	journalMu.Lock()
			92	+	defer journalMu.Unlock()
			93	+	if journalErr == nil {
			94	+	return nil
			95	+	}
			96	+	return fmt.Errorf("persist resident child transcript: %w", journalErr)
			97	+	}
			98	+	// usedSteps is the step high-water mark of this accepted turn. The sink
			99	+	// runs on the runner goroutine, so it needs no lock; the mark bounds the
			100	+	// continuation after a compaction recovery so an explicit parent limit
			101	+	// is never replenished.
			102	+	usedSteps := 0
			103	+	sink := func(event core.AgentEvent) {
			104	+	if turnStart, ok := event.(core.EvTurnStart); ok && turnStart.Step > usedSteps {
			105	+	usedSteps = turnStart.Step
			106	+	}
90		if err := journal.RecordAgentEvent(event); err != nil {	107		if err := journal.RecordAgentEvent(event); err != nil {
91		journalMu.Lock()	108		journalMu.Lock()
92		if journalErr == nil {	109		if journalErr == nil {

@@ -95,44 +112,133 @@ func newResidentChildRunner(args Args, spec subagents.ResidentChildSpec, journal
95		}	112		}
96		journalMu.Unlock()	113		journalMu.Unlock()
97		}	114		}
98	-	})	115	+	}
99	-	journalMu.Lock()	116	+	err := agent.Prompt(turnCtx, prompt, nil, sink)
100	-	defer journalMu.Unlock()	117	+	if journalFailure := checkJournal(); journalFailure != nil {
101	-	if journalErr != nil {	118	+	return journalFailure
102	-	return fmt.Errorf("persist resident child transcript: %w", journalErr)	119	+	}
			120	+	if err != nil && turnCtx.Err() == nil && provider.IsContextOverflowError(err) {
			121	+	err = recoverResidentContextOverflow(turnCtx, agent, journal, sink, limit, usedSteps, err)
			122	+	if journalFailure := checkJournal(); journalFailure != nil {
			123	+	return journalFailure
			124	+	}
103		}	125		}
104		return err	126		return err
105		}, nil	127		}, nil
106		}	128		}
107			129
108	-	// residentContextNudgePercent is the single context-usage threshold for the	130	+	// A compaction recovery never rewrites a transcript with nothing worth
109	-	// resident reminder. It is a completion aid, not a cost guard: execution	131	+	// summarizing: the child task and the current exchange must survive verbatim.
110	-	// stays bounded through the finite MaxSteps default.	132	+	const (
111	-	const residentContextNudgePercent = 85	133	+	residentContextRecoveryMinimumMessages = 4
			134	+	residentContextRecoveryKeepTail = 2
			135	+	)
			136	+
			137	+	// recoverResidentContextOverflow gives one accepted resident turn a second
			138	+	// chance after the provider rejected it for exceeding the model context
			139	+	// window. It compacts the transcript, journals the replacement transcript, and
			140	+	// continues the same turn. Recovery happens at most once per call, and any
			141	+	// compaction or persistence failure restores the exact pre-compaction
			142	+	// transcript before returning the original error.
			143	+	func recoverResidentContextOverflow(ctx context.Context, agent core.Agent, journal subagents.ResidentJournal, sink func(core.AgentEvent), limit, usedSteps int, overflowErr error) error {
			144	+	if agent == nil \|\| journal == nil {
			145	+	return overflowErr
			146	+	}
			147	+	before := agent.Messages()
			148	+	if len(before) < residentContextRecoveryMinimumMessages {
			149	+	// Summarizing a transcript this short would replace the instruction
			150	+	// rather than shrink prior work.
			151	+	return overflowErr
			152	+	}
			153	+	if _, err := agent.CompactWithEvents(ctx, residentContextRecoveryKeepTail, sink); err != nil {
			154	+	agent.SetMessages(before)
			155	+	return errors.Join(overflowErr, fmt.Errorf("compact resident child transcript: %w", err))
			156	+	}
			157	+	if err := journal.RecordCompacted(residentCheckpointMessages(agent.Messages())); err != nil {
			158	+	// Memory must not run ahead of durable history: a later resume would
			159	+	// otherwise replay the pre-compaction transcript.
			160	+	agent.SetMessages(before)
			161	+	return errors.Join(overflowErr, fmt.Errorf("persist resident child compaction: %w", err))
			162	+	}
			163	+	// The summarization request's usage describes the pre-compaction
			164	+	// transcript. It must not arm the next reminder, which would report a
			165	+	// false percentage for the compacted transcript the continuation sends.
			166	+	agent.SeedLastTurnUsage(provider.Usage{})
			167	+	if limit > 0 {
			168	+	remaining := limit - usedSteps
			169	+	if remaining <= 0 {
			170	+	// An explicit parent limit is already spent. The checkpoint stays
			171	+	// durable so an explicit resume starts from the compacted
			172	+	// transcript, but this turn is not continued.
			173	+	return overflowErr
			174	+	}
			175	+	agent.MaxSteps = remaining
			176	+	defer func() { agent.MaxSteps = limit }()
			177	+	}
			178	+	return agent.Continue(ctx, sink)
			179	+	}
			180	+
			181	+	// residentCheckpointMessages builds the durable replacement transcript of a
			182	+	// compaction. Host-authored internal context is dropped: it belongs to the
			183	+	// live request, and a replayed reminder would survive as a stale instruction
			184	+	// that the resumed child would still obey.
			185	+	func residentCheckpointMessages(messages []provider.Message) []provider.Message {
			186	+	checkpoint := make([]provider.Message, 0, len(messages))
			187	+	for _, message := range messages {
			188	+	if core.IsInternalContextMessage(message) {
			189	+	continue
			190	+	}
			191	+	checkpoint = append(checkpoint, message)
			192	+	}
			193	+	return checkpoint
			194	+	}
			195	+
			196	+	// residentContextReminderBands are the context-usage percentages that arm the
			197	+	// resident reminder. Each band fires at most once per child runner, so a child
			198	+	// whose prompt keeps growing toward the window is told to wrap up again
			199	+	// without turning the reminder into a budget.
			200	+	var residentContextReminderBands = [...]int{85, 90, 95}
112			201
113		func configureResidentContextNudge(agent *core.Agent, contextMax int) {	202		func configureResidentContextNudge(agent *core.Agent, contextMax int) {
114		if agent == nil {	203		if agent == nil {
115		return	204		return
116		}	205		}
117	-	// notified is owned by the runner closure and never reset from Resume	206	+	// deliveredBand is owned by the runner closure and never reset from
118	-	// or journal paths. The injected developer-context message persists via	207	+	// Resume or journal paths. The injected developer-context message persists
119	-	// appendDynamicContext and is not replayed verbatim from the child	208	+	// via appendDynamicContext and is not replayed verbatim from the child
120	-	// journal after restart; re-nudging after restart is acceptable.	209	+	// journal after restart; restarting a child rests the ladder.
121	-	var notified bool	210	+	deliveredBand := -1
122		agent.BeforeTurnContext = func(_ context.Context, _ int) (bool, string, string) {	211		agent.BeforeTurnContext = func(_ context.Context, _ int) (bool, string, string) {
123	-	if notified \|\| contextMax <= 0 {	212	+	if contextMax <= 0 {
124		return true, "", ""	213		return true, "", ""
125		}	214		}
126		usage := agent.LastTurnUsage()	215		usage := agent.LastTurnUsage()
127		used := usage.InputTokens + usage.CacheReadTokens + usage.CacheWriteTokens	216		used := usage.InputTokens + usage.CacheReadTokens + usage.CacheWriteTokens
128	-	if used100 < contextMaxresidentContextNudgePercent {	217	+	band := residentContextReminderBand(used, contextMax)
			218	+	if band <= deliveredBand {
129		return true, "", ""	219		return true, "", ""
130		}	220		}
131	-	notified = true	221	+	deliveredBand = band
132		return true, "", fmt.Sprintf("Context usage is at %d%% of the model window. Finish the current task, report results, limits, and remaining verifications, and do not start broad new work.", used*100/contextMax)	222		return true, "", fmt.Sprintf("Context usage is at %d%% of the model window. Finish the current task, report results, limits, and remaining verifications, and do not start broad new work.", used*100/contextMax)
133		}	223		}
134		}	224		}
135			225
			226	+	// residentContextReminderBand returns the highest armed band crossed by the
			227	+	// current prompt size, or -1 while usage stays below the first band. A direct
			228	+	// jump to a higher band delivers one reminder for that band.
			229	+	func residentContextReminderBand(used, contextMax int) int {
			230	+	if contextMax <= 0 {
			231	+	return -1
			232	+	}
			233	+	band := -1
			234	+	for _, candidate := range residentContextReminderBands {
			235	+	if used100 >= contextMaxcandidate {
			236	+	band = candidate
			237	+	}
			238	+	}
			239	+	return band
			240	+	}
			241	+
136		// residentChildArgs applies the complete durable child specification without	242		// residentChildArgs applies the complete durable child specification without
137		// rediscovering a mutable profile file. That preserves the profile inheritance	243		// rediscovering a mutable profile file. That preserves the profile inheritance
138		// decision accepted at spawn time across explicit resume and restart.	244		// decision accepted at spawn time across explicit resume and restart.

M packages/agent/subagent_factory_test.go +596 -79


@@ -7,6 +7,9 @@ import (
7	7		"fmt"
8	8		"net/http"
9	9		"net/http/httptest"
	10	+	"os"
	11	+	"path/filepath"
	12	+	"reflect"
10	13		"strings"
11	14		"sync"
12	15		"sync/atomic"

@@ -18,65 +21,87 @@ import (
18	21		"github.com/bnema/zut/packages/provider"
19	22		)
20	23
21		-	func TestResidentChildRunnerDefaultsToFiniteMaxSteps(t *testing.T) {
22		-	for _, tc := range []struct {
23		-	name string
24		-	maxSteps int
25		-	want int
26		-	}{
27		-	{name: "unlimited parent defaults to bounded child", maxSteps: 0, want: 50},
28		-	{name: "explicit parent limit preserved", maxSteps: 30, want: 30},
29		-	} {
30		-	t.Run(tc.name, func(t *testing.T) {
31		-	var requests atomic.Int32
32		-	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
33		-	requests.Add(1)
34		-	w.Header().Set("content-type", "text/event-stream")
35		-	writeOpenAIChunk(t, w, map[string]any{
36		-	"choices": []any{map[string]any{
37		-	"index": 0,
38		-	"delta": map[string]any{"tool_calls": []any{map[string]any{
39		-	"index": 0, "id": "call-loop", "type": "function",
40		-	"function": map[string]string{"name": "read", "arguments": `{"path":"x"}`},
41		-	}}},
42		-	"finish_reason": "tool_calls",
43		-	}},
44		-	"usage": map[string]int{"prompt_tokens": 10, "completion_tokens": 1},
45		-	})
46		-	}))
47		-	defer server.Close()
48		-
49		-	spec := subagents.ResidentChildSpec{
50		-	ID: "maxsteps-child", SessionID: "maxsteps-session", Provider: "openai", Model: "gpt-4o",
51		-	Tools: []string{"read"},
52		-	}
53		-	journal, err := subagents.OpenResidentJournal(t.TempDir(), spec.ID)
54		-	if err != nil {
55		-	t.Fatal(err)
56		-	}
57		-	t.Cleanup(func() { _ = journal.Close() })
58		-	if err := journal.Accept(spec, "review"); err != nil {
59		-	t.Fatal(err)
60		-	}
61		-	runner, err := newResidentChildRunner(Args{
62		-	Provider: "openai", Model: "gpt-4o", APIKey: "synthetic", BaseURL: server.URL,
63		-	MaxSteps: tc.maxSteps, CWD: t.TempDir(), NoContextFiles: true, NoSkill: true,
64		-	}, spec, journal)
65		-	if err != nil {
66		-	t.Fatal(err)
67		-	}
68		-	err = runner(t.Context(), "review")
69		-	if !errors.Is(err, core.ErrMaxSteps) {
70		-	t.Fatalf("runner error = %v, want step-limit stop", err)
71		-	}
72		-	if !strings.Contains(err.Error(), fmt.Sprintf("%d", tc.want)) {
73		-	t.Fatalf("runner error = %v, want limit %d", err, tc.want)
74		-	}
75		-	if got := requests.Load(); got != int32(tc.want) {
76		-	t.Fatalf("provider requests = %d, want %d", got, tc.want)
77		-	}
	24	+	func TestResidentChildRunnerWithoutStepLimitRunsPastFormerCap(t *testing.T) {
	25	+	server, requests := residentToolLoopServer(t, 51)
	26	+	runner, _, _ := newResidentChildTestRunner(t, "unlimited-child", Args{
	27	+	Provider: "openai", Model: "gpt-4o", APIKey: "synthetic", BaseURL: server.URL,
	28	+	CWD: t.TempDir(), NoContextFiles: true, NoSkill: true,
	29	+	})
	30	+	if err := runner(t.Context(), "review"); err != nil {
	31	+	t.Fatalf("runner error = %v, want success past the former 50-step cap", err)
	32	+	}
	33	+	if got := requests.Load(); got != 52 {
	34	+	t.Fatalf("provider requests = %d, want 51 tool turns and one answer", got)
	35	+	}
	36	+	}
	37	+
	38	+	func TestResidentChildRunnerEnforcesExplicitParentStepLimit(t *testing.T) {
	39	+	server, requests := residentToolLoopServer(t, 200)
	40	+	runner, _, _ := newResidentChildTestRunner(t, "limited-child", Args{
	41	+	Provider: "openai", Model: "gpt-4o", APIKey: "synthetic", BaseURL: server.URL,
	42	+	MaxSteps: 3, CWD: t.TempDir(), NoContextFiles: true, NoSkill: true,
	43	+	})
	44	+	err := runner(t.Context(), "review")
	45	+	if !errors.Is(err, core.ErrMaxSteps) \|\| !strings.Contains(err.Error(), "3") {
	46	+	t.Fatalf("runner error = %v, want step limit 3", err)
	47	+	}
	48	+	if got := requests.Load(); got != 3 {
	49	+	t.Fatalf("provider requests = %d, want 3", got)
	50	+	}
	51	+	}
	52	+
	53	+	// residentToolLoopServer answers the first toolCalls requests with a distinct
	54	+	// read call and then returns a substantive final answer.
	55	+	func residentToolLoopServer(t testing.T, toolCalls int) (httptest.Server, *atomic.Int32) {
	56	+	t.Helper()
	57	+	var requests atomic.Int32
	58	+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
	59	+	number := requests.Add(1)
	60	+	w.Header().Set("content-type", "text/event-stream")
	61	+	if int(number) <= toolCalls {
	62	+	writeOpenAIChunk(t, w, map[string]any{
	63	+	"choices": []any{map[string]any{
	64	+	"index": 0,
	65	+	"delta": map[string]any{"tool_calls": []any{map[string]any{
	66	+	"index": 0, "id": fmt.Sprintf("call-%d", number), "type": "function",
	67	+	"function": map[string]string{"name": "read", "arguments": `{"path":"x"}`},
	68	+	}}},
	69	+	"finish_reason": "tool_calls",
	70	+	}},
	71	+	"usage": map[string]int{"prompt_tokens": 10, "completion_tokens": 1},
	72	+	})
	73	+	return
	74	+	}
	75	+	writeOpenAIChunk(t, w, map[string]any{
	76	+	"choices": []any{map[string]any{"index": 0, "delta": map[string]string{"content": "final result"}, "finish_reason": "stop"}},
	77	+	"usage": map[string]int{"prompt_tokens": 10, "completion_tokens": 1},
78	78		})
	79	+	}))
	80	+	t.Cleanup(server.Close)
	81	+	return server, &requests
	82	+	}
	83	+
	84	+	// newResidentChildTestRunner builds one journaled resident child runner with
	85	+	// its own accepted spec.
	86	+	func newResidentChildTestRunner(t testing.T, childID string, args Args) (subagents.ResidentTurnRunner, subagents.ResidentJournal, subagents.ResidentChildSpec) {
	87	+	t.Helper()
	88	+	spec := subagents.ResidentChildSpec{
	89	+	ID: childID, SessionID: childID + "-session", Provider: "openai", Model: "gpt-4o",
	90	+	Tools: []string{"read"},
	91	+	}
	92	+	journal, err := subagents.OpenResidentJournal(t.TempDir(), spec.ID)
	93	+	if err != nil {
	94	+	t.Fatal(err)
79	95		}
	96	+	t.Cleanup(func() { _ = journal.Close() })
	97	+	if err := journal.Accept(spec, "review"); err != nil {
	98	+	t.Fatal(err)
	99	+	}
	100	+	runner, err := newResidentChildRunner(args, spec, journal)
	101	+	if err != nil {
	102	+	t.Fatal(err)
	103	+	}
	104	+	return runner, journal, spec
80	105		}
81	106
82	107		func TestResidentChildRunnerInjectsContextReminderIntoNextRequest(t *testing.T) {

@@ -202,7 +227,7 @@ func TestResidentChildRegistryUsesExactToolListAndForbidsDelegation(t *testing.T
202	227		}
203	228		}
204	229
205		-	func TestConfigureResidentContextNudgeRemindsOnceAtThreshold(t *testing.T) {
	230	+	func TestConfigureResidentContextNudgeRemindsOncePerBand(t *testing.T) {
206	231		agent := core.NewAgent(nil, "model", "system", core.Registry{"read": nil})
207	232		configureResidentContextNudge(agent, 1_000)
208	233

@@ -222,16 +247,52 @@ func TestConfigureResidentContextNudgeRemindsOnceAtThreshold(t *testing.T) {
222	247		return contextText
223	248		}
224	249
225		-	// Below the threshold: silent, including cache-heavy usage.
	250	+	// Below the first band: silent, including cache-heavy usage.
226	251		check(provider.Usage{InputTokens: 849}, false)
227	252		check(provider.Usage{InputTokens: 400, CacheReadTokens: 400, CacheWriteTokens: 49}, false)
228		-	// First crossing: exactly one reminder carrying the computed percent.
	253	+	// Each band fires once, carrying the computed percent of the crossing turn.
229	254		first := check(provider.Usage{InputTokens: 850}, true)
230	255		if !strings.Contains(first, "85%") {
231	256		t.Fatalf("reminder percent = %q, want 85%%", first)
232	257		}
233		-	// Later turns stay silent even as usage grows.
234		-	check(provider.Usage{InputTokens: 970}, false)
	258	+	check(provider.Usage{InputTokens: 860}, false)
	259	+	second := check(provider.Usage{InputTokens: 900}, true)
	260	+	if !strings.Contains(second, "90%") {
	261	+	t.Fatalf("reminder percent = %q, want 90%%", second)
	262	+	}
	263	+	check(provider.Usage{InputTokens: 920}, false)
	264	+	third := check(provider.Usage{InputTokens: 950}, true)
	265	+	if !strings.Contains(third, "95%") {
	266	+	t.Fatalf("reminder percent = %q, want 95%%", third)
	267	+	}
	268	+	// Past the last band the ladder is spent.
	269	+	check(provider.Usage{InputTokens: 999}, false)
	270	+	}
	271	+
	272	+	func TestConfigureResidentContextNudgeIgnoresShrinkingUsage(t *testing.T) {
	273	+	agent := core.NewAgent(nil, "model", "system", core.Registry{"read": nil})
	274	+	configureResidentContextNudge(agent, 1_000)
	275	+
	276	+	check := func(usage provider.Usage, wantReminder bool) string {
	277	+	t.Helper()
	278	+	agent.SeedLastTurnUsage(usage)
	279	+	allowed, reason, contextText := agent.BeforeTurnContext(t.Context(), 1)
	280	+	if !allowed \|\| reason != "" {
	281	+	t.Fatalf("nudge blocked turn: allowed=%t reason=%q", allowed, reason)
	282	+	}
	283	+	if wantReminder != (contextText != "") {
	284	+	t.Fatalf("reminder for usage %#v = %q, want reminder=%t", usage, contextText, wantReminder)
	285	+	}
	286	+	return contextText
	287	+	}
	288	+
	289	+	// A shrinking prompt never delivers a band it did not cross.
	290	+	check(provider.Usage{InputTokens: 700}, false)
	291	+	check(provider.Usage{InputTokens: 500}, false)
	292	+	jumped := check(provider.Usage{InputTokens: 910}, true)
	293	+	if !strings.Contains(jumped, "91%") {
	294	+	t.Fatalf("reminder percent = %q, want 91%%", jumped)
	295	+	}
235	296		}
236	297
237	298		func TestConfigureResidentContextNudgeHandlesDirectJumpAndUnknownWindow(t *testing.T) {

@@ -400,25 +461,10 @@ func TestResidentChildRunnerResumesWithRetainedHistory(t *testing.T) {
400	461
401	462		func newResidentTestRunner(t testing.T, baseURL string) (subagents.ResidentTurnRunner, subagents.ResidentJournal) {
402	463		t.Helper()
403		-	spec := subagents.ResidentChildSpec{
404		-	ID: "resident-child", SessionID: "resident-session", Provider: "openai", Model: "gpt-4o",
405		-	Tools: []string{"read"},
406		-	}
407		-	journal, err := subagents.OpenResidentJournal(t.TempDir(), spec.ID)
408		-	if err != nil {
409		-	t.Fatal(err)
410		-	}
411		-	t.Cleanup(func() { _ = journal.Close() })
412		-	if err := journal.Accept(spec, "review"); err != nil {
413		-	t.Fatal(err)
414		-	}
415		-	runner, err := newResidentChildRunner(Args{
	464	+	runner, journal, _ := newResidentChildTestRunner(t, "resident-child", Args{
416	465		Provider: "openai", Model: "gpt-4o", APIKey: "synthetic", BaseURL: baseURL,
417	466		CWD: t.TempDir(), NoContextFiles: true, NoSkill: true,
418		-	}, spec, journal)
419		-	if err != nil {
420		-	t.Fatal(err)
421		-	}
	467	+	})
422	468		return runner, journal
423	469		}
424	470

@@ -484,3 +530,474 @@ func TestResidentChildArgsDoNotForwardCLIKeyAcrossProviders(t *testing.T) {
484	530		t.Fatalf("same-provider child API key = %q, want parent key", next.APIKey)
485	531		}
486	532		}
	533	+
	534	+	// residentOverflowScript scripts the provider responses of a child that hits
	535	+	// the context window: leading tool-call turns, one overflow, one compaction
	536	+	// summary, then continued tool turns or the final text.
	537	+	type residentOverflowScript struct {
	538	+	toolCallRequests int
	539	+	continuationToolCalls int
	540	+	secondOverflow bool
	541	+	finalText string
	542	+	onOverflow func()
	543	+	// toolUsageTokens is the prompt size reported by tool-call responses; zero
	544	+	// reports a small prompt.
	545	+	toolUsageTokens int
	546	+	// summaryUsageTokens, when positive, is the prompt size reported by the
	547	+	// compaction summary response.
	548	+	summaryUsageTokens int
	549	+	// continuationUsageTokens is the prompt size reported by tool turns after a
	550	+	// successful compaction; zero reuses toolUsageTokens.
	551	+	continuationUsageTokens int
	552	+
	553	+	requests atomic.Int32
	554	+	mu sync.Mutex
	555	+	bodies []string
	556	+	}
	557	+
	558	+	func (s residentOverflowScript) start(t testing.T) *httptest.Server {
	559	+	t.Helper()
	560	+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
	561	+	number := int(s.requests.Add(1))
	562	+	var body map[string]any
	563	+	if err := json.NewDecoder(r.Body).Decode(&body); err != nil {
	564	+	t.Errorf("decode request: %v", err)
	565	+	}
	566	+	encoded, _ := json.Marshal(body["messages"])
	567	+	s.mu.Lock()
	568	+	s.bodies = append(s.bodies, string(encoded))
	569	+	s.mu.Unlock()
	570	+
	571	+	overflowAt := s.toolCallRequests + 1
	572	+	switch {
	573	+	case number <= s.toolCallRequests:
	574	+	w.Header().Set("content-type", "text/event-stream")
	575	+	writeOpenAIChunk(t, w, residentToolCallChunk(number, s.toolUsageTokens))
	576	+	case number == overflowAt:
	577	+	if s.onOverflow != nil {
	578	+	s.onOverflow()
	579	+	}
	580	+	w.WriteHeader(http.StatusRequestEntityTooLarge)
	581	+	_, _ = w.Write([]byte(`{"error":{"message":"input exceeds the context window","code":"context_length_exceeded"}}`))
	582	+	case number == overflowAt+1:
	583	+	w.Header().Set("content-type", "text/event-stream")
	584	+	writeOpenAIChunk(t, w, residentSummaryChunk(s.summaryUsageTokens))
	585	+	case s.secondOverflow:
	586	+	w.WriteHeader(http.StatusRequestEntityTooLarge)
	587	+	_, _ = w.Write([]byte(`{"error":{"message":"input exceeds the context window","code":"context_length_exceeded"}}`))
	588	+	case number < overflowAt+2+s.continuationToolCalls:
	589	+	w.Header().Set("content-type", "text/event-stream")
	590	+	continuationUsage := s.continuationUsageTokens
	591	+	if continuationUsage <= 0 {
	592	+	continuationUsage = s.toolUsageTokens
	593	+	}
	594	+	writeOpenAIChunk(t, w, residentToolCallChunk(number, continuationUsage))
	595	+	default:
	596	+	w.Header().Set("content-type", "text/event-stream")
	597	+	writeOpenAIChunk(t, w, residentFinalChunk(s.finalText))
	598	+	}
	599	+	}))
	600	+	t.Cleanup(server.Close)
	601	+	return server
	602	+	}
	603	+
	604	+	func residentToolCallChunk(number, usageTokens int) map[string]any {
	605	+	if usageTokens <= 0 {
	606	+	usageTokens = 10
	607	+	}
	608	+	return map[string]any{
	609	+	"choices": []any{map[string]any{
	610	+	"index": 0,
	611	+	"delta": map[string]any{"tool_calls": []any{map[string]any{
	612	+	"index": 0, "id": fmt.Sprintf("call-%d", number), "type": "function",
	613	+	"function": map[string]string{"name": "read", "arguments": `{"path":"x"}`},
	614	+	}}},
	615	+	"finish_reason": "tool_calls",
	616	+	}},
	617	+	"usage": map[string]int{"prompt_tokens": usageTokens, "completion_tokens": 1},
	618	+	}
	619	+	}
	620	+
	621	+	func residentSummaryChunk(usageTokens int) map[string]any {
	622	+	chunk := map[string]any{
	623	+	"choices": []any{map[string]any{"index": 0, "delta": map[string]string{"content": "compacted digest"}, "finish_reason": "stop"}},
	624	+	}
	625	+	if usageTokens > 0 {
	626	+	chunk["usage"] = map[string]int{"prompt_tokens": usageTokens, "completion_tokens": 2}
	627	+	}
	628	+	return chunk
	629	+	}
	630	+
	631	+	func residentFinalChunk(text string) map[string]any {
	632	+	return map[string]any{
	633	+	"choices": []any{map[string]any{"index": 0, "delta": map[string]string{"content": text}, "finish_reason": "stop"}},
	634	+	"usage": map[string]int{"prompt_tokens": 12, "completion_tokens": 5},
	635	+	}
	636	+	}
	637	+
	638	+	func (s *residentOverflowScript) capturedBodies() []string {
	639	+	s.mu.Lock()
	640	+	defer s.mu.Unlock()
	641	+	return append([]string(nil), s.bodies...)
	642	+	}
	643	+
	644	+	func residentArguments(t *testing.T, baseURL string) Args {
	645	+	t.Helper()
	646	+	return Args{
	647	+	Provider: "openai", Model: "gpt-4o", APIKey: "synthetic", BaseURL: baseURL,
	648	+	CWD: t.TempDir(), NoContextFiles: true, NoSkill: true,
	649	+	}
	650	+	}
	651	+
	652	+	func countResidentCheckpoints(t testing.T, journal subagents.ResidentJournal) int {
	653	+	t.Helper()
	654	+	records, err := subagents.ReadResidentJournal(filepath.Join(journal.Dir(), "transcript.jsonl"))
	655	+	if err != nil {
	656	+	t.Fatal(err)
	657	+	}
	658	+	checkpoints := 0
	659	+	for _, record := range records {
	660	+	if record.Type == "child.compacted" {
	661	+	checkpoints++
	662	+	}
	663	+	}
	664	+	return checkpoints
	665	+	}
	666	+
	667	+	func TestResidentChildRunnerRecoversFromContextOverflowOnce(t *testing.T) {
	668	+	script := &residentOverflowScript{toolCallRequests: 2, finalText: "overflow recovered result"}
	669	+	server := script.start(t)
	670	+	runner, journal, spec := newResidentChildTestRunner(t, "overflow-child", residentArguments(t, server.URL))
	671	+	if err := journal.RecordTurnStarted(spec, "turn-1"); err != nil {
	672	+	t.Fatal(err)
	673	+	}
	674	+	runErr := runner(t.Context(), "review")
	675	+	if err := journal.RecordTurnFinished(spec, "turn-1", runErr); err != nil {
	676	+	t.Fatal(err)
	677	+	}
	678	+	if runErr != nil {
	679	+	t.Fatalf("runner error = %v, want one recovered turn", runErr)
	680	+	}
	681	+	if got := script.requests.Load(); got != 5 {
	682	+	t.Fatalf("provider requests = %d, want 3 prompt, 1 compaction, and 1 continuation", got)
	683	+	}
	684	+	result, err := journal.Result()
	685	+	if err != nil {
	686	+	t.Fatal(err)
	687	+	}
	688	+	if result.State != subagents.ResidentIdle \|\| result.Summary != "overflow recovered result" {
	689	+	t.Fatalf("result = %#v, want the recovered final text", result)
	690	+	}
	691	+	records, err := subagents.ReadResidentJournal(filepath.Join(journal.Dir(), "transcript.jsonl"))
	692	+	if err != nil {
	693	+	t.Fatal(err)
	694	+	}
	695	+	checkpoints := 0
	696	+	for _, record := range records {
	697	+	if record.Type != "child.compacted" {
	698	+	continue
	699	+	}
	700	+	checkpoints++
	701	+	if record.TurnID != "turn-1" {
	702	+	t.Fatalf("checkpoint turn ID = %q, want the active accepted turn", record.TurnID)
	703	+	}
	704	+	}
	705	+	if checkpoints != 1 {
	706	+	t.Fatalf("compaction checkpoints = %d, want 1", checkpoints)
	707	+	}
	708	+	// Resume replay starts from the checkpoint: the digest replaces the
	709	+	// pre-overflow tool exchange and the continuation request carried it.
	710	+	messages, err := subagents.ReadResidentTranscriptMessages(journal.Dir())
	711	+	if err != nil {
	712	+	t.Fatal(err)
	713	+	}
	714	+	if len(messages) != 3 \|\| !strings.HasPrefix(residentMessageText(messages[0]), "## Context Summary (compacted)") {
	715	+	t.Fatalf("resumed transcript = %#v, want the checkpoint plus the kept tail", messages)
	716	+	}
	717	+	bodies := script.capturedBodies()
	718	+	if len(bodies) != 5 \|\| !strings.Contains(bodies[4], "compacted digest") {
	719	+	t.Fatalf("continuation request = %q, want the compacted transcript", bodies[len(bodies)-1])
	720	+	}
	721	+	if strings.Contains(bodies[4], "call-1") {
	722	+	t.Fatalf("continuation request kept the dropped tool exchange: %q", bodies[4])
	723	+	}
	724	+	}
	725	+
	726	+	func TestResidentChildRunnerDoesNotRecoverTwicePerTurn(t *testing.T) {
	727	+	script := &residentOverflowScript{toolCallRequests: 2, secondOverflow: true, finalText: "unreached"}
	728	+	server := script.start(t)
	729	+	runner, journal, _ := newResidentChildTestRunner(t, "double-overflow", residentArguments(t, server.URL))
	730	+	err := runner(t.Context(), "review")
	731	+	if err == nil \|\| !provider.IsContextOverflowError(err) {
	732	+	t.Fatalf("runner error = %v, want the second overflow to stay terminal", err)
	733	+	}
	734	+	if got := script.requests.Load(); got != 5 {
	735	+	t.Fatalf("provider requests = %d, want no second compaction", got)
	736	+	}
	737	+	if checkpoints := countResidentCheckpoints(t, journal); checkpoints != 1 {
	738	+	t.Fatalf("compaction checkpoints = %d, want 1", checkpoints)
	739	+	}
	740	+	}
	741	+
	742	+	func TestResidentChildRunnerDoesNotReplenishExplicitStepLimit(t *testing.T) {
	743	+	script := &residentOverflowScript{toolCallRequests: 3, finalText: "unreached"}
	744	+	server := script.start(t)
	745	+	args := residentArguments(t, server.URL)
	746	+	args.MaxSteps = 4
	747	+	runner, journal, _ := newResidentChildTestRunner(t, "limited-overflow", args)
	748	+	err := runner(t.Context(), "review")
	749	+	if err == nil \|\| !provider.IsContextOverflowError(err) {
	750	+	t.Fatalf("runner error = %v, want the overflow", err)
	751	+	}
	752	+	if got := script.requests.Load(); got != 5 {
	753	+	t.Fatalf("provider requests = %d, want the compaction attempt only", got)
	754	+	}
	755	+	// The checkpoint is still journaled so an explicit resume starts small,
	756	+	// but the spent allowance forbids continuing this turn.
	757	+	if checkpoints := countResidentCheckpoints(t, journal); checkpoints != 1 {
	758	+	t.Fatalf("compaction checkpoints = %d, want 1", checkpoints)
	759	+	}
	760	+	for _, body := range script.capturedBodies() {
	761	+	if strings.Contains(body, "compacted digest") {
	762	+	t.Fatalf("recovery continued after the step allowance was spent: %q", body)
	763	+	}
	764	+	}
	765	+	}
	766	+
	767	+	func TestResidentChildRunnerFailsTurnWhenCheckpointCannotPersist(t *testing.T) {
	768	+	root := t.TempDir()
	769	+	spec := subagents.ResidentChildSpec{
	770	+	ID: "persist-failure", SessionID: "persist-failure-session", Provider: "openai", Model: "gpt-4o",
	771	+	Tools: []string{"read"},
	772	+	}
	773	+	journal, err := subagents.OpenResidentJournal(root, spec.ID)
	774	+	if err != nil {
	775	+	t.Fatal(err)
	776	+	}
	777	+	t.Cleanup(func() { _ = journal.Close() })
	778	+	if err := journal.Accept(spec, "review"); err != nil {
	779	+	t.Fatal(err)
	780	+	}
	781	+	// The journal is closed at the exact overflow boundary, so the recovery
	782	+	// cannot persist its checkpoint.
	783	+	script := &residentOverflowScript{
	784	+	toolCallRequests: 2,
	785	+	finalText: "unreached",
	786	+	onOverflow: func() { _ = journal.Close() },
	787	+	}
	788	+	server := script.start(t)
	789	+	runner, err := newResidentChildRunner(residentArguments(t, server.URL), spec, journal)
	790	+	if err != nil {
	791	+	t.Fatal(err)
	792	+	}
	793	+	if err := runner(t.Context(), "review"); err == nil {
	794	+	t.Fatal("runner recovered without persisting its checkpoint")
	795	+	}
	796	+	if got := script.requests.Load(); got != 4 {
	797	+	t.Fatalf("provider requests = %d, want the compaction attempt only", got)
	798	+	}
	799	+	data, readErr := os.ReadFile(filepath.Join(journal.Dir(), "transcript.jsonl"))
	800	+	if readErr != nil {
	801	+	t.Fatal(readErr)
	802	+	}
	803	+	if strings.Contains(string(data), `"type":"child.compacted"`) {
	804	+	t.Fatal("a failed recovery recorded a checkpoint")
	805	+	}
	806	+
	807	+	// A later reconstructed child resumes the pre-compaction transcript.
	808	+	reopened, err := subagents.OpenResidentJournal(root, spec.ID)
	809	+	if err != nil {
	810	+	t.Fatal(err)
	811	+	}
	812	+	t.Cleanup(func() { _ = reopened.Close() })
	813	+	runner, err = newResidentChildRunner(residentArguments(t, server.URL), spec, reopened)
	814	+	if err != nil {
	815	+	t.Fatal(err)
	816	+	}
	817	+	if err := runner(t.Context(), "verify"); err != nil {
	818	+	t.Fatalf("follow-up error = %v", err)
	819	+	}
	820	+	bodies := script.capturedBodies()
	821	+	if last := bodies[len(bodies)-1]; !strings.Contains(last, "call-1") {
	822	+	t.Fatalf("follow-up request = %q, want the pre-compaction transcript", last)
	823	+	}
	824	+	}
	825	+
	826	+	func TestRecoverResidentContextOverflowRestoresTranscript(t *testing.T) {
	827	+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
	828	+	w.Header().Set("content-type", "text/event-stream")
	829	+	writeOpenAIChunk(t, w, map[string]any{
	830	+	"choices": []any{map[string]any{"index": 0, "delta": map[string]string{"content": "compacted digest"}, "finish_reason": "stop"}},
	831	+	"usage": map[string]int{"prompt_tokens": 5, "completion_tokens": 2},
	832	+	})
	833	+	}))
	834	+	defer server.Close()
	835	+	resolved, err := Resolve(residentArguments(t, server.URL), true)
	836	+	if err != nil {
	837	+	t.Fatal(err)
	838	+	}
	839	+	agent := core.NewAgent(resolved.NewClient(), resolved.Model, "", resolved.ToolRegistry)
	840	+	before := []provider.Message{
	841	+	{Role: provider.RoleUser, Content: []provider.Content{provider.TextBlock{Text: "task"}}},
	842	+	{Role: provider.RoleAssistant, Content: []provider.Content{provider.ToolCallBlock{ID: "call-1", Name: "read", Arguments: json.RawMessage(`{"path":"x"}`)}}},
	843	+	{Role: provider.RoleTool, Content: []provider.Content{provider.ToolResultBlock{CallID: "call-1", Content: []provider.Content{provider.TextBlock{Text: "missing"}}}}},
	844	+	{Role: provider.RoleUser, Content: []provider.Content{provider.TextBlock{Text: "keep working"}}},
	845	+	}
	846	+	agent.SetMessages(before)
	847	+
	848	+	journal, err := subagents.OpenResidentJournal(t.TempDir(), "restore-child")
	849	+	if err != nil {
	850	+	t.Fatal(err)
	851	+	}
	852	+	if err := journal.Accept(subagents.ResidentChildSpec{ID: "restore-child", SessionID: "restore-session", Provider: "openai", Model: "gpt-4o"}, "task"); err != nil {
	853	+	t.Fatal(err)
	854	+	}
	855	+	if err := journal.Close(); err != nil {
	856	+	t.Fatal(err)
	857	+	}
	858	+
	859	+	overflow := errors.New("openai: http 413: context_length_exceeded")
	860	+	err = recoverResidentContextOverflow(t.Context(), agent, journal, func(core.AgentEvent) {}, 0, 4, overflow)
	861	+	if !errors.Is(err, overflow) \|\| !strings.Contains(err.Error(), "persist resident child compaction") {
	862	+	t.Fatalf("recovery error = %v, want a checkpoint persistence failure", err)
	863	+	}
	864	+	if !reflect.DeepEqual(agent.Messages(), before) {
	865	+	t.Fatalf("transcript = %#v, want the pre-compaction transcript restored", agent.Messages())
	866	+	}
	867	+	}
	868	+
	869	+	func residentMessageText(message provider.Message) string {
	870	+	var text strings.Builder
	871	+	for _, content := range message.Content {
	872	+	if block, ok := content.(provider.TextBlock); ok {
	873	+	text.WriteString(block.Text)
	874	+	}
	875	+	}
	876	+	return text.String()
	877	+	}
	878	+
	879	+	func TestResidentChildRunnerIgnoresCompactionUsageInReminder(t *testing.T) {
	880	+	// The summarization request reports the pre-compaction prompt size. If that
	881	+	// gauge armed the next reminder, the first continuation step would claim the
	882	+	// compacted transcript sits at the top of the window. The continuation's own
	883	+	// usage must still be able to arm a truthful reminder.
	884	+	script := &residentOverflowScript{
	885	+	toolCallRequests: 2,
	886	+	continuationToolCalls: 1,
	887	+	finalText: "overflow recovered result",
	888	+	summaryUsageTokens: 1 << 20,
	889	+	continuationUsageTokens: 1 << 20,
	890	+	}
	891	+	server := script.start(t)
	892	+	runner, _, _ := newResidentChildTestRunner(t, "usage-child", residentArguments(t, server.URL))
	893	+	if err := runner(t.Context(), "review"); err != nil {
	894	+	t.Fatalf("runner error = %v, want one recovered turn", err)
	895	+	}
	896	+	bodies := script.capturedBodies()
	897	+	if len(bodies) != 6 {
	898	+	t.Fatalf("provider requests = %d, want 3 prompt, 1 compaction, and 2 continued steps", len(bodies))
	899	+	}
	900	+	if strings.Contains(bodies[4], "Context usage is at") {
	901	+	t.Fatalf("first continuation step carries a reminder from the compaction request: %q", bodies[4])
	902	+	}
	903	+	if !strings.Contains(bodies[5], "Context usage is at") {
	904	+	t.Fatalf("second continuation step lost the reminder: %q", bodies[5])
	905	+	}
	906	+	}
	907	+
	908	+	func TestResidentCheckpointMessagesDropsOnlyHostContext(t *testing.T) {
	909	+	internal := provider.Message{
	910	+	Role: provider.RoleDeveloper,
	911	+	Meta: map[string]string{"internal_context": "true"},
	912	+	Content: []provider.Content{provider.TextBlock{Text: "Context usage is at 96% of the model window."}},
	913	+	}
	914	+	importedNote := provider.Message{
	915	+	Role: provider.RoleDeveloper,
	916	+	Content: []provider.Content{provider.TextBlock{Text: "imported developer note"}},
	917	+	}
	918	+	task := provider.Message{Role: provider.RoleUser, Content: []provider.Content{provider.TextBlock{Text: "task"}}}
	919	+	kept := residentCheckpointMessages([]provider.Message{internal, importedNote, task})
	920	+	if len(kept) != 2 \|\| kept[0].Role != provider.RoleDeveloper \|\| kept[1].Role != provider.RoleUser {
	921	+	t.Fatalf("checkpoint messages = %#v, want host context dropped and other messages kept", kept)
	922	+	}
	923	+	if !core.IsInternalContextMessage(internal) \|\| core.IsInternalContextMessage(importedNote) {
	924	+	t.Fatal("IsInternalContextMessage misclassified a developer message")
	925	+	}
	926	+	}
	927	+
	928	+	func TestResidentChildRunnerKeepsHostContextOutOfCheckpoint(t *testing.T) {
	929	+	// Tool turns report a prompt size past the top band, so the child holds a
	930	+	// host reminder when the overflow arrives.
	931	+	script := &residentOverflowScript{
	932	+	toolCallRequests: 2,
	933	+	finalText: "overflow recovered result",
	934	+	toolUsageTokens: 1 << 20,
	935	+	}
	936	+	server := script.start(t)
	937	+	runner, journal, _ := newResidentChildTestRunner(t, "checkpoint-context", residentArguments(t, server.URL))
	938	+	if err := runner(t.Context(), "review"); err != nil {
	939	+	t.Fatalf("runner error = %v, want one recovered turn", err)
	940	+	}
	941	+	messages, err := subagents.ReadResidentTranscriptMessages(journal.Dir())
	942	+	if err != nil {
	943	+	t.Fatal(err)
	944	+	}
	945	+	joined := joinedResidentText(messages)
	946	+	if !strings.Contains(joined, "compacted digest") {
	947	+	t.Fatalf("resumed transcript = %q, want the checkpoint digest", joined)
	948	+	}
	949	+	if strings.Contains(joined, "Context usage is at") {
	950	+	t.Fatalf("checkpoint replayed a stale host reminder: %q", joined)
	951	+	}
	952	+	}
	953	+
	954	+	func TestResidentChildRunnerKeepsRemainingStepAllowanceAcrossRecovery(t *testing.T) {
	955	+	// Two tool turns and the overflow leave two of the five allowed steps. The
	956	+	// continuation must receive exactly those, not a fresh limit.
	957	+	script := &residentOverflowScript{
	958	+	toolCallRequests: 2,
	959	+	continuationToolCalls: 10,
	960	+	finalText: "unreached",
	961	+	}
	962	+	server := script.start(t)
	963	+	args := residentArguments(t, server.URL)
	964	+	args.MaxSteps = 5
	965	+	runner, journal, _ := newResidentChildTestRunner(t, "partial-allowance", args)
	966	+	err := runner(t.Context(), "review")
	967	+	if !errors.Is(err, core.ErrMaxSteps) {
	968	+	t.Fatalf("runner error = %v, want the continuation to exhaust the remaining allowance", err)
	969	+	}
	970	+	if got := script.requests.Load(); got != 6 {
	971	+	t.Fatalf("provider requests = %d, want 2 tool turns, the overflow, the compaction, and 2 more steps", got)
	972	+	}
	973	+	if checkpoints := countResidentCheckpoints(t, journal); checkpoints != 1 {
	974	+	t.Fatalf("compaction checkpoints = %d, want 1", checkpoints)
	975	+	}
	976	+	}
	977	+
	978	+	func TestResidentChildRunnerDoesNotCompactWithoutHistory(t *testing.T) {
	979	+	// The overflow arrives on the first request, so the transcript is too short
	980	+	// to summarize without replacing the task itself.
	981	+	script := &residentOverflowScript{toolCallRequests: 0, finalText: "unreached"}
	982	+	server := script.start(t)
	983	+	runner, journal, _ := newResidentChildTestRunner(t, "short-overflow", residentArguments(t, server.URL))
	984	+	err := runner(t.Context(), "review")
	985	+	if err == nil \|\| !provider.IsContextOverflowError(err) \|\| !strings.Contains(err.Error(), "context_length_exceeded") {
	986	+	t.Fatalf("runner error = %v, want the original overflow", err)
	987	+	}
	988	+	if got := script.requests.Load(); got != 1 {
	989	+	t.Fatalf("provider requests = %d, want no compaction attempt", got)
	990	+	}
	991	+	if checkpoints := countResidentCheckpoints(t, journal); checkpoints != 0 {
	992	+	t.Fatalf("compaction checkpoints = %d, want none without history", checkpoints)
	993	+	}
	994	+	}
	995	+
	996	+	func joinedResidentText(messages []provider.Message) string {
	997	+	var text strings.Builder
	998	+	for _, message := range messages {
	999	+	text.WriteString(residentMessageText(message))
	1000	+	text.WriteString("\n")
	1001	+	}
	1002	+	return text.String()
	1003	+	}


