synergyspec-selfevolving 2.1.0 → 2.1.2

package/dist/core/self-evolution/episode-orchestrator.js

@@ -27,6 +27,18 @@ import { reconcilePrediction, summarizeCalibration, } from './policy/prediction-
 export async function captureMainArm(opts) {
     const sample = opts.report.fitnessSample;
     const facts = sample?.trajectoryFacts;
+    // ④ Observable degrade: a verified:false arm — whether because NO
+    // observed-trajectory facts were captured, OR facts exist but the runner's
+    // pass/fail was not derivable (`facts.verified !== true`) — is surfaced on
+    // stderr so it is never SILENT. A wedged/missing/unextractable runner is the
+    // most common loop-stall cause, and a silent false reads identically to a real
+    // miss. The arm's recorded `verified` collapses to false in BOTH cases (see the
+    // `objective.verified` below), so EVERY verified:false arm warns exactly once;
+    // a genuinely verified arm (`facts.verified === true`) stays quiet.
+    if (!facts || facts.verified !== true) {
+        // eslint-disable-next-line no-console
+        console.warn(`[episode-orchestrator] observed grading unavailable for change "${opts.changeName}" — recording verified:false (observed run not verified)`);
+    }
     // Honesty: prefer the OBSERVED pass rate (a real runner ran), else the
     // authored test-report summary; null when neither parsed (never fabricated) —
     // the exact precedence learn.ts uses to compute the loss.
@@ -80,14 +92,30 @@ export async function captureMainArm(opts) {
  * Whether the episode SKIPS the rollback decision + 演进智能体 EVOLVING AGENT:
  * the judge 弃权 abstained (no nameable gap), found no gaps, OR returned the
  * ⑤ `insufficient-signal` verdict (within the A/A noise floor, or a blocked
- * tamper). All three mean "do not evolve on this episode".
+ * tamper). These mean "do not evolve on this episode".
+ *
+ * EXCEPTION (cold-start bootstrap): `insufficient-signal` is honored ONLY as a
+ * genuine can't-tell — a within-noise-floor result (the baseline ran) or a blocked
+ * tamper. On a baseline-skipped episode (policyVersions.baseline === null) there is
+ * no comparison to be uncertain about, so a stray `insufficient-signal` emitted
+ * alongside real gaps (and no tamper) must NOT block: the first v0→v1 evolution has
+ * to be reachable from absolute signal, or a fresh target stays at v0 forever. This
+ * is defense-in-depth behind {@link deriveSingleSampleVerdict}, which already drops
+ * a volunteered verdict to `undefined` on a baseline skip.
  */
 function shouldSkipEvolution(diagnosis) {
     if (diagnosis === null)
         return true;
-    return (diagnosis.abstained ||
-        diagnosis.gaps.length === 0 ||
-        diagnosis.verdict === 'insufficient-signal');
+    if (diagnosis.abstained || diagnosis.gaps.length === 0)
+        return true;
+    if (diagnosis.verdict === 'insufficient-signal') {
+        const baselineSkipped = diagnosis.policyVersions?.baseline === null;
+        const tamper = diagnosis.integrity?.testTamperSuspected ?? false;
+        if (baselineSkipped && !tamper)
+            return false;
+        return true;
+    }
+    return false;
 }
 /**
  * Count the consecutive trailing rolled-back episodes in the 版本账本 ledger.
@@ -152,6 +180,17 @@ function deriveEpisodeId(changeName, now) {
         .replace(/-{2,}/g, '-')
         .replace(/^-+|-+$/g, '');
 }
+/**
+ * Build the `terminalError` note for a thrown step. A timeout reads identically
+ * to a hard crash on disk otherwise, so a message that names a host-agent timeout
+ * (the spawn timeout puts `headless agent timed out after Nms` into stderr → the
+ * error message) is PREFIXED with a `timeout:` marker. A timed-out episode is
+ * then distinguishable from a genuine crash in episode.json. Pure.
+ */
+function terminalErrorLabel(err) {
+    const msg = err instanceof Error ? err.message : String(err);
+    return /timed out/i.test(msg) ? `timeout: ${msg}` : msg;
+}
 /**
  * Run ONE episode through the loop in the strict, durably-persisted order
  * documented at the top of this module. See {@link RunEpisodeResult}.
