@open-mercato/cezar 0.9.1 → 0.9.2-pr772.1071

package/dist/workflows/run.js

@@ -2,29 +2,36 @@ import { randomUUID } from 'node:crypto';
 import { spawn } from 'node:child_process';
 import { existsSync, mkdirSync, readFileSync, readdirSync, rmSync, writeFileSync } from 'node:fs';
 import { dirname, join } from 'node:path';
-import { parseAskMarker, stripAskMarker } from '../core/ask.js';
+import { parseAskMarkerResult, stripAskMarker, } from '../core/ask.js';
 import { onUsage, registerRunProcess, unregisterRunProcess } from '../core/process-usage.js';
+import { parseUsageLimit } from '../core/usage-limit.js';
 import { createRunner } from '../core/runner-factory.js';
 import { modelConflictsWithRunner } from '../core/model-presets.js';
+import { AGENT_MODELS_LOCKED_ERROR, agentModelsLocked } from '../core/agent-model-policy.js';
 import { ModelIdentityError, formatModelIdentity, normalizeModelForBackend, } from '../core/model-identity.js';
 import { HANDOFF_ONLY_INSTRUCTIONS, HANDOFF_INSTRUCTIONS, appendHandoffHeartbeat, followupsEnabled, handoffPath, seedHandoffFile, } from '../handoff.js';
 import { todosPath } from '../todos.js';
 import { discoverSkills } from '../skills.js';
 import { materializeSkillDir } from '../skills-remote.js';
 import { seedAgentConfigLocalLayer } from '../agent-config/seed.js';
+import { readAgentModelProvider } from '../agent-config/models.js';
 import { loadConfig, resolveWorktreeRetention } from '../config.js';
 import { autosaveCommit, createWorktree, resolveBaseRef, worktreeDiff, worktreeShortstat } from '../git-worktree.js';
-import { getRepoInfo } from '../server/git.js';
+import { getHeadCommit, getRepoInfo } from '../server/git.js';
 import { loadWorkflows } from './load.js';
 import { reclaimWorktrees, rematerializeReclaimedWorktree } from '../runs/retention.js';
 import { extractTaskRefs, refineTaskRefs, titleRefNumber } from '../runs/task-refs.js';
 import { parseTaskMarkers, stripTaskMarkers } from '../runs/task-markers.js';
 import { autoNamingActive, generateRunName, liveTitleUpdatesEnabled, postValidateTitle } from '../runs/auto-name.js';
 import { reviewGateEnabled } from '../runs/review-gate.js';
+import { resolveProfileEnvForRoot } from '../workspace/agent-profiles.js';
 import { WorkspaceSemaphore } from '../workspace/semaphore.js';
 import { UiEventSink } from '../runs/ui-event-sink.js';
 import { chainStepNote, DEFAULT_ALLOWED_TOOLS, stepKind } from './types.js';
 const CHECK_OUTPUT_CAP = 20_000;
+async function configuredModelProvider(backend, repoRoot) {
+    return readAgentModelProvider(backend, repoRoot).catch(() => undefined);
+}
 /** An interactive session that hears nothing from the user closes itself. */
 export const IDLE_TIMEOUT_MS = 15 * 60_000;
 /**
@@ -77,6 +84,18 @@ function emitAskRequested(sink, ask) {
     sink.handle({ type: 'ask.requested', requestId, questions: ask.questions });
     return requestId;
 }
+/** A persisted, non-fatal explanation for protocol-shaped text that could not
+ * become an ask card. Never include the raw payload in this diagnostic. */
+function askMarkerRejection(result) {
+    if (result.kind === 'invalid-json') {
+        return 'structured question ignored — CEZ:ASK payload is not valid JSON';
+    }
+    if (result.kind !== 'invalid-structure')
+        return undefined;
+    const issue = result.issues[0];
+    const location = issue?.path.length ? ` at ${issue.path.join('.')}` : '';
+    return `structured question ignored — CEZ:ASK payload failed validation${location}${issue ? `: ${issue.message}` : ''}`;
+}
 /** Periodic "cezar autosave" commit in the task worktree (spec 006). */
 export const AUTOSAVE_INTERVAL_MS = 90_000;
 /** The periodic autosave timer is opt-in (#471): off, a task branch carries only the
@@ -89,6 +108,96 @@ export function periodicAutosaveEnabled(env = process.env) {
 /** Safety cap on autonomous auto-continues per run — stops a stuck agent from nudging forever. */
 const MAX_AUTO_CONTINUES = 40;
 const AUTONOMOUS_NUDGE = 'Continue working autonomously until the task is fully complete. Do not ask me for confirmation or clarification — make reasonable assumptions and proceed. When everything is done, end the session with your done signal.';
+const MONITORING_WAKE_NUDGE = 'Re-check the downstream work you were monitoring. Continue toward the task goal; emit CEZ:MONITORING again only if it is still pending.';
+/**
+ * Auto-resume after a provider usage limit (spec 2026-08-03-auto-resume-after-usage-limit).
+ *
+ * The wait is the provider's own reset instant plus this grace: resuming AT the boundary races the
+ * provider's clock (and its rounding), and one failed resume costs the whole window over again.
+ * Thirty seconds is cheap next to five hours and long enough to be past any sane skew.
+ */
+export const AUTO_RESUME_GRACE_MS = 30_000;
+/**
+ * Consecutive automatic resumes allowed without a human turn. A resume can only fire after a real
+ * reset instant, so this is not a throttle — it is the backstop for the pathological case (a
+ * provider that answers "limit reached, retry now" in a loop), and it is deliberately generous
+ * enough to sit through a couple of days of five-hour windows.
+ */
+export const MAX_AUTO_RESUMES = 12;
+/**
+ * How long a missed deadline stays worth acting on. The promise is "we pick this up when the
+ * window reopens" — kept across a restart or an overnight close, which is the case the feature
+ * exists for. A day later it is no longer that promise: the user has moved on, and a task
+ * springing back to life is a surprise rather than a service. Such a deadline is retired with a
+ * note instead of fired, so the only tasks a sweep can revive are ones someone is still waiting on.
+ */
+export const AUTO_RESUME_MISSED_WINDOW_MS = 24 * 60 * 60_000;
+/**
+ * How often the queue checks that it is not wedged.
+ *
+ * A hold is the only thing in the engine that can make an idle queue CORRECT, so it is also the
+ * only thing that can make a wedged one look correct. This tick is the way out: cheap (a few
+ * in-memory checks), unref'd, and it only ever acts when idling has no justification left.
+ */
+export const QUEUE_WATCHDOG_MS = 60_000;
+/** Shared empty holds for the common "nothing is held" pump — avoids allocating per sweep. */
+const NO_HOLDS = { deadline: new Set(), inFlight: new Set() };
+/**
+ * May this run start, given what its account is holding?
+ *
+ * The two kinds of hold bind different work, and getting that wrong has produced a bug in each
+ * direction (spec 2026-08-03-auto-resume-after-usage-limit):
+ *
+ *  - a `deadline` hold means the window is KNOWN shut until an instant, so it blocks everything
+ *    on that account — resumes included. Exempting them let four resumes fire at once and
+ *    re-limit one after another, which is the stampede wearing a different hat.
+ *  - an `inFlight` hold means a resume is testing the window right now and nothing is proven, so
+ *    it blocks fresh work but not other resumes. Blocking those deadlocked a live workspace.
+ */
+function accountHeldFor(run, holds, fallbackRunner) {
+    const key = runAccountKey(run, fallbackRunner);
+    if (holds.deadline.has(key))
+        return true;
+    return holds.inFlight.has(key) && !resumeInFlight(run);
+}
+/**
+ * Which agent ACCOUNT a run's work runs on — the thing a provider usage limit actually closes
+ * (spec 2026-08-03-auto-resume-after-usage-limit).
+ *
+ * Backend plus agent account, because those are the two axes a limit is scoped to: a Claude
+ * limit must never stall a Codex task, and a second Claude login is a second budget. A record
+ * that names no runner has not started yet and will take the configured default, which is what
+ * `fallbackRunner` carries; a run that HAS started always carries its resolved runner (execute
+ * persists it), and only started runs can be holding.
+ */
+export function runAccountKey(run, fallbackRunner) {
+    return `${run.runner ?? fallbackRunner}:${run.agentProfile ?? 'default'}`;
+}
+/**
+ * Is this run an automatic resume that has not completed a turn yet?
+ *
+ * Such a run is the work the reopened window is FOR, so the hold must never apply to it — not
+ * its own, and not another resume's. Two resumes that hold each other is a deadlock the queue
+ * cannot recover from: both sit `queued` with a counter and no deadline, each waiting for the
+ * other to prove a window neither will ever get to test. That is the shape a live run produced
+ * — two scheduled tasks fired, both went `queued`, and nothing in the workspace moved again.
+ *
+ * The hold exists to stop NEW work walking into a closed window. A resume is not new work.
+ */
+function resumeInFlight(run) {
+    return (run.autoResumeAttempts !== undefined && (run.status === 'queued' || run.status === 'running'));
+}
+const AUTO_RESUME_PROMPT = 'The provider usage limit that interrupted this task has reset. Read the handoff file (CEZ_HANDOFF_FILE) to recover context, then continue the task from where you left off.';
+/**
+ * The wake instant as a human reads it — local, to the SECOND, with the zone named. The
+ * transcript line is what someone scanning a stalled task actually reads, and "18:41" is not
+ * enough to tell a wait that is nearly over from one that just started; the machine-readable ISO
+ * copy lives on `RunRecord.autoResumeAt`. Server-side formatting is honest here because cezar is
+ * local-first: the process and the browser reading it are the same machine.
+ */
+function formatWakeInstant(at) {
+    return new Intl.DateTimeFormat(undefined, { dateStyle: 'medium', timeStyle: 'long' }).format(at);
+}
 /**
  * The effective "extra" system prompt for a run (spec §protocol v2, R2 2.3):
  * the per-run override (`POST /api/runs` `systemPrompt`) REPLACES the
@@ -171,6 +280,7 @@ const VARIANT_HINTS = {
     B: 'Approach hint: prefer the minimal, surgical change.',
     C: 'Approach hint: prefer a thorough, structural approach.',
 };
+const RESTART_CONTINUATION_PROMPT = 'The cezar process restarted while you were working on this task. Read the handoff file (CEZ_HANDOFF_FILE) to recover context, then continue the task from where you left off.';
 /**
  * The mini workflow engine: executes a `WorkflowDef` against a repo, one step
  * at a time, persisting every event to the RunStore (which the SSE endpoints
@@ -200,7 +310,13 @@ export class RunManager {
     // timeout already bounds how long a session can sit open. Invariant:
     // `waiting ⊆ active` — always cleared together via dropActive().
     waiting = new Set();
+    /** Durable monitoring subset. Only the configured number receives the waiting-slot exemption. */
+    monitoring = new Set();
     pendingJobs = new Map();
+    /** Interrupted agent turns recovered after a process restart. Unlike an
+     *  explicit user Continue, these are bulk scheduler work and must re-enter
+     *  through `pump()` so both workspace and per-project caps are honored. */
+    pendingContinuations = new Map();
     /** Per-run image counter behind `pasted-<n>` / `screenshot-<n>` (#472). Lives on
      *  the manager rather than the `ActiveRun` so a *queued* run — which has no
      *  `ActiveRun` at all — can persist attachments. Seeded lazily from disk. */
@@ -208,6 +324,12 @@ export class RunManager {
     /** Messages that landed in the dequeue → session-open gap (#472), flushed as
      *  ordinary follow-up turns the moment the session opens. In-memory only. */
     deferredMessages = new Map();
+    /** Armed usage-limit resumes, keyed by run id (spec
+     *  2026-08-03-auto-resume-after-usage-limit). The DEADLINE itself lives on the record
+     *  (`autoResumeAt`) — this map holds only the process-local timer, so a restart rebuilds it
+     *  from the record rather than losing the wait. Runs here are `failed` and therefore NOT in
+     *  `active`, which is why the timer cannot live on an `ActiveRun` like the monitoring one. */
+    autoResumeTimers = new Map();
     pumping = false;
     /** A pump that arrived while one was in flight — replayed by `pump()`'s own
      *  loop so a slot freed mid-sweep is never a lost wakeup. */
@@ -226,6 +348,13 @@ export class RunManager {
     /** Unsubscribe handle for the constructor's `onUsage` subscription — released
      *  by dispose() so a torn-down manager stops receiving sampler ticks. */
     offUsage;
+    /** The stalled-queue watchdog (see `rescueStalledQueue`). */
+    queueWatchdog;
+    /** Set by the watchdog for exactly one sweep: ignore the usage-limit hold and make progress. */
+    forceNextPump = false;
+    /** Runs the watchdog started despite the hold. The spawn-time gate (`requeueWhileHeld`) would
+     *  otherwise hand them straight back and the rescue would undo itself in a millisecond. */
+    forceStarted = new Set();
     /** The workspace-wide parallel-cap semaphore + cached resource config
      *  (spec 2026-07-20, step 2.5). Boot constructs ONE and every manager shares
      *  it; the private fallback keeps single-manager callers and tests working. */
@@ -242,10 +371,13 @@ export class RunManager {
             busySlots: () => this.busySlots(),
             pump: () => this.pump(),
             oldestQueuedAt: () => this.oldestQueuedAt(),
+            accountHolds: () => this.accountHolds(),
         });
         // Memory guard (#memory-guard): the shared process-tree sampler already ticks ~every 2 s for
         // the runs table; piggyback on it to enforce the per-task memory ceiling.
         this.offUsage = onUsage((snapshot) => void this.enforceMemoryLimit(snapshot));
+        this.queueWatchdog = setInterval(() => void this.rescueStalledQueue(), QUEUE_WATCHDOG_MS);
+        this.queueWatchdog.unref?.();
     }
     /**
      * Release everything this manager owns without touching run records
@@ -260,17 +392,23 @@ export class RunManager {
     dispose() {
         this.offUsage();
         this.offSemaphore();
-        for (const state of this.active.values()) {
+        clearInterval(this.queueWatchdog);
+        for (const [runId, state] of this.active) {
             this.clearIdleTimer(state);
+            this.clearMonitoringWakeTimer(state, runId);
             this.clearAutosaveTimer(state);
             state.releaseRepoRoot?.();
             state.releaseRepoRoot = undefined;
         }
+        for (const timer of this.autoResumeTimers.values())
+            clearTimeout(timer);
+        this.autoResumeTimers.clear();
         this.active.clear();
         this.waiting.clear();
         this.starting.clear();
         this.queue.length = 0;
         this.pendingJobs.clear();
+        this.pendingContinuations.clear();
         this.memoryPausing.clear();
         this.lastNamerKey.clear();
     }
@@ -338,13 +476,49 @@ export class RunManager {
             CEZ_TODOS_FILE: generateFollowups ? todosPath(this.dataDir) : '',
         };
     }
+    /**
+     * `agentEnv` plus the agent-account variable for the profile this STEP runs under (spec
+     * 2026-07-29-agent-profiles), and the id it resolved to so the caller can record it.
+     *
+     * Resolved per step, not per run, because a workflow can mix backends: an override naming a
+     * Claude account says nothing about which Codex account a codex step should use. Resolution
+     * order, most specific first:
+     *
+     *   1. the step's ALREADY-RECORDED `profileId` — a resume or Continue must reattach to the
+     *      account that created the session, whatever the project has since been switched to;
+     *   2. the run's composer override, but only for steps on the run's own runner;
+     *   3. the project's stored selection, and failing that the discovered default.
+     *
+     * Read fresh every time. `~/.cezar/config.json` is shared by every cezar process on this
+     * machine, so a cached snapshot is a staleness bug, and one small JSON read is free next to
+     * spawning a CLI. Never throws: an unreadable home degrades to the default profile, which is
+     * exactly the behaviour that predates profiles.
+     */
+    async agentEnvForStep(runId, backend, options = {}) {
+        const run = this.store.getRun(runId);
+        const profileId = options.recordedProfileId
+            ?? (backend === (run?.runner ?? 'claude') ? run?.agentProfile : undefined);
+        const resolved = await resolveProfileEnvForRoot(this.repoRoot, backend, profileId);
+        return {
+            env: { ...this.agentEnv(runId, options.generateFollowups), ...resolved.env },
+            profileId: resolved.profile.id,
+        };
+    }
     startRun(workflow, input, group) {
+        // Sanitize at the manager boundary so CLI runs, workflows, variants, and
+        // direct callers cannot bypass the HTTP policy.
+        const effectiveInput = agentModelsLocked(this.repoRoot)
+            ? { ...input, model: undefined }
+            : input;
         const run = this.store.createRun({
             title: makeRunTitle(input.task, workflow) + (group ? ` (${group.variant})` : ''),
             workflow: workflow.name,
             task: input.task,
-            model: input.model,
+            model: effectiveInput.model,
             runner: input.runner,
+            // The composer's per-task account (spec 2026-07-29-agent-profiles). Persisted at creation
+            // so a queued run picks it up at dequeue and every later resume reads the same answer.
+            agentProfile: input.agentProfile,
             // The global inbox is the ceiling on the per-run flag (#471). Enforced here rather than
             // at the HTTP route because `cezar run`, the inbox's own "▶ Run" and variants all reach
             // startRun directly — a route-level gate would leave those writing todos.json.
@@ -354,6 +528,9 @@ export class RunManager {
             // auto-nudge reads `input.autonomous` (`execute`), but the record is the
             // only source those after-the-fact consumers have.
             autonomous: input.autonomous === true,
+            // Persist the explicit opt-out so queued-run restart recovery and the
+            // session Git routes can distinguish it from a removed isolated worktree.
+            worktree: !group && input.worktree === false ? false : undefined,
             groupId: group?.groupId,
             variant: group?.variant,
             steps: workflow.steps.map((s) => ({ id: s.id, name: s.name ?? s.id, kind: stepKind(s) })),
@@ -388,7 +565,7 @@ export class RunManager {
         // above shows instantly; the namer's short title replaces it when (and if)
         // the model answers. Never awaited, never fails the run.
         void this.autoNameRun(run.id, skillHint, input.task);
-        this.pendingJobs.set(run.id, { workflow, input });
+        this.pendingJobs.set(run.id, { workflow, input: effectiveInput });
         this.queue.push(run.id);
         void this.pump();
         return run;
@@ -405,7 +582,7 @@ export class RunManager {
         return VARIANT_LETTERS.slice(0, Math.min(Math.max(count, 1), VARIANT_LETTERS.length)).map((variant) => {
             const hint = VARIANT_HINTS[variant];
             const task = hint ? `${input.task}\n\n${hint}` : input.task;
-            return this.startRun(workflow, { ...input, task }, { groupId, variant });
+            return this.startRun(workflow, { ...input, task, worktree: undefined }, { groupId, variant });
         });
     }
     /**
@@ -418,7 +595,9 @@ export class RunManager {
      * `maxParallel`, including when other projects saturate the cap.
      */
     busySlots() {
-        return this.active.size + this.starting.size - this.waiting.size;
+        const ordinaryWaiting = this.waiting.size - this.monitoring.size;
+        const exemptMonitoring = Math.min(this.monitoring.size, this.semaphore.maxMonitoringSessions());
+        return this.active.size + this.starting.size - ordinaryWaiting - exemptMonitoring;
     }
     /** Epoch ms of this manager's oldest queued run (the semaphore's fairness
      *  key when a freed slot is broadcast), or null when nothing is queued.
@@ -454,6 +633,8 @@ export class RunManager {
      * root (spec 006 degradation rule), which is always the tighter bound.
      */
     async pump() {
+        this.reconcileMonitoringWakeTimers();
+        this.reconcileAutoResumes();
         // A pump requested while one is in flight can't just be dropped: the
         // in-flight pass may already have read capacity (it awaits `getRepoInfo`
         // before the first check), so a slot freed in that window would be lost
@@ -478,15 +659,69 @@ export class RunManager {
                 const capacity = () => this.semaphore.busy() < maxParallel &&
                     this.busySlots() < projectMax &&
                     (repo !== null || this.busySlots() < 1);
+                // The usage-limit hold (spec 2026-08-03-auto-resume-after-usage-limit).
+                //
+                // A limit closes an ACCOUNT, not a run — so starting the next queued task walks it into
+                // the same wall. Measured before this gate existed: eight tasks under `maxParallel: 2`
+                // all failed within 517 ms, each spawning a CLI (and, outside worktree-opt-out mode, a
+                // worktree and a branch) only to be marked `scheduled`. The cap was respected at every
+                // instant and was no brake at all, because a doomed run lives ~200 ms.
+                //
+                // So: while any run on an account is waiting out a limit, nothing new starts on THAT
+                // account. Other accounts (a second login, a different backend) keep running — the hold
+                // is keyed, not global. The set is derived from the durable records rather than tracked
+                // separately, which is what makes it survive a restart, expire on its own, and lift the
+                // instant a user cancels a resume.
+                // The watchdog's one-shot override — read and cleared here, so a forced sweep never
+                // leaks into the next ordinary one.
+                const forced = this.forceNextPump;
+                this.forceNextPump = false;
+                const holds = this.queue.length > 0 && !forced ? this.semaphore.accountHolds() : NO_HOLDS;
+                const anyHold = holds.deadline.size > 0 || holds.inFlight.size > 0;
+                // Only pay for the config read when something is actually held: a queued record may name
+                // no runner, and then the account it would use is the configured default.
+                const defaultRunner = anyHold ? (await loadConfig(this.repoRoot)).defaultRunner : undefined;
                 while (this.queue.length > 0 && capacity()) {
-                    const runId = this.queue.shift();
+                    // FIFO among the runs that CAN start; a held one keeps its place in the queue rather
+                    // than being dequeued and re-queued (which would churn its position and its record).
+                    const next = !anyHold
+                        ? 0
+                        : this.queue.findIndex((id) => {
+                            const queued = this.store.getRun(id);
+                            return !queued || !accountHeldFor(queued, holds, defaultRunner ?? 'claude');
+                        });
+                    if (next === -1)
+                        break; // everything queued is waiting on a held account
+                    const runId = this.queue.splice(next, 1)[0];
                     if (!runId)
                         break;
+                    // A forced sweep has to reach the spawn: the gate inside `execute` asks the same
+                    // question and would send this run straight back to the queue.