@@ -7,6 +7,9 @@ import (
7		"fmt"	7		"fmt"
8		"net/http"	8		"net/http"
9		"net/http/httptest"	9		"net/http/httptest"
			10	+	"os"
			11	+	"path/filepath"
			12	+	"reflect"
10		"strings"	13		"strings"
11		"sync"	14		"sync"
12		"sync/atomic"	15		"sync/atomic"

@@ -18,65 +21,87 @@ import (
18		"github.com/bnema/zut/packages/provider"	21		"github.com/bnema/zut/packages/provider"
19		)	22		)
20			23
21	-	func TestResidentChildRunnerDefaultsToFiniteMaxSteps(t *testing.T) {	24	+	func TestResidentChildRunnerWithoutStepLimitRunsPastFormerCap(t *testing.T) {
22	-	for _, tc := range []struct {	25	+	server, requests := residentToolLoopServer(t, 51)
23	-	name string	26	+	runner, _, _ := newResidentChildTestRunner(t, "unlimited-child", Args{
24	-	maxSteps int	27	+	Provider: "openai", Model: "gpt-4o", APIKey: "synthetic", BaseURL: server.URL,
25	-	want int	28	+	CWD: t.TempDir(), NoContextFiles: true, NoSkill: true,
26	-	}{	29	+	})
27	-	{name: "unlimited parent defaults to bounded child", maxSteps: 0, want: 50},	30	+	if err := runner(t.Context(), "review"); err != nil {
28	-	{name: "explicit parent limit preserved", maxSteps: 30, want: 30},	31	+	t.Fatalf("runner error = %v, want success past the former 50-step cap", err)
29	-	} {	32	+	}
30	-	t.Run(tc.name, func(t *testing.T) {	33	+	if got := requests.Load(); got != 52 {
31	-	var requests atomic.Int32	34	+	t.Fatalf("provider requests = %d, want 51 tool turns and one answer", got)
32	-	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {	35	+	}
33	-	requests.Add(1)	36	+	}
34	-	w.Header().Set("content-type", "text/event-stream")	37	+
35	-	writeOpenAIChunk(t, w, map[string]any{	38	+	func TestResidentChildRunnerEnforcesExplicitParentStepLimit(t *testing.T) {
36	-	"choices": []any{map[string]any{	39	+	server, requests := residentToolLoopServer(t, 200)
37	-	"index": 0,	40	+	runner, _, _ := newResidentChildTestRunner(t, "limited-child", Args{
38	-	"delta": map[string]any{"tool_calls": []any{map[string]any{	41	+	Provider: "openai", Model: "gpt-4o", APIKey: "synthetic", BaseURL: server.URL,
39	-	"index": 0, "id": "call-loop", "type": "function",	42	+	MaxSteps: 3, CWD: t.TempDir(), NoContextFiles: true, NoSkill: true,
40	-	"function": map[string]string{"name": "read", "arguments": `{"path":"x"}`},	43	+	})
41	-	}}},	44	+	err := runner(t.Context(), "review")
42	-	"finish_reason": "tool_calls",	45	+	if !errors.Is(err, core.ErrMaxSteps) \|\| !strings.Contains(err.Error(), "3") {
43	-	}},	46	+	t.Fatalf("runner error = %v, want step limit 3", err)
44	-	"usage": map[string]int{"prompt_tokens": 10, "completion_tokens": 1},	47	+	}
45	-	})	48	+	if got := requests.Load(); got != 3 {
46	-	}))	49	+	t.Fatalf("provider requests = %d, want 3", got)
47	-	defer server.Close()	50	+	}
48	-		51	+	}
49	-	spec := subagents.ResidentChildSpec{	52	+
50	-	ID: "maxsteps-child", SessionID: "maxsteps-session", Provider: "openai", Model: "gpt-4o",	53	+	// residentToolLoopServer answers the first toolCalls requests with a distinct
51	-	Tools: []string{"read"},	54	+	// read call and then returns a substantive final answer.
52	-	}	55	+	func residentToolLoopServer(t testing.T, toolCalls int) (httptest.Server, *atomic.Int32) {
53	-	journal, err := subagents.OpenResidentJournal(t.TempDir(), spec.ID)	56	+	t.Helper()
54	-	if err != nil {	57	+	var requests atomic.Int32
55	-	t.Fatal(err)	58	+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
56	-	}	59	+	number := requests.Add(1)
57	-	t.Cleanup(func() { _ = journal.Close() })	60	+	w.Header().Set("content-type", "text/event-stream")
58	-	if err := journal.Accept(spec, "review"); err != nil {	61	+	if int(number) <= toolCalls {
59	-	t.Fatal(err)	62	+	writeOpenAIChunk(t, w, map[string]any{
60	-	}	63	+	"choices": []any{map[string]any{
61	-	runner, err := newResidentChildRunner(Args{	64	+	"index": 0,
62	-	Provider: "openai", Model: "gpt-4o", APIKey: "synthetic", BaseURL: server.URL,	65	+	"delta": map[string]any{"tool_calls": []any{map[string]any{
63	-	MaxSteps: tc.maxSteps, CWD: t.TempDir(), NoContextFiles: true, NoSkill: true,	66	+	"index": 0, "id": fmt.Sprintf("call-%d", number), "type": "function",
64	-	}, spec, journal)	67	+	"function": map[string]string{"name": "read", "arguments": `{"path":"x"}`},
65	-	if err != nil {	68	+	}}},
66	-	t.Fatal(err)	69	+	"finish_reason": "tool_calls",
67	-	}	70	+	}},
68	-	err = runner(t.Context(), "review")	71	+	"usage": map[string]int{"prompt_tokens": 10, "completion_tokens": 1},
69	-	if !errors.Is(err, core.ErrMaxSteps) {	72	+	})
70	-	t.Fatalf("runner error = %v, want step-limit stop", err)	73	+	return
71	-	}	74	+	}
72	-	if !strings.Contains(err.Error(), fmt.Sprintf("%d", tc.want)) {	75	+	writeOpenAIChunk(t, w, map[string]any{
73	-	t.Fatalf("runner error = %v, want limit %d", err, tc.want)	76	+	"choices": []any{map[string]any{"index": 0, "delta": map[string]string{"content": "final result"}, "finish_reason": "stop"}},
74	-	}	77	+	"usage": map[string]int{"prompt_tokens": 10, "completion_tokens": 1},
75	-	if got := requests.Load(); got != int32(tc.want) {
76	-	t.Fatalf("provider requests = %d, want %d", got, tc.want)
77	-	}
78		})	78		})
			79	+	}))
			80	+	t.Cleanup(server.Close)
			81	+	return server, &requests
			82	+	}
			83	+
			84	+	// newResidentChildTestRunner builds one journaled resident child runner with
			85	+	// its own accepted spec.
			86	+	func newResidentChildTestRunner(t testing.T, childID string, args Args) (subagents.ResidentTurnRunner, subagents.ResidentJournal, subagents.ResidentChildSpec) {
			87	+	t.Helper()
			88	+	spec := subagents.ResidentChildSpec{
			89	+	ID: childID, SessionID: childID + "-session", Provider: "openai", Model: "gpt-4o",
			90	+	Tools: []string{"read"},
			91	+	}
			92	+	journal, err := subagents.OpenResidentJournal(t.TempDir(), spec.ID)
			93	+	if err != nil {
			94	+	t.Fatal(err)
79		}	95		}
			96	+	t.Cleanup(func() { _ = journal.Close() })
			97	+	if err := journal.Accept(spec, "review"); err != nil {
			98	+	t.Fatal(err)
			99	+	}
			100	+	runner, err := newResidentChildRunner(args, spec, journal)
			101	+	if err != nil {
			102	+	t.Fatal(err)
			103	+	}
			104	+	return runner, journal, spec
80		}	105		}
81			106
82		func TestResidentChildRunnerInjectsContextReminderIntoNextRequest(t *testing.T) {	107		func TestResidentChildRunnerInjectsContextReminderIntoNextRequest(t *testing.T) {

@@ -202,7 +227,7 @@ func TestResidentChildRegistryUsesExactToolListAndForbidsDelegation(t *testing.T
202		}	227		}
203		}	228		}
204			229
205	-	func TestConfigureResidentContextNudgeRemindsOnceAtThreshold(t *testing.T) {	230	+	func TestConfigureResidentContextNudgeRemindsOncePerBand(t *testing.T) {
206		agent := core.NewAgent(nil, "model", "system", core.Registry{"read": nil})	231		agent := core.NewAgent(nil, "model", "system", core.Registry{"read": nil})
207		configureResidentContextNudge(agent, 1_000)	232		configureResidentContextNudge(agent, 1_000)
208			233

@@ -222,16 +247,52 @@ func TestConfigureResidentContextNudgeRemindsOnceAtThreshold(t *testing.T) {
222		return contextText	247		return contextText
223		}	248		}
224			249
225	-	// Below the threshold: silent, including cache-heavy usage.	250	+	// Below the first band: silent, including cache-heavy usage.
226		check(provider.Usage{InputTokens: 849}, false)	251		check(provider.Usage{InputTokens: 849}, false)
227		check(provider.Usage{InputTokens: 400, CacheReadTokens: 400, CacheWriteTokens: 49}, false)	252		check(provider.Usage{InputTokens: 400, CacheReadTokens: 400, CacheWriteTokens: 49}, false)
228	-	// First crossing: exactly one reminder carrying the computed percent.	253	+	// Each band fires once, carrying the computed percent of the crossing turn.
229		first := check(provider.Usage{InputTokens: 850}, true)	254		first := check(provider.Usage{InputTokens: 850}, true)
230		if !strings.Contains(first, "85%") {	255		if !strings.Contains(first, "85%") {
231		t.Fatalf("reminder percent = %q, want 85%%", first)	256		t.Fatalf("reminder percent = %q, want 85%%", first)
232		}	257		}
233	-	// Later turns stay silent even as usage grows.	258	+	check(provider.Usage{InputTokens: 860}, false)
234	-	check(provider.Usage{InputTokens: 970}, false)	259	+	second := check(provider.Usage{InputTokens: 900}, true)
			260	+	if !strings.Contains(second, "90%") {
			261	+	t.Fatalf("reminder percent = %q, want 90%%", second)
			262	+	}
			263	+	check(provider.Usage{InputTokens: 920}, false)
			264	+	third := check(provider.Usage{InputTokens: 950}, true)
			265	+	if !strings.Contains(third, "95%") {
			266	+	t.Fatalf("reminder percent = %q, want 95%%", third)
			267	+	}
			268	+	// Past the last band the ladder is spent.
			269	+	check(provider.Usage{InputTokens: 999}, false)
			270	+	}
			271	+
			272	+	func TestConfigureResidentContextNudgeIgnoresShrinkingUsage(t *testing.T) {
			273	+	agent := core.NewAgent(nil, "model", "system", core.Registry{"read": nil})
			274	+	configureResidentContextNudge(agent, 1_000)
			275	+
			276	+	check := func(usage provider.Usage, wantReminder bool) string {
			277	+	t.Helper()
			278	+	agent.SeedLastTurnUsage(usage)
			279	+	allowed, reason, contextText := agent.BeforeTurnContext(t.Context(), 1)
			280	+	if !allowed \|\| reason != "" {
			281	+	t.Fatalf("nudge blocked turn: allowed=%t reason=%q", allowed, reason)
			282	+	}
			283	+	if wantReminder != (contextText != "") {
			284	+	t.Fatalf("reminder for usage %#v = %q, want reminder=%t", usage, contextText, wantReminder)
			285	+	}
			286	+	return contextText
			287	+	}
			288	+
			289	+	// A shrinking prompt never delivers a band it did not cross.
			290	+	check(provider.Usage{InputTokens: 700}, false)
			291	+	check(provider.Usage{InputTokens: 500}, false)
			292	+	jumped := check(provider.Usage{InputTokens: 910}, true)
			293	+	if !strings.Contains(jumped, "91%") {
			294	+	t.Fatalf("reminder percent = %q, want 91%%", jumped)
			295	+	}
235		}	296		}
236			297
237		func TestConfigureResidentContextNudgeHandlesDirectJumpAndUnknownWindow(t *testing.T) {	298		func TestConfigureResidentContextNudgeHandlesDirectJumpAndUnknownWindow(t *testing.T) {

@@ -400,25 +461,10 @@ func TestResidentChildRunnerResumesWithRetainedHistory(t *testing.T) {
400			461
401		func newResidentTestRunner(t testing.T, baseURL string) (subagents.ResidentTurnRunner, subagents.ResidentJournal) {	462		func newResidentTestRunner(t testing.T, baseURL string) (subagents.ResidentTurnRunner, subagents.ResidentJournal) {
402		t.Helper()	463		t.Helper()
403	-	spec := subagents.ResidentChildSpec{	464	+	runner, journal, _ := newResidentChildTestRunner(t, "resident-child", Args{
404	-	ID: "resident-child", SessionID: "resident-session", Provider: "openai", Model: "gpt-4o",
405	-	Tools: []string{"read"},
406	-	}
407	-	journal, err := subagents.OpenResidentJournal(t.TempDir(), spec.ID)
408	-	if err != nil {
409	-	t.Fatal(err)
410	-	}
411	-	t.Cleanup(func() { _ = journal.Close() })
412	-	if err := journal.Accept(spec, "review"); err != nil {
413	-	t.Fatal(err)
414	-	}
415	-	runner, err := newResidentChildRunner(Args{
416		Provider: "openai", Model: "gpt-4o", APIKey: "synthetic", BaseURL: baseURL,	465		Provider: "openai", Model: "gpt-4o", APIKey: "synthetic", BaseURL: baseURL,
417		CWD: t.TempDir(), NoContextFiles: true, NoSkill: true,	466		CWD: t.TempDir(), NoContextFiles: true, NoSkill: true,
418	-	}, spec, journal)	467	+	})
419	-	if err != nil {
420	-	t.Fatal(err)
421	-	}
422		return runner, journal	468		return runner, journal
423		}	469		}
424			470

@@ -484,3 +530,474 @@ func TestResidentChildArgsDoNotForwardCLIKeyAcrossProviders(t *testing.T) {
484		t.Fatalf("same-provider child API key = %q, want parent key", next.APIKey)	530		t.Fatalf("same-provider child API key = %q, want parent key", next.APIKey)
485		}	531		}
486		}	532		}
			533	+
			534	+	// residentOverflowScript scripts the provider responses of a child that hits
			535	+	// the context window: leading tool-call turns, one overflow, one compaction
			536	+	// summary, then continued tool turns or the final text.
			537	+	type residentOverflowScript struct {
			538	+	toolCallRequests int
			539	+	continuationToolCalls int
			540	+	secondOverflow bool
			541	+	finalText string
			542	+	onOverflow func()
			543	+	// toolUsageTokens is the prompt size reported by tool-call responses; zero
			544	+	// reports a small prompt.
			545	+	toolUsageTokens int
			546	+	// summaryUsageTokens, when positive, is the prompt size reported by the
			547	+	// compaction summary response.
			548	+	summaryUsageTokens int
			549	+	// continuationUsageTokens is the prompt size reported by tool turns after a
			550	+	// successful compaction; zero reuses toolUsageTokens.
			551	+	continuationUsageTokens int
			552	+
			553	+	requests atomic.Int32
			554	+	mu sync.Mutex
			555	+	bodies []string
			556	+	}
			557	+
			558	+	func (s residentOverflowScript) start(t testing.T) *httptest.Server {
			559	+	t.Helper()
			560	+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
			561	+	number := int(s.requests.Add(1))
			562	+	var body map[string]any
			563	+	if err := json.NewDecoder(r.Body).Decode(&body); err != nil {
			564	+	t.Errorf("decode request: %v", err)
			565	+	}
			566	+	encoded, _ := json.Marshal(body["messages"])
			567	+	s.mu.Lock()
			568	+	s.bodies = append(s.bodies, string(encoded))
			569	+	s.mu.Unlock()
			570	+
			571	+	overflowAt := s.toolCallRequests + 1
			572	+	switch {
			573	+	case number <= s.toolCallRequests:
			574	+	w.Header().Set("content-type", "text/event-stream")
			575	+	writeOpenAIChunk(t, w, residentToolCallChunk(number, s.toolUsageTokens))
			576	+	case number == overflowAt:
			577	+	if s.onOverflow != nil {
			578	+	s.onOverflow()
			579	+	}
			580	+	w.WriteHeader(http.StatusRequestEntityTooLarge)
			581	+	_, _ = w.Write([]byte(`{"error":{"message":"input exceeds the context window","code":"context_length_exceeded"}}`))
			582	+	case number == overflowAt+1:
			583	+	w.Header().Set("content-type", "text/event-stream")
			584	+	writeOpenAIChunk(t, w, residentSummaryChunk(s.summaryUsageTokens))
			585	+	case s.secondOverflow:
			586	+	w.WriteHeader(http.StatusRequestEntityTooLarge)
			587	+	_, _ = w.Write([]byte(`{"error":{"message":"input exceeds the context window","code":"context_length_exceeded"}}`))
			588	+	case number < overflowAt+2+s.continuationToolCalls:
			589	+	w.Header().Set("content-type", "text/event-stream")
			590	+	continuationUsage := s.continuationUsageTokens
			591	+	if continuationUsage <= 0 {
			592	+	continuationUsage = s.toolUsageTokens
			593	+	}
			594	+	writeOpenAIChunk(t, w, residentToolCallChunk(number, continuationUsage))
			595	+	default:
			596	+	w.Header().Set("content-type", "text/event-stream")
			597	+	writeOpenAIChunk(t, w, residentFinalChunk(s.finalText))
			598	+	}
			599	+	}))
			600	+	t.Cleanup(server.Close)
			601	+	return server
			602	+	}
			603	+
			604	+	func residentToolCallChunk(number, usageTokens int) map[string]any {
			605	+	if usageTokens <= 0 {
			606	+	usageTokens = 10
			607	+	}
			608	+	return map[string]any{
			609	+	"choices": []any{map[string]any{
			610	+	"index": 0,
			611	+	"delta": map[string]any{"tool_calls": []any{map[string]any{
			612	+	"index": 0, "id": fmt.Sprintf("call-%d", number), "type": "function",
			613	+	"function": map[string]string{"name": "read", "arguments": `{"path":"x"}`},
			614	+	}}},
			615	+	"finish_reason": "tool_calls",
			616	+	}},
			617	+	"usage": map[string]int{"prompt_tokens": usageTokens, "completion_tokens": 1},
			618	+	}
			619	+	}
			620	+
			621	+	func residentSummaryChunk(usageTokens int) map[string]any {
			622	+	chunk := map[string]any{
			623	+	"choices": []any{map[string]any{"index": 0, "delta": map[string]string{"content": "compacted digest"}, "finish_reason": "stop"}},
			624	+	}
			625	+	if usageTokens > 0 {
			626	+	chunk["usage"] = map[string]int{"prompt_tokens": usageTokens, "completion_tokens": 2}
			627	+	}
			628	+	return chunk
			629	+	}
			630	+
			631	+	func residentFinalChunk(text string) map[string]any {
			632	+	return map[string]any{
			633	+	"choices": []any{map[string]any{"index": 0, "delta": map[string]string{"content": text}, "finish_reason": "stop"}},
			634	+	"usage": map[string]int{"prompt_tokens": 12, "completion_tokens": 5},
			635	+	}
			636	+	}
			637	+
			638	+	func (s *residentOverflowScript) capturedBodies() []string {
			639	+	s.mu.Lock()
			640	+	defer s.mu.Unlock()
			641	+	return append([]string(nil), s.bodies...)
			642	+	}
			643	+
			644	+	func residentArguments(t *testing.T, baseURL string) Args {
			645	+	t.Helper()
			646	+	return Args{
			647	+	Provider: "openai", Model: "gpt-4o", APIKey: "synthetic", BaseURL: baseURL,
			648	+	CWD: t.TempDir(), NoContextFiles: true, NoSkill: true,
			649	+	}
			650	+	}
			651	+
			652	+	func countResidentCheckpoints(t testing.T, journal subagents.ResidentJournal) int {
			653	+	t.Helper()
			654	+	records, err := subagents.ReadResidentJournal(filepath.Join(journal.Dir(), "transcript.jsonl"))
			655	+	if err != nil {
			656	+	t.Fatal(err)
			657	+	}
			658	+	checkpoints := 0
			659	+	for _, record := range records {
			660	+	if record.Type == "child.compacted" {
			661	+	checkpoints++
			662	+	}
			663	+	}
			664	+	return checkpoints
			665	+	}
			666	+
			667	+	func TestResidentChildRunnerRecoversFromContextOverflowOnce(t *testing.T) {
			668	+	script := &residentOverflowScript{toolCallRequests: 2, finalText: "overflow recovered result"}
			669	+	server := script.start(t)
			670	+	runner, journal, spec := newResidentChildTestRunner(t, "overflow-child", residentArguments(t, server.URL))
			671	+	if err := journal.RecordTurnStarted(spec, "turn-1"); err != nil {
			672	+	t.Fatal(err)
			673	+	}
			674	+	runErr := runner(t.Context(), "review")
			675	+	if err := journal.RecordTurnFinished(spec, "turn-1", runErr); err != nil {
			676	+	t.Fatal(err)
			677	+	}
			678	+	if runErr != nil {
			679	+	t.Fatalf("runner error = %v, want one recovered turn", runErr)
			680	+	}
			681	+	if got := script.requests.Load(); got != 5 {
			682	+	t.Fatalf("provider requests = %d, want 3 prompt, 1 compaction, and 1 continuation", got)
			683	+	}
			684	+	result, err := journal.Result()
			685	+	if err != nil {
			686	+	t.Fatal(err)
			687	+	}
			688	+	if result.State != subagents.ResidentIdle \|\| result.Summary != "overflow recovered result" {
			689	+	t.Fatalf("result = %#v, want the recovered final text", result)
			690	+	}
			691	+	records, err := subagents.ReadResidentJournal(filepath.Join(journal.Dir(), "transcript.jsonl"))
			692	+	if err != nil {
			693	+	t.Fatal(err)
			694	+	}
			695	+	checkpoints := 0
			696	+	for _, record := range records {
			697	+	if record.Type != "child.compacted" {
			698	+	continue
			699	+	}
			700	+	checkpoints++
			701	+	if record.TurnID != "turn-1" {
			702	+	t.Fatalf("checkpoint turn ID = %q, want the active accepted turn", record.TurnID)
			703	+	}
			704	+	}
			705	+	if checkpoints != 1 {
			706	+	t.Fatalf("compaction checkpoints = %d, want 1", checkpoints)
			707	+	}
			708	+	// Resume replay starts from the checkpoint: the digest replaces the
			709	+	// pre-overflow tool exchange and the continuation request carried it.
			710	+	messages, err := subagents.ReadResidentTranscriptMessages(journal.Dir())
			711	+	if err != nil {
			712	+	t.Fatal(err)
			713	+	}
			714	+	if len(messages) != 3 \|\| !strings.HasPrefix(residentMessageText(messages[0]), "## Context Summary (compacted)") {
			715	+	t.Fatalf("resumed transcript = %#v, want the checkpoint plus the kept tail", messages)
			716	+	}
			717	+	bodies := script.capturedBodies()
			718	+	if len(bodies) != 5 \|\| !strings.Contains(bodies[4], "compacted digest") {
			719	+	t.Fatalf("continuation request = %q, want the compacted transcript", bodies[len(bodies)-1])
			720	+	}
			721	+	if strings.Contains(bodies[4], "call-1") {
			722	+	t.Fatalf("continuation request kept the dropped tool exchange: %q", bodies[4])
			723	+	}
			724	+	}
			725	+
			726	+	func TestResidentChildRunnerDoesNotRecoverTwicePerTurn(t *testing.T) {
			727	+	script := &residentOverflowScript{toolCallRequests: 2, secondOverflow: true, finalText: "unreached"}
			728	+	server := script.start(t)
			729	+	runner, journal, _ := newResidentChildTestRunner(t, "double-overflow", residentArguments(t, server.URL))
			730	+	err := runner(t.Context(), "review")
			731	+	if err == nil \|\| !provider.IsContextOverflowError(err) {
			732	+	t.Fatalf("runner error = %v, want the second overflow to stay terminal", err)
			733	+	}
			734	+	if got := script.requests.Load(); got != 5 {
			735	+	t.Fatalf("provider requests = %d, want no second compaction", got)
			736	+	}
			737	+	if checkpoints := countResidentCheckpoints(t, journal); checkpoints != 1 {
			738	+	t.Fatalf("compaction checkpoints = %d, want 1", checkpoints)
			739	+	}
			740	+	}
			741	+
			742	+	func TestResidentChildRunnerDoesNotReplenishExplicitStepLimit(t *testing.T) {
			743	+	script := &residentOverflowScript{toolCallRequests: 3, finalText: "unreached"}
			744	+	server := script.start(t)
			745	+	args := residentArguments(t, server.URL)
			746	+	args.MaxSteps = 4
			747	+	runner, journal, _ := newResidentChildTestRunner(t, "limited-overflow", args)
			748	+	err := runner(t.Context(), "review")
			749	+	if err == nil \|\| !provider.IsContextOverflowError(err) {
			750	+	t.Fatalf("runner error = %v, want the overflow", err)
			751	+	}
			752	+	if got := script.requests.Load(); got != 5 {
			753	+	t.Fatalf("provider requests = %d, want the compaction attempt only", got)
			754	+	}
			755	+	// The checkpoint is still journaled so an explicit resume starts small,
			756	+	// but the spent allowance forbids continuing this turn.
			757	+	if checkpoints := countResidentCheckpoints(t, journal); checkpoints != 1 {
			758	+	t.Fatalf("compaction checkpoints = %d, want 1", checkpoints)
			759	+	}
			760	+	for _, body := range script.capturedBodies() {
			761	+	if strings.Contains(body, "compacted digest") {
			762	+	t.Fatalf("recovery continued after the step allowance was spent: %q", body)
			763	+	}
			764	+	}
			765	+	}
			766	+
			767	+	func TestResidentChildRunnerFailsTurnWhenCheckpointCannotPersist(t *testing.T) {
			768	+	root := t.TempDir()
			769	+	spec := subagents.ResidentChildSpec{
			770	+	ID: "persist-failure", SessionID: "persist-failure-session", Provider: "openai", Model: "gpt-4o",
			771	+	Tools: []string{"read"},
			772	+	}
			773	+	journal, err := subagents.OpenResidentJournal(root, spec.ID)
			774	+	if err != nil {
			775	+	t.Fatal(err)
			776	+	}
			777	+	t.Cleanup(func() { _ = journal.Close() })
			778	+	if err := journal.Accept(spec, "review"); err != nil {
			779	+	t.Fatal(err)
			780	+	}
			781	+	// The journal is closed at the exact overflow boundary, so the recovery
			782	+	// cannot persist its checkpoint.
			783	+	script := &residentOverflowScript{
			784	+	toolCallRequests: 2,
			785	+	finalText: "unreached",
			786	+	onOverflow: func() { _ = journal.Close() },
			787	+	}
			788	+	server := script.start(t)
			789	+	runner, err := newResidentChildRunner(residentArguments(t, server.URL), spec, journal)
			790	+	if err != nil {
			791	+	t.Fatal(err)
			792	+	}
			793	+	if err := runner(t.Context(), "review"); err == nil {
			794	+	t.Fatal("runner recovered without persisting its checkpoint")
			795	+	}
			796	+	if got := script.requests.Load(); got != 4 {
			797	+	t.Fatalf("provider requests = %d, want the compaction attempt only", got)
			798	+	}
			799	+	data, readErr := os.ReadFile(filepath.Join(journal.Dir(), "transcript.jsonl"))
			800	+	if readErr != nil {
			801	+	t.Fatal(readErr)
			802	+	}
			803	+	if strings.Contains(string(data), `"type":"child.compacted"`) {
			804	+	t.Fatal("a failed recovery recorded a checkpoint")
			805	+	}
			806	+
			807	+	// A later reconstructed child resumes the pre-compaction transcript.
			808	+	reopened, err := subagents.OpenResidentJournal(root, spec.ID)
			809	+	if err != nil {
			810	+	t.Fatal(err)
			811	+	}
			812	+	t.Cleanup(func() { _ = reopened.Close() })
			813	+	runner, err = newResidentChildRunner(residentArguments(t, server.URL), spec, reopened)
			814	+	if err != nil {
			815	+	t.Fatal(err)
			816	+	}
			817	+	if err := runner(t.Context(), "verify"); err != nil {
			818	+	t.Fatalf("follow-up error = %v", err)
			819	+	}
			820	+	bodies := script.capturedBodies()
			821	+	if last := bodies[len(bodies)-1]; !strings.Contains(last, "call-1") {
			822	+	t.Fatalf("follow-up request = %q, want the pre-compaction transcript", last)
			823	+	}
			824	+	}
			825	+
			826	+	func TestRecoverResidentContextOverflowRestoresTranscript(t *testing.T) {
			827	+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
			828	+	w.Header().Set("content-type", "text/event-stream")
			829	+	writeOpenAIChunk(t, w, map[string]any{
			830	+	"choices": []any{map[string]any{"index": 0, "delta": map[string]string{"content": "compacted digest"}, "finish_reason": "stop"}},
			831	+	"usage": map[string]int{"prompt_tokens": 5, "completion_tokens": 2},
			832	+	})
			833	+	}))
			834	+	defer server.Close()
			835	+	resolved, err := Resolve(residentArguments(t, server.URL), true)
			836	+	if err != nil {
			837	+	t.Fatal(err)
			838	+	}
			839	+	agent := core.NewAgent(resolved.NewClient(), resolved.Model, "", resolved.ToolRegistry)
			840	+	before := []provider.Message{
			841	+	{Role: provider.RoleUser, Content: []provider.Content{provider.TextBlock{Text: "task"}}},
			842	+	{Role: provider.RoleAssistant, Content: []provider.Content{provider.ToolCallBlock{ID: "call-1", Name: "read", Arguments: json.RawMessage(`{"path":"x"}`)}}},
			843	+	{Role: provider.RoleTool, Content: []provider.Content{provider.ToolResultBlock{CallID: "call-1", Content: []provider.Content{provider.TextBlock{Text: "missing"}}}}},
			844	+	{Role: provider.RoleUser, Content: []provider.Content{provider.TextBlock{Text: "keep working"}}},
			845	+	}
			846	+	agent.SetMessages(before)
			847	+
			848	+	journal, err := subagents.OpenResidentJournal(t.TempDir(), "restore-child")
			849	+	if err != nil {
			850	+	t.Fatal(err)
			851	+	}
			852	+	if err := journal.Accept(subagents.ResidentChildSpec{ID: "restore-child", SessionID: "restore-session", Provider: "openai", Model: "gpt-4o"}, "task"); err != nil {
			853	+	t.Fatal(err)
			854	+	}
			855	+	if err := journal.Close(); err != nil {
			856	+	t.Fatal(err)
			857	+	}
			858	+
			859	+	overflow := errors.New("openai: http 413: context_length_exceeded")
			860	+	err = recoverResidentContextOverflow(t.Context(), agent, journal, func(core.AgentEvent) {}, 0, 4, overflow)
			861	+	if !errors.Is(err, overflow) \|\| !strings.Contains(err.Error(), "persist resident child compaction") {
			862	+	t.Fatalf("recovery error = %v, want a checkpoint persistence failure", err)
			863	+	}
			864	+	if !reflect.DeepEqual(agent.Messages(), before) {
			865	+	t.Fatalf("transcript = %#v, want the pre-compaction transcript restored", agent.Messages())
			866	+	}
			867	+	}
			868	+
			869	+	func residentMessageText(message provider.Message) string {
			870	+	var text strings.Builder
			871	+	for _, content := range message.Content {
			872	+	if block, ok := content.(provider.TextBlock); ok {
			873	+	text.WriteString(block.Text)
			874	+	}
			875	+	}
			876	+	return text.String()
			877	+	}
			878	+
			879	+	func TestResidentChildRunnerIgnoresCompactionUsageInReminder(t *testing.T) {
			880	+	// The summarization request reports the pre-compaction prompt size. If that
			881	+	// gauge armed the next reminder, the first continuation step would claim the
			882	+	// compacted transcript sits at the top of the window. The continuation's own
			883	+	// usage must still be able to arm a truthful reminder.
			884	+	script := &residentOverflowScript{
			885	+	toolCallRequests: 2,
			886	+	continuationToolCalls: 1,
			887	+	finalText: "overflow recovered result",
			888	+	summaryUsageTokens: 1 << 20,
			889	+	continuationUsageTokens: 1 << 20,
			890	+	}
			891	+	server := script.start(t)
			892	+	runner, _, _ := newResidentChildTestRunner(t, "usage-child", residentArguments(t, server.URL))
			893	+	if err := runner(t.Context(), "review"); err != nil {
			894	+	t.Fatalf("runner error = %v, want one recovered turn", err)
			895	+	}
			896	+	bodies := script.capturedBodies()
			897	+	if len(bodies) != 6 {
			898	+	t.Fatalf("provider requests = %d, want 3 prompt, 1 compaction, and 2 continued steps", len(bodies))
			899	+	}
			900	+	if strings.Contains(bodies[4], "Context usage is at") {
			901	+	t.Fatalf("first continuation step carries a reminder from the compaction request: %q", bodies[4])
			902	+	}
			903	+	if !strings.Contains(bodies[5], "Context usage is at") {
			904	+	t.Fatalf("second continuation step lost the reminder: %q", bodies[5])
			905	+	}
			906	+	}
			907	+
			908	+	func TestResidentCheckpointMessagesDropsOnlyHostContext(t *testing.T) {
			909	+	internal := provider.Message{
			910	+	Role: provider.RoleDeveloper,
			911	+	Meta: map[string]string{"internal_context": "true"},
			912	+	Content: []provider.Content{provider.TextBlock{Text: "Context usage is at 96% of the model window."}},
			913	+	}
			914	+	importedNote := provider.Message{
			915	+	Role: provider.RoleDeveloper,
			916	+	Content: []provider.Content{provider.TextBlock{Text: "imported developer note"}},
			917	+	}
			918	+	task := provider.Message{Role: provider.RoleUser, Content: []provider.Content{provider.TextBlock{Text: "task"}}}
			919	+	kept := residentCheckpointMessages([]provider.Message{internal, importedNote, task})
			920	+	if len(kept) != 2 \|\| kept[0].Role != provider.RoleDeveloper \|\| kept[1].Role != provider.RoleUser {
			921	+	t.Fatalf("checkpoint messages = %#v, want host context dropped and other messages kept", kept)
			922	+	}
			923	+	if !core.IsInternalContextMessage(internal) \|\| core.IsInternalContextMessage(importedNote) {
			924	+	t.Fatal("IsInternalContextMessage misclassified a developer message")
			925	+	}
			926	+	}
			927	+
			928	+	func TestResidentChildRunnerKeepsHostContextOutOfCheckpoint(t *testing.T) {
			929	+	// Tool turns report a prompt size past the top band, so the child holds a
			930	+	// host reminder when the overflow arrives.
			931	+	script := &residentOverflowScript{
			932	+	toolCallRequests: 2,
			933	+	finalText: "overflow recovered result",
			934	+	toolUsageTokens: 1 << 20,
			935	+	}
			936	+	server := script.start(t)
			937	+	runner, journal, _ := newResidentChildTestRunner(t, "checkpoint-context", residentArguments(t, server.URL))
			938	+	if err := runner(t.Context(), "review"); err != nil {
			939	+	t.Fatalf("runner error = %v, want one recovered turn", err)
			940	+	}
			941	+	messages, err := subagents.ReadResidentTranscriptMessages(journal.Dir())
			942	+	if err != nil {
			943	+	t.Fatal(err)
			944	+	}
			945	+	joined := joinedResidentText(messages)
			946	+	if !strings.Contains(joined, "compacted digest") {
			947	+	t.Fatalf("resumed transcript = %q, want the checkpoint digest", joined)
			948	+	}
			949	+	if strings.Contains(joined, "Context usage is at") {
			950	+	t.Fatalf("checkpoint replayed a stale host reminder: %q", joined)
			951	+	}
			952	+	}
			953	+
			954	+	func TestResidentChildRunnerKeepsRemainingStepAllowanceAcrossRecovery(t *testing.T) {
			955	+	// Two tool turns and the overflow leave two of the five allowed steps. The
			956	+	// continuation must receive exactly those, not a fresh limit.
			957	+	script := &residentOverflowScript{
			958	+	toolCallRequests: 2,
			959	+	continuationToolCalls: 10,
			960	+	finalText: "unreached",
			961	+	}
			962	+	server := script.start(t)
			963	+	args := residentArguments(t, server.URL)
			964	+	args.MaxSteps = 5
			965	+	runner, journal, _ := newResidentChildTestRunner(t, "partial-allowance", args)
			966	+	err := runner(t.Context(), "review")
			967	+	if !errors.Is(err, core.ErrMaxSteps) {
			968	+	t.Fatalf("runner error = %v, want the continuation to exhaust the remaining allowance", err)
			969	+	}
			970	+	if got := script.requests.Load(); got != 6 {
			971	+	t.Fatalf("provider requests = %d, want 2 tool turns, the overflow, the compaction, and 2 more steps", got)
			972	+	}
			973	+	if checkpoints := countResidentCheckpoints(t, journal); checkpoints != 1 {
			974	+	t.Fatalf("compaction checkpoints = %d, want 1", checkpoints)
			975	+	}
			976	+	}
			977	+
			978	+	func TestResidentChildRunnerDoesNotCompactWithoutHistory(t *testing.T) {
			979	+	// The overflow arrives on the first request, so the transcript is too short
			980	+	// to summarize without replacing the task itself.
			981	+	script := &residentOverflowScript{toolCallRequests: 0, finalText: "unreached"}
			982	+	server := script.start(t)
			983	+	runner, journal, _ := newResidentChildTestRunner(t, "short-overflow", residentArguments(t, server.URL))
			984	+	err := runner(t.Context(), "review")
			985	+	if err == nil \|\| !provider.IsContextOverflowError(err) \|\| !strings.Contains(err.Error(), "context_length_exceeded") {
			986	+	t.Fatalf("runner error = %v, want the original overflow", err)
			987	+	}
			988	+	if got := script.requests.Load(); got != 1 {
			989	+	t.Fatalf("provider requests = %d, want no compaction attempt", got)
			990	+	}
			991	+	if checkpoints := countResidentCheckpoints(t, journal); checkpoints != 0 {
			992	+	t.Fatalf("compaction checkpoints = %d, want none without history", checkpoints)
			993	+	}
			994	+	}
			995	+
			996	+	func joinedResidentText(messages []provider.Message) string {
			997	+	var text strings.Builder
			998	+	for _, message := range messages {
			999	+	text.WriteString(residentMessageText(message))
			1000	+	text.WriteString("\n")
			1001	+	}
			1002	+	return text.String()
			1003	+	}

M packages/agent/subagents/resident_history.go +72 -7


@@ -29,13 +29,18 @@ const (
29	29		// raw message and tool payloads are preserved so callers can render complete
30	30		// arguments/results without relying on a lossy text projection.
31	31		type ResidentHistoryItem struct {
32		-	Type string `json:"type"`
33		-	Time time.Time `json:"time"`
34		-	Message json.RawMessage `json:"message,omitempty"`
35		-	ToolID string `json:"tool_id,omitempty"`
36		-	ToolName string `json:"tool_name,omitempty"`
37		-	ToolArgs json.RawMessage `json:"tool_args,omitempty"`
38		-	ToolResult json.RawMessage `json:"tool_result,omitempty"`
	32	+	Type string `json:"type"`
	33	+	Time time.Time `json:"time"`
	34	+	Message json.RawMessage `json:"message,omitempty"`
	35	+	// Messages carries a compaction checkpoint: the complete replacement
	36	+	// transcript of a child that recovered from a context overflow. It is
	37	+	// produced only for residentRecordCompacted items, which the UI pager does
	38	+	// not expose.
	39	+	Messages []json.RawMessage `json:"messages,omitempty"`
	40	+	ToolID string `json:"tool_id,omitempty"`
	41	+	ToolName string `json:"tool_name,omitempty"`
	42	+	ToolArgs json.RawMessage `json:"tool_args,omitempty"`
	43	+	ToolResult json.RawMessage `json:"tool_result,omitempty"`
39	44		}
40	45
41	46		// ResidentHistoryPage is a recent-first, bounded page of finalized history.

@@ -273,6 +278,9 @@ func residentHistoryItem(record residentRecord) (ResidentHistoryItem, bool, erro
273	278		}
274	279		item.ToolID, item.ToolResult = record.ToolID, append(json.RawMessage(nil), record.ToolResult...)
275	280		default:
	281	+	// Lifecycle records and compaction checkpoints are not history items:
	282	+	// the pager keeps showing the append-only log, while resume replay
	283	+	// starts from the newest checkpoint (see ReadResidentTranscriptMessages).
276	284		return ResidentHistoryItem{}, false, nil
277	285		}
278	286		return item, true, nil

@@ -297,15 +305,49 @@ func residentHistoryPageStart(entries []residentHistoryEntry, limit int) int {
297	305
298	306		func cloneResidentHistoryItem(item ResidentHistoryItem) ResidentHistoryItem {
299	307		item.Message = append(json.RawMessage(nil), item.Message...)
	308	+	item.Messages = cloneRawMessages(item.Messages)
300	309		item.ToolArgs = append(json.RawMessage(nil), item.ToolArgs...)
301	310		item.ToolResult = append(json.RawMessage(nil), item.ToolResult...)
302	311		return item
303	312		}
304	313
	314	+	func cloneRawMessages(messages []json.RawMessage) []json.RawMessage {
	315	+	if messages == nil {
	316	+	return nil
	317	+	}
	318	+	cloned := make([]json.RawMessage, len(messages))
	319	+	for index, message := range messages {
	320	+	cloned[index] = append(json.RawMessage(nil), message...)
	321	+	}
	322	+	return cloned
	323	+	}
	324	+
	325	+	// decodeResidentCompactedMessages decodes one compaction checkpoint. The
	326	+	// checkpoint is the complete replacement transcript, so an empty or malformed
	327	+	// payload is corruption rather than an empty transcript.
	328	+	func decodeResidentCompactedMessages(encoded []json.RawMessage) ([]provider.Message, error) {
	329	+	if len(encoded) == 0 {
	330	+	return nil, errors.New("empty compaction checkpoint")
	331	+	}
	332	+	messages := make([]provider.Message, 0, len(encoded))
	333	+	for _, raw := range encoded {
	334	+	message, err := core.DecodeMessageJSON(raw)
	335	+	if err != nil {
	336	+	return nil, fmt.Errorf("decode compaction checkpoint message: %w", err)
	337	+	}
	338	+	messages = append(messages, message)
	339	+	}
	340	+	return messages, nil
	341	+	}
	342	+
305	343		// ResidentHistoryMessages converts finalized journal items into the same
306	344		// provider-neutral message structures used by tui.View. Tool-call records
307	345		// duplicate calls already embedded in an assistant message, so they are only
308	346		// synthesized when the assistant record is absent (for recovery resilience).
	347	+	//
	348	+	// Items are one ordered log, not independent groups: a checkpoint item
	349	+	// replaces every message accumulated before it. A bounded history page never
	350	+	// contains one, because the pager excludes checkpoints.
309	351		func ResidentHistoryMessages(items []ResidentHistoryItem) ([]provider.Message, error) {
310	352		messages := make([]provider.Message, 0, len(items))
311	353		calls := make(map[string]struct{})

@@ -322,6 +364,24 @@ func ResidentHistoryMessages(items []ResidentHistoryItem) ([]provider.Message, e
322	364		}
323	365		}
324	366		messages = append(messages, message)
	367	+	case residentRecordCompacted:
	368	+	// A checkpoint replaces the transcript accumulated so far: the
	369	+	// resumed model must see the compacted view, not the full
	370	+	// append-only log. Tool-call bookkeeping starts over from it so a
	371	+	// call dropped by compaction is not re-synthesized below.
	372	+	checkpoint, err := decodeResidentCompactedMessages(item.Messages)
	373	+	if err != nil {
	374	+	return nil, fmt.Errorf("resident history: %w", err)
	375	+	}
	376	+	messages = checkpoint
	377	+	calls = make(map[string]struct{})
	378	+	for _, message := range checkpoint {
	379	+	for _, content := range message.Content {
	380	+	if call, ok := content.(provider.ToolCallBlock); ok {
	381	+	calls[call.ID] = struct{}{}
	382	+	}
	383	+	}
	384	+	}
325	385		case residentRecordToolCall:
326	386		if _, exists := calls[item.ToolID]; exists {
327	387		continue

@@ -378,6 +438,11 @@ func ReadResidentTranscriptMessages(dir string) ([]provider.Message, error) {
378	438		}
379	439		items := make([]ResidentHistoryItem, 0, len(records))
380	440		for _, record := range records {
	441	+	if record.Type == residentRecordCompacted {
	442	+	// The pager hides checkpoints, but resume replay starts from one.
	443	+	items = append(items, ResidentHistoryItem{Type: record.Type, Time: record.Time, Messages: cloneRawMessages(record.Messages)})
	444	+	continue
	445	+	}
381	446		item, include, err := residentHistoryItem(record)
382	447		if err != nil {
383	448		return nil, err


@@ -29,13 +29,18 @@ const (
29		// raw message and tool payloads are preserved so callers can render complete	29		// raw message and tool payloads are preserved so callers can render complete
30		// arguments/results without relying on a lossy text projection.	30		// arguments/results without relying on a lossy text projection.
31		type ResidentHistoryItem struct {	31		type ResidentHistoryItem struct {
32	-	Type string `json:"type"`	32	+	Type string `json:"type"`
33	-	Time time.Time `json:"time"`	33	+	Time time.Time `json:"time"`
34	-	Message json.RawMessage `json:"message,omitempty"`	34	+	Message json.RawMessage `json:"message,omitempty"`
35	-	ToolID string `json:"tool_id,omitempty"`	35	+	// Messages carries a compaction checkpoint: the complete replacement
36	-	ToolName string `json:"tool_name,omitempty"`	36	+	// transcript of a child that recovered from a context overflow. It is
37	-	ToolArgs json.RawMessage `json:"tool_args,omitempty"`	37	+	// produced only for residentRecordCompacted items, which the UI pager does
38	-	ToolResult json.RawMessage `json:"tool_result,omitempty"`	38	+	// not expose.
			39	+	Messages []json.RawMessage `json:"messages,omitempty"`
			40	+	ToolID string `json:"tool_id,omitempty"`
			41	+	ToolName string `json:"tool_name,omitempty"`
			42	+	ToolArgs json.RawMessage `json:"tool_args,omitempty"`
			43	+	ToolResult json.RawMessage `json:"tool_result,omitempty"`
39		}	44		}
40			45
41		// ResidentHistoryPage is a recent-first, bounded page of finalized history.	46		// ResidentHistoryPage is a recent-first, bounded page of finalized history.

@@ -273,6 +278,9 @@ func residentHistoryItem(record residentRecord) (ResidentHistoryItem, bool, erro
273		}	278		}
274		item.ToolID, item.ToolResult = record.ToolID, append(json.RawMessage(nil), record.ToolResult...)	279		item.ToolID, item.ToolResult = record.ToolID, append(json.RawMessage(nil), record.ToolResult...)
275		default:	280		default:
			281	+	// Lifecycle records and compaction checkpoints are not history items:
			282	+	// the pager keeps showing the append-only log, while resume replay
			283	+	// starts from the newest checkpoint (see ReadResidentTranscriptMessages).
276		return ResidentHistoryItem{}, false, nil	284		return ResidentHistoryItem{}, false, nil
277		}	285		}
278		return item, true, nil	286		return item, true, nil

@@ -297,15 +305,49 @@ func residentHistoryPageStart(entries []residentHistoryEntry, limit int) int {
297			305
298		func cloneResidentHistoryItem(item ResidentHistoryItem) ResidentHistoryItem {	306		func cloneResidentHistoryItem(item ResidentHistoryItem) ResidentHistoryItem {
299		item.Message = append(json.RawMessage(nil), item.Message...)	307		item.Message = append(json.RawMessage(nil), item.Message...)
			308	+	item.Messages = cloneRawMessages(item.Messages)
300		item.ToolArgs = append(json.RawMessage(nil), item.ToolArgs...)	309		item.ToolArgs = append(json.RawMessage(nil), item.ToolArgs...)
301		item.ToolResult = append(json.RawMessage(nil), item.ToolResult...)	310		item.ToolResult = append(json.RawMessage(nil), item.ToolResult...)
302		return item	311		return item
303		}	312		}
304			313
			314	+	func cloneRawMessages(messages []json.RawMessage) []json.RawMessage {
			315	+	if messages == nil {
			316	+	return nil
			317	+	}
			318	+	cloned := make([]json.RawMessage, len(messages))
			319	+	for index, message := range messages {
			320	+	cloned[index] = append(json.RawMessage(nil), message...)
			321	+	}
			322	+	return cloned
			323	+	}
			324	+
			325	+	// decodeResidentCompactedMessages decodes one compaction checkpoint. The
			326	+	// checkpoint is the complete replacement transcript, so an empty or malformed
			327	+	// payload is corruption rather than an empty transcript.
			328	+	func decodeResidentCompactedMessages(encoded []json.RawMessage) ([]provider.Message, error) {
			329	+	if len(encoded) == 0 {
			330	+	return nil, errors.New("empty compaction checkpoint")
			331	+	}
			332	+	messages := make([]provider.Message, 0, len(encoded))
			333	+	for _, raw := range encoded {
			334	+	message, err := core.DecodeMessageJSON(raw)
			335	+	if err != nil {
			336	+	return nil, fmt.Errorf("decode compaction checkpoint message: %w", err)
			337	+	}
			338	+	messages = append(messages, message)
			339	+	}
			340	+	return messages, nil
			341	+	}
			342	+
305		// ResidentHistoryMessages converts finalized journal items into the same	343		// ResidentHistoryMessages converts finalized journal items into the same
306		// provider-neutral message structures used by tui.View. Tool-call records	344		// provider-neutral message structures used by tui.View. Tool-call records
307		// duplicate calls already embedded in an assistant message, so they are only	345		// duplicate calls already embedded in an assistant message, so they are only
308		// synthesized when the assistant record is absent (for recovery resilience).	346		// synthesized when the assistant record is absent (for recovery resilience).
			347	+	//
			348	+	// Items are one ordered log, not independent groups: a checkpoint item
			349	+	// replaces every message accumulated before it. A bounded history page never
			350	+	// contains one, because the pager excludes checkpoints.
309		func ResidentHistoryMessages(items []ResidentHistoryItem) ([]provider.Message, error) {	351		func ResidentHistoryMessages(items []ResidentHistoryItem) ([]provider.Message, error) {
310		messages := make([]provider.Message, 0, len(items))	352		messages := make([]provider.Message, 0, len(items))
311		calls := make(map[string]struct{})	353		calls := make(map[string]struct{})

@@ -322,6 +364,24 @@ func ResidentHistoryMessages(items []ResidentHistoryItem) ([]provider.Message, e
322		}	364		}
323		}	365		}
324		messages = append(messages, message)	366		messages = append(messages, message)
			367	+	case residentRecordCompacted:
			368	+	// A checkpoint replaces the transcript accumulated so far: the
			369	+	// resumed model must see the compacted view, not the full
			370	+	// append-only log. Tool-call bookkeeping starts over from it so a
			371	+	// call dropped by compaction is not re-synthesized below.
			372	+	checkpoint, err := decodeResidentCompactedMessages(item.Messages)
			373	+	if err != nil {
			374	+	return nil, fmt.Errorf("resident history: %w", err)
			375	+	}
			376	+	messages = checkpoint
			377	+	calls = make(map[string]struct{})
			378	+	for _, message := range checkpoint {
			379	+	for _, content := range message.Content {
			380	+	if call, ok := content.(provider.ToolCallBlock); ok {
			381	+	calls[call.ID] = struct{}{}
			382	+	}
			383	+	}
			384	+	}
325		case residentRecordToolCall:	385		case residentRecordToolCall:
326		if _, exists := calls[item.ToolID]; exists {	386		if _, exists := calls[item.ToolID]; exists {
327		continue	387		continue

@@ -378,6 +438,11 @@ func ReadResidentTranscriptMessages(dir string) ([]provider.Message, error) {
378		}	438		}
379		items := make([]ResidentHistoryItem, 0, len(records))	439		items := make([]ResidentHistoryItem, 0, len(records))
380		for _, record := range records {	440		for _, record := range records {
			441	+	if record.Type == residentRecordCompacted {
			442	+	// The pager hides checkpoints, but resume replay starts from one.
			443	+	items = append(items, ResidentHistoryItem{Type: record.Type, Time: record.Time, Messages: cloneRawMessages(record.Messages)})
			444	+	continue
			445	+	}
381		item, include, err := residentHistoryItem(record)	446		item, include, err := residentHistoryItem(record)
382		if err != nil {	447		if err != nil {
383		return nil, err	448		return nil, err

M packages/agent/subagents/resident_history_test.go +154 -0


@@ -5,6 +5,7 @@ import (
5	5		"encoding/json"
6	6		"os"
7	7		"path/filepath"
	8	+	"strings"
8	9		"testing"
9	10
10	11		"github.com/bnema/zut/packages/core"

@@ -244,3 +245,156 @@ func TestResidentManagerHistoryRejectsPathTraversal(t *testing.T) {
244	245		}
245	246		}
246	247		}
	248	+
	249	+	func TestReadResidentTranscriptMessagesReplaysCompactionCheckpoint(t *testing.T) {
	250	+	root := t.TempDir()
	251	+	journal, err := OpenResidentJournal(root, "compacted")
	252	+	if err != nil {
	253	+	t.Fatal(err)
	254	+	}
	255	+	spec := ResidentChildSpec{ID: "compacted", SessionID: "compacted-session", Provider: "openai", Model: "gpt-5"}
	256	+	if err := journal.Accept(spec, "task"); err != nil {
	257	+	t.Fatal(err)
	258	+	}
	259	+	if err := journal.RecordAgentEvent(core.EvUserMessage{Message: provider.Message{Role: provider.RoleUser, Content: []provider.Content{provider.TextBlock{Text: "old question"}}}}); err != nil {
	260	+	t.Fatal(err)
	261	+	}
	262	+	if err := journal.RecordAgentEvent(core.EvToolCall{ID: "dropped-call", Name: "bash", Args: json.RawMessage(`{"command":"pwd"}`)}); err != nil {
	263	+	t.Fatal(err)
	264	+	}
	265	+	if err := journal.RecordAgentEvent(core.EvToolResult{ID: "dropped-call", Result: core.ToolResult{Content: []provider.Content{provider.TextBlock{Text: "/repo"}}}}); err != nil {
	266	+	t.Fatal(err)
	267	+	}
	268	+	if err := journal.RecordTurnStarted(spec, "turn-1"); err != nil {
	269	+	t.Fatal(err)
	270	+	}
	271	+	if err := journal.RecordCompacted([]provider.Message{
	272	+	{Role: provider.RoleUser, Content: []provider.Content{provider.TextBlock{Text: "## Context Summary (compacted)\n\ndigest"}}},
	273	+	}); err != nil {
	274	+	t.Fatal(err)
	275	+	}
	276	+	if err := journal.RecordAgentEvent(core.EvAssistantMessage{Message: provider.Message{Role: provider.RoleAssistant, Content: []provider.Content{provider.TextBlock{Text: "after checkpoint"}}}}); err != nil {
	277	+	t.Fatal(err)
	278	+	}
	279	+	transcript := filepath.Join(journal.Dir(), residentTranscriptName)
	280	+	if err := journal.Close(); err != nil {
	281	+	t.Fatal(err)
	282	+	}
	283	+
	284	+	records, err := ReadResidentJournal(transcript)
	285	+	if err != nil {
	286	+	t.Fatal(err)
	287	+	}
	288	+	var checkpoints []residentRecord
	289	+	for _, record := range records {
	290	+	if record.Type == residentRecordCompacted {
	291	+	checkpoints = append(checkpoints, record)
	292	+	}
	293	+	}
	294	+	if len(checkpoints) != 1 \|\| checkpoints[0].TurnID != "turn-1" \|\| len(checkpoints[0].Messages) != 1 {
	295	+	t.Fatalf("checkpoints = %#v, want one attributed to the active turn", checkpoints)
	296	+	}
	297	+
	298	+	dir := filepath.Join(root, spec.ID)
	299	+	messages, err := ReadResidentTranscriptMessages(dir)
	300	+	if err != nil {
	301	+	t.Fatal(err)
	302	+	}
	303	+	if len(messages) != 2 \|\| messages[0].Role != provider.RoleUser \|\| messages[1].Role != provider.RoleAssistant {
	304	+	t.Fatalf("resumed messages = %#v, want the checkpoint followed by the later reply", messages)
	305	+	}
	306	+	if text := joinedResidentMessageText(messages); strings.Contains(text, "old question") {
	307	+	t.Fatalf("resume replayed pre-checkpoint content: %q", text)
	308	+	}
	309	+	if !strings.Contains(joinedResidentMessageText(messages[:1]), "Context Summary (compacted)") {
	310	+	t.Fatalf("resume lost the checkpoint summary: %#v", messages[0])
	311	+	}
	312	+	for _, message := range messages {
	313	+	for _, content := range message.Content {
	314	+	if call, ok := content.(provider.ToolCallBlock); ok {
	315	+	t.Fatalf("resume re-synthesized compacted tool call %q", call.ID)
	316	+	}
	317	+	}
	318	+	}
	319	+
	320	+	// The UI pager keeps the append-only view: a checkpoint is not a history
	321	+	// item, so paging still reaches the pre-checkpoint log.
	322	+	page, err := ReadResidentHistoryPage(dir, "", 10)
	323	+	if err != nil {
	324	+	t.Fatal(err)
	325	+	}
	326	+	if page.OlderCursor == "" {
	327	+	t.Fatal("page has no cursor for the pre-checkpoint log")
	328	+	}
	329	+	older, err := ReadResidentHistoryPage(dir, page.OlderCursor, 10)
	330	+	if err != nil {
	331	+	t.Fatal(err)
	332	+	}
	333	+	if len(older.Items) != 3 \|\| older.Items[0].Type != residentRecordUser \|\| !strings.Contains(string(older.Items[0].Message), "old question") {
	334	+	t.Fatalf("older page = %#v, want the append-only log before the checkpoint", older.Items)
	335	+	}
	336	+	for _, item := range append(append([]ResidentHistoryItem(nil), page.Items...), older.Items...) {
	337	+	if item.Type == residentRecordCompacted {
	338	+	t.Fatalf("pager exposed a compaction checkpoint: %#v", item)
	339	+	}
	340	+	}
	341	+	}
	342	+
	343	+	func TestReadResidentTranscriptMessagesRejectsInvalidCheckpoint(t *testing.T) {
	344	+	for _, tc := range []struct {
	345	+	name string
	346	+	line string
	347	+	}{
	348	+	{name: "empty", line: `{"version":3,"type":"child.compacted","time":"2026-01-01T00:00:00Z"}`},
	349	+	{name: "undecodable", line: `{"version":3,"type":"child.compacted","time":"2026-01-01T00:00:00Z","messages":["not a message"]}`},
	350	+	} {
	351	+	t.Run(tc.name, func(t *testing.T) {
	352	+	root := t.TempDir()
	353	+	journal, err := OpenResidentJournal(root, "bad-checkpoint")
	354	+	if err != nil {
	355	+	t.Fatal(err)
	356	+	}
	357	+	spec := ResidentChildSpec{ID: "bad-checkpoint", SessionID: "session", Provider: "openai", Model: "gpt-5"}
	358	+	if err := journal.Accept(spec, "task"); err != nil {
	359	+	t.Fatal(err)
	360	+	}
	361	+	if err := journal.RecordAgentEvent(core.EvUserMessage{Message: provider.Message{Role: provider.RoleUser, Content: []provider.Content{provider.TextBlock{Text: "question"}}}}); err != nil {
	362	+	t.Fatal(err)
	363	+	}
	364	+	dir := journal.Dir()
	365	+	if err := journal.Close(); err != nil {
	366	+	t.Fatal(err)
	367	+	}
	368	+	path := filepath.Join(dir, residentTranscriptName)
	369	+	file, err := os.OpenFile(path, os.O_APPEND\|os.O_WRONLY, 0o600)
	370	+	if err != nil {
	371	+	t.Fatal(err)
	372	+	}
	373	+	if _, err := file.WriteString(tc.line + "\n"); err != nil {
	374	+	t.Fatal(err)
	375	+	}
	376	+	if err := file.Close(); err != nil {
	377	+	t.Fatal(err)
	378	+	}
	379	+	if _, err := ReadResidentTranscriptMessages(dir); err == nil {
	380	+	t.Fatal("invalid checkpoint was accepted")
	381	+	}
	382	+	if _, err := ReconcileResidentJournal(dir); err == nil {
	383	+	t.Fatal("reconciliation accepted an invalid checkpoint")
	384	+	}
	385	+	})
	386	+	}
	387	+	}
	388	+
	389	+	func joinedResidentMessageText(messages []provider.Message) string {
	390	+	var text strings.Builder
	391	+	for _, message := range messages {
	392	+	for _, content := range message.Content {
	393	+	if block, ok := content.(provider.TextBlock); ok {
	394	+	text.WriteString(block.Text)
	395	+	text.WriteString("\n")
	396	+	}
	397	+	}
	398	+	}
	399	+	return text.String()
	400	+	}


@@ -5,6 +5,7 @@ import (
5	"encoding/json"	5		"encoding/json"
6	"os"	6		"os"
7	"path/filepath"	7		"path/filepath"
		8	+	"strings"
8	"testing"	9		"testing"
9		10
10	"github.com/bnema/zut/packages/core"	11		"github.com/bnema/zut/packages/core"

@@ -244,3 +245,156 @@ func TestResidentManagerHistoryRejectsPathTraversal(t *testing.T) {
244	}	245		}
245	}	246		}
246	}	247		}
		248	+
		249	+	func TestReadResidentTranscriptMessagesReplaysCompactionCheckpoint(t *testing.T) {
		250	+	root := t.TempDir()
		251	+	journal, err := OpenResidentJournal(root, "compacted")
		252	+	if err != nil {
		253	+	t.Fatal(err)
		254	+	}
		255	+	spec := ResidentChildSpec{ID: "compacted", SessionID: "compacted-session", Provider: "openai", Model: "gpt-5"}
		256	+	if err := journal.Accept(spec, "task"); err != nil {
		257	+	t.Fatal(err)
		258	+	}
		259	+	if err := journal.RecordAgentEvent(core.EvUserMessage{Message: provider.Message{Role: provider.RoleUser, Content: []provider.Content{provider.TextBlock{Text: "old question"}}}}); err != nil {
		260	+	t.Fatal(err)
		261	+	}
		262	+	if err := journal.RecordAgentEvent(core.EvToolCall{ID: "dropped-call", Name: "bash", Args: json.RawMessage(`{"command":"pwd"}`)}); err != nil {
		263	+	t.Fatal(err)
		264	+	}
		265	+	if err := journal.RecordAgentEvent(core.EvToolResult{ID: "dropped-call", Result: core.ToolResult{Content: []provider.Content{provider.TextBlock{Text: "/repo"}}}}); err != nil {
		266	+	t.Fatal(err)
		267	+	}
		268	+	if err := journal.RecordTurnStarted(spec, "turn-1"); err != nil {
		269	+	t.Fatal(err)
		270	+	}
		271	+	if err := journal.RecordCompacted([]provider.Message{
		272	+	{Role: provider.RoleUser, Content: []provider.Content{provider.TextBlock{Text: "## Context Summary (compacted)\n\ndigest"}}},
		273	+	}); err != nil {
		274	+	t.Fatal(err)
		275	+	}
		276	+	if err := journal.RecordAgentEvent(core.EvAssistantMessage{Message: provider.Message{Role: provider.RoleAssistant, Content: []provider.Content{provider.TextBlock{Text: "after checkpoint"}}}}); err != nil {
		277	+	t.Fatal(err)
		278	+	}
		279	+	transcript := filepath.Join(journal.Dir(), residentTranscriptName)
		280	+	if err := journal.Close(); err != nil {
		281	+	t.Fatal(err)
		282	+	}
		283	+
		284	+	records, err := ReadResidentJournal(transcript)
		285	+	if err != nil {
		286	+	t.Fatal(err)
		287	+	}
		288	+	var checkpoints []residentRecord
		289	+	for _, record := range records {
		290	+	if record.Type == residentRecordCompacted {
		291	+	checkpoints = append(checkpoints, record)
		292	+	}
		293	+	}
		294	+	if len(checkpoints) != 1 \|\| checkpoints[0].TurnID != "turn-1" \|\| len(checkpoints[0].Messages) != 1 {
		295	+	t.Fatalf("checkpoints = %#v, want one attributed to the active turn", checkpoints)
		296	+	}
		297	+
		298	+	dir := filepath.Join(root, spec.ID)
		299	+	messages, err := ReadResidentTranscriptMessages(dir)
		300	+	if err != nil {
		301	+	t.Fatal(err)
		302	+	}
		303	+	if len(messages) != 2 \|\| messages[0].Role != provider.RoleUser \|\| messages[1].Role != provider.RoleAssistant {
		304	+	t.Fatalf("resumed messages = %#v, want the checkpoint followed by the later reply", messages)
		305	+	}
		306	+	if text := joinedResidentMessageText(messages); strings.Contains(text, "old question") {
		307	+	t.Fatalf("resume replayed pre-checkpoint content: %q", text)
		308	+	}
		309	+	if !strings.Contains(joinedResidentMessageText(messages[:1]), "Context Summary (compacted)") {
		310	+	t.Fatalf("resume lost the checkpoint summary: %#v", messages[0])
		311	+	}
		312	+	for _, message := range messages {
		313	+	for _, content := range message.Content {
		314	+	if call, ok := content.(provider.ToolCallBlock); ok {
		315	+	t.Fatalf("resume re-synthesized compacted tool call %q", call.ID)
		316	+	}
		317	+	}
		318	+	}
		319	+
		320	+	// The UI pager keeps the append-only view: a checkpoint is not a history
		321	+	// item, so paging still reaches the pre-checkpoint log.
		322	+	page, err := ReadResidentHistoryPage(dir, "", 10)
		323	+	if err != nil {
		324	+	t.Fatal(err)
		325	+	}
		326	+	if page.OlderCursor == "" {
		327	+	t.Fatal("page has no cursor for the pre-checkpoint log")
		328	+	}
		329	+	older, err := ReadResidentHistoryPage(dir, page.OlderCursor, 10)
		330	+	if err != nil {
		331	+	t.Fatal(err)
		332	+	}
		333	+	if len(older.Items) != 3 \|\| older.Items[0].Type != residentRecordUser \|\| !strings.Contains(string(older.Items[0].Message), "old question") {
		334	+	t.Fatalf("older page = %#v, want the append-only log before the checkpoint", older.Items)
		335	+	}
		336	+	for _, item := range append(append([]ResidentHistoryItem(nil), page.Items...), older.Items...) {
		337	+	if item.Type == residentRecordCompacted {
		338	+	t.Fatalf("pager exposed a compaction checkpoint: %#v", item)
		339	+	}
		340	+	}
		341	+	}
		342	+
		343	+	func TestReadResidentTranscriptMessagesRejectsInvalidCheckpoint(t *testing.T) {
		344	+	for _, tc := range []struct {
		345	+	name string
		346	+	line string
		347	+	}{
		348	+	{name: "empty", line: `{"version":3,"type":"child.compacted","time":"2026-01-01T00:00:00Z"}`},
		349	+	{name: "undecodable", line: `{"version":3,"type":"child.compacted","time":"2026-01-01T00:00:00Z","messages":["not a message"]}`},
		350	+	} {
		351	+	t.Run(tc.name, func(t *testing.T) {
		352	+	root := t.TempDir()
		353	+	journal, err := OpenResidentJournal(root, "bad-checkpoint")
		354	+	if err != nil {
		355	+	t.Fatal(err)
		356	+	}
		357	+	spec := ResidentChildSpec{ID: "bad-checkpoint", SessionID: "session", Provider: "openai", Model: "gpt-5"}
		358	+	if err := journal.Accept(spec, "task"); err != nil {
		359	+	t.Fatal(err)
		360	+	}
		361	+	if err := journal.RecordAgentEvent(core.EvUserMessage{Message: provider.Message{Role: provider.RoleUser, Content: []provider.Content{provider.TextBlock{Text: "question"}}}}); err != nil {
		362	+	t.Fatal(err)
		363	+	}
		364	+	dir := journal.Dir()
		365	+	if err := journal.Close(); err != nil {
		366	+	t.Fatal(err)
		367	+	}
		368	+	path := filepath.Join(dir, residentTranscriptName)
		369	+	file, err := os.OpenFile(path, os.O_APPEND\|os.O_WRONLY, 0o600)
		370	+	if err != nil {
		371	+	t.Fatal(err)
		372	+	}
		373	+	if _, err := file.WriteString(tc.line + "\n"); err != nil {
		374	+	t.Fatal(err)
		375	+	}
		376	+	if err := file.Close(); err != nil {
		377	+	t.Fatal(err)
		378	+	}
		379	+	if _, err := ReadResidentTranscriptMessages(dir); err == nil {
		380	+	t.Fatal("invalid checkpoint was accepted")
		381	+	}
		382	+	if _, err := ReconcileResidentJournal(dir); err == nil {
		383	+	t.Fatal("reconciliation accepted an invalid checkpoint")
		384	+	}
		385	+	})
		386	+	}
		387	+	}
		388	+
		389	+	func joinedResidentMessageText(messages []provider.Message) string {
		390	+	var text strings.Builder
		391	+	for _, message := range messages {
		392	+	for _, content := range message.Content {
		393	+	if block, ok := content.(provider.TextBlock); ok {
		394	+	text.WriteString(block.Text)
		395	+	text.WriteString("\n")
		396	+	}
		397	+	}
		398	+	}
		399	+	return text.String()
		400	+	}

M packages/agent/subagents/resident_journal.go +105 -19


@@ -36,6 +36,7 @@ const (
36	36		residentRecordAssistant = "message.assistant"
37	37		residentRecordToolCall = "tool.call"
38	38		residentRecordToolResult = "tool.result"
	39	+	residentRecordCompacted = "child.compacted"
39	40		residentRecordUsage = "usage"
40	41		residentJournalVersion = 3
41	42		residentJournalVersionV2 = 2

@@ -121,15 +122,19 @@ type residentRecord struct {
121	122		PatchRef string `json:"patch_ref,omitempty"`
122	123		ChangedFiles []string `json:"changed_files,omitempty"`
123	124		Message json.RawMessage `json:"message,omitempty"`
124		-	ToolID string `json:"tool_id,omitempty"`
125		-	ToolName string `json:"tool_name,omitempty"`
126		-	ToolArgs json.RawMessage `json:"tool_args,omitempty"`
127		-	ToolResult json.RawMessage `json:"tool_result,omitempty"`
128		-	Usage *provider.Usage `json:"usage,omitempty"`
129		-	ContextUsed int `json:"context_used,omitempty"`
130		-	ContextMax int `json:"context_max,omitempty"`
131		-	Subscription bool `json:"subscription,omitempty"`
132		-	raw json.RawMessage `json:"-"`
	125	+	// Messages carries a full replacement transcript. It is used only by
	126	+	// child.compacted, the durable checkpoint written after an overflow
	127	+	// recovery compacts the child transcript.
	128	+	Messages []json.RawMessage `json:"messages,omitempty"`
	129	+	ToolID string `json:"tool_id,omitempty"`
	130	+	ToolName string `json:"tool_name,omitempty"`
	131	+	ToolArgs json.RawMessage `json:"tool_args,omitempty"`
	132	+	ToolResult json.RawMessage `json:"tool_result,omitempty"`
	133	+	Usage *provider.Usage `json:"usage,omitempty"`
	134	+	ContextUsed int `json:"context_used,omitempty"`
	135	+	ContextMax int `json:"context_max,omitempty"`
	136	+	Subscription bool `json:"subscription,omitempty"`
	137	+	raw json.RawMessage `json:"-"`
133	138		}
134	139
135	140		// RecordAgentEvent persists finalized provider-neutral transcript events. It

@@ -231,6 +236,40 @@ func (j *ResidentJournal) RecordAgentEvent(event core.AgentEvent) error {
231	236		return err
232	237		}
233	238
	239	+	// RecordCompacted persists the replacement transcript produced by a
	240	+	// context-overflow recovery. The checkpoint is additive: journals without one
	241	+	// replay exactly as before, and resume replay starts from the newest
	242	+	// checkpoint instead of the full append-only log.
	243	+	func (j *ResidentJournal) RecordCompacted(messages []provider.Message) error {
	244	+	if j == nil {
	245	+	return errors.New("resident journal: unavailable")
	246	+	}
	247	+	if len(messages) == 0 {
	248	+	return errors.New("resident journal: empty compaction checkpoint")
	249	+	}
	250	+	encoded := make([]json.RawMessage, 0, len(messages))
	251	+	for _, message := range messages {
	252	+	data, err := json.Marshal(message)
	253	+	if err != nil {
	254	+	return fmt.Errorf("resident journal encode compacted message: %w", err)
	255	+	}
	256	+	encoded = append(encoded, data)
	257	+	}
	258	+	j.mu.Lock()
	259	+	defer j.mu.Unlock()
	260	+	if j.file == nil {
	261	+	return errors.New("resident journal: closed")
	262	+	}
	263	+	record := residentRecord{
	264	+	Version: residentJournalVersion,
	265	+	Type: residentRecordCompacted,
	266	+	Time: time.Now().UTC(),
	267	+	TurnID: j.activeTurnID,
	268	+	Messages: encoded,
	269	+	}
	270	+	return j.appendSync(record)
	271	+	}
	272	+
234	273		// ResidentMetadata is a bounded, rebuildable projection. The transcript is
235	274		// the authority for acceptance; callers must never infer acceptance solely
236	275		// from this file.

@@ -285,6 +324,12 @@ type ResidentJournal struct {
285	324		dir string
286	325		file *os.File
287	326		lease *residentLease
	327	+	// activeTurnID is the resident turn currently executing. It attributes
	328	+	// records that the runner writes without a turn parameter, such as a
	329	+	// compaction checkpoint, and is diagnostics only: no reader consumes a
	330	+	// checkpoint's turn ID. RecordTurnStarted owns the value and a terminal
	331	+	// boundary clears it. It is guarded by mu.
	332	+	activeTurnID string
288	333		}
289	334
290	335		// ResidentUsageSnapshot is the bounded usage projection shared by durable and

@@ -581,6 +626,18 @@ func (j *ResidentJournal) recordTurnBoundary(spec ResidentChildSpec, state Resid
581	626		if err := j.appendSync(record); err != nil {
582	627		return err
583	628		}
	629	+	// Track the accepted turn that is currently executing so records written
	630	+	// without an explicit turn parameter, such as a compaction checkpoint, stay
	631	+	// attributable. A terminal boundary clears it; a late checkpoint from a
	632	+	// finished turn can therefore not be attributed to the next one.
	633	+	switch record.Type {
	634	+	case residentRecordTurnStarted:
	635	+	j.activeTurnID = record.TurnID
	636	+	case residentRecordTurnFinished, residentRecordInterrupted:
	637	+	if j.activeTurnID == record.TurnID {
	638	+	j.activeTurnID = ""
	639	+	}
	640	+	}
584	641		return writeResidentMetadata(j.dir, j.metadata(spec, state, record.Time))
585	642		}
586	643

@@ -617,7 +674,9 @@ func (j *ResidentJournal) appendSync(record residentRecord) error {
617	674		if err != nil {
618	675		return fmt.Errorf("resident journal encode: %w", err)
619	676		}
620		-	if len(data) > residentMaxRecordBytes {
	677	+	if len(data)+1 > residentMaxRecordBytes {
	678	+	// Readers reject the whole record line, newline included, so the writer
	679	+	// applies the same bound instead of emitting a line it cannot read back.
621	680		return errors.New("resident journal: record too large")
622	681		}
623	682		data = append(data, '\n')

@@ -912,6 +971,32 @@ func reconcileOwnedResidentJournal(journal *ResidentJournal) (ResidentMetadata,
912	971		return ResidentMetadata{}, errors.New("resident journal: duplicate tool result")
913	972		}
914	973		toolResults[record.ToolID] = struct{}{}
	974	+	case residentRecordCompacted:
	975	+	checkpoint, err := decodeResidentCompactedMessages(record.Messages)
	976	+	if err != nil {
	977	+	return ResidentMetadata{}, fmt.Errorf("resident journal: invalid compaction checkpoint: %w", err)
	978	+	}
	979	+	// The checkpoint replaces everything before it, so tool bookkeeping
	980	+	// restarts from the compacted transcript. Without this, a call that
	981	+	// compaction dropped would look dangling and gain a synthetic
	982	+	// interrupted result appended after the checkpoint, which resume
	983	+	// replay would then surface as an orphan.
	984	+	toolCalls = make(map[string]struct{})
	985	+	toolResults = make(map[string]struct{})
	986	+	for _, message := range checkpoint {
	987	+	for _, content := range message.Content {
	988	+	switch block := content.(type) {
	989	+	case provider.ToolCallBlock:
	990	+	if block.ID != "" {
	991	+	toolCalls[block.ID] = struct{}{}
	992	+	}
	993	+	case provider.ToolResultBlock:
	994	+	if block.CallID != "" {
	995	+	toolResults[block.CallID] = struct{}{}
	996	+	}
	997	+	}
	998	+	}
	999	+	}
915	1000		case residentRecordUsage:
916	1001		if record.Usage == nil {
917	1002		return ResidentMetadata{}, errors.New("resident journal: invalid usage record")

@@ -928,18 +1013,19 @@ func reconcileOwnedResidentJournal(journal *ResidentJournal) (ResidentMetadata,
928	1013		Usage: usage.Usage, ContextUsed: usage.ContextUsed, ContextMax: usage.ContextMax, Subscription: usage.Subscription,
929	1014		}
930	1015		needsInterruption := state == ResidentQueued \|\| state == ResidentRunning
931		-	needsToolRepair := len(toolCalls) != len(toolResults)
932		-	if needsInterruption \|\| needsToolRepair {
	1016	+	// Dangling calls, not a tool-count mismatch, are the repair trigger: a
	1017	+	// checkpoint may legitimately carry a result without repeating its call.
	1018	+	dangling := make([]string, 0, len(toolCalls))
	1019	+	for toolID := range toolCalls {
	1020	+	if _, finished := toolResults[toolID]; !finished {
	1021	+	dangling = append(dangling, toolID)
	1022	+	}
	1023	+	}
	1024	+	sort.Strings(dangling)
	1025	+	if needsInterruption \|\| len(dangling) > 0 {
933	1026		// A recorded repair is new observable work. A no-op reconciliation,
934	1027		// however, must retain the last durable lifecycle time.
935	1028		metadata.UpdatedAt = time.Now().UTC()
936		-	dangling := make([]string, 0, len(toolCalls))
937		-	for toolID := range toolCalls {
938		-	if _, finished := toolResults[toolID]; !finished {
939		-	dangling = append(dangling, toolID)
940		-	}
941		-	}
942		-	sort.Strings(dangling)
943	1029		for _, toolID := range dangling {
944	1030		result, marshalErr := json.Marshal(core.ToolResult{IsError: true, Content: []provider.Content{provider.TextBlock{Text: residentInterruptedText}}})
945	1031		if marshalErr != nil {


@@ -36,6 +36,7 @@ const (
36		residentRecordAssistant = "message.assistant"	36		residentRecordAssistant = "message.assistant"
37		residentRecordToolCall = "tool.call"	37		residentRecordToolCall = "tool.call"
38		residentRecordToolResult = "tool.result"	38		residentRecordToolResult = "tool.result"
			39	+	residentRecordCompacted = "child.compacted"
39		residentRecordUsage = "usage"	40		residentRecordUsage = "usage"
40		residentJournalVersion = 3	41		residentJournalVersion = 3
41		residentJournalVersionV2 = 2	42		residentJournalVersionV2 = 2

@@ -121,15 +122,19 @@ type residentRecord struct {
121		PatchRef string `json:"patch_ref,omitempty"`	122		PatchRef string `json:"patch_ref,omitempty"`
122		ChangedFiles []string `json:"changed_files,omitempty"`	123		ChangedFiles []string `json:"changed_files,omitempty"`
123		Message json.RawMessage `json:"message,omitempty"`	124		Message json.RawMessage `json:"message,omitempty"`
124	-	ToolID string `json:"tool_id,omitempty"`	125	+	// Messages carries a full replacement transcript. It is used only by
125	-	ToolName string `json:"tool_name,omitempty"`	126	+	// child.compacted, the durable checkpoint written after an overflow
126	-	ToolArgs json.RawMessage `json:"tool_args,omitempty"`	127	+	// recovery compacts the child transcript.
127	-	ToolResult json.RawMessage `json:"tool_result,omitempty"`	128	+	Messages []json.RawMessage `json:"messages,omitempty"`
128	-	Usage *provider.Usage `json:"usage,omitempty"`	129	+	ToolID string `json:"tool_id,omitempty"`
129	-	ContextUsed int `json:"context_used,omitempty"`	130	+	ToolName string `json:"tool_name,omitempty"`
130	-	ContextMax int `json:"context_max,omitempty"`	131	+	ToolArgs json.RawMessage `json:"tool_args,omitempty"`
131	-	Subscription bool `json:"subscription,omitempty"`	132	+	ToolResult json.RawMessage `json:"tool_result,omitempty"`
132	-	raw json.RawMessage `json:"-"`	133	+	Usage *provider.Usage `json:"usage,omitempty"`
			134	+	ContextUsed int `json:"context_used,omitempty"`
			135	+	ContextMax int `json:"context_max,omitempty"`
			136	+	Subscription bool `json:"subscription,omitempty"`
			137	+	raw json.RawMessage `json:"-"`
133		}	138		}
134			139
135		// RecordAgentEvent persists finalized provider-neutral transcript events. It	140		// RecordAgentEvent persists finalized provider-neutral transcript events. It

@@ -231,6 +236,40 @@ func (j *ResidentJournal) RecordAgentEvent(event core.AgentEvent) error {
231		return err	236		return err
232		}	237		}
233			238
			239	+	// RecordCompacted persists the replacement transcript produced by a
			240	+	// context-overflow recovery. The checkpoint is additive: journals without one
			241	+	// replay exactly as before, and resume replay starts from the newest
			242	+	// checkpoint instead of the full append-only log.
			243	+	func (j *ResidentJournal) RecordCompacted(messages []provider.Message) error {
			244	+	if j == nil {
			245	+	return errors.New("resident journal: unavailable")
			246	+	}
			247	+	if len(messages) == 0 {
			248	+	return errors.New("resident journal: empty compaction checkpoint")
			249	+	}
			250	+	encoded := make([]json.RawMessage, 0, len(messages))
			251	+	for _, message := range messages {
			252	+	data, err := json.Marshal(message)
			253	+	if err != nil {
			254	+	return fmt.Errorf("resident journal encode compacted message: %w", err)
			255	+	}
			256	+	encoded = append(encoded, data)
			257	+	}
			258	+	j.mu.Lock()
			259	+	defer j.mu.Unlock()
			260	+	if j.file == nil {
			261	+	return errors.New("resident journal: closed")
			262	+	}
			263	+	record := residentRecord{
			264	+	Version: residentJournalVersion,
			265	+	Type: residentRecordCompacted,
			266	+	Time: time.Now().UTC(),
			267	+	TurnID: j.activeTurnID,
			268	+	Messages: encoded,
			269	+	}
			270	+	return j.appendSync(record)
			271	+	}
			272	+
234		// ResidentMetadata is a bounded, rebuildable projection. The transcript is	273		// ResidentMetadata is a bounded, rebuildable projection. The transcript is
235		// the authority for acceptance; callers must never infer acceptance solely	274		// the authority for acceptance; callers must never infer acceptance solely
236		// from this file.	275		// from this file.

@@ -285,6 +324,12 @@ type ResidentJournal struct {
285		dir string	324		dir string
286		file *os.File	325		file *os.File
287		lease *residentLease	326		lease *residentLease
			327	+	// activeTurnID is the resident turn currently executing. It attributes
			328	+	// records that the runner writes without a turn parameter, such as a
			329	+	// compaction checkpoint, and is diagnostics only: no reader consumes a
			330	+	// checkpoint's turn ID. RecordTurnStarted owns the value and a terminal
			331	+	// boundary clears it. It is guarded by mu.
			332	+	activeTurnID string
288		}	333		}
289			334
290		// ResidentUsageSnapshot is the bounded usage projection shared by durable and	335		// ResidentUsageSnapshot is the bounded usage projection shared by durable and

@@ -581,6 +626,18 @@ func (j *ResidentJournal) recordTurnBoundary(spec ResidentChildSpec, state Resid
581		if err := j.appendSync(record); err != nil {	626		if err := j.appendSync(record); err != nil {
582		return err	627		return err
583		}	628		}
			629	+	// Track the accepted turn that is currently executing so records written
			630	+	// without an explicit turn parameter, such as a compaction checkpoint, stay
			631	+	// attributable. A terminal boundary clears it; a late checkpoint from a
			632	+	// finished turn can therefore not be attributed to the next one.
			633	+	switch record.Type {
			634	+	case residentRecordTurnStarted:
			635	+	j.activeTurnID = record.TurnID
			636	+	case residentRecordTurnFinished, residentRecordInterrupted:
			637	+	if j.activeTurnID == record.TurnID {
			638	+	j.activeTurnID = ""
			639	+	}
			640	+	}
584		return writeResidentMetadata(j.dir, j.metadata(spec, state, record.Time))	641		return writeResidentMetadata(j.dir, j.metadata(spec, state, record.Time))
585		}	642		}
586			643

@@ -617,7 +674,9 @@ func (j *ResidentJournal) appendSync(record residentRecord) error {
617		if err != nil {	674		if err != nil {
618		return fmt.Errorf("resident journal encode: %w", err)	675		return fmt.Errorf("resident journal encode: %w", err)
619		}	676		}
620	-	if len(data) > residentMaxRecordBytes {	677	+	if len(data)+1 > residentMaxRecordBytes {
			678	+	// Readers reject the whole record line, newline included, so the writer
			679	+	// applies the same bound instead of emitting a line it cannot read back.
621		return errors.New("resident journal: record too large")	680		return errors.New("resident journal: record too large")
622		}	681		}
623		data = append(data, '\n')	682		data = append(data, '\n')

@@ -912,6 +971,32 @@ func reconcileOwnedResidentJournal(journal *ResidentJournal) (ResidentMetadata,
912		return ResidentMetadata{}, errors.New("resident journal: duplicate tool result")	971		return ResidentMetadata{}, errors.New("resident journal: duplicate tool result")
913		}	972		}
914		toolResults[record.ToolID] = struct{}{}	973		toolResults[record.ToolID] = struct{}{}
			974	+	case residentRecordCompacted:
			975	+	checkpoint, err := decodeResidentCompactedMessages(record.Messages)
			976	+	if err != nil {
			977	+	return ResidentMetadata{}, fmt.Errorf("resident journal: invalid compaction checkpoint: %w", err)
			978	+	}
			979	+	// The checkpoint replaces everything before it, so tool bookkeeping
			980	+	// restarts from the compacted transcript. Without this, a call that
			981	+	// compaction dropped would look dangling and gain a synthetic
			982	+	// interrupted result appended after the checkpoint, which resume
			983	+	// replay would then surface as an orphan.
			984	+	toolCalls = make(map[string]struct{})
			985	+	toolResults = make(map[string]struct{})
			986	+	for _, message := range checkpoint {
			987	+	for _, content := range message.Content {
			988	+	switch block := content.(type) {
			989	+	case provider.ToolCallBlock:
			990	+	if block.ID != "" {
			991	+	toolCalls[block.ID] = struct{}{}
			992	+	}
			993	+	case provider.ToolResultBlock:
			994	+	if block.CallID != "" {
			995	+	toolResults[block.CallID] = struct{}{}
			996	+	}
			997	+	}
			998	+	}
			999	+	}
915		case residentRecordUsage:	1000		case residentRecordUsage:
916		if record.Usage == nil {	1001		if record.Usage == nil {
917		return ResidentMetadata{}, errors.New("resident journal: invalid usage record")	1002		return ResidentMetadata{}, errors.New("resident journal: invalid usage record")

@@ -928,18 +1013,19 @@ func reconcileOwnedResidentJournal(journal *ResidentJournal) (ResidentMetadata,
928		Usage: usage.Usage, ContextUsed: usage.ContextUsed, ContextMax: usage.ContextMax, Subscription: usage.Subscription,	1013		Usage: usage.Usage, ContextUsed: usage.ContextUsed, ContextMax: usage.ContextMax, Subscription: usage.Subscription,
929		}	1014		}
930		needsInterruption := state == ResidentQueued \|\| state == ResidentRunning	1015		needsInterruption := state == ResidentQueued \|\| state == ResidentRunning
931	-	needsToolRepair := len(toolCalls) != len(toolResults)	1016	+	// Dangling calls, not a tool-count mismatch, are the repair trigger: a
932	-	if needsInterruption \|\| needsToolRepair {	1017	+	// checkpoint may legitimately carry a result without repeating its call.
			1018	+	dangling := make([]string, 0, len(toolCalls))
			1019	+	for toolID := range toolCalls {
			1020	+	if _, finished := toolResults[toolID]; !finished {
			1021	+	dangling = append(dangling, toolID)
			1022	+	}
			1023	+	}
			1024	+	sort.Strings(dangling)
			1025	+	if needsInterruption \|\| len(dangling) > 0 {
933		// A recorded repair is new observable work. A no-op reconciliation,	1026		// A recorded repair is new observable work. A no-op reconciliation,
934		// however, must retain the last durable lifecycle time.	1027		// however, must retain the last durable lifecycle time.
935		metadata.UpdatedAt = time.Now().UTC()	1028		metadata.UpdatedAt = time.Now().UTC()
936	-	dangling := make([]string, 0, len(toolCalls))
937	-	for toolID := range toolCalls {
938	-	if _, finished := toolResults[toolID]; !finished {
939	-	dangling = append(dangling, toolID)
940	-	}
941	-	}
942	-	sort.Strings(dangling)
943		for _, toolID := range dangling {	1029		for _, toolID := range dangling {
944		result, marshalErr := json.Marshal(core.ToolResult{IsError: true, Content: []provider.Content{provider.TextBlock{Text: residentInterruptedText}}})	1030		result, marshalErr := json.Marshal(core.ToolResult{IsError: true, Content: []provider.Content{provider.TextBlock{Text: residentInterruptedText}}})
945		if marshalErr != nil {	1031		if marshalErr != nil {

M packages/agent/subagents/resident_journal_test.go +334 -0


@@ -1055,3 +1055,337 @@ func TestReconcileResidentJournalRejectsConflictingToolPairs(t *testing.T) {
1055	1055		})
1056	1056		}
1057	1057		}
	1058	+
	1059	+	func TestRecordCompactedRejectsOversizedCheckpoint(t *testing.T) {
	1060	+	root := t.TempDir()
	1061	+	journal, err := OpenResidentJournal(root, "oversized")
	1062	+	if err != nil {
	1063	+	t.Fatal(err)
	1064	+	}
	1065	+	spec := ResidentChildSpec{ID: "oversized", SessionID: "session", Provider: "openai", Model: "gpt-5"}
	1066	+	if err := journal.Accept(spec, "task"); err != nil {
	1067	+	t.Fatal(err)
	1068	+	}
	1069	+	if err := journal.RecordAgentEvent(core.EvUserMessage{Message: provider.Message{Role: provider.RoleUser, Content: []provider.Content{provider.TextBlock{Text: "question"}}}}); err != nil {
	1070	+	t.Fatal(err)
	1071	+	}
	1072	+	oversized := []provider.Message{{
	1073	+	Role: provider.RoleUser,
	1074	+	Content: []provider.Content{provider.TextBlock{Text: strings.Repeat("x", residentMaxRecordBytes)}},
	1075	+	}}
	1076	+	if err := journal.RecordCompacted(oversized); err == nil \|\| !strings.Contains(err.Error(), "too large") {
	1077	+	t.Fatalf("RecordCompacted error = %v, want oversize rejection", err)
	1078	+	}
	1079	+	transcript := filepath.Join(journal.Dir(), residentTranscriptName)
	1080	+	if err := journal.Close(); err != nil {
	1081	+	t.Fatal(err)
	1082	+	}
	1083	+	records, err := ReadResidentJournal(transcript)
	1084	+	if err != nil {
	1085	+	t.Fatalf("journal unreadable after an oversized checkpoint: %v", err)
	1086	+	}
	1087	+	for _, record := range records {
	1088	+	if record.Type == residentRecordCompacted {
	1089	+	t.Fatal("oversized checkpoint was appended")
	1090	+	}
	1091	+	}
	1092	+	if _, err := ReadResidentTranscriptMessages(filepath.Join(root, spec.ID)); err != nil {
	1093	+	t.Fatalf("transcript unreadable after an oversized checkpoint: %v", err)
	1094	+	}
	1095	+	}
	1096	+
	1097	+	func TestReconcileResidentJournalKeepsToolBookkeepingAcrossCheckpoint(t *testing.T) {
	1098	+	root := t.TempDir()
	1099	+	journal, err := OpenResidentJournal(root, "checkpoint-repair")
	1100	+	if err != nil {
	1101	+	t.Fatal(err)
	1102	+	}
	1103	+	spec := ResidentChildSpec{ID: "checkpoint-repair", SessionID: "session", Provider: "openai", Model: "gpt-5"}
	1104	+	if err := journal.Accept(spec, "task"); err != nil {
	1105	+	t.Fatal(err)
	1106	+	}
	1107	+	if err := journal.RecordAgentEvent(core.EvUserMessage{Message: provider.Message{Role: provider.RoleUser, Content: []provider.Content{provider.TextBlock{Text: "question"}}}}); err != nil {
	1108	+	t.Fatal(err)
	1109	+	}
	1110	+	// The call never produced a result and compaction dropped it from the
	1111	+	// replacement transcript.
	1112	+	if err := journal.RecordAgentEvent(core.EvToolCall{ID: "dropped-call", Name: "bash", Args: json.RawMessage(`{"command":"pwd"}`)}); err != nil {
	1113	+	t.Fatal(err)
	1114	+	}
	1115	+	if err := journal.AcceptFollowUp(spec, "turn-1", "follow up"); err != nil {
	1116	+	t.Fatal(err)
	1117	+	}
	1118	+	if err := journal.RecordTurnStarted(spec, "turn-1"); err != nil {
	1119	+	t.Fatal(err)
	1120	+	}
	1121	+	if err := journal.RecordCompacted([]provider.Message{
	1122	+	{Role: provider.RoleUser, Content: []provider.Content{provider.TextBlock{Text: "## Context Summary (compacted)\n\ndigest"}}},
	1123	+	}); err != nil {
	1124	+	t.Fatal(err)
	1125	+	}
	1126	+	if err := journal.RecordTurnFinished(spec, "turn-1", nil); err != nil {
	1127	+	t.Fatal(err)
	1128	+	}
	1129	+	transcript := filepath.Join(journal.Dir(), residentTranscriptName)
	1130	+	if err := journal.Close(); err != nil {
	1131	+	t.Fatal(err)
	1132	+	}
	1133	+
	1134	+	metadata, err := ReconcileResidentJournal(filepath.Join(root, spec.ID))
	1135	+	if err != nil {
	1136	+	t.Fatal(err)
	1137	+	}
	1138	+	if metadata.State != ResidentIdle {
	1139	+	t.Fatalf("reconciled state = %q, want idle", metadata.State)
	1140	+	}
	1141	+	records, err := ReadResidentJournal(transcript)
	1142	+	if err != nil {
	1143	+	t.Fatal(err)
	1144	+	}
	1145	+	for _, record := range records {
	1146	+	if record.Type == residentRecordToolResult {
	1147	+	t.Fatalf("reconciliation repaired a call dropped by compaction: %#v", record)
	1148	+	}
	1149	+	}
	1150	+	messages, err := ReadResidentTranscriptMessages(filepath.Join(root, spec.ID))
	1151	+	if err != nil {
	1152	+	t.Fatal(err)
	1153	+	}
	1154	+	if len(messages) != 1 \|\| messages[0].Role != provider.RoleUser {
	1155	+	t.Fatalf("resumed messages = %#v, want only the checkpoint summary", messages)
	1156	+	}
	1157	+	}
	1158	+
	1159	+	func TestRecordCompactedAttributesTheStartedTurn(t *testing.T) {
	1160	+	root := t.TempDir()
	1161	+	journal, err := OpenResidentJournal(root, "turn-attribution")
	1162	+	if err != nil {
	1163	+	t.Fatal(err)
	1164	+	}
	1165	+	spec := ResidentChildSpec{ID: "turn-attribution", SessionID: "session", Provider: "openai", Model: "gpt-5"}
	1166	+	if err := journal.Accept(spec, "task"); err != nil {
	1167	+	t.Fatal(err)
	1168	+	}
	1169	+	if err := journal.RecordTurnStarted(spec, "turn-1"); err != nil {
	1170	+	t.Fatal(err)
	1171	+	}
	1172	+	// A follow-up accepted while turn-1 is still running must not steal the
	1173	+	// checkpoint's turn attribution.
	1174	+	if err := journal.AcceptFollowUp(spec, "turn-2", "queued follow up"); err != nil {
	1175	+	t.Fatal(err)
	1176	+	}
	1177	+	if err := journal.RecordCompacted([]provider.Message{
	1178	+	{Role: provider.RoleUser, Content: []provider.Content{provider.TextBlock{Text: "## Context Summary (compacted)\n\ndigest"}}},
	1179	+	}); err != nil {
	1180	+	t.Fatal(err)
	1181	+	}
	1182	+	transcript := filepath.Join(journal.Dir(), residentTranscriptName)
	1183	+	if err := journal.Close(); err != nil {
	1184	+	t.Fatal(err)
	1185	+	}
	1186	+	records, err := ReadResidentJournal(transcript)
	1187	+	if err != nil {
	1188	+	t.Fatal(err)
	1189	+	}
	1190	+	for _, record := range records {
	1191	+	if record.Type != residentRecordCompacted {
	1192	+	continue
	1193	+	}
	1194	+	if record.TurnID != "turn-1" {
	1195	+	t.Fatalf("checkpoint turn ID = %q, want the executing turn", record.TurnID)
	1196	+	}
	1197	+	return
	1198	+	}
	1199	+	t.Fatal("journal has no compaction checkpoint")
	1200	+	}
	1201	+
	1202	+	func TestReconcileResidentJournalIsIdempotentForCheckpointResults(t *testing.T) {
	1203	+	root := t.TempDir()
	1204	+	journal, err := OpenResidentJournal(root, "idempotent-checkpoint")
	1205	+	if err != nil {
	1206	+	t.Fatal(err)
	1207	+	}
	1208	+	spec := ResidentChildSpec{ID: "idempotent-checkpoint", SessionID: "session", Provider: "openai", Model: "gpt-5"}
	1209	+	if err := journal.Accept(spec, "task"); err != nil {
	1210	+	t.Fatal(err)
	1211	+	}
	1212	+	if err := journal.AcceptFollowUp(spec, "turn-1", "follow up"); err != nil {
	1213	+	t.Fatal(err)
	1214	+	}
	1215	+	if err := journal.RecordTurnStarted(spec, "turn-1"); err != nil {
	1216	+	t.Fatal(err)
	1217	+	}
	1218	+	// A checkpoint can carry a result whose call was compacted away. That is
	1219	+	// not repairable work, so reconciliation must stay a no-op.
	1220	+	if err := journal.RecordCompacted([]provider.Message{{
	1221	+	Role: provider.RoleTool,
	1222	+	Content: []provider.Content{provider.ToolResultBlock{
	1223	+	CallID: "compacted-call",
	1224	+	Content: []provider.Content{provider.TextBlock{Text: "kept result"}},
	1225	+	}},
	1226	+	}}); err != nil {
	1227	+	t.Fatal(err)
	1228	+	}
	1229	+	if err := journal.RecordTurnFinished(spec, "turn-1", nil); err != nil {
	1230	+	t.Fatal(err)
	1231	+	}
	1232	+	if err := journal.Close(); err != nil {
	1233	+	t.Fatal(err)
	1234	+	}
	1235	+	first, err := ReconcileResidentJournal(filepath.Join(root, spec.ID))
	1236	+	if err != nil {
	1237	+	t.Fatal(err)
	1238	+	}
	1239	+	second, err := ReconcileResidentJournal(filepath.Join(root, spec.ID))
	1240	+	if err != nil {
	1241	+	t.Fatal(err)
	1242	+	}
	1243	+	if !first.UpdatedAt.Equal(second.UpdatedAt) {
	1244	+	t.Fatalf("reconciliation rewrote the lifecycle time: %v then %v", first.UpdatedAt, second.UpdatedAt)
	1245	+	}
	1246	+	if second.State != ResidentIdle {
	1247	+	t.Fatalf("reconciled state = %q, want idle", second.State)
	1248	+	}
	1249	+	}
	1250	+
	1251	+	func TestAppendSyncEnforcesTheReaderRecordBound(t *testing.T) {
	1252	+	root := t.TempDir()
	1253	+	journal, err := OpenResidentJournal(root, "record-bound")
	1254	+	if err != nil {
	1255	+	t.Fatal(err)
	1256	+	}
	1257	+	spec := ResidentChildSpec{ID: "record-bound", SessionID: "session", Provider: "openai", Model: "gpt-5"}
	1258	+	if err := journal.Accept(spec, "task"); err != nil {
	1259	+	t.Fatal(err)
	1260	+	}
	1261	+	_, base := residentAssistantRecordSize(t, "")
	1262	+	if base >= residentMaxRecordBytes-1 {
	1263	+	t.Fatalf("empty assistant record = %d bytes, too large for a boundary check", base)
	1264	+	}
	1265	+	// A JSON record of max-1 bytes produces a max-byte line with its newline,
	1266	+	// which is exactly what the readers accept.
	1267	+	padding := residentMaxRecordBytes - 1 - base
	1268	+	borderline, borderlineSize := residentAssistantRecordSize(t, strings.Repeat("x", padding))
	1269	+	if borderlineSize != residentMaxRecordBytes-1 {
	1270	+	t.Fatalf("borderline record = %d bytes, want %d", borderlineSize, residentMaxRecordBytes-1)
	1271	+	}
	1272	+	if err := journal.appendSync(borderline); err != nil {
	1273	+	t.Fatalf("appendSync rejected a reader-legal record: %v", err)
	1274	+	}
	1275	+	oversized, oversizedSize := residentAssistantRecordSize(t, strings.Repeat("x", padding+1))
	1276	+	if oversizedSize != residentMaxRecordBytes {
	1277	+	t.Fatalf("oversized record = %d bytes, want %d", oversizedSize, residentMaxRecordBytes)
	1278	+	}
	1279	+	if err := journal.appendSync(oversized); err == nil \|\| !strings.Contains(err.Error(), "too large") {
	1280	+	t.Fatalf("appendSync error = %v, want the reader record bound", err)
	1281	+	}
	1282	+	transcript := filepath.Join(journal.Dir(), residentTranscriptName)
	1283	+	if err := journal.Close(); err != nil {
	1284	+	t.Fatal(err)
	1285	+	}
	1286	+	records, err := ReadResidentJournal(transcript)
	1287	+	if err != nil {
	1288	+	t.Fatalf("journal unreadable after a boundary record: %v", err)
	1289	+	}
	1290	+	if len(records) != 2 {
	1291	+	t.Fatalf("records = %d, want the accepted child and the borderline message", len(records))
	1292	+	}
	1293	+	}
	1294	+
	1295	+	func residentAssistantRecordSize(t *testing.T, text string) (residentRecord, int) {
	1296	+	t.Helper()
	1297	+	message, err := json.Marshal(provider.Message{
	1298	+	Role: provider.RoleAssistant,
	1299	+	Content: []provider.Content{provider.TextBlock{Text: text}},
	1300	+	})
	1301	+	if err != nil {
	1302	+	t.Fatal(err)
	1303	+	}
	1304	+	record := residentRecord{Version: residentJournalVersion, Type: residentRecordAssistant, Message: message}
	1305	+	data, err := json.Marshal(record)
	1306	+	if err != nil {
	1307	+	t.Fatal(err)
	1308	+	}
	1309	+	return record, len(data)
	1310	+	}
	1311	+
	1312	+	func TestReconcileResidentJournalRepairsDanglingCheckpointCall(t *testing.T) {
	1313	+	root := t.TempDir()
	1314	+	journal, err := OpenResidentJournal(root, "checkpoint-dangling")
	1315	+	if err != nil {
	1316	+	t.Fatal(err)
	1317	+	}
	1318	+	spec := ResidentChildSpec{ID: "checkpoint-dangling", SessionID: "session", Provider: "openai", Model: "gpt-5"}
	1319	+	if err := journal.Accept(spec, "task"); err != nil {
	1320	+	t.Fatal(err)
	1321	+	}
	1322	+	if err := journal.AcceptFollowUp(spec, "turn-1", "follow up"); err != nil {
	1323	+	t.Fatal(err)
	1324	+	}
	1325	+	if err := journal.RecordTurnStarted(spec, "turn-1"); err != nil {
	1326	+	t.Fatal(err)
	1327	+	}
	1328	+	// The checkpoint kept a call whose result never arrived.
	1329	+	if err := journal.RecordCompacted([]provider.Message{{
	1330	+	Role: provider.RoleAssistant,
	1331	+	Content: []provider.Content{provider.ToolCallBlock{
	1332	+	ID: "compacted-call",
	1333	+	Name: "bash",
	1334	+	Arguments: json.RawMessage(`{"command":"pwd"}`),
	1335	+	}},
	1336	+	}}); err != nil {
	1337	+	t.Fatal(err)
	1338	+	}
	1339	+	if err := journal.RecordTurnFinished(spec, "turn-1", nil); err != nil {
	1340	+	t.Fatal(err)
	1341	+	}
	1342	+	transcript := filepath.Join(journal.Dir(), residentTranscriptName)
	1343	+	if err := journal.Close(); err != nil {
	1344	+	t.Fatal(err)
	1345	+	}
	1346	+	dir := filepath.Join(root, spec.ID)
	1347	+	if _, err := ReconcileResidentJournal(dir); err != nil {
	1348	+	t.Fatal(err)
	1349	+	}
	1350	+	records, err := ReadResidentJournal(transcript)
	1351	+	if err != nil {
	1352	+	t.Fatal(err)
	1353	+	}
	1354	+	repaired := 0
	1355	+	for _, record := range records {
	1356	+	if record.Type != residentRecordToolResult \|\| record.ToolID != "compacted-call" {
	1357	+	continue
	1358	+	}
	1359	+	if !isSyntheticInterruption(record.ToolResult) {
	1360	+	t.Fatalf("repaired result = %s, want the interruption result", record.ToolResult)
	1361	+	}
	1362	+	repaired++
	1363	+	}
	1364	+	if repaired != 1 {
	1365	+	t.Fatalf("repaired results = %d, want exactly one synthetic result", repaired)
	1366	+	}
	1367	+	// Once the repair is durable, later reconciliations must be stable no-ops.
	1368	+	second, err := ReconcileResidentJournal(dir)
	1369	+	if err != nil {
	1370	+	t.Fatal(err)
	1371	+	}
	1372	+	third, err := ReconcileResidentJournal(dir)
	1373	+	if err != nil {
	1374	+	t.Fatal(err)
	1375	+	}
	1376	+	if !second.UpdatedAt.Equal(third.UpdatedAt) {
	1377	+	t.Fatalf("reconciliation kept rewriting the lifecycle time: %v then %v", second.UpdatedAt, third.UpdatedAt)
	1378	+	}
	1379	+	if records, err = ReadResidentJournal(transcript); err != nil {
	1380	+	t.Fatal(err)
	1381	+	}
	1382	+	results := 0
	1383	+	for _, record := range records {
	1384	+	if record.Type == residentRecordToolResult && record.ToolID == "compacted-call" {
	1385	+	results++
	1386	+	}
	1387	+	}
	1388	+	if results != 1 {
	1389	+	t.Fatalf("tool results = %d, want the single repaired result", results)
	1390	+	}
	1391	+	}


@@ -1055,3 +1055,337 @@ func TestReconcileResidentJournalRejectsConflictingToolPairs(t *testing.T) {
1055	})	1055		})
1056	}	1056		}
1057	}	1057		}
		1058	+
		1059	+	func TestRecordCompactedRejectsOversizedCheckpoint(t *testing.T) {
		1060	+	root := t.TempDir()
		1061	+	journal, err := OpenResidentJournal(root, "oversized")
		1062	+	if err != nil {
		1063	+	t.Fatal(err)
		1064	+	}
		1065	+	spec := ResidentChildSpec{ID: "oversized", SessionID: "session", Provider: "openai", Model: "gpt-5"}
		1066	+	if err := journal.Accept(spec, "task"); err != nil {
		1067	+	t.Fatal(err)
		1068	+	}
		1069	+	if err := journal.RecordAgentEvent(core.EvUserMessage{Message: provider.Message{Role: provider.RoleUser, Content: []provider.Content{provider.TextBlock{Text: "question"}}}}); err != nil {
		1070	+	t.Fatal(err)
		1071	+	}
		1072	+	oversized := []provider.Message{{
		1073	+	Role: provider.RoleUser,
		1074	+	Content: []provider.Content{provider.TextBlock{Text: strings.Repeat("x", residentMaxRecordBytes)}},
		1075	+	}}
		1076	+	if err := journal.RecordCompacted(oversized); err == nil \|\| !strings.Contains(err.Error(), "too large") {
		1077	+	t.Fatalf("RecordCompacted error = %v, want oversize rejection", err)
		1078	+	}
		1079	+	transcript := filepath.Join(journal.Dir(), residentTranscriptName)
		1080	+	if err := journal.Close(); err != nil {
		1081	+	t.Fatal(err)
		1082	+	}
		1083	+	records, err := ReadResidentJournal(transcript)
		1084	+	if err != nil {
		1085	+	t.Fatalf("journal unreadable after an oversized checkpoint: %v", err)
		1086	+	}
		1087	+	for _, record := range records {
		1088	+	if record.Type == residentRecordCompacted {
		1089	+	t.Fatal("oversized checkpoint was appended")
		1090	+	}
		1091	+	}
		1092	+	if _, err := ReadResidentTranscriptMessages(filepath.Join(root, spec.ID)); err != nil {
		1093	+	t.Fatalf("transcript unreadable after an oversized checkpoint: %v", err)
		1094	+	}
		1095	+	}
		1096	+
		1097	+	func TestReconcileResidentJournalKeepsToolBookkeepingAcrossCheckpoint(t *testing.T) {
		1098	+	root := t.TempDir()
		1099	+	journal, err := OpenResidentJournal(root, "checkpoint-repair")
		1100	+	if err != nil {
		1101	+	t.Fatal(err)
		1102	+	}
		1103	+	spec := ResidentChildSpec{ID: "checkpoint-repair", SessionID: "session", Provider: "openai", Model: "gpt-5"}
		1104	+	if err := journal.Accept(spec, "task"); err != nil {
		1105	+	t.Fatal(err)
		1106	+	}
		1107	+	if err := journal.RecordAgentEvent(core.EvUserMessage{Message: provider.Message{Role: provider.RoleUser, Content: []provider.Content{provider.TextBlock{Text: "question"}}}}); err != nil {
		1108	+	t.Fatal(err)
		1109	+	}
		1110	+	// The call never produced a result and compaction dropped it from the
		1111	+	// replacement transcript.
		1112	+	if err := journal.RecordAgentEvent(core.EvToolCall{ID: "dropped-call", Name: "bash", Args: json.RawMessage(`{"command":"pwd"}`)}); err != nil {
		1113	+	t.Fatal(err)
		1114	+	}
		1115	+	if err := journal.AcceptFollowUp(spec, "turn-1", "follow up"); err != nil {
		1116	+	t.Fatal(err)
		1117	+	}
		1118	+	if err := journal.RecordTurnStarted(spec, "turn-1"); err != nil {
		1119	+	t.Fatal(err)
		1120	+	}
		1121	+	if err := journal.RecordCompacted([]provider.Message{
		1122	+	{Role: provider.RoleUser, Content: []provider.Content{provider.TextBlock{Text: "## Context Summary (compacted)\n\ndigest"}}},
		1123	+	}); err != nil {
		1124	+	t.Fatal(err)
		1125	+	}
		1126	+	if err := journal.RecordTurnFinished(spec, "turn-1", nil); err != nil {
		1127	+	t.Fatal(err)
		1128	+	}
		1129	+	transcript := filepath.Join(journal.Dir(), residentTranscriptName)
		1130	+	if err := journal.Close(); err != nil {
		1131	+	t.Fatal(err)
		1132	+	}
		1133	+
		1134	+	metadata, err := ReconcileResidentJournal(filepath.Join(root, spec.ID))
		1135	+	if err != nil {
		1136	+	t.Fatal(err)
		1137	+	}
		1138	+	if metadata.State != ResidentIdle {
		1139	+	t.Fatalf("reconciled state = %q, want idle", metadata.State)
		1140	+	}
		1141	+	records, err := ReadResidentJournal(transcript)
		1142	+	if err != nil {
		1143	+	t.Fatal(err)
		1144	+	}
		1145	+	for _, record := range records {
		1146	+	if record.Type == residentRecordToolResult {
		1147	+	t.Fatalf("reconciliation repaired a call dropped by compaction: %#v", record)
		1148	+	}
		1149	+	}
		1150	+	messages, err := ReadResidentTranscriptMessages(filepath.Join(root, spec.ID))
		1151	+	if err != nil {
		1152	+	t.Fatal(err)
		1153	+	}
		1154	+	if len(messages) != 1 \|\| messages[0].Role != provider.RoleUser {
		1155	+	t.Fatalf("resumed messages = %#v, want only the checkpoint summary", messages)
		1156	+	}
		1157	+	}
		1158	+
		1159	+	func TestRecordCompactedAttributesTheStartedTurn(t *testing.T) {
		1160	+	root := t.TempDir()
		1161	+	journal, err := OpenResidentJournal(root, "turn-attribution")
		1162	+	if err != nil {
		1163	+	t.Fatal(err)
		1164	+	}
		1165	+	spec := ResidentChildSpec{ID: "turn-attribution", SessionID: "session", Provider: "openai", Model: "gpt-5"}
		1166	+	if err := journal.Accept(spec, "task"); err != nil {
		1167	+	t.Fatal(err)
		1168	+	}
		1169	+	if err := journal.RecordTurnStarted(spec, "turn-1"); err != nil {
		1170	+	t.Fatal(err)
		1171	+	}
		1172	+	// A follow-up accepted while turn-1 is still running must not steal the
		1173	+	// checkpoint's turn attribution.
		1174	+	if err := journal.AcceptFollowUp(spec, "turn-2", "queued follow up"); err != nil {
		1175	+	t.Fatal(err)
		1176	+	}
		1177	+	if err := journal.RecordCompacted([]provider.Message{
		1178	+	{Role: provider.RoleUser, Content: []provider.Content{provider.TextBlock{Text: "## Context Summary (compacted)\n\ndigest"}}},
		1179	+	}); err != nil {
		1180	+	t.Fatal(err)
		1181	+	}
		1182	+	transcript := filepath.Join(journal.Dir(), residentTranscriptName)
		1183	+	if err := journal.Close(); err != nil {
		1184	+	t.Fatal(err)
		1185	+	}
		1186	+	records, err := ReadResidentJournal(transcript)
		1187	+	if err != nil {
		1188	+	t.Fatal(err)
		1189	+	}
		1190	+	for _, record := range records {
		1191	+	if record.Type != residentRecordCompacted {
		1192	+	continue
		1193	+	}
		1194	+	if record.TurnID != "turn-1" {
		1195	+	t.Fatalf("checkpoint turn ID = %q, want the executing turn", record.TurnID)
		1196	+	}
		1197	+	return
		1198	+	}
		1199	+	t.Fatal("journal has no compaction checkpoint")
		1200	+	}
		1201	+
		1202	+	func TestReconcileResidentJournalIsIdempotentForCheckpointResults(t *testing.T) {
		1203	+	root := t.TempDir()
		1204	+	journal, err := OpenResidentJournal(root, "idempotent-checkpoint")
		1205	+	if err != nil {
		1206	+	t.Fatal(err)
		1207	+	}
		1208	+	spec := ResidentChildSpec{ID: "idempotent-checkpoint", SessionID: "session", Provider: "openai", Model: "gpt-5"}
		1209	+	if err := journal.Accept(spec, "task"); err != nil {
		1210	+	t.Fatal(err)
		1211	+	}
		1212	+	if err := journal.AcceptFollowUp(spec, "turn-1", "follow up"); err != nil {
		1213	+	t.Fatal(err)
		1214	+	}
		1215	+	if err := journal.RecordTurnStarted(spec, "turn-1"); err != nil {
		1216	+	t.Fatal(err)
		1217	+	}
		1218	+	// A checkpoint can carry a result whose call was compacted away. That is
		1219	+	// not repairable work, so reconciliation must stay a no-op.
		1220	+	if err := journal.RecordCompacted([]provider.Message{{
		1221	+	Role: provider.RoleTool,
		1222	+	Content: []provider.Content{provider.ToolResultBlock{
		1223	+	CallID: "compacted-call",
		1224	+	Content: []provider.Content{provider.TextBlock{Text: "kept result"}},
		1225	+	}},
		1226	+	}}); err != nil {
		1227	+	t.Fatal(err)
		1228	+	}
		1229	+	if err := journal.RecordTurnFinished(spec, "turn-1", nil); err != nil {
		1230	+	t.Fatal(err)
		1231	+	}
		1232	+	if err := journal.Close(); err != nil {
		1233	+	t.Fatal(err)
		1234	+	}
		1235	+	first, err := ReconcileResidentJournal(filepath.Join(root, spec.ID))
		1236	+	if err != nil {
		1237	+	t.Fatal(err)
		1238	+	}
		1239	+	second, err := ReconcileResidentJournal(filepath.Join(root, spec.ID))
		1240	+	if err != nil {
		1241	+	t.Fatal(err)
		1242	+	}
		1243	+	if !first.UpdatedAt.Equal(second.UpdatedAt) {
		1244	+	t.Fatalf("reconciliation rewrote the lifecycle time: %v then %v", first.UpdatedAt, second.UpdatedAt)
		1245	+	}
		1246	+	if second.State != ResidentIdle {
		1247	+	t.Fatalf("reconciled state = %q, want idle", second.State)
		1248	+	}
		1249	+	}
		1250	+
		1251	+	func TestAppendSyncEnforcesTheReaderRecordBound(t *testing.T) {
		1252	+	root := t.TempDir()
		1253	+	journal, err := OpenResidentJournal(root, "record-bound")
		1254	+	if err != nil {
		1255	+	t.Fatal(err)
		1256	+	}
		1257	+	spec := ResidentChildSpec{ID: "record-bound", SessionID: "session", Provider: "openai", Model: "gpt-5"}
		1258	+	if err := journal.Accept(spec, "task"); err != nil {
		1259	+	t.Fatal(err)
		1260	+	}
		1261	+	_, base := residentAssistantRecordSize(t, "")
		1262	+	if base >= residentMaxRecordBytes-1 {
		1263	+	t.Fatalf("empty assistant record = %d bytes, too large for a boundary check", base)
		1264	+	}
		1265	+	// A JSON record of max-1 bytes produces a max-byte line with its newline,
		1266	+	// which is exactly what the readers accept.
		1267	+	padding := residentMaxRecordBytes - 1 - base
		1268	+	borderline, borderlineSize := residentAssistantRecordSize(t, strings.Repeat("x", padding))
		1269	+	if borderlineSize != residentMaxRecordBytes-1 {
		1270	+	t.Fatalf("borderline record = %d bytes, want %d", borderlineSize, residentMaxRecordBytes-1)
		1271	+	}
		1272	+	if err := journal.appendSync(borderline); err != nil {
		1273	+	t.Fatalf("appendSync rejected a reader-legal record: %v", err)
		1274	+	}
		1275	+	oversized, oversizedSize := residentAssistantRecordSize(t, strings.Repeat("x", padding+1))
		1276	+	if oversizedSize != residentMaxRecordBytes {
		1277	+	t.Fatalf("oversized record = %d bytes, want %d", oversizedSize, residentMaxRecordBytes)
		1278	+	}
		1279	+	if err := journal.appendSync(oversized); err == nil \|\| !strings.Contains(err.Error(), "too large") {
		1280	+	t.Fatalf("appendSync error = %v, want the reader record bound", err)
		1281	+	}
		1282	+	transcript := filepath.Join(journal.Dir(), residentTranscriptName)
		1283	+	if err := journal.Close(); err != nil {
		1284	+	t.Fatal(err)
		1285	+	}
		1286	+	records, err := ReadResidentJournal(transcript)
		1287	+	if err != nil {
		1288	+	t.Fatalf("journal unreadable after a boundary record: %v", err)
		1289	+	}
		1290	+	if len(records) != 2 {
		1291	+	t.Fatalf("records = %d, want the accepted child and the borderline message", len(records))
		1292	+	}
		1293	+	}
		1294	+
		1295	+	func residentAssistantRecordSize(t *testing.T, text string) (residentRecord, int) {
		1296	+	t.Helper()
		1297	+	message, err := json.Marshal(provider.Message{
		1298	+	Role: provider.RoleAssistant,
		1299	+	Content: []provider.Content{provider.TextBlock{Text: text}},
		1300	+	})
		1301	+	if err != nil {
		1302	+	t.Fatal(err)
		1303	+	}
		1304	+	record := residentRecord{Version: residentJournalVersion, Type: residentRecordAssistant, Message: message}
		1305	+	data, err := json.Marshal(record)
		1306	+	if err != nil {
		1307	+	t.Fatal(err)
		1308	+	}
		1309	+	return record, len(data)
		1310	+	}
		1311	+
		1312	+	func TestReconcileResidentJournalRepairsDanglingCheckpointCall(t *testing.T) {
		1313	+	root := t.TempDir()
		1314	+	journal, err := OpenResidentJournal(root, "checkpoint-dangling")
		1315	+	if err != nil {
		1316	+	t.Fatal(err)
		1317	+	}
		1318	+	spec := ResidentChildSpec{ID: "checkpoint-dangling", SessionID: "session", Provider: "openai", Model: "gpt-5"}
		1319	+	if err := journal.Accept(spec, "task"); err != nil {
		1320	+	t.Fatal(err)
		1321	+	}
		1322	+	if err := journal.AcceptFollowUp(spec, "turn-1", "follow up"); err != nil {
		1323	+	t.Fatal(err)
		1324	+	}
		1325	+	if err := journal.RecordTurnStarted(spec, "turn-1"); err != nil {
		1326	+	t.Fatal(err)
		1327	+	}
		1328	+	// The checkpoint kept a call whose result never arrived.
		1329	+	if err := journal.RecordCompacted([]provider.Message{{
		1330	+	Role: provider.RoleAssistant,
		1331	+	Content: []provider.Content{provider.ToolCallBlock{
		1332	+	ID: "compacted-call",
		1333	+	Name: "bash",
		1334	+	Arguments: json.RawMessage(`{"command":"pwd"}`),
		1335	+	}},
		1336	+	}}); err != nil {
		1337	+	t.Fatal(err)
		1338	+	}
		1339	+	if err := journal.RecordTurnFinished(spec, "turn-1", nil); err != nil {
		1340	+	t.Fatal(err)
		1341	+	}
		1342	+	transcript := filepath.Join(journal.Dir(), residentTranscriptName)
		1343	+	if err := journal.Close(); err != nil {
		1344	+	t.Fatal(err)
		1345	+	}
		1346	+	dir := filepath.Join(root, spec.ID)
		1347	+	if _, err := ReconcileResidentJournal(dir); err != nil {
		1348	+	t.Fatal(err)
		1349	+	}
		1350	+	records, err := ReadResidentJournal(transcript)
		1351	+	if err != nil {
		1352	+	t.Fatal(err)
		1353	+	}
		1354	+	repaired := 0
		1355	+	for _, record := range records {
		1356	+	if record.Type != residentRecordToolResult \|\| record.ToolID != "compacted-call" {
		1357	+	continue
		1358	+	}
		1359	+	if !isSyntheticInterruption(record.ToolResult) {
		1360	+	t.Fatalf("repaired result = %s, want the interruption result", record.ToolResult)
		1361	+	}
		1362	+	repaired++
		1363	+	}
		1364	+	if repaired != 1 {
		1365	+	t.Fatalf("repaired results = %d, want exactly one synthetic result", repaired)
		1366	+	}
		1367	+	// Once the repair is durable, later reconciliations must be stable no-ops.
		1368	+	second, err := ReconcileResidentJournal(dir)
		1369	+	if err != nil {
		1370	+	t.Fatal(err)
		1371	+	}
		1372	+	third, err := ReconcileResidentJournal(dir)
		1373	+	if err != nil {
		1374	+	t.Fatal(err)
		1375	+	}
		1376	+	if !second.UpdatedAt.Equal(third.UpdatedAt) {
		1377	+	t.Fatalf("reconciliation kept rewriting the lifecycle time: %v then %v", second.UpdatedAt, third.UpdatedAt)
		1378	+	}
		1379	+	if records, err = ReadResidentJournal(transcript); err != nil {
		1380	+	t.Fatal(err)
		1381	+	}
		1382	+	results := 0
		1383	+	for _, record := range records {
		1384	+	if record.Type == residentRecordToolResult && record.ToolID == "compacted-call" {
		1385	+	results++
		1386	+	}
		1387	+	}
		1388	+	if results != 1 {
		1389	+	t.Fatalf("tool results = %d, want the single repaired result", results)
		1390	+	}
		1391	+	}

M packages/core/dynamic_context.go +9 -0


@@ -80,6 +80,15 @@ func isInternalContextMessage(message provider.Message) bool {
80	80		return message.Role == provider.RoleDeveloper && message.Meta[internalContextMarker] == "true"
81	81		}
82	82
	83	+	// IsInternalContextMessage reports whether message is host-authored context
	84	+	// injected by the agent loop, such as a time snapshot or an extension
	85	+	// reminder, rather than finalized conversation history. Hosts that persist a
	86	+	// replacement transcript must exclude these messages: their value belongs to
	87	+	// the live request and replaying a stale copy would contradict it.
	88	+	func IsInternalContextMessage(message provider.Message) bool {
	89	+	return isInternalContextMessage(message)
	90	+	}
	91	+
83	92		func latestInternalContext(messages []provider.Message) (*provider.Message, []provider.Message) {
84	93		filtered := make([]provider.Message, 0, len(messages))
85	94		var latest *provider.Message


@@ -80,6 +80,15 @@ func isInternalContextMessage(message provider.Message) bool {
80	return message.Role == provider.RoleDeveloper && message.Meta[internalContextMarker] == "true"	80		return message.Role == provider.RoleDeveloper && message.Meta[internalContextMarker] == "true"
81	}	81		}
82		82
		83	+	// IsInternalContextMessage reports whether message is host-authored context
		84	+	// injected by the agent loop, such as a time snapshot or an extension
		85	+	// reminder, rather than finalized conversation history. Hosts that persist a
		86	+	// replacement transcript must exclude these messages: their value belongs to
		87	+	// the live request and replaying a stale copy would contradict it.
		88	+	func IsInternalContextMessage(message provider.Message) bool {
		89	+	return isInternalContextMessage(message)
		90	+	}
		91	+
83	func latestInternalContext(messages []provider.Message) (*provider.Message, []provider.Message) {	92		func latestInternalContext(messages []provider.Message) (*provider.Message, []provider.Message) {
84	filtered := make([]provider.Message, 0, len(messages))	93		filtered := make([]provider.Message, 0, len(messages))
85	var latest *provider.Message	94		var latest *provider.Message