@@ -251,149 +290,180 @@ async function runEpisodeAfterCreate(opts) {
         objective: opts.mainArm.objective,
     });
     await advanceEpisodeStage({ repoRoot, episodeId, stage: 'main-arm-captured' });
-    // ── d: CRITIC AGENT（基线智能体 baseline agent）or skip ───────────────────────
-    const shouldCritic = await shouldRunCriticAgent({ repoRoot, targetId });
-    if (shouldCritic.run && shouldCritic.baselineVersion !== null) {
-        // runCriticAgent advances the episode to 'baseline-arm-captured'.
-        await runCriticAgent({
-            repoRoot,
-            targetId,
-            changeName: opts.changeName,
-            episodeId,
-            baselineVersion: shouldCritic.baselineVersion,
-            ...(opts.critic?.baselineMode ? { baselineMode: opts.critic.baselineMode } : {}),
-            spawn: opts.spawn,
-        });
-    }
-    else {
-        baselineSkipped = true;
-        await advanceEpisodeStage({
-            repoRoot,
-            episodeId,
-            stage: 'baseline-skipped',
-            patch: { baselineSkippedReason: shouldCritic.reason },
-        });
-    }
-    // ── e: 奖励智能体 REWARD AGENT — score + diagnosis.json + advance 'scored' ────
-    // ④ Compute the tamper hint first (unless tamperCheck is 'off'); the ensemble
-    // injects it into the judge prompt + the diagnosis integrity, and BLOCK mode
-    // forces an insufficient-signal verdict (no extra spawns at the default
-    // samples=1, flag-only).
-    const tamperMode = opts.reward?.tamperCheck ?? 'flag';
-    const integrityHint = tamperMode === 'off' ? null : await detectTestTamper({ changeDirPath: opts.changeDirPath });
-    const reward = await runRewardAgentEnsemble({
-        repoRoot,
-        episodeId,
-        spawn: opts.spawn,
-        ...(opts.reward ? { reward: opts.reward } : {}),
-        integrityHint,
-    });
-    const diagnosis = reward.diagnosis;
-    advantage = diagnosis.advantage;
-    // 预测校准: settle the PRIOR 'evolve' step's checkable prediction against this
-    // episode's measured arm delta (verifiable per-metric main−baseline, NOT the
-    // judge's reward) and log the residual. Best-effort — a calibration miss must
-    // never fail the episode.
+    // Steps d–g spawn the three agents (any of which may THROW — a wedged/crashed
+    // host CLI, a never-validating reward output, an evolving-agent invocation
+    // error). A thrown agent step must record a DURABLE terminal 'errored' stage
+    // (with the error text) so the episode is never orphaned at 'kept'/'scored'
+    // (indistinguishable from a still-running episode — ses_1330/1331). The lock
+    // release stays in runEpisode's finally; the re-throw below reaches it.
     try {
-        await reconcilePrediction({ repoRoot, targetId, episodeId });
-    }
-    catch {
-        // best-effort: the prediction-reconcile ledger is advisory only
-    }
-    // ── f: DECISION (every step durably persisted before the next) ───────────────
-    if (shouldSkipEvolution(diagnosis)) {
-        // 弃权 abstained / no nameable gap / insufficient-signal → no rollback
-        // decision; SKIP evolution. A BLOCKED tamper records a reject-buffer entry so
-        // the gamed attempt is visible to future episodes' 演进智能体.
-        if (tamperMode === 'block' &&
-            diagnosis.integrity?.testTamperSuspected &&
-            diagnosis.verdict === 'insufficient-signal') {
-            const head = await currentPolicyVersion(repoRoot, targetId);
-            await appendRejectBufferEntry(repoRoot, {
-                schemaVersion: 1,
-                at: new Date().toISOString(),
-                episodeId,
-                targetId,
-                // No version moved (the main arm's tampered tests are not a policy edit);
-                // record at the current head so the entry is informational, not a rollback.
-                fromVersion: head ?? 0,
-                toVersion: head ?? 0,
-                advantage: diagnosis.advantage,
-                rewardMain: diagnosis.rewardMain,
-                rewardBaseline: diagnosis.rewardBaseline,
-                textualGradientTried: diagnosis.textualGradient ?? '',
-                editSummary: buildRejectEditSummary(diagnosis),
-                reason: 'tamper-suspected',
-            });
-        }
-        decision = 'abstained';
-        await advanceEpisodeStage({ repoRoot, episodeId, stage: 'abstained' });
-    }
-    else {
-        const badAdvantage = advantage !== null && advantage < threshold;
-        const ep = await readEpisode(repoRoot, episodeId);
-        const headBeforeRollback = await currentPolicyVersion(repoRoot, targetId);
-        // Resolve the rollback target: the policy the CRITIC AGENT reran
-        // (`policyVersionBaseline`) when it is a valid EARLIER version, else the
-        // version immediately before the head (the prior good policy the bad edit
-        // advanced past). `rollbackPolicyVersion` requires `toVersion < head`.
-        const rollbackTarget = resolveRollbackTarget(ep.policyVersionBaseline, headBeforeRollback);
-        if (badAdvantage && rollbackTarget !== null && headBeforeRollback !== null) {
-            // (i) ROLLBACK first — durable on disk. `advantage` is non-null inside the
-            // badAdvantage branch; `?? undefined` satisfies the optional `number` param.
-            await rollbackPolicyVersion({
+        // ── d: CRITIC AGENT（基线智能体 baseline agent）or skip ───────────────────────
+        const shouldCritic = await shouldRunCriticAgent({ repoRoot, targetId });
+        if (shouldCritic.run && shouldCritic.baselineVersion !== null) {
+            // runCriticAgent advances the episode to 'baseline-arm-captured'.
+            await runCriticAgent({
                 repoRoot,
                 targetId,
+                changeName: opts.changeName,
                 episodeId,
-                toVersion: rollbackTarget,
-                advantage: advantage ?? undefined,
+                baselineVersion: shouldCritic.baselineVersion,
+                ...(opts.critic?.baselineMode ? { baselineMode: opts.critic.baselineMode } : {}),
+                spawn: opts.spawn,
+                ...(opts.agentTimeoutMs !== undefined ? { timeoutMs: opts.agentTimeoutMs } : {}),
+                ...(opts.harness ? { harness: opts.harness } : {}),
             });
-            // (ii) THEN append the 否决缓冲 reject-buffer entry — durable on disk —
-            // BEFORE the 演进智能体 EVOLVING AGENT is even called, so the entry written
-            // THIS episode is in its fresh-from-disk prompt.
-            const rejectEntry = {
-                schemaVersion: 1,
-                at: new Date().toISOString(),
-                episodeId,
-                targetId,
-                // fromVersion = the version we rolled back TO (the prior good policy);
-                // toVersion = the (now rolled-back) version the rejected edit reached.
-                fromVersion: rollbackTarget,
-                toVersion: headBeforeRollback,
-                advantage,
-                rewardMain: diagnosis.rewardMain,
-                rewardBaseline: diagnosis.rewardBaseline,
-                textualGradientTried: diagnosis.textualGradient ?? '',
-                editSummary: buildRejectEditSummary(diagnosis),
-                reason: 'bad-advantage',
-            };
-            await appendRejectBufferEntry(repoRoot, rejectEntry);
-            decision = 'rolled-back';
-            await advanceEpisodeStage({ repoRoot, episodeId, stage: 'rolled-back' });
         }
         else {
-            // Good advantage, OR no earlier version to roll back to (e.g. head is v0):
-            // keep the current head.
-            decision = 'kept';
-            await advanceEpisodeStage({ repoRoot, episodeId, stage: 'kept' });
+            baselineSkipped = true;
+            await advanceEpisodeStage({
+                repoRoot,
+                episodeId,
+                stage: 'baseline-skipped',
+                patch: { baselineSkippedReason: shouldCritic.reason },
+            });
         }
-        // ── g: ONLY AFTER (f) persisted 'rolled-back'/'kept' ───────────────────────
-        // runEvolvingAgent reads the reject-buffer FRESH from disk (the entry just
-        // written THIS episode is in its prompt). Never parallelized with (f).
-        // 步长: after a rollback, shrink the edit budget (smaller step after a step
-        // that lost ground). 预测校准: pass the proposer's recent prediction record.
-        const scheduledBudget = decision === 'rolled-back'
-            ? scheduledEditBudget(await readPolicyLedger(repoRoot, targetId), editBudget)
-            : editBudget;
-        const calibrationNote = await summarizeCalibration(repoRoot, targetId);
-        evolution = await runEvolvingAgent({
+        // ── e: 奖励智能体 REWARD AGENT — score + diagnosis.json + advance 'scored' ────
+        // ④ Compute the tamper hint first (unless tamperCheck is 'off'); the ensemble
+        // injects it into the judge prompt + the diagnosis integrity, and BLOCK mode
+        // forces an insufficient-signal verdict (no extra spawns at the default
+        // samples=1, flag-only).
+        const tamperMode = opts.reward?.tamperCheck ?? 'flag';
+        const integrityHint = tamperMode === 'off' ? null : await detectTestTamper({ changeDirPath: opts.changeDirPath });
+        const reward = await runRewardAgentEnsemble({
             repoRoot,
             episodeId,
-            targetId,
-            editBudget: scheduledBudget,
-            ...(calibrationNote ? { calibrationNote } : {}),
             spawn: opts.spawn,
+            ...(opts.reward ? { reward: opts.reward } : {}),
+            integrityHint,
+            ...(opts.agentTimeoutMs !== undefined ? { timeoutMs: opts.agentTimeoutMs } : {}),
+            ...(opts.harness ? { harness: opts.harness } : {}),
         });
+        const diagnosis = reward.diagnosis;
+        advantage = diagnosis.advantage;
+        // 预测校准: settle the PRIOR 'evolve' step's checkable prediction against this
+        // episode's measured arm delta (verifiable per-metric main−baseline, NOT the
+        // judge's reward) and log the residual. Best-effort — a calibration miss must
+        // never fail the episode.
+        try {
+            await reconcilePrediction({ repoRoot, targetId, episodeId });
+        }
+        catch {
+            // best-effort: the prediction-reconcile ledger is advisory only
+        }
+        // ── f: DECISION (every step durably persisted before the next) ───────────────
+        if (shouldSkipEvolution(diagnosis)) {
+            // 弃权 abstained / no nameable gap / insufficient-signal → no rollback
+            // decision; SKIP evolution. A BLOCKED tamper records a reject-buffer entry so
+            // the gamed attempt is visible to future episodes' 演进智能体.
+            if (tamperMode === 'block' &&
+                diagnosis.integrity?.testTamperSuspected &&
+                diagnosis.verdict === 'insufficient-signal') {
+                const head = await currentPolicyVersion(repoRoot, targetId);
+                await appendRejectBufferEntry(repoRoot, {
+                    schemaVersion: 1,
+                    at: new Date().toISOString(),
+                    episodeId,
+                    targetId,
+                    // No version moved (the main arm's tampered tests are not a policy edit);
+                    // record at the current head so the entry is informational, not a rollback.
+                    fromVersion: head ?? 0,
+                    toVersion: head ?? 0,
+                    advantage: diagnosis.advantage,
+                    rewardMain: diagnosis.rewardMain,
+                    rewardBaseline: diagnosis.rewardBaseline,
+                    textualGradientTried: diagnosis.textualGradient ?? '',
+                    editSummary: buildRejectEditSummary(diagnosis),
+                    reason: 'tamper-suspected',
+                });
+            }
+            decision = 'abstained';
+            await advanceEpisodeStage({ repoRoot, episodeId, stage: 'abstained' });
+        }
+        else {
+            const badAdvantage = advantage !== null && advantage < threshold;
+            const ep = await readEpisode(repoRoot, episodeId);
+            const headBeforeRollback = await currentPolicyVersion(repoRoot, targetId);
+            // Resolve the rollback target: the policy the CRITIC AGENT reran
+            // (`policyVersionBaseline`) when it is a valid EARLIER version, else the
+            // version immediately before the head (the prior good policy the bad edit
+            // advanced past). `rollbackPolicyVersion` requires `toVersion < head`.
+            const rollbackTarget = resolveRollbackTarget(ep.policyVersionBaseline, headBeforeRollback);
+            if (badAdvantage && rollbackTarget !== null && headBeforeRollback !== null) {
+                // (i) ROLLBACK first — durable on disk. `advantage` is non-null inside the
+                // badAdvantage branch; `?? undefined` satisfies the optional `number` param.
+                await rollbackPolicyVersion({
+                    repoRoot,
+                    targetId,
+                    episodeId,
+                    toVersion: rollbackTarget,
+                    advantage: advantage ?? undefined,
+                });
+                // (ii) THEN append the 否决缓冲 reject-buffer entry — durable on disk —
+                // BEFORE the 演进智能体 EVOLVING AGENT is even called, so the entry written
+                // THIS episode is in its fresh-from-disk prompt.
+                const rejectEntry = {
+                    schemaVersion: 1,
+                    at: new Date().toISOString(),
+                    episodeId,
+                    targetId,
+                    // fromVersion = the version we rolled back TO (the prior good policy);
+                    // toVersion = the (now rolled-back) version the rejected edit reached.
+                    fromVersion: rollbackTarget,
+                    toVersion: headBeforeRollback,
+                    advantage,
+                    rewardMain: diagnosis.rewardMain,
+                    rewardBaseline: diagnosis.rewardBaseline,
+                    textualGradientTried: diagnosis.textualGradient ?? '',
+                    editSummary: buildRejectEditSummary(diagnosis),
+                    reason: 'bad-advantage',
+                };
+                await appendRejectBufferEntry(repoRoot, rejectEntry);
+                decision = 'rolled-back';
+                await advanceEpisodeStage({ repoRoot, episodeId, stage: 'rolled-back' });
+            }
+            else {
+                // Good advantage, OR no earlier version to roll back to (e.g. head is v0):
+                // keep the current head.
+                decision = 'kept';
+                await advanceEpisodeStage({ repoRoot, episodeId, stage: 'kept' });
+            }
+            // ── g: ONLY AFTER (f) persisted 'rolled-back'/'kept' ───────────────────────
+            // runEvolvingAgent reads the reject-buffer FRESH from disk (the entry just
+            // written THIS episode is in its prompt). Never parallelized with (f).
+            // 步长: after a rollback, shrink the edit budget (smaller step after a step
+            // that lost ground). 预测校准: pass the proposer's recent prediction record.
+            const scheduledBudget = decision === 'rolled-back'
+                ? scheduledEditBudget(await readPolicyLedger(repoRoot, targetId), editBudget)
+                : editBudget;
+            const calibrationNote = await summarizeCalibration(repoRoot, targetId);
+            evolution = await runEvolvingAgent({
+                repoRoot,
+                episodeId,
+                targetId,
+                editBudget: scheduledBudget,
+                ...(calibrationNote ? { calibrationNote } : {}),
+                spawn: opts.spawn,
+                ...(opts.agentTimeoutMs !== undefined ? { timeoutMs: opts.agentTimeoutMs } : {}),
+                ...(opts.harness ? { harness: opts.harness } : {}),
+            });
+        }
+    }
+    catch (err) {
+        // A thrown agent step (d–g) records a DURABLE terminal 'errored' stage so the
+        // episode is never orphaned at 'kept'/'scored' (which is indistinguishable
+        // from a still-running episode — the ses_1330/1331 wsgidav orphan). The
+        // 'errored' stage is reachable from EVERY non-terminal stage; the patch
+        // carries the error text (prefixed `timeout:` when the throw was a host-agent
+        // timeout, so a timed-out episode is distinguishable from a hard crash on
+        // disk). Best-effort: a failed record write must not mask the original throw,
+        // which still propagates to runEpisode's finally (lock release).
+        await advanceEpisodeStage({
+            repoRoot,
+            episodeId,
+            stage: 'errored',
+            patch: { terminalError: terminalErrorLabel(err) },
+        }).catch(() => { });
+        throw err;
     }
     // ── h (stage half): advance 'closed' (best-effort) ───────────────────────────
     await closeEpisodeBestEffort(repoRoot, episodeId);
@@ -401,11 +471,25 @@ async function runEpisodeAfterCreate(opts) {
     return { episodeId, baselineSkipped, advantage, decision, evolution, newPolicyVersion };
 }
 /**
- * Advance the episode to 'closed' from whatever terminal-ish stage it reached
- * (evolved | evolution-refused | abstained), best-effort: a stage that cannot
- * legally reach 'closed' (e.g. the evolving agent was not-spawned, leaving the
- * episode at 'kept'/'rolled-back') is left as-is rather than throwing, so the
- * close never masks the real episode outcome.
+ * Advance the episode to 'closed' from whatever terminal-ish stage it reached,
+ * best-effort.
+ *
+ * Closable stages:
+ *   - evolved | evolution-refused | abstained — the 演进智能体 EVOLVING AGENT
+ *     reached a definite outcome (or the judge 弃权 abstained), the normal close.
+ *   - kept | rolled-back — the 演进智能体 returned not-spawned (its diagnosis
+ *     abstained-after-gap-check, no gaps, or the target resolved to no editable
+ *     local files), so the episode never advanced past the decision. By the time
+ *     this runs (AFTER runEvolvingAgent returned), a stage still at 'kept'/
+ *     'rolled-back' can ONLY mean not-spawned — a success advances 'evolved', a
+ *     refusal advances 'evolution-refused', and a throw is caught upstream and
+ *     records 'errored' + rethrows so this close is never reached. So a leftover
+ *     kept/rolled-back at close time IS the finished-nothing-to-evolve case and
+ *     must close, not rest forever at a non-terminal stage (the exact ambiguity
+ *     the 'errored' stage was meant to remove).
+ *
+ * Any other (genuinely non-closable) stage is left as-is rather than throwing, so
+ * the close never masks the real episode outcome.
  */
 async function closeEpisodeBestEffort(repoRoot, episodeId) {
     const ep = await readEpisode(repoRoot, episodeId);
@@ -413,6 +497,9 @@ async function closeEpisodeBestEffort(repoRoot, episodeId) {
         'evolved',
         'evolution-refused',
         'abstained',
+        // not-spawned 演进智能体 leaves the episode here — close the finished episode.
+        'kept',
+        'rolled-back',
     ]);
     if (closable.has(ep.stage)) {
         await advanceEpisodeStage({ repoRoot, episodeId, stage: 'closed' });
@@ -501,6 +588,16 @@ async function ensureRejectBufferEntry(repoRoot, opts) {
  *   - 'scored'                    → run the decision (f) then the 演进智能体 (g).
  *   - 'rolled-back' / 'kept'      → run the 演进智能体 EVOLVING AGENT (g) then close.
  *   - 'evolved'/'evolution-refused'/'abstained' → close.
+ *   - 'errored'                   → RE-DRIVE from the last GOOD pre-error stage
+ *                                   (an episode may have errored on a TRANSIENT
+ *                                   cause — a one-off git/analyzer/agent timeout).
+ *                                   The pre-error stage is the last `stageHistory`
+ *                                   entry that is NOT 'errored'; when it is one of
+ *                                   {'scored','rolled-back','kept'} (the
+ *                                   resume-entry stages) we advance errored → that
+ *                                   stage and fall through to the normal dispatch.
+ *                                   Otherwise the pre-error stage is not
+ *                                   auto-resumable and the episode is reported as-is.
  *   - earlier stages              → not auto-resumable here (the arms / reward
  *                                   agent need their own re-entry); reported as-is.
  *
@@ -516,78 +613,135 @@ export async function resumeEpisode(opts) {
     const resumedFrom = ep.stage;
     const targetId = ep.targetId;
     let evolution = null;
-    if (ep.stage === 'scored') {
-        // Re-run the decision (f) from the on-disk diagnosis, then (g).
-        const diagnosis = await readDiagnosisForResume(repoRoot, episodeId);
-        if (shouldSkipEvolution(diagnosis)) {
-            await advanceEpisodeStage({ repoRoot, episodeId, stage: 'abstained' });
+    // The effective stage we dispatch on. Normally the episode's current stage;
+    // for an 'errored' episode we attempt to RE-DRIVE from the last good pre-error
+    // stage (a transient git/analyzer/agent failure should be retryable via an
+    // operator resume). 'errored' stays terminal for every OTHER caller — only this
+    // resume path may re-drive it, via the errored → {scored,rolled-back,kept}
+    // transitions the stage machine allows ONLY for operator recovery.
+    let stage = ep.stage;
+    if (ep.stage === 'errored') {
+        const preError = [...ep.stageHistory]
+            .reverse()
+            .find((h) => h.stage !== 'errored')?.stage;
+        if (preError === 'scored' ||
+            preError === 'rolled-back' ||
+            preError === 'kept') {
+            // Re-open the errored episode at its last auto-resumable stage, then fall
+            // through to the normal dispatch for that stage.
+            await advanceEpisodeStage({ repoRoot, episodeId, stage: preError });
+            stage = preError;
         }
-        else if (diagnosis) {
-            // (shouldSkipEvolution returns true for null, so diagnosis is non-null here)
-            const advantage = diagnosis.advantage;
-            const badAdvantage = advantage !== null && advantage < threshold;
-            // Crash-resume dedup: a 'rollback' ledger entry whose episodeId is THIS
-            // episode means runEpisode already applied the rollback before the host
-            // crashed (the rollback head is monotonic — re-calling rollbackPolicyVersion
-            // would stack a SECOND, duplicate rollback version). When present, reuse its
-            // recorded version axis and SKIP the re-rollback; only ensure the
-            // reject-buffer entry + the 'rolled-back' stage advance complete.
-            const ledger = await readPolicyLedger(repoRoot, targetId);
-            const priorRollback = ledger.find((e) => e.action === 'rollback' && e.episodeId === episodeId);
-            if (priorRollback) {
-                // The prior rollback already advanced to `priorRollback.version`, rolling
-                // FORWARD to the content of the version immediately before the rejected
-                // edit's head. Reconstruct the reject-buffer axis from that entry:
-                //   toVersion   = the (rolled-back) version the rejected edit reached
-                //               = priorRollback.version - 1 (the head before the rollback)
-                //   fromVersion = the prior good policy restored (one before that head)
-                const toVersion = priorRollback.version - 1;
-                const fromVersion = resolveRollbackTarget(ep.policyVersionBaseline, toVersion);
-                await ensureRejectBufferEntry(repoRoot, {
-                    episodeId,
-                    targetId,
-                    fromVersion: fromVersion ?? toVersion,
-                    toVersion,
-                    advantage,
-                    diagnosis,
-                });
-                await advanceEpisodeStage({ repoRoot, episodeId, stage: 'rolled-back' });
+        // Otherwise the pre-error stage is not auto-resumable (e.g. a reward throw at
+        // 'baseline-skipped'); leave the episode at 'errored' and report it as-is.
+    }
+    // The decision (f) + 演进智能体 EVOLVING AGENT (g) re-runs below can THROW — a
+    // wedged/crashed host CLI (CanonicalProposerInvocationError), a timeout, or an
+    // observed-GREEN gate throw. UNCAUGHT, that leaves the episode DURABLY stuck at
+    // a non-terminal stage ('scored'/'rolled-back'/'kept' — the orphan state fix ❷
+    // eliminates for runEpisode). Record the SAME terminal 'errored' stage here
+    // (the transition map already allows scored/rolled-back/kept → 'errored'), then
+    // re-throw. Resume holds NO in-flight lock, so this is a durable-stage fix, not
+    // a leak fix. Best-effort write: a failed record must not mask the original throw.
+    try {
+        if (stage === 'scored') {
+            // Re-run the decision (f) from the on-disk diagnosis, then (g).
+            const diagnosis = await readDiagnosisForResume(repoRoot, episodeId);
+            if (shouldSkipEvolution(diagnosis)) {
+                await advanceEpisodeStage({ repoRoot, episodeId, stage: 'abstained' });
             }
-            else {
-                const headBeforeRollback = await currentPolicyVersion(repoRoot, targetId);
-                const rollbackTarget = resolveRollbackTarget(ep.policyVersionBaseline, headBeforeRollback);
-                if (badAdvantage && rollbackTarget !== null && headBeforeRollback !== null) {
-                    await rollbackPolicyVersion({
-                        repoRoot,
-                        targetId,
-                        episodeId,
-                        toVersion: rollbackTarget,
-                        advantage: