@ours.network/fleet 0.19.0-nightly.9 → 1.0.1

package/dist/runner.d.ts

@@ -35,6 +35,15 @@ export interface RunnerDeps {
 }
 /** Environment injected only into the managed harness process. */
 export declare function managedFleetProxyEnv(role: ResolvedRole, stateDir: string): Record<string, string>;
+/**
+ * The environment a managed harness child actually receives, checked at the one
+ * point where it is composed. `role.env` deliberately wins over harness prep,
+ * which is exactly how a stale fleet-wide model pin used to outrank the model
+ * the role was spawned with — so the model pin is verified here rather than
+ * trusted, and a disagreement stops the launch instead of being reported as a
+ * success (see src/model-env.ts).
+ */
+export declare function harnessChildEnv(role: ResolvedRole, launchEnv: Record<string, string> | undefined, stateDir: string): Record<string, string>;
 /**
  * Record who owns wake delivery for this run. Returning true means a fleet
  * monitor is taking ownership back from a native harness and must start at the
@@ -61,6 +70,23 @@ export declare function readExitRecord(path: string): ExitRecord | null;
 export declare const RESTART_LEDGER_FILE = ".restart-ledger.json";
 /** Consecutive immediate failures tolerated before the agent is held down. */
 export declare const RESTART_FAIL_THRESHOLD = 5;
+/**
+ * How the previous supervisor process ended.
+ *
+ * `abrupt` is the case the ledger used to miss entirely: an OOM-kill or any
+ * other external signal takes the supervisor down before it can write anything,
+ * the service manager restarts the unit, and every durable indicator still
+ * describes the run that died. A health check reading them reported "no
+ * restarts" for a role that had died and come back.
+ */
+export interface TerminationRecord {
+    class: 'clean' | 'abrupt' | 'unknown';
+    detail: string;
+    /** When the SURVIVING process observed it, not when it happened. */
+    observedAt: string;
+    /** Start time of the run that ended, when it was recorded. */
+    runStartedAt?: string;
+}
 export interface RestartLedger {
     version: 1;
     consecutiveImmediateFailures: number;
@@ -72,7 +98,29 @@ export interface RestartLedger {
     updatedAt: string;
     /** When the circuit opened, for the held-down status line. */
     openedAt?: string;
+    /** How the previous supervisor process ended, including abnormal exits. */
+    lastTermination?: TerminationRecord;
+    /** Supervisor processes that died without closing their run marker. */
+    abruptTerminations?: number;
+    /** Start of the supervisor run that owns this state directory now. */
+    supervisorStartedAt?: string;
 }
+/**
+ * Carried across a supervisor process's life so its successor can tell an
+ * orderly exit from a kill. Present on disk == "a supervisor believed it was
+ * running"; the next start finding one that is not its own is proof the
+ * previous process died without getting to write anything.
+ */
+export declare const RUN_MARKER_FILE = ".supervisor-run.json";
+/**
+ * Claim this state directory for the current supervisor process and report how
+ * the previous one ended. Runs BEFORE the first attempt, which is the whole
+ * point: after an abrupt kill nothing else writes until an attempt finishes,
+ * and an attempt can take minutes.
+ */
+export declare function claimSupervisorRun(dir: string, startedAt: string, pid?: number): TerminationRecord;
+/** Orderly exit: the successor must not read this run as a kill. */
+export declare function releaseSupervisorRun(dir: string): void;
 /** Bounded exponential backoff for the nth consecutive immediate failure. */
 export declare function backoffFor(consecutiveFailures: number): number;
 /** Read a role's restart ledger; a missing or corrupt one starts clean. */

package/dist/runner.js

@@ -25,6 +25,7 @@ import { RoleTurnArbiter } from './session/arbiter.js';
 import { ScheduledLoopManager, } from './loops/manager.js';
 import { FLEET_PROXY_CALLER_ENV, FLEET_PROXY_STATE_DIR_ENV, inheritCallerSpawnDefaults, } from './fleet-proxy.js';
 import { effectivePermissionMode } from './permissions.js';
+import { assertModelPinReachesChild, effectiveRoleModel, repinModelEnv } from './model-env.js';
 import { archiveTempState, markTempSupervisorActive, requestedTempStopReason, } from './temp-lifecycle.js';
 const defaultDeps = () => ({
     tmux: new Tmux(),
@@ -60,14 +61,32 @@ const defaultDeps = () => ({
     },
 });
 const MONITOR_OWNER_FILE = '.monitor-owner';
+/** Fleet roles consume the operator-owned daemon; a role session never starts it. */
+const FLEET_OURS_AUTOSTART = '0';
 /** Environment injected only into the managed harness process. */
 export function managedFleetProxyEnv(role, stateDir) {
     return {
         ...(role.env ?? {}),
+        // This must win over both inherited/configured auto-start. ACP agents run
+        // directly rather than through ours-codex, so the runner owns this fence.
+        OURS_AUTOSTART: FLEET_OURS_AUTOSTART,
         [FLEET_PROXY_STATE_DIR_ENV]: stateDir,
         [FLEET_PROXY_CALLER_ENV]: role.name,
     };
 }
+/**
+ * The environment a managed harness child actually receives, checked at the one
+ * point where it is composed. `role.env` deliberately wins over harness prep,
+ * which is exactly how a stale fleet-wide model pin used to outrank the model
+ * the role was spawned with — so the model pin is verified here rather than
+ * trusted, and a disagreement stops the launch instead of being reported as a
+ * success (see src/model-env.ts).
+ */
+export function harnessChildEnv(role, launchEnv, stateDir) {
+    const env = { ...(launchEnv ?? {}), ...managedFleetProxyEnv(role, stateDir) };
+    assertModelPinReachesChild(role, env);
+    return env;
+}
 /**
  * Execute a typed proxy request in the caller's supervisor. Dynamic imports
  * avoid a runner↔spawn initialization cycle (spawn imports runner constants).
@@ -105,13 +124,19 @@ async function executeManagedSpawn(caller, configPath, requested, log) {
         statePath,
         harness: preview.harness,
         session: preview.session,
-        ...(preview.model ? { model: preview.model } : {}),
+        // Read back from the resolved environment, not from the request: the banner
+        // must name the model the child will run, not the one that was asked for.
+        ...(effectiveRoleModel(preview) ? { model: effectiveRoleModel(preview) } : {}),
         monitor: { mode: preview.monitor.mode, interrupt: preview.monitor.interrupt },
+        permissionMode: effectivePermissionMode(preview),
         inherited,
         creationActionId,
     };
     log(`[${caller.name}] managed fleet proxy spawned ${result.lifetime} role ${result.role} `
-        + `harness=${result.harness} session=${result.session}`);
+        + `harness=${result.harness} session=${result.session} `
+        + `model=${result.model ?? '(harness default)'} `
+        + `permission=${result.permissionMode.fleetMode} `
+        + `native=${result.permissionMode.nativeMode}`);
     return result;
 }
 /**
@@ -141,6 +166,10 @@ export function recordMonitorOwner(dir, owner) {
 export function buildPaneCommand(launch, roleEnv, exitStatusPath, paneArgv = launch.argv) {
     const env = {
         PATH: process.env.PATH ?? '', COLORTERM: 'truecolor', ...launch.env, ...(roleEnv ?? {}),
+        // Tmux roles have the same daemon-client boundary as ACP roles. Keep this
+        // last so neither harness preparation nor a role env block can take over
+        // the shared daemon lifecycle.
+        OURS_AUTOSTART: FLEET_OURS_AUTOSTART,
     };
     // Interactive panes should advertise colour even when the supervisor itself
     // was launched with NO_COLOR. A role may still deliberately opt back in to
@@ -195,7 +224,7 @@ export function readExitRecord(path) {
     catch { /* fall through to unknown */ }
     return { version: 1, class: 'unknown', detail: `unreadable exit record: ${raw.slice(0, 120)}` };
 }
-// ─── Restart-loop containment (3.2) ──────────────────────────────────────────
+// ─── Restart-loop containment ──────────────────────────────────────────
 //
 // The child-session restart loop used to BE the service manager: systemd's
 // `Restart=always RestartSec=2` and launchd's `KeepAlive`. Neither can count,
@@ -219,6 +248,66 @@ const emptyLedger = () => ({
     circuit: 'closed',
     updatedAt: new Date(0).toISOString(),
 });
+/**
+ * Carried across a supervisor process's life so its successor can tell an
+ * orderly exit from a kill. Present on disk == "a supervisor believed it was
+ * running"; the next start finding one that is not its own is proof the
+ * previous process died without getting to write anything.
+ */
+export const RUN_MARKER_FILE = '.supervisor-run.json';
+function readRunMarker(dir) {
+    try {
+        const raw = JSON.parse(readFileSync(join(dir, RUN_MARKER_FILE), 'utf8'));
+        return raw.version === 1 && typeof raw.pid === 'number' && typeof raw.startedAt === 'string'
+            ? raw : undefined;
+    }
+    catch {
+        return undefined;
+    }
+}
+/**
+ * Claim this state directory for the current supervisor process and report how
+ * the previous one ended. Runs BEFORE the first attempt, which is the whole
+ * point: after an abrupt kill nothing else writes until an attempt finishes,
+ * and an attempt can take minutes.
+ */
+export function claimSupervisorRun(dir, startedAt, pid = process.pid) {
+    const previous = readRunMarker(dir);
+    const termination = previous && previous.pid !== pid
+        ? {
+            class: 'abrupt',
+            detail: `supervisor pid ${previous.pid} left an open run marker; `
+                + 'it was terminated without an orderly exit (signal, OOM-kill, or host reset)',
+            observedAt: startedAt,
+            runStartedAt: previous.startedAt,
+        }
+        : previous
+            ? { class: 'unknown', detail: 'run marker belongs to this process', observedAt: startedAt }
+            : { class: 'clean', detail: 'no previous run marker', observedAt: startedAt };
+    try {
+        mkdirSync(dir, { recursive: true });
+        writeFileSync(join(dir, RUN_MARKER_FILE), JSON.stringify({ version: 1, pid, startedAt }, null, 2) + '\n');
+    }
+    catch { /* diagnostics must never take the role down */ }
+    return termination;
+}
+/** Orderly exit: the successor must not read this run as a kill. */
+export function releaseSupervisorRun(dir) {
+    try {
+        rmSync(join(dir, RUN_MARKER_FILE), { force: true });
+    }
+    catch { /* best effort */ }
+}
+/**
+ * Fields that describe THIS process's history rather than the current failure
+ * streak. Clearing the streak (recovery, an operator `up`, an approved model
+ * transition) must not erase the record that the role died and came back.
+ */
+const carriedForward = (previous) => ({
+    ...(previous.lastTermination ? { lastTermination: previous.lastTermination } : {}),
+    ...(previous.abruptTerminations ? { abruptTerminations: previous.abruptTerminations } : {}),
+    ...(previous.supervisorStartedAt ? { supervisorStartedAt: previous.supervisorStartedAt } : {}),
+});
 /** Bounded exponential backoff for the nth consecutive immediate failure. */
 export function backoffFor(consecutiveFailures) {
     if (consecutiveFailures <= 0)
@@ -253,7 +342,10 @@ export function writeRestartLedger(dir, ledger) {
 export function resetRestartLedger(dir) {
     if (!existsSync(dir))
         return;
-    writeRestartLedger(dir, { ...emptyLedger(), updatedAt: new Date().toISOString() });
+    const previous = readRestartLedger(dir);
+    writeRestartLedger(dir, {
+        ...emptyLedger(), ...carriedForward(previous), updatedAt: new Date().toISOString(),
+    });
 }
 /** Filename spawnTemp writes into a temp agent dir to carry the fleet start-stagger. */
 export const START_STAGGER_FILE = '.start-stagger-ms';
@@ -384,11 +476,18 @@ export async function runOnce(name, opts = {}, partialDeps = {}) {
     const effectiveModel = effectiveModelForRole(dir, role);
     if (effectiveModel !== role.model) {
         deps.log(`[${name}] model recovery drift: declared=${role.model ?? '(none)'} effective=${effectiveModel}`);
-        role = { ...role, model: effectiveModel };
+        // The env pin has to move with it. A down-shift that changed only
+        // `role.model` was reported as a model change while the child kept running
+        // the model that had just failed, because the pin is what the harness reads.
+        role = { ...role, model: effectiveModel, env: repinModelEnv(role, effectiveModel) };
     }
     if (modelRecoveryHeld(dir))
         throw new Error(`[${name}] model chain exhausted — held down until config changes or recovery reset`);
     const adapter = getAdapter(role.harness);
+    // Say the running model out loud, once, from the resolved environment. The
+    // spawn banner is a claim made before the process exists; this is the log line
+    // that can be checked against the session afterwards.
+    deps.log(`[${name}] model: ${effectiveRoleModel(role) ?? '(harness default)'}`);
     mkdirSync(dir, { recursive: true });
     const rotation = rotateWorklog(join(dir, 'WORKLOG.md'), role.worklog);
     if (rotation.deferred)
@@ -403,8 +502,12 @@ export async function runOnce(name, opts = {}, partialDeps = {}) {
     const exitFile = join(dir, '.exit-status');
     const booted = existsSync(bootedFile);
     const mode = booted && adapter.supportsResume ? 'resume' : 'fresh';
-    if (mode === 'fresh')
-        writeFileSync(bootedFile, '');
+    // Stamp EVERY attempt, not just the first. `.booted` used to be written only
+    // on the fresh path, so after a restart — including one the supervisor never
+    // saw, like an OOM-kill — its mtime still read the original boot and any
+    // health check reading it reported "no restarts". The existence test above
+    // already ran, so rewriting cannot change the fresh/resume decision.
+    writeFileSync(bootedFile, `${new Date(deps.now()).toISOString()} ${mode}\n`);
     const runCwd = role.cwd && existsSync(role.cwd) ? role.cwd : dir;
     const prep = await adapter.prepareSession(role, { stateDir: dir, runCwd });
     const sessionBackend = role.session ?? 'tmux';
@@ -416,11 +519,11 @@ export async function runOnce(name, opts = {}, partialDeps = {}) {
         })()
         : adapter.buildLaunch(role, mode, { sessionId }, prep);
     // Isolation is additive: only roles that declare `isolation:` are wrapped. The
-    // env prefix + exit capture in buildPaneCommand stay host-side (see §5.3).
+    // env prefix + exit capture in buildPaneCommand stay host-side.
     let wrappedArgv = launch.argv;
     if (role.isolation) {
-        // Start with the SAME durable context config validation and doctor judged
-        // (5.2), then add the selected launch's exact runtime closure. Those paths
+        // Start with the same durable context that config validation and doctor judged,
+        // then add the selected launch's exact runtime closure. Those paths
         // still pass through resolveIsolation's canonical blocklist enforcement.
         const runtime = resolveLaunchRuntime(launch.argv);
         launch = { ...launch, argv: runtime.argv };
@@ -440,8 +543,8 @@ export async function runOnce(name, opts = {}, partialDeps = {}) {
             rmSync(degradedMarker, { force: true });
         }
         wrappedArgv = sel.backend.wrap(launch.argv, policy, ctx);
-        // Resource caps wrap the sandbox from OUTSIDE, at the pane's own cgroup scope
-        // (§5.4). Applies even when the sandbox degraded to none.
+        // Resource caps wrap the sandbox from outside, at the pane's own cgroup scope.
+        // This applies even when the sandbox degraded to none.
         const { argv: rprefix, warnings } = resourceArgs(policy.resources, deps.cpuDelegated());
         for (const w of warnings)
             deps.log(`[${name}] WARNING ${w}`);
@@ -462,7 +565,7 @@ export async function runOnce(name, opts = {}, partialDeps = {}) {
             await deps.sleep(wait);
         }
     }
-    // Supervisor mail monitor (design §1): prime the notification cursor at the
+    // Prime the supervisor mail monitor's notification cursor at the
     // stream tip BEFORE the session launches so no arrival is missed during boot
     // (backlog before the tip is the SessionStart hook's job). Native-mode roles
     // leave wake ownership to the harness. Temp snapshots predating `monitor:` are
@@ -525,12 +628,23 @@ export async function runOnce(name, opts = {}, partialDeps = {}) {
             name,
             argv: wrappedArgv,
             cwd: runCwd,
-            env: { ...launch.env, ...managedFleetProxyEnv(role, dir) },
+            env: harnessChildEnv(role, launch.env, dir),
             stateDir: dir,
             mode,
             permissions: perms,
             modeId: adapter.acpPermissionModeId?.(role),
+            // The role's declared MCP servers, and the bundled agent's `_meta`
+            // vocabulary for the options it takes no flag for. Both come from the
+            // ADAPTER and from `prep`: the ACP launch cannot carry `prep.argv`, so this
+            // is the route by which harness_options that used to be silently dropped
+            // for an ACP role actually reach the session.
+            mcpServers: adapter.acpMcpServers?.(role),
+            sessionMeta: adapter.acpSessionMeta?.(role, prep),
             permissionMode: effectivePermissionMode(role),
+            // Provenance travels with the exact ACP launch. Keeping it out of a
+            // role-only adapter hook prevents a PATH fallback or resolver skew from
+            // claiming metadata trust for an argv it did not authenticate.
+            permissionMetadataSource: launch.permissionMetadataSource,
             log: deps.log,
         });
         pid = acpSession.pid;
@@ -928,100 +1042,138 @@ export async function runSupervised(name, opts = {}, partialDeps = {}, attempt =
     mkdirSync(dir, { recursive: true });
     const shouldStop = deps.shouldStop ?? (() => false);
     const stamp = () => new Date(deps.now()).toISOString();
-    while (!shouldStop()) {
-        let ledger = readRestartLedger(dir);
-        try {
-            const configPath = resolveConfigPath(dir, opts.configPath);
-            const role = findRole(loadConfig(configPath), name);
-            reconcileModelRecovery(dir, role, stamp());
-            if (modelRecoveryHeld(dir)) {
+    // Record how the PREVIOUS supervisor process ended before doing anything
+    // else. An external kill writes nothing itself, and the next ledger write is
+    // an attempt away — which is why an OOM-kill used to leave every durable
+    // indicator describing the run that died.
+    const startedAt = stamp();
+    const termination = claimSupervisorRun(dir, startedAt);
+    {
+        const previous = readRestartLedger(dir);
+        const abrupt = (previous.abruptTerminations ?? 0) + (termination.class === 'abrupt' ? 1 : 0);
+        writeRestartLedger(dir, {
+            ...previous,
+            lastTermination: termination,
+            abruptTerminations: abrupt,
+            supervisorStartedAt: startedAt,
+            updatedAt: startedAt,
+        });
+        if (termination.class === 'abrupt')
+            deps.log(`[${name}] previous supervisor run (started ${termination.runStartedAt}) `
+                + `ended abruptly: ${termination.detail}; abrupt terminations recorded: ${abrupt}`);
+    }
+    try {
+        while (!shouldStop()) {
+            let ledger = readRestartLedger(dir);
+            try {
+                const configPath = resolveConfigPath(dir, opts.configPath);
+                const role = findRole(loadConfig(configPath), name);
+                reconcileModelRecovery(dir, role, stamp());
+                if (modelRecoveryHeld(dir)) {
+                    await deps.sleep(HELD_DOWN_POLL_MS);
+                    continue;
+                }
+            }
+            catch {
+                // Normal attempt path reports config errors through the restart circuit.
+            }
+            if (ledger.circuit === 'open') {
+                // Held down. Stay alive — exiting would hand the role straight back to
+                // the service manager — and watch for an operator reset.
                 await deps.sleep(HELD_DOWN_POLL_MS);
                 continue;
             }
-        }
-        catch {
-            // Normal attempt path reports config errors through the restart circuit.
-        }
-        if (ledger.circuit === 'open') {
-            // Held down. Stay alive — exiting would hand the role straight back to
-            // the service manager — and watch for an operator reset.
-            await deps.sleep(HELD_DOWN_POLL_MS);
-            continue;
-        }
-        let result;
-        try {
-            // Service-manager boot and automatic retry bypass the operator-facing
-            // up/restart commands. Reconcile again immediately before every real
-            // permanent harness attempt; the operation is idempotent and releases its
-            // provisioning lease before the agent or owner channel binds.
-            if (attempt === runOnce) {
-                const configPath = resolveConfigPath(dir, opts.configPath);
-                const role = findRole(loadConfig(configPath), name);
-                await reconcilePermanentRoleIdentities(role, undefined, deps.log);
+            let result;
+            try {
+                // Service-manager boot and automatic retry bypass the operator-facing
+                // up/restart commands. Reconcile again immediately before every real
+                // permanent harness attempt; the operation is idempotent and releases its
+                // provisioning lease before the agent or owner channel binds.
+                if (attempt === runOnce) {
+                    const configPath = resolveConfigPath(dir, opts.configPath);
+                    const role = findRole(loadConfig(configPath), name);
+                    await reconcilePermanentRoleIdentities(role, undefined, deps.log);
+                }
+                result = await attempt(name, { configPath: opts.configPath, allowResumeRotation: !ledger.resumeDiscarded }, deps);
             }
-            result = await attempt(name, { configPath: opts.configPath, allowResumeRotation: !ledger.resumeDiscarded }, deps);
-        }
-        catch (e) {
-            // A session that could not even start is an immediate failure like any
-            // other; it must count, or an unstartable role loops forever.
-            result = {
-                elapsedSecs: 0,
-                exit: { version: 1, class: 'unknown', detail: e instanceof Error ? e.message : String(e) },
-                rotated: false,
-                mode: 'fresh',
-            };
-        }
-        // Re-read: the attempt itself may have taken minutes, and an operator may
-        // have reset the ledger meanwhile.
-        ledger = readRestartLedger(dir);
-        if (result.modelRecovery === 'advance') {
-            writeRestartLedger(dir, {
-                ...emptyLedger(),
-                lastReason: 'approved model-chain transition',
-                updatedAt: stamp(),
-            });
-            continue;
-        }
-        if (result.modelRecovery === 'hold') {
-            await deps.sleep(HELD_DOWN_POLL_MS);
-            continue;
-        }
-        const fastFailSecs = fastFailSecsFor(name, opts.configPath);
-        const immediate = result.elapsedSecs < fastFailSecs;
-        if (!immediate) {
-            // A session that ran for a while is not a restart loop, whatever ended it.
-            writeRestartLedger(dir, {
-                ...emptyLedger(),
-                lastReason: result.exit.detail,
+            catch (e) {
+                // A session that could not even start is an immediate failure like any
+                // other; it must count, or an unstartable role loops forever.
+                result = {
+                    elapsedSecs: 0,
+                    exit: { version: 1, class: 'unknown', detail: e instanceof Error ? e.message : String(e) },
+                    rotated: false,
+                    mode: 'fresh',
+                };
+            }
+            // Re-read: the attempt itself may have taken minutes, and an operator may
+            // have reset the ledger meanwhile.
+            ledger = readRestartLedger(dir);
+            if (result.modelRecovery === 'advance') {
+                writeRestartLedger(dir, {
+                    ...emptyLedger(),
+                    ...carriedForward(ledger),
+                    lastReason: 'approved model-chain transition',
+                    updatedAt: stamp(),
+                });
+                continue;
+            }
+            if (result.modelRecovery === 'hold') {
+                await deps.sleep(HELD_DOWN_POLL_MS);
+                continue;
+            }
+            const fastFailSecs = fastFailSecsFor(name, opts.configPath);
+            // The fast-fail boundary starts a recovery episode; it must not also be
+            // the boundary that declares recovery successful. Otherwise alternating
+            // 19s and 20s deaths erase one another forever. Require the configured
+            // number of fast-fail windows to survive before closing an active streak.
+            // This hysteresis stays adapter-relative (100s for the current 20s/5-attempt
+            // policy) and still lets a genuinely sustained session reset the breaker.
+            const stableRecoverySecs = fastFailSecs * RESTART_FAIL_THRESHOLD;
+            const recoveryFailed = result.elapsedSecs < fastFailSecs
+                || (ledger.consecutiveImmediateFailures > 0 && result.elapsedSecs < stableRecoverySecs);
+            if (!recoveryFailed) {
+                // A session that ran for a while is not a restart loop, whatever ended it.
+                writeRestartLedger(dir, {
+                    ...emptyLedger(),
+                    ...carriedForward(ledger),
+                    lastReason: result.exit.detail,
+                    updatedAt: stamp(),
+                });
+                continue;
+            }
+            const failures = ledger.consecutiveImmediateFailures + 1;
+            const reason = `${result.exit.detail} after ${result.elapsedSecs.toFixed(1)}s`;
+            const next = {
+                version: 1,
+                ...carriedForward(ledger),
+                consecutiveImmediateFailures: failures,
+                lastReason: reason,
+                nextDelayMs: backoffFor(failures),
+                resumeDiscarded: ledger.resumeDiscarded || result.rotated,
+                circuit: failures >= RESTART_FAIL_THRESHOLD ? 'open' : 'closed',
                 updatedAt: stamp(),
-            });
-            continue;
-        }
-        const failures = ledger.consecutiveImmediateFailures + 1;
-        const reason = `${result.exit.detail} after ${result.elapsedSecs.toFixed(1)}s`;
-        const next = {
-            version: 1,
-            consecutiveImmediateFailures: failures,
-            lastReason: reason,
-            nextDelayMs: backoffFor(failures),
-            resumeDiscarded: ledger.resumeDiscarded || result.rotated,
-            circuit: failures >= RESTART_FAIL_THRESHOLD ? 'open' : 'closed',
-            updatedAt: stamp(),
-        };
-        if (next.circuit === 'open') {
-            next.openedAt = stamp();
-            next.nextDelayMs = 0;
+            };
+            if (next.circuit === 'open') {
+                next.openedAt = stamp();
+                next.nextDelayMs = 0;
+                writeRestartLedger(dir, next);
+                deps.log(`[${name}] HELD DOWN after ${failures} immediate failures at ${next.openedAt} — ` +
+                    `${reason}; the agent will not be restarted until: ours-fleet restart ${name}`);
+                continue;
+            }
             writeRestartLedger(dir, next);
-            deps.log(`[${name}] HELD DOWN after ${failures} immediate failures at ${next.openedAt} — ` +
-                `${reason}; the agent will not be restarted until: ours-fleet restart ${name}`);
-            continue;
+            deps.log(`[${name}] immediate failure ${failures}/${RESTART_FAIL_THRESHOLD} (${reason}) ` +
+                `-> backing off ${next.nextDelayMs}ms`);
+            await deps.sleep(next.nextDelayMs);
         }
-        writeRestartLedger(dir, next);
-        deps.log(`[${name}] immediate failure ${failures}/${RESTART_FAIL_THRESHOLD} (${reason}) ` +
-            `-> backing off ${next.nextDelayMs}ms`);
-        await deps.sleep(next.nextDelayMs);
+        return readRestartLedger(dir);
+    }
+    finally {
+        // Only an orderly return through this loop clears the marker; a signal or
+        // an OOM-kill leaves it, which is exactly how the successor detects them.
+        releaseSupervisorRun(dir);
     }
-    return readRestartLedger(dir);
 }
 /**
  * How short an attempt has to be to count as immediate. The role's harness