+                    if (forced)
+                        this.forceStarted.add(runId);
                     const job = this.pendingJobs.get(runId);
+                    const continuation = this.pendingContinuations.get(runId);
                     this.pendingJobs.delete(runId);
-                    if (!job)
+                    this.pendingContinuations.delete(runId);
+                    if (!job && !continuation)
                         continue;
                     this.starting.add(runId);
+                    if (continuation) {
+                        const hydrated = this.hydrateQueuedContinuation(runId, continuation);
+                        void this.runContinuation(runId, hydrated.stepId, hydrated.sessionId, hydrated.backend, hydrated.prompt, hydrated.images, hydrated.persistedImages, hydrated.persistedAttachments).catch((err) => {
+                            const message = err instanceof Error ? err.message : String(err);
+                            this.store.updateRun(runId, {
+                                status: 'failed',
+                                error: `continue crashed: ${message}`,
+                                finishedAt: new Date().toISOString(),
+                            });
+                            this.starting.delete(runId);
+                            this.dropActive(runId);
+                        });
+                        continue;
+                    }
+                    if (!job)
+                        continue;
                     // Rebuild the prompt from the store at the last instant (#472), so an edit
                     // or a stacked message that landed while the run waited is honored. Entered
                     // in the same synchronous tick as the `pendingJobs.delete` above, so no
@@ -514,6 +749,87 @@ export class RunManager {
             this.pumping = false;
         }
     }
+    /**
+     * Make one `queued` RECORD executable again — the engine half a queued run needs but does not
+     * persist (`pendingJobs` / `pendingContinuations` are process-local, the record is not).
+     *
+     * Two callers, one path: boot recovery re-adopts everything the previous process was holding,
+     * and the queue watchdog re-adopts anything the running process has somehow lost. A queued
+     * record with no work item behind it is invisible to `pump()` and would sit there for good,
+     * which is the worst failure this engine has — the task is neither running nor failed, just
+     * silently never going to happen.
+     *
+     * A continuation is reconstructed first: its executable details are gone, but the pending
+     * `continue-N` step and the session before it are durable, which is enough. Otherwise the
+     * workflow is revived from the record. A run that can be neither is failed loudly rather than
+     * left in the queue as a ghost.
+     */
+    async reviveQueuedRun(run, reason) {
+        const queuedContinuation = [...run.steps]
+            .reverse()
+            .find((step) => step.status === 'pending' && step.id.startsWith('continue-'));
+        const sessionStep = queuedContinuation
+            ? [...run.steps].reverse().find((step) => step.id !== queuedContinuation.id && step.sessionId)
+            : undefined;
+        if (queuedContinuation && sessionStep?.sessionId) {
+            const backend = run.runner ?? 'claude';
+            const sessionBackend = sessionStep.backend ?? backend;
+            this.pendingContinuations.set(run.id, {
+                stepId: queuedContinuation.id,
+                sessionId: sessionBackend === backend ? sessionStep.sessionId : undefined,
+                backend,
+                prompt: RESTART_CONTINUATION_PROMPT,
+                images: [],
+            });
+            this.queue.push(run.id);
+            this.store.appendEvent(run.id, {
+                type: 'lifecycle',
+                message: `${reason} — interrupted continuation re-queued`,
+            });
+            return;
+        }
+        const workflow = await this.reviveWorkflow(run);
+        if (!workflow) {
+            this.store.updateRun(run.id, {
+                status: 'failed',
+                error: 'interrupted — workflow definition not recoverable after a restart',
+                finishedAt: new Date().toISOString(),
+            });
+            this.store.appendEvent(run.id, {
+                type: 'lifecycle',
+                message: `${reason} — workflow definition not recoverable, task failed`,
+            });
+            return;
+        }
+        // Re-apply the inbox ceiling (#471). `execute()` gates again at spawn time, so the agent is
+        // safe either way — but a run queued while the inbox was on and recovered after it was
+        // switched off would otherwise keep echoing `generateFollowups: true` on a run that
+        // demonstrably produced none. Normalize the record, the way startRun does.
+        const generateFollowups = followupsEnabled() ? run.generateFollowups : false;
+        if (generateFollowups !== run.generateFollowups) {
+            this.store.updateRun(run.id, { generateFollowups });
+        }
+        this.pendingJobs.set(run.id, {
+            workflow,
+            // Folded through the same helper `pump()` uses (#472) so a restart carries the stack.
+            // Idempotent: hydration always composes from `run.task` + the stack, never from an
+            // already-folded `input.task`, so re-hydrating at dequeue yields the same string.
+            input: this.hydrateQueuedInput(run.id, {
+                task: run.task,
+                model: run.model,
+                runner: run.runner,
+                generateFollowups,
+                // Re-thread autonomy (#489): the rebuilt input feeds `execute`, whose mid-run auto-nudge
+                // reads `input.autonomous`. Without this a recovered autonomous run would run
+                // non-autonomously and later wrongly park at `review`.
+                autonomous: run.autonomous,
+                // Preserve an explicit worktree opt-out across a queued restart.
+                worktree: run.worktree,
+            }),
+        });
+        this.queue.push(run.id);
+        this.store.appendEvent(run.id, { type: 'lifecycle', message: `${reason} — task re-queued` });
+    }
     /**
      * Startup recovery (#367) — re-adopt runs that were live when the previous
      * cezar process exited (requires the store opened with `keepLive`):
@@ -532,49 +848,7 @@ export class RunManager {
             .sort((a, b) => a.createdAt.localeCompare(b.createdAt));
         for (const run of live) {
             if (run.status === 'queued') {
-                const workflow = await this.reviveWorkflow(run);
-                if (workflow) {
-                    // Re-apply the inbox ceiling (#471). `execute()` gates again at spawn time, so the
-                    // agent is safe either way — but a run queued while the inbox was on and recovered
-                    // after it was switched off would otherwise keep echoing `generateFollowups: true`
-                    // on a run that demonstrably produced none. Normalize the record, the way startRun
-                    // does, so the stored answer matches what actually happens.
-                    const generateFollowups = followupsEnabled() ? run.generateFollowups : false;
-                    if (generateFollowups !== run.generateFollowups) {
-                        this.store.updateRun(run.id, { generateFollowups });
-                    }
-                    this.pendingJobs.set(run.id, {
-                        workflow,
-                        // Folded through the same helper `pump()` uses (#472) so a restart
-                        // carries the stack. Idempotent: hydration always composes from
-                        // `run.task` + the stack, never from an already-folded `input.task`,
-                        // so re-hydrating at dequeue yields the same string, not a doubled one.
-                        input: this.hydrateQueuedInput(run.id, {
-                            task: run.task,
-                            model: run.model,
-                            runner: run.runner,
-                            generateFollowups,
-                            // Re-thread autonomy (#489): the rebuilt input feeds `execute`,
-                            // whose mid-run auto-nudge reads `input.autonomous`. Without this a
-                            // recovered queued autonomous run would run non-autonomously (no
-                            // auto-nudge) and later wrongly park at `review`.
-                            autonomous: run.autonomous,
-                        }),
-                    });
-                    this.queue.push(run.id);
-                    this.store.appendEvent(run.id, { type: 'lifecycle', message: 'cezar restarted — task re-queued' });
-                }
-                else {
-                    this.store.updateRun(run.id, {
-                        status: 'failed',
-                        error: 'interrupted — workflow definition not recoverable after a restart',
-                        finishedAt: new Date().toISOString(),
-                    });
-                    this.store.appendEvent(run.id, {
-                        type: 'lifecycle',
-                        message: 'cezar restarted — workflow definition not recoverable, task failed',
-                    });
-                }
+                await this.reviveQueuedRun(run, 'cezar restarted');
                 continue;
             }
             if (run.status === 'waiting') {
@@ -605,8 +879,8 @@ export class RunManager {
                 currentStepId: undefined,
             });
             const resumed = this.continueRun(run.id, {
-                text: 'The cezar process restarted while you were working on this task. Read the handoff file (CEZ_HANDOFF_FILE) to recover context, then continue the task from where you left off.',
-            });
+                text: RESTART_CONTINUATION_PROMPT,
+            }, true);
             this.store.appendEvent(run.id, {
                 type: 'lifecycle',
                 message: resumed.ok
@@ -614,14 +888,21 @@ export class RunManager {
                     : `cezar restarted — could not resume the interrupted task (${resumed.error ?? 'unknown'})`,
             });
         }
+        // Re-arm usage-limit resumes (spec 2026-08-03-auto-resume-after-usage-limit): the wait is
+        // routinely longer than a cezar session, so the deadline is durable and the timer is rebuilt
+        // from it. `pump()` reconciles again on every sweep, so this is the fast path, not the only
+        // one — see `reconcileAutoResumes`.
+        this.reconcileAutoResumes();
         void this.pump();
     }
     /** The persisted definition when it looks sane, else the catalog by name. */
     async reviveWorkflow(run) {
+        // "Looks sane" is the STORE's job now: it parses `workflowDef` against the definition schema
+        // and `.catch`es a def that no longer fits to `undefined`, so anything present here already
+        // has the `steps` array the old inline `Array.isArray` check was asking for.
         const def = run.workflowDef;
-        if (def && Array.isArray(def.steps)) {
+        if (def)
             return def;
-        }
         const { workflows } = await loadWorkflows(this.repoRoot);
         return workflows.find((w) => w.name === run.workflow) ?? null;
     }
@@ -632,12 +913,26 @@ export class RunManager {
         if (state)
             state.releaseRepoRoot = undefined;
         this.waiting.delete(runId);
+        this.monitoring.delete(runId);
+        if (state)
+            this.clearMonitoringWakeTimer(state, runId);
         this.active.delete(runId);
         this.memoryPausing.delete(runId);
         this.lastNamerKey.delete(runId);
+        this.forceStarted.delete(runId);
         // The run's slot is gone from busySlots() as of the deletes above — hand it
         // to the workspace's oldest queued run, in ANY project. Every terminal path
         // funnels through here, so this one call covers them all.
+        // Same reasoning as retention below — every terminal path funnels through here, so the
+        // usage-limit question ("did this run stop because the account is out of window, and when
+        // does that window reopen?") is asked once, in one place, off the record the failing path
+        // has already written. Nothing to do for any other outcome.
+        //
+        // BEFORE releasing the slot, and that order is the whole point: `releaseSlot` pumps every
+        // manager, and a pump reads the hold off the records. Publishing the schedule afterwards
+        // left a window — measured as exactly one extra task — where the queue saw a free slot and
+        // an account that looked healthy, and started work that was already doomed.
+        this.scheduleAutoResumeIfLimited(runId);
         this.releaseSlot();
         // A run leaving the active registry is a terminal transition (done/review/
         // failed/cancelled) — the one moment the finished-worktree count can grow.
@@ -646,6 +941,333 @@ export class RunManager {
         // the lifecycle.
         void this.enforceRetention();
     }
+    // ---- usage-limit auto-resume (spec 2026-08-03-auto-resume-after-usage-limit) --------------
+    /**
+     * A run just failed: if the provider said "usage limit, back at T", promise to resume it at
+     * `T + AUTO_RESUME_GRACE_MS` instead of leaving the task dead until someone notices.
+     *
+     * Every refusal below is silent-but-honest — the run stays `failed` with its Continue button,
+     * which is exactly the pre-feature behavior — except the safety cap, which says so on the
+     * transcript, because a run that stops resuming itself needs to explain why.
+     */
+    scheduleAutoResumeIfLimited(runId) {
+        if (this.autoResumeTimers.has(runId))
+            return; // already promised
+        const run = this.store.getRun(runId);
+        if (!run || run.status !== 'failed')
+            return;
+        // Archiving IS resigning from a task. Reviving one because a window happened to reopen would
+        // be the feature working against the clearest signal the user can give it.
+        if (run.archived)
+            return;
+        const limit = parseUsageLimit(run.error);
+        if (!limit)
+            return;
+        if (!this.semaphore.autoResumeOnUsageLimit())
+            return;
+        // No session to resume = nothing this feature can do; `continueRun` would refuse anyway.
+        if (!run.steps.some((step) => step.sessionId))
+            return;
+        const attempts = run.autoResumeAttempts ?? 0;
+        if (attempts >= MAX_AUTO_RESUMES) {
+            this.store.appendEvent(runId, {
+                type: 'note',
+                message: `automatic resume cap reached (${MAX_AUTO_RESUMES}) — continue this task manually`,
+            });
+            return;
+        }
+        const wakeAt = new Date(limit.resetAt.getTime() + AUTO_RESUME_GRACE_MS);
+        this.armAutoResume(runId, wakeAt.getTime());
+        this.store.appendEvent(runId, {
+            type: 'lifecycle',
+            message: `usage limit reached — resuming automatically at ${formatWakeInstant(wakeAt)}`,
+        });
+    }
+    /** Publish the deadline on the record (the cockpit's only source) and arm the timer for it. */
+    armAutoResume(runId, deadline) {
+        this.store.updateRun(runId, { autoResumeAt: new Date(deadline).toISOString() });
+        const timer = setTimeout(() => this.fireAutoResume(runId), Math.max(0, deadline - Date.now()));
+        timer.unref?.();
+        this.autoResumeTimers.set(runId, timer);
+    }
+    /**
+     * The window has reopened. Re-check the record synchronously — hours may have passed, and the
+     * user may have continued, deleted or cancelled the run in them — then hand the resume to the
+     * ordinary queued-continuation path so it obeys both concurrency caps like any other work.
+     */
+    fireAutoResume(runId) {
+        this.autoResumeTimers.delete(runId);
+        const run = this.store.getRun(runId);
+        if (!run || run.status !== 'failed' || !run.autoResumeAt)
+            return;
+        // Belt and braces against the one gap `reconcileAutoResumes` cannot close: the setting going
+        // off in the window between the last pump and this tick.
+        if (!this.semaphore.autoResumeOnUsageLimit()) {
+            this.clearAutoResume(runId);
+            return;
+        }
+        const attempts = (run.autoResumeAttempts ?? 0) + 1;
+        // `continueRun` retires the pending resume (timer + record fields) on the way in — this is a
+        // resume, not a user turn, so the counter is put back straight after.
+        const resumed = this.continueRun(runId, { text: AUTO_RESUME_PROMPT }, true);
+        if (!resumed.ok) {
+            // Refusals happen before `continueRun` retires anything, so the deadline is still on the
+            // record — and a deadline in the past is a promise the cockpit keeps displaying and the
+            // engine will never keep. Retire it here instead, and say why.
+            this.clearAutoResume(runId);
+            this.store.appendEvent(runId, {
+                type: 'note',
+                message: `automatic resume could not start — ${resumed.error ?? 'unknown'}`,
+            });
+            return;
+        }
+        this.store.updateRun(runId, { autoResumeAttempts: attempts });
+        this.store.appendEvent(runId, {
+            type: 'lifecycle',
+            message: `usage limit reset — resuming automatically (${attempts}/${MAX_AUTO_RESUMES})`,
+        });
+        // A deferred continuation only ENQUEUES itself; the queue moves when something pumps it, and
+        // `recover()` — the other deferring caller — pumps once after its whole bulk sweep. A timer
+        // firing on its own has no such follow-up, so without this the resumed run sits at `queued`
+        // until some unrelated run happens to finish. This is the pump for it.
+        void this.pump();
+    }
+    /**
+     * Make the armed timers agree with the records and the current setting. Runs on every `pump()`
+     * — which is where a settings change lands (a config PUT refreshes the shared semaphore, which
+     * pumps every manager) — and once from `recover()`.
+     *
+     * It is a RECONCILE rather than a one-shot restore because the deadline is durable state and
+     * the timer is not: a restart, a rebuilt project context, a manager disposed mid-wait, or a
+     * refusal all leave a record promising a resume that no timer is holding. Rebuilding from the
+     * record covers every one of those at once — the alternative is a hint counting down to a time
+     * that has already passed, which is exactly the failure this method exists to make impossible.
+     *
+     * Cheap: an in-memory scan, and arming is skipped for every run already held.
+     */
+    reconcileAutoResumes() {
+        if (!this.semaphore.autoResumeOnUsageLimit()) {
+            // Sweep the RECORDS, not the timer map. A record promising a resume that no timer is
+            // holding is the exact population this method exists for, and it is also the one the
+            // setting can be switched off in front of: cezar restarted while it was off, the config
+            // was hand-edited, or the project context was disposed mid-wait. Retiring only the armed
+            // timers leaves such a record with a live `autoResumeAt`, which `accountHolds()` reads as
+            // a deadline hold — so nothing new starts on that account, `rescueStalledQueue` treats the
+            // phantom appointment as a legitimate reason to sit still, and the cockpit shows a
+            // `scheduled` row for a resume that will never come. `clearAutoResume` covers the armed
+            // ones too, so this one loop is the whole cancellation.
+            const pending = new Set([
+                ...this.autoResumeTimers.keys(),
+                ...this.store.listRuns().filter((run) => run.autoResumeAt !== undefined).map((run) => run.id),
+            ]);
+            for (const runId of pending) {
+                this.clearAutoResume(runId);
+                this.store.appendEvent(runId, {
+                    type: 'note',
+                    message: 'automatic resume cancelled — auto-resume is switched off',
+                });
+            }
+            return;
+        }
+        for (const run of this.store.listRuns()) {
+            if (run.status !== 'failed' || !run.autoResumeAt)
+                continue;
+            if (this.autoResumeTimers.has(run.id))
+                continue;
+            const deadline = Date.parse(run.autoResumeAt);
+            // A deadline that is unreadable, belongs to a run that has spent its cap, or belongs to a
+            // task the user has archived is retired rather than re-armed: it can only mislead. One
+            // that has just passed arms at zero — the window is open, which is the point.
+            if (run.archived
+                || !Number.isFinite(deadline)
+                || (run.autoResumeAttempts ?? 0) >= MAX_AUTO_RESUMES) {
+                this.store.updateRun(run.id, { autoResumeAt: undefined });
+                continue;
+            }
+            // …and one missed by more than a day is retired loudly: reviving a task from another era
+            // is a surprise, not a service, and this is what keeps a sweep from resurrecting every
+            // limit-stopped task a user has long since walked away from.
+            if (Date.now() - deadline > AUTO_RESUME_MISSED_WINDOW_MS) {
+                this.store.updateRun(run.id, { autoResumeAt: undefined });
+                this.store.appendEvent(run.id, {
+                    type: 'note',
+                    message: 'automatic resume expired — its window reopened over a day ago; continue this task manually',
+                });
+                continue;
+            }
+            this.armAutoResume(run.id, deadline);
+        }
+    }
+    /**
+     * Hand a run that has not spawned anything back to the queue, when the account it would run on
+     * went into a usage-limit hold (spec 2026-08-03-auto-resume-after-usage-limit).
+     *
+     * The dequeue-time gate in `pump()` cannot be the only one: a run can sit between dequeue and
+     * spawn for a long time — an in-place run waiting for the exclusive repo-root lease is the
+     * measured case — and the account can close in that gap. This is the last honest moment to
+     * refuse, because everything after it costs a real agent turn.
+     *
+     * "Untouched" is the contract: the run has created no session and no worktree, so it goes back
+     * as plain `queued` with its `startedAt` cleared, and `pump()` will pick it up when the window
+     * reopens. Returns true when the caller must abandon the run.
+     */
+    requeueWhileHeld(runId, workflow, input, runner, state) {
+        const run = this.store.getRun(runId);
+        if (!run || run.status === 'cancelled' || state?.cancelled)
+            return false;
+        // The watchdog sent this one through. Checked, never consumed: the spawn path asks this
+        // question TWICE — here at the top of `execute`, and again after the exclusive repo-root
+        // lease is granted — so a one-shot flag would clear at the first gate and let the second one
+        // hand an in-place run straight back, re-wedging the queue the rescue had just freed.
+        // `dropActive` retires the entry on every terminal path, so the set still cleans itself up.
+        if (this.forceStarted.has(runId))
+            return false;
+        if (!accountHeldFor({ ...run, runner }, this.semaphore.accountHolds(), runner))
+            return false;
+        state?.releaseRepoRoot?.();
+        if (state)
+            state.releaseRepoRoot = undefined;
+        this.pendingJobs.set(runId, { workflow, input });
+        this.queue.push(runId);
+        this.store.updateRun(runId, { status: 'queued', startedAt: undefined, currentStepId: undefined });
+        this.store.appendEvent(runId, {
+            type: 'note',
+            message: 'held in the queue — this agent account is waiting out a usage limit',
+        });
+        this.dropActive(runId);
+        return true;
+    }
+    /**
+     * The failsafe: a queue must never be able to wedge.
+     *
+     * Everything else in this file makes an idle queue CORRECT under some condition — a slot cap, a
+     * repo-root lease, and now a usage-limit hold. That is also what makes a wedged queue look
+     * correct, and the hold has already produced one in the field: two resumes fired together, each
+     * holding the account the other was waiting on, and the whole workspace stopped with every task
+     * `queued`. That specific bug is fixed and tested, but "the queue stopped and nothing will ever
+     * restart it" is too expensive a failure mode to leave resting on any single fix being right.
+     *
+     * The test is deliberately about JUSTIFICATION rather than about any particular bug: idling is
+     * legitimate while work is running (here or in another project), or while a real appointment is
+     * still ahead — a scheduled resume that will fire and pump on its own. Anything else is a
+     * queue with work in it, nothing running anywhere, and no event coming to wake it. That gets one
+     * forced sweep, which starts work under the ordinary caps and lets the account's real state
+     * re-assert itself: if the window truly is shut, that task meets the limit and re-establishes an
+     * honest hold, with a real deadline behind it this time.
+     *
+     * Public so a test can drive the wedge directly instead of waiting out the interval.
+     */
+    async rescueStalledQueue(now = Date.now()) {
+        // First, the worst shape: a record that says `queued` while the engine holds no job, no
+        // continuation and no queue entry for it. `pump()` cannot see such a run — it iterates the
+        // queue, and this one is not in it — so nothing will ever start it. Re-adopt it through the
+        // same path boot recovery uses.
+        for (const run of this.store.listRuns()) {
+            if (run.status !== 'queued')
+                continue;
+            if (this.active.has(run.id) || this.starting.has(run.id))
+                continue;
+            if (this.pendingJobs.has(run.id) || this.pendingContinuations.has(run.id))
+                continue;
+            if (this.queue.includes(run.id))
+                continue;
+            console.warn(`[cez] queue watchdog: re-adopting queued run ${run.id} the engine had lost`);
+            await this.reviveQueuedRun(run, 'queue watchdog');
+        }
+        if (this.queue.length === 0)
+            return;
+        if (this.busySlots() > 0 || this.starting.size > 0)
+            return;
+        if (this.semaphore.busy() > 0)
+            return;
+        // A future deadline is a real reason to sit still: that timer will fire and pump.
+        for (const run of this.store.listRuns()) {
+            if (run.status !== 'failed' || !run.autoResumeAt)
+                continue;
+            const deadline = Date.parse(run.autoResumeAt);
+            if (Number.isFinite(deadline) && deadline > now)
+                return;
+        }
+        if (this.semaphore.accountHolds().inFlight.size === 0) {
+            // Not the hold, then — some other wakeup went missing. An ordinary pump is the whole fix,
+            // and it is idempotent, so this stays quiet.
+            void this.pump();
+            return;
+        }
+        console.warn('[cez] queue watchdog: work is queued, nothing is running, and the usage-limit hold has no'
+            + ' deadline behind it — starting the next task anyway');
+        this.forceNextPump = true;
+        void this.pump();
+    }
+    /**
+     * The accounts this project is currently holding: one key per run parked on a usage-limit
+     * resume that has not come due yet (spec 2026-08-03-auto-resume-after-usage-limit).
+     *
+     * Published to the shared semaphore so the hold spans PROJECTS — one Claude account can be
+     * driving tasks in three repos, and a limit closes it for all of them. Derived from the
+     * records on every ask rather than tracked as state: a deadline that passes, a resume that
+     * fires, a cancel, an archive and a delete all lift the hold with no bookkeeping.
+     *
+     * Deliberately excludes a deadline that has already passed — that run is about to resume, and
+     * holding the queue for it would only stall the very work the window reopened for.
+     */
+    accountHolds(now = Date.now()) {
+        const deadline = new Set();
+        const inFlight = new Set();
+        for (const run of this.store.listRuns()) {
+            // A holding run always carries the runner it actually ran on, so the fallback is unused
+            // here — it is spelled out rather than `!` so a future record shape degrades, not throws.
+            const key = () => runAccountKey(run, run.runner ?? 'claude');
+            if (run.status === 'failed' && run.autoResumeAt) {
+                const at = Date.parse(run.autoResumeAt);
+                if (Number.isFinite(at) && at > now)
+                    deadline.add(key());
+            }
+            else if (resumeInFlight(run)) {
+                inFlight.add(key());
+            }
+        }
+        return { deadline, inFlight };
+    }
+    /**
+     * The PER-TASK off switch (`DELETE /api/v1/runs/:id/auto-resume`, and the archive route):
+     * stop resuming THIS task, without touching the workspace setting or any other task.
+     *
+     * Idempotent — a run with nothing pending answers the same way, because "this task will not
+     * resume itself" is equally true either way. Returns false only when the run does not exist,
+     * which is the route's 404.
+     */
+    cancelAutoResume(runId) {
+        const run = this.store.getRun(runId);
+        if (!run)
+            return false;
+        const pending = run.autoResumeAt !== undefined || this.autoResumeTimers.has(runId);
+        this.clearAutoResume(runId);
+        if (pending) {
+            this.store.appendEvent(runId, {
+                type: 'note',
+                message: 'automatic resume cancelled for this task',
+            });
+            // This run may have been the last thing holding its account's queue — nothing else will
+            // notice, since the hold is derived and its release is not an event.
+            void this.pump();
+        }
+        return true;
+    }
+    /** Retire a pending resume — timer, deadline and counter. The counter goes too because every
+     *  caller is a fresh epoch: a human Continue, or a resume that re-stamps its own count. */
+    clearAutoResume(runId) {
+        const timer = this.autoResumeTimers.get(runId);
+        if (timer)
+            clearTimeout(timer);
+        this.autoResumeTimers.delete(runId);
+        const run = this.store.getRun(runId);
+        if (!run)
+            return;
+        if (run.autoResumeAt !== undefined || run.autoResumeAttempts !== undefined) {
+            this.store.updateRun(runId, { autoResumeAt: undefined, autoResumeAttempts: undefined });
+        }
+    }
     /** Reclaim finished worktrees beyond the keep-limit (#483) — directory only,
      *  `cez/<id8>` branch kept. Best-effort; a failure never affects run
      *  lifecycle. `review`/live runs are excluded by the selector. */
@@ -723,6 +1345,7 @@ export class RunManager {
         if (queuedAt >= 0) {
             this.queue.splice(queuedAt, 1);
             this.pendingJobs.delete(runId);
+            this.pendingContinuations.delete(runId);
             this.store.updateRun(runId, { status: 'cancelled', finishedAt: new Date().toISOString() });
             this.store.appendEvent(runId, { type: 'lifecycle', message: 'cancelled while queued' });
             return true;
@@ -763,35 +1386,13 @@ export class RunManager {
             .map((part) => part.trim())
             .filter((part) => part.length > 0)
             .join('\n\n');
-        const readImages = (urls, kind) => {
-            const images = [];
-            for (const url of urls) {
-                const name = url.split('/').pop();
-                if (!name || name.includes('..') || name.includes('/') || name.includes('\\'))
-                    continue;
-                try {
-                    const data = readFileSync(join(this.dataDir, 'runs', `${runId}-images`, name));
-                    images.push({
-                        type: 'image',
-                        source: { type: 'base64', media_type: mediaTypeFor(name), data: data.toString('base64') },
-                    });
-                }
-                catch {
-                    // Degrade, never fail the boot (AGENTS.md): the user deleted `.ai/cezar/`
-                    // or the file is unreadable — start with the text and say which image went.
-                    this.store.appendEvent(runId, {
-                        type: 'note',
-                        message: `${kind} attachment ${name} could not be read — starting without it`,
-                    });
-                }
-            }
-            return images;
-        };
         // Keep the original in-memory blocks for a live process (including the
         // best-effort case where persistence failed). Recovery has no such copy,
         // so rebuild it from the durable task-image URLs.
-        const images = input.images?.length ? input.images : readImages(run.taskImages ?? [], 'task');
-        const stackedImages = readImages(stack.flatMap((m) => m.images ?? []), 'queued');
+        const images = input.images?.length
+            ? input.images
+            : this.readPersistedImages(runId, run.taskImages ?? [], 'task').blocks;
+        const stackedImages = this.readPersistedImages(runId, stack.flatMap((m) => m.images ?? []), 'queued').blocks;
         return {
             ...input,
             task,
@@ -799,15 +1400,67 @@ export class RunManager {
             ...(stackedImages.length ? { stackedImages } : { stackedImages: undefined }),
         };
     }
+    /** Apply edits and messages made while a restart continuation waits for
+     * capacity. The durable record remains the source of truth, just as it is for
+     * an ordinary queued workflow (#472), so a second restart reconstructs and
+     * hydrates the same amendments instead of dropping them. */
+    hydrateQueuedContinuation(runId, continuation) {
+        const run = this.store.getRun(runId);
+        if (!run) {
+            return { ...continuation, persistedImages: [], persistedAttachments: [] };
+        }
+        const stack = run.queuedMessages ?? [];
+        const amendedTask = [run.task, ...stack.map((message) => message.text)]
+            .map((part) => part.trim())
+            .filter((part) => part.length > 0)
+            .join('\n\n');
+        const prompt = amendedTask
+            ? `${continuation.prompt}\n\nCurrent task and queued updates:\n\n${amendedTask}`
+            : continuation.prompt;
+        const persisted = this.readPersistedImages(runId, stack.flatMap((message) => message.images ?? []), 'queued');
+        return {
+            ...continuation,
+            prompt,
+            persistedImages: persisted.blocks,
+            persistedAttachments: persisted.attachments,
+        };
+    }
+    readPersistedImages(runId, urls, kind) {
+        const blocks = [];
+        const attachments = [];
+        for (const url of urls) {
+            const name = url.split('/').pop();
+            if (!name || name.includes('..') || name.includes('/') || name.includes('\\'))
+                continue;
+            const path = join(this.dataDir, 'runs', `${runId}-images`, name);
+            try {
+                const data = readFileSync(path);
+                blocks.push({
+                    type: 'image',
+                    source: { type: 'base64', media_type: mediaTypeFor(name), data: data.toString('base64') },
+                });
+                attachments.push({ name, url, path });
+            }
+            catch {
+                // Degrade, never fail the boot (AGENTS.md): the user deleted `.ai/cezar/`
+                // or the file is unreadable — start with the text and say which image went.
+                this.store.appendEvent(runId, {
+                    type: 'note',
+                    message: `${kind} attachment ${name} could not be read — starting without it`,
+                });
+            }
+        }
+        return { blocks, attachments };
+    }
     /**
      * Still waiting for a slot? Checked against the engine's own queue rather than
      * the record's `status` (#472): the record is written by `execute()` a tick
      * after `pump()` dequeues, so a status read can see `queued` for a run that has
-     * already started. `pendingJobs` is deleted synchronously at dequeue, so it is
-     * the authoritative answer for "can this prompt still be amended".
+     * already started. The pending maps are deleted synchronously at dequeue, so
+     * they are the authoritative answer for "can this prompt still be amended".
      */
     isQueued(runId) {
-        return this.pendingJobs.has(runId);
+        return this.pendingJobs.has(runId) || this.pendingContinuations.has(runId);
     }
     /** Split `ContentBlock[]` into the persisted shape a stacked message holds. */
     toQueuedMessage(runId, content) {
@@ -987,6 +1640,18 @@ export class RunManager {
      * then offers "Continue" instead.
      */
     sendMessage(runId, content) {
+        const delivered = this.deliverMessage(runId, content, true);
+        if (delivered) {
+            const state = this.active.get(runId);
+            if (state)
+                state.monitoringWakeups = 0;
+            this.store.updateRun(runId, { monitoringWakeCapReached: undefined });
+        }
+        return delivered;
+    }
+    /** Shared live-session delivery. Synthetic scheduler prompts reuse lifecycle
+     * bookkeeping without masquerading as user-authored transcript messages. */
+    deliverMessage(runId, content, userAuthored) {
         const state = this.active.get(runId);
         if (!state?.session?.open || state.cancelled)
             return false;
@@ -997,27 +1662,32 @@ export class RunManager {
         // Persist the attached images so the thread can render them (not just count them) — the same
         // on-disk store + `/images/` route the agent's own screenshots use. `pasted` prefix marks
         // these as user attachments (vs. agent tool screenshots) on disk (#357).
-        const persisted = content
+        const persisted = userAuthored ? content
             .filter((b) => b.type === 'image')
             .map((b) => this.persistImage(runId, b.source.media_type, b.source.data, 'pasted'))
-            .filter((saved) => saved !== null);
+            .filter((saved) => saved !== null) : [];
         const images = persisted.map((saved) => saved.url);
-        this.store.appendEvent(runId, {
-            type: 'user-message',
-            stepId: state.currentStepId,
-            text,
-            imageCount: content.filter((b) => b.type === 'image').length,
-            images,
-        });
+        if (userAuthored) {
+            this.store.appendEvent(runId, {
+                type: 'user-message',
+                stepId: state.currentStepId,
+                text,
+                imageCount: content.filter((b) => b.type === 'image').length,
+                images,
+            });
+        }
         // Tell the agent where the pasted files live on disk (#357): the base64 blocks below still
         // ride along so the model can *view* them, but a real path is what lets it *operate* on them
         // (save, `cp`, attach to a GitHub issue/PR) — and it's the only usable reference on backends
         // (codex, opencode) that drop image blocks entirely before reaching the model.
-        const deliverable = persisted.length ? [...content, pastedAttachmentsNote(persisted)] : content;
+        const expanded = userAuthored ? expandRegistrySlashSkill(content, state.skills ?? []) : content;
+        const deliverable = persisted.length ? [...expanded, pastedAttachmentsNote(persisted)] : expanded;
         const delivered = state.session.sendMessage(deliverable);
         if (delivered) {
             this.clearIdleTimer(state);
+            this.clearMonitoringWakeTimer(state, runId);
             this.waiting.delete(runId); // resumed — the run counts against slots again
+            this.monitoring.delete(runId);
             // Clear any `monitoring` activity — the agent is actively working again
             // (spec 2026-07-18-subagent-monitoring-status, #490).
             this.store.updateRun(runId, { status: 'running', activity: undefined });
@@ -1054,7 +1724,13 @@ export class RunManager {
      * behaves exactly like an interactive step: `waiting` after each turn,
      * messages via sendMessage, closed by finish/idle/cancel.
      */
-    continueRun(runId, opts = {}) {
+    continueRun(runId, opts = {}, 
+    /** Restart recovery may discover several interrupted tasks at once. Those
+     *  continuations are queued; an explicit user Continue remains immediate. */
+    deferForCapacity = false) {
+        if (agentModelsLocked(this.repoRoot) && opts.model?.trim()) {
+            return { ok: false, error: AGENT_MODELS_LOCKED_ERROR };
+        }
         if (this.active.has(runId))
             return { ok: false, error: 'run is still active' };
         const run = this.store.getRun(runId);
@@ -1105,10 +1781,34 @@ export class RunManager {
                         : {}),
             });
         }
+        // Everything that could refuse this continuation has now passed, so a pending usage-limit
+        // resume is superseded either way: this IS that resume (it re-stamps its own counter), or a
+        // human got there first — and then the counter starts over, because the cap only exists to
+        // bound UNATTENDED resumes.
+        this.clearAutoResume(runId);
         const continuations = run.steps.filter((s) => s.id.startsWith('continue-')).length;
         const stepId = `continue-${continuations + 1}`;
         this.store.addStep(runId, { id: stepId, name: 'Continue', kind: 'agent' });
-        void this.runContinuation(runId, stepId, resume ? sessionStep.sessionId : undefined, targetRunner, opts.text?.trim() || 'Continue.', opts.images ?? []).catch((err) => {
+        const prompt = opts.text?.trim() || 'Continue.';
+        const images = opts.images ?? [];
+        if (deferForCapacity) {
+            this.pendingContinuations.set(runId, {
+                stepId,
+                sessionId: resume ? sessionStep.sessionId : undefined,
+                backend: targetRunner,
+                prompt,
+                images,
+            });
+            this.queue.push(runId);
+            this.store.updateRun(runId, {
+                status: 'queued',
+                error: undefined,
+                finishedAt: undefined,
+                currentStepId: undefined,
+            });
+            return { ok: true };
+        }
+        void this.runContinuation(runId, stepId, resume ? sessionStep.sessionId : undefined, targetRunner, prompt, images).catch((err) => {
             const message = err instanceof Error ? err.message : String(err);
             this.store.updateRun(runId, {
                 status: 'failed',
@@ -1123,7 +1823,11 @@ export class RunManager {
     /** Screenshots pasted into the follow-up composer — delivered with the
      *  reopened session's opening message, exactly like a live-session
      *  message's attachments. */
-    images = []) {
+    images = [], 
+    /** Queued-message screenshots were persisted when they were enqueued and
+     *  reconstructed at dequeue. Keep them separate from fresh `images` so
+     *  opening a recovered continuation does not persist duplicate files. */
+    persistedImages = [], persistedAttachments = []) {
         // Continuation runs in the task's worktree when it still exists (spec
         // 006) — the resumed session sees exactly what the original run left.
         // Retention (#483) may have reclaimed this run's worktree directory while
@@ -1141,6 +1845,7 @@ export class RunManager {
             : this.repoRoot;
         const state = { cancelled: false, interrupt: () => undefined, cwd };
         this.active.set(runId, state);
+        this.starting.delete(runId);
         if (state.cwd === this.repoRoot) {
             this.store.appendEvent(runId, {
                 type: 'note',
@@ -1180,19 +1885,22 @@ export class RunManager {
         // images rather than a bare count, and handed to the agent BOTH as base64 blocks (so it can
         // view them) and as absolute paths appended to the prompt (so it can operate on them — and
         // because codex/opencode drop image blocks before they reach the model).
-        const attachments = images
+        const freshAttachments = images
             .filter((b) => b.type === 'image')
             .map((b) => this.persistImage(runId, b.source.media_type, b.source.data, 'pasted'))
             .filter((saved) => saved !== null);
+        const openingImages = [...images, ...persistedImages];
+        const attachments = [...freshAttachments, ...persistedAttachments];
         this.store.appendEvent(runId, {
             type: 'user-message',
             stepId,
             text: prompt,
-            imageCount: images.filter((b) => b.type === 'image').length,
+            imageCount: openingImages.filter((b) => b.type === 'image').length,
             ...(attachments.length ? { images: attachments.map((saved) => saved.url) } : {}),
         });
         let stepCost = 0;
         let turnText = '';
+        let sessionError;
         const sink = this.makeUiSink(runId, stepId);
         const onEvent = (event) => {
             if (event.type === 'image') {
@@ -1209,6 +1917,13 @@ export class RunManager {
                 return;
             }
             this.store.appendEvent(runId, { ...event, stepId });
+            if (event.type === 'error') {
+                sessionError ??= event.message;
+                state.session?.interrupt();
+                return;
+            }
+            if (sessionError)
+                return;
             if (event.type === 'session') {
                 this.store.updateStep(runId, stepId, { sessionId: event.sessionId, backend });
             }
@@ -1229,9 +1944,13 @@ export class RunManager {
                 const done = sessionOpen && DONE_MARKER_RE.test(turnText.trimEnd());
                 // `CEZ:ASK` → the user is genuinely blocked; wins over `CEZ:MONITORING`
                 // (a pending question is always attention), loses to `CEZ:DONE` (#473).
-                const ask = sessionOpen && !done ? parseAskMarker(turnText) : null;
+                const askResult = sessionOpen && !done ? parseAskMarkerResult(turnText) : undefined;
+                const ask = askResult?.kind === 'valid' ? askResult.request : null;
+                const askRejection = askResult ? askMarkerRejection(askResult) : undefined;
                 const monitoring = sessionOpen && !done && !ask && MONITORING_MARKER_RE.test(turnText.trimEnd());
                 turnText = '';
+                if (askRejection)
+                    this.store.appendEvent(runId, { type: 'note', message: askRejection, stepId });
                 if (done) {
                     // Goal achieved (agent contract, #347) — same as in runAgentStep.
                     this.store.appendEvent(runId, { type: 'lifecycle', message: 'goal achieved — session closed' });
@@ -1267,16 +1986,30 @@ export class RunManager {
                         if (monitoring) {
                             this.store.updateRun(runId, { status: 'running', activity: 'monitoring' });
                             this.store.updateStep(runId, stepId, { status: 'running' });
+                            this.monitoring.add(runId);
+                            this.clearIdleTimer(state);
+                            this.armMonitoringWakeTimer(runId, state);
                         }
                         else {
                             this.store.updateRun(runId, { status: 'waiting', activity: undefined });
                             this.store.updateStep(runId, stepId, { status: 'waiting' });
+                            this.monitoring.delete(runId);
+                            this.clearMonitoringWakeTimer(state, runId);
                         }
                         this.waiting.add(runId);
-                        this.armIdleTimer(runId, state);
+                        if (!monitoring)
+                            this.armIdleTimer(runId, state);
                         this.releaseSlot();
                     }
                 }
+                // A turn that completed is the ONLY evidence the provider's window actually reopened, so
+                // it is what retires the consecutive-resume counter — which in turn releases the account
+                // hold for every other task queued behind it (spec
+                // 2026-08-03-auto-resume-after-usage-limit). `settleSuccess` does the same for a run that
+                // finishes outright; this covers the far more common "parked for the user" ending.
+                if (this.store.getRun(runId)?.autoResumeAttempts !== undefined) {
+                    this.store.updateRun(runId, { autoResumeAttempts: undefined });
+                }
                 appendHandoffHeartbeat(this.dataDir, runId, `turn complete — status=${monitoring ? 'monitoring' : sessionOpen ? 'waiting' : 'running'}`);
             }
         };
@@ -1291,7 +2024,7 @@ export class RunManager {
         // instead of `opus`). Fail loud here too rather than let the backend pick a default.
         let continueModel;
         try {
-            const normalized = normalizeModelForBackend(continueBackend, record?.model);
+            const normalized = normalizeModelForBackend(continueBackend, agentModelsLocked(this.repoRoot) ? undefined : record?.model, { configuredProvider: await configuredModelProvider(continueBackend, state.cwd) });
             continueModel = normalized?.backendModel;
             this.store.updateRun(runId, {
                 modelIdentity: normalized ? formatModelIdentity(normalized.identity) : undefined,
@@ -1320,18 +2053,31 @@ export class RunManager {
             this.dropActive(runId);
             return;
         }
+        // Resuming reattaches to a session that lives inside ONE account's config dir, so the
+        // continuation must run under the account that created it — not whatever the project has
+        // been switched to since. The owning step is the one carrying this session id.
+        const resumedProfileId = sessionId === undefined
+            ? undefined
+            : record?.steps.find((s) => s.sessionId === sessionId)?.profileId;
+        const continueProfile = await this.agentEnvForStep(runId, continueBackend, {
+            generateFollowups,
+            recordedProfileId: resumedProfileId,
+        });
+        this.store.updateStep(runId, stepId, { profileId: continueProfile.profileId });
         const runner = createRunner(continueBackend);
+        state.currentStepId = stepId;
+        this.beginUsageInvocation(runId, state, stepId);
         const session = runner.startSession({
             // The Continue step is a fresh agent session on the same run — the
             // run's extra system prompt (already resolved at execute time and
             // echoed on the record) rides along with the handoff contract.
             systemPrompt: composeSystemPrompt(record?.systemPrompt, generateFollowups ? HANDOFF_INSTRUCTIONS : HANDOFF_ONLY_INSTRUCTIONS),
             userPrompt: attachments.length ? `${prompt}\n\n${pastedAttachmentsText(attachments)}` : prompt,
-            ...(images.length ? { images } : {}),
+            ...(openingImages.length ? { images: openingImages } : {}),
             cwd: state.cwd,
             allowedTools: DEFAULT_ALLOWED_TOOLS,
             additionalDirectories: [join(this.dataDir, 'runs')],
-            env: this.agentEnv(runId, generateFollowups),
+            env: continueProfile.env,
             model: continueModel,
             sessionId,
             resume: sessionId !== undefined,
@@ -1340,13 +2086,14 @@ export class RunManager {
         state.session = session;
         state.sessionEverOpened = true;
         this.flushDeferred(runId);
-        state.currentStepId = stepId;
         state.interrupt = () => session.interrupt();
         if (session.pid !== undefined)
             registerRunProcess(runId, session.pid);
         const finishedAt = () => new Date().toISOString();
         try {
             await session.result;
+            if (sessionError)
+                throw new Error(sessionError);
             sink.sessionEnded(state.cancelled ? 'cancelled' : 'end_turn');
             if (state.cancelled) {
                 this.store.updateStep(runId, stepId, { status: 'cancelled', finishedAt: finishedAt() });
@@ -1399,6 +2146,12 @@ export class RunManager {
         // the config default. Per-step `runner` can still override it below.
         const config = await loadConfig(this.repoRoot);
         const taskBackend = input.runner ?? config.defaultRunner;
+        // The account may have gone into a usage-limit hold since this run was dequeued — the queue
+        // gate cannot be the only one, because dequeue is not the moment of no return. Nothing has
+        // happened yet here, so the run goes back to the queue untouched (spec
+        // 2026-08-03-auto-resume-after-usage-limit).
+        if (this.requeueWhileHeld(runId, workflow, input, taskBackend))
+            return;
         // Extra system prompt (R2 2.3): POST override > config default; echoed on
         // the record so the UI/API can show what the run actually used.
         const extraSystemPrompt = resolveExtraSystemPrompt(input.systemPrompt, config.systemPrompt);
@@ -1408,7 +2161,7 @@ export class RunManager {
         // can still override below); the authoritative fail-loud gate is at spawn.
         let modelIdentity;
         try {
-            const normalized = normalizeModelForBackend(taskBackend, input.model);
+            const normalized = normalizeModelForBackend(taskBackend, agentModelsLocked(this.repoRoot) ? undefined : input.model, { configuredProvider: await configuredModelProvider(taskBackend, this.repoRoot) });
             modelIdentity = normalized ? formatModelIdentity(normalized.identity) : undefined;
         }
         catch {
@@ -1429,11 +2182,20 @@ export class RunManager {
         // established; only explicit opt-out and non-Git modes run in place.
         const repo = await getRepoInfo(this.repoRoot);
         if (repo && input.worktree === false) {
-            // Composer opt-out: run in the repo working tree, no branch/worktree
-            // (read-only skills — review, summarize — that never need isolation).
+            // Composer opt-out: run in the repo working tree, no branch/worktree. The
+            // repository-root lease serializes these runs so workflows cannot overlap.
+            // Pin the starting commit: the session's Changes and Commits views use it
+            // as their stable lower bound while reading the current working copy.
+            const startingCommit = await getHeadCommit(repo.root);
+            if (startingCommit)
+                this.store.updateRun(runId, { baseBranch: startingCommit });
             emit({ type: 'note', message: 'worktree off — running in the repo working tree' });
         }
         else if (repo) {
+            emit({
+                type: 'note',
+                message: `worktree on — using an isolated task worktree (${input.worktree === true ? 'explicit request' : 'default'})`,
+            });
             // Fork from the configured base branch (config.json `baseBranch`, e.g.
             // `develop`) — also the target of the eventual draft PR. Unresolvable
             // (typo, not fetched) → note + the currently checked-out branch.
@@ -1501,6 +2263,13 @@ export class RunManager {
             // below breaks on `cancelled` before touching the tree and settles the
             // run through the usual path.
             await this.acquireRepoRoot(runId, state);
+            // THE window that matters for an in-place run. Waiting for the exclusive tree can take
+            // minutes, and a run parked on that lease holds no slot (#347) — so the queue keeps
+            // advancing behind it and the dequeue-time gate is long past. Measured with five in-place
+            // tasks and `maxParallel: 2`: four of them started. Re-ask here, where the very next thing
+            // is a spawn, and hand the run back to the queue if the account closed meanwhile.
+            if (this.requeueWhileHeld(runId, workflow, input, taskBackend, state))
+                return;
         }
         // Handoff journal (spec 007) — seeded after the worktree exists so the
         // header can name the branch. Idempotent: an existing file stays as-is.
@@ -1508,6 +2277,7 @@ export class RunManager {
         if (seeded)
             seedHandoffFile(this.dataDir, seeded);
         const skills = await discoverSkills(this.repoRoot);
+        state.skills = skills;
         const retriesUsed = new Map();
         let checkFailure = null;
         let runError = null;
@@ -1687,6 +2457,7 @@ export class RunManager {
         const startTokens = stepRecord?.tokensUsed ?? 0;
         let stepCost = stepRecord?.costUsd ?? 0;
         let turnText = '';
+        let sessionError;
         const sink = this.makeUiSink(runId, step.id);
         const onEvent = (event) => {
             if (event.type === 'image') {
@@ -1703,6 +2474,13 @@ export class RunManager {
                 return;
             }
             emit({ ...event, stepId: step.id });
+            if (event.type === 'error') {
+                sessionError ??= event.message;
+                state.session?.interrupt();
+                return;
+            }
+            if (sessionError)
+                return;
             if (event.type === 'session') {
                 // Codex/OpenCode mint their own session id — persist it so resume works.
                 this.store.updateStep(runId, step.id, { sessionId: event.sessionId, backend });
@@ -1723,13 +2501,17 @@ export class RunManager {
                 const done = interactive && sessionOpen && DONE_MARKER_RE.test(turnText.trimEnd());
                 // `CEZ:ASK` → the user is blocked; wins over `CEZ:MONITORING`, loses to
                 // `CEZ:DONE` (#473).
-                const ask = interactive && sessionOpen && !done ? parseAskMarker(turnText) : null;
+                const askResult = interactive && sessionOpen && !done ? parseAskMarkerResult(turnText) : undefined;
+                const ask = askResult?.kind === 'valid' ? askResult.request : null;
+                const askRejection = askResult ? askMarkerRejection(askResult) : undefined;
                 const monitoring = interactive &&
                     sessionOpen &&
                     !done &&
                     !ask &&
                     MONITORING_MARKER_RE.test(turnText.trimEnd());
                 turnText = '';
+                if (askRejection)
+                    emit({ type: 'note', stepId: step.id, message: askRejection });
                 if (done) {
                     // Goal achieved (agent contract, #347): close the session instead
                     // of parking at `waiting` — the run completes and frees its slot.
@@ -1752,15 +2534,25 @@ export class RunManager {
                     if (monitoring) {
                         this.store.updateRun(runId, { status: 'running', activity: 'monitoring' });
                         this.store.updateStep(runId, step.id, { status: 'running' });
+                        this.monitoring.add(runId);
+                        this.clearIdleTimer(state);
+                        this.armMonitoringWakeTimer(runId, state);
                     }
                     else {
                         this.store.updateRun(runId, { status: 'waiting', activity: undefined });
                         this.store.updateStep(runId, step.id, { status: 'waiting' });
+                        this.monitoring.delete(runId);
+                        this.clearMonitoringWakeTimer(state, runId);
                     }
                     this.waiting.add(runId);
-                    this.armIdleTimer(runId, state);
+                    if (!monitoring)
+                        this.armIdleTimer(runId, state);
                     this.releaseSlot(); // the freed slot can start a queued run right away — in any project
                 }
+                // The window is proven open — see the twin in `runContinuation`.
+                if (this.store.getRun(runId)?.autoResumeAttempts !== undefined) {
+                    this.store.updateRun(runId, { autoResumeAttempts: undefined });
+                }
                 // Cez's own heartbeat — the handoff stays current even when the
                 // agent forgets to write (spec 007).
                 appendHandoffHeartbeat(this.dataDir, runId, `turn complete — status=${monitoring ? 'monitoring' : waiting ? 'waiting' : 'running'}`);
@@ -1773,7 +2565,7 @@ export class RunManager {
         // instead of letting the backend silently substitute its default.
         let backendModel;
         try {
-            const normalized = normalizeModelForBackend(stepBackend, step.model ?? input.model);
+            const normalized = normalizeModelForBackend(stepBackend, agentModelsLocked(this.repoRoot) ? undefined : step.model ?? input.model, { configuredProvider: await configuredModelProvider(stepBackend, state.cwd) });
             backendModel = normalized?.backendModel;
             // Persist the identity of what ACTUALLY runs (#405, review M1). The run-start echo
             // (line ~993) is best-effort from `taskBackend`/`input.model`; a per-step `runner`/`model`
@@ -1788,8 +2580,17 @@ export class RunManager {
                 return err.message;
             throw err;
         }
+        // Which agent account this step spawns under, and — recorded on the step before the spawn —
+        // which one its session belongs to. `sessionId` and `profileId` are a pair: a resume that
+        // reads the wrong account's config dir finds no session and silently starts a fresh one.
+        const stepProfile = await this.agentEnvForStep(runId, stepBackend, {
+            generateFollowups: followupsEnabled() && input.generateFollowups !== false,
+        });
+        this.store.updateStep(runId, step.id, { profileId: stepProfile.profileId });
         const runner = createRunner(stepBackend);
         let session;
+        state.currentStepId = step.id;
+        this.beginUsageInvocation(runId, state, step.id);
         try {
             session = runner.startSession({
                 // Skill body, then the run's extra prompt (POST override or config
@@ -1804,7 +2605,7 @@ export class RunManager {
                 bashAllowlist: step.bashAllowlist,
                 // The handoff file lives outside the worktree — grant access.
                 additionalDirectories: [join(this.dataDir, 'runs')],
-                env: this.agentEnv(runId, followupsEnabled() && input.generateFollowups !== false),
+                env: stepProfile.env,
                 model: backendModel,
                 sessionId,
                 // Interactive sessions have no wall clock — the idle timer rules.
@@ -1815,6 +2616,7 @@ export class RunManager {
             });
         }
         catch (err) {
+            state.currentStepId = undefined;
             return err instanceof Error ? err.message : String(err);
         }
         state.session = session;
@@ -1826,6 +2628,10 @@ export class RunManager {
             registerRunProcess(runId, session.pid);
         try {
             const result = await session.result;
+            if (sessionError) {
+                sink.sessionEnded('error', sessionError);
+                return sessionError;
+            }
             // v2 counterpart of v1's `done` (spec: the mappers leave session-close
             // events to the RunManager — only it knows how the session settled).
             sink.sessionEnded(state.cancelled ? 'cancelled' : 'end_turn');
@@ -1840,6 +2646,9 @@ export class RunManager {
         finally {
             this.recordUsagePeaks(runId);
             this.clearIdleTimer(state);
+            this.monitoring.delete(runId);
+            this.waiting.delete(runId);
+            this.clearMonitoringWakeTimer(state, runId);
             state.session = undefined;
             state.currentStepId = undefined;
             state.interrupt = () => undefined;
@@ -1863,16 +2672,90 @@ export class RunManager {
     /** Native backend asks arrive before turn-end. Persist and park immediately
      * so the cockpit shows attention and the run releases its workspace slot. */
     handleRunnerUiEvent(runId, state, sink, event) {
+        this.recordUsageUiEvent(runId, state, event);
         sink.handle(event);
         if (event.type !== 'ask.requested' || state.cancelled)
             return;
         this.clearIdleTimer(state);
+        this.monitoring.delete(runId);
+        this.clearMonitoringWakeTimer(state, runId);
         this.waiting.add(runId);
         this.store.updateRun(runId, { status: 'waiting', activity: undefined });
         if (state.currentStepId)
             this.store.updateStep(runId, state.currentStepId, { status: 'waiting' });
         this.releaseSlot();
     }
+    /** Persist the invocation checkpoint before launching a runner. A throw or
+     * process exit before `turn.started` therefore leaves a durable mismatch. */
+    beginUsageInvocation(runId, state, stepId) {
+        const step = this.store.getRun(runId)?.steps.find((candidate) => candidate.id === stepId);
+        if (!step)
+            return;
+        const epoch = (step.usageInvocationEpoch ?? 0) + 1;
+        this.persistUsageCheckpoint(runId, stepId, {
+            usageInvocationEpoch: epoch,
+            usageInvocationsStarted: (step.usageInvocationsStarted ?? 0) + 1,
+        });
+        state.usageInvocation = {
+            stepId,
+            epoch,
+            observed: false,
+            startedTurns: new Set(),
+            recordedTurns: new Set(),
+        };
+    }
+    /** Fold backend-neutral completed-turn usage into the current step exactly
+     * once. Invocation/turn counters are written before the event reaches the
+     * NDJSON sink so crashes cannot preserve a falsely complete subtotal. */
+    recordUsageUiEvent(runId, state, event) {
+        const invocation = state.usageInvocation;
+        if (!invocation)
+            return;
+        const step = this.store.getRun(runId)?.steps.find((candidate) => candidate.id === invocation.stepId);
+        if (!step)
+            return;
+        if (event.type === 'turn.started') {
+            if (invocation.startedTurns.has(event.turnId))
+                return;
+            invocation.startedTurns.add(event.turnId);
+            const firstObservedTurn = !invocation.observed;
+            invocation.observed = true;
+            this.persistUsageCheckpoint(runId, invocation.stepId, {
+                usageTurnsStarted: (step.usageTurnsStarted ?? 0) + 1,
+                ...(firstObservedTurn
+                    ? { usageInvocationsObserved: (step.usageInvocationsObserved ?? 0) + 1 }
+                    : {}),
+            });
+            return;
+        }
+        if (event.type !== 'turn.completed')
+            return;
+        if (!invocation.startedTurns.has(event.turnId) || invocation.recordedTurns.has(event.turnId))
+            return;
+        const input = event.usage?.input;
+        const output = event.usage?.output;
+        if (typeof input !== 'number' ||
+            !Number.isFinite(input) ||
+            input < 0 ||
+            typeof output !== 'number' ||
+            !Number.isFinite(output) ||
+            output < 0) {
+            return;
+        }
+        invocation.recordedTurns.add(event.turnId);
+        this.persistUsageCheckpoint(runId, invocation.stepId, {
+            inputTokens: (step.inputTokens ?? 0) + input,
+            outputTokens: (step.outputTokens ?? 0) + output,
+            usageTurnsRecorded: (step.usageTurnsRecorded ?? 0) + 1,
+        });
+    }
+    /** Usage completeness is a crash boundary, unlike high-frequency token
+     * snapshots: the checkpoint must reach `runs.json` before the runner starts
+     * or the matching UI event is persisted and forwarded. */
+    persistUsageCheckpoint(runId, stepId, patch) {
+        this.store.updateStep(runId, stepId, patch);
+        this.store.flush();
+    }
     /**
      * Turn-end bookkeeping (#389), shared by `runAgentStep` and
      * `runContinuation` — called (fire-and-forget) from every `turn-end` event:
@@ -1939,7 +2822,12 @@ export class RunManager {
             // deliberately NEVER a title source; see maybeRefreshTitle below. The
             // one exception is an explicit CEZ:TITLE declaration (applied above).
             if (run.worktreePath && existsSync(run.worktreePath)) {
-                const stat = await worktreeShortstat(run.worktreePath, run.baseBranch ?? 'HEAD');
+                // `taskBranch` is what keeps this number *this task's* (#751): a review/QA run
+                // repoints the worktree onto the branch under review, and without the branch to
+                // compare HEAD against, the stat would claim that whole branch's diff.
+                const stat = await worktreeShortstat(run.worktreePath, run.baseBranch ?? 'HEAD', {
+                    taskBranch: run.branch,
+                });
                 if (stat)
                     this.store.updateRun(runId, { diffStat: stat });
                 else
@@ -2044,6 +2932,10 @@ export class RunManager {
             status: review ? 'review' : 'done',
             finishedAt: new Date().toISOString(),
             currentStepId: undefined,
+            // A run that got all the way to a settled turn is not in a limit loop, so the resume
+            // counter starts over — otherwise a task that legitimately met the limit once a week would
+            // creep toward the cap forever and stop resuming for no reason anyone could see.
+            autoResumeAttempts: undefined,
         });
         this.store.appendEvent(runId, {
             type: 'lifecycle',
@@ -2096,7 +2988,10 @@ export class RunManager {
                     throw err;
                 }
                 this.queuedImageSeq.set(runId, seq);
-                return { name, url: `/api/runs/${runId}/images/${name}`, path };
+                // Versioned, because that is the only surface served now. The cockpit still upgrades
+                // the unversioned URLs sitting in OLD transcripts when it renders them
+                // (`resolveApiUrl`), but a URL minted today must be fetchable as written.
+                return { name, url: `/api/v1/runs/${runId}/images/${name}`, path };
             }
             return null;
         }
@@ -2123,6 +3018,68 @@ export class RunManager {
             state.idleTimer = undefined;
         }
     }
+    reconcileMonitoringWakeTimers() {
+        for (const runId of this.monitoring) {
+            const state = this.active.get(runId);
+            if (state)
+                this.armMonitoringWakeTimer(runId, state);
+        }
+    }
+    armMonitoringWakeTimer(runId, state) {
+        const minutes = this.semaphore.monitoringWakeIntervalMinutes();
+        if (minutes === null) {
+            this.clearMonitoringWakeTimer(state, runId);
+            return;
+        }
+        if ((state.monitoringWakeups ?? 0) >= MAX_AUTO_CONTINUES) {
+            this.clearMonitoringWakeTimer(state, runId);
+            if (!this.store.getRun(runId)?.monitoringWakeCapReached) {
+                this.store.updateRun(runId, { monitoringWakeCapReached: true });
+                this.store.appendEvent(runId, {
+                    type: 'note',
+                    message: `automatic monitoring wake-up cap reached (${MAX_AUTO_CONTINUES}); session remains parked`,
+                });
+            }
+            return;
+        }
+        if (state.monitoringWakeTimer && state.monitoringWakeIntervalMinutes === minutes)
+            return;
+        this.clearMonitoringWakeTimer(state, runId);
+        state.monitoringWakeIntervalMinutes = minutes;
+        this.store.updateRun(runId, { monitoringWakeCapReached: undefined });
+        const deadline = Date.now() + minutes * 60_000;
+        this.store.updateRun(runId, { monitoringWakeAt: new Date(deadline).toISOString() });
+        state.monitoringWakeTimer = setTimeout(() => {
+            state.monitoringWakeTimer = undefined;
+            this.store.updateRun(runId, { monitoringWakeAt: undefined });
+            if (!this.monitoring.has(runId) || !state.session?.open || state.cancelled)
+                return;
+            const wakeups = state.monitoringWakeups ?? 0;
+            if (wakeups >= MAX_AUTO_CONTINUES) {
+                this.store.updateRun(runId, { monitoringWakeCapReached: true });
+                this.store.appendEvent(runId, {
+                    type: 'note',
+                    message: `automatic monitoring wake-up cap reached (${MAX_AUTO_CONTINUES}); session remains parked`,
+                });
+                return;
+            }
+            state.monitoringWakeups = wakeups + 1;
+            this.store.appendEvent(runId, {
+                type: 'note',
+                message: `automatic monitoring wake-up (${state.monitoringWakeups}/${MAX_AUTO_CONTINUES})`,
+            });
+            this.deliverMessage(runId, [{ type: 'text', text: MONITORING_WAKE_NUDGE }], false);
+        }, Math.max(0, deadline - Date.now()));
+        state.monitoringWakeTimer.unref?.();
+    }
+    clearMonitoringWakeTimer(state, runId) {
+        if (state.monitoringWakeTimer)
+            clearTimeout(state.monitoringWakeTimer);
+        state.monitoringWakeTimer = undefined;
+        state.monitoringWakeIntervalMinutes = undefined;
+        if (runId)
+            this.store.updateRun(runId, { monitoringWakeAt: undefined });
+    }
     /** Autosave-commit the worktree every 90 s while the run lives (spec 006).
      *  Opt-in via CEZ_AUTOSAVE=1 (#471) — see periodicAutosaveEnabled. */
     armAutosave(state) {
@@ -2244,4 +3201,35 @@ export function skillSystemPrompt(skill) {
     lines.push('', 'Skill instructions:', skill.body.trim());
     return lines.join('\n');
 }
+/**
+ * Expand a registry-backed slash skill in a live chat message before it reaches
+ * a backend. Claude otherwise intercepts an unknown leading slash command, and
+ * Codex/OpenCode have no native slash-skill lookup at all (#676).
+ *
+ * Only the first text block is eligible, only at character zero, and unknown
+ * commands pass through byte-for-byte. The caller persists the original user
+ * content before applying this delivery-only rewrite.
+ */
+export function expandRegistrySlashSkill(content, skills) {
+    const textIndex = content.findIndex((block) => block.type === 'text');
+    if (textIndex < 0)
+        return content;
+    const block = content[textIndex];
+    if (!block || block.type !== 'text')
+        return content;
+    const match = /^\/([A-Za-z0-9][A-Za-z0-9._-]*)(?=\s|$)/.exec(block.text);
+    if (!match)
+        return content;
+    const skill = skills.find((candidate) => candidate.name === match[1]);
+    if (!skill)
+        return content;
+    const request = block.text.slice(match[0].length).trim();
+    const replacement = {
+        type: 'text',
+        text: request ? `${skillSystemPrompt(skill)}\n\nUser request:\n${request}` : skillSystemPrompt(skill),
+    };
+    const expanded = [...content];
+    expanded[textIndex] = replacement;
+    return expanded;
+}
 //# sourceMappingURL=run.js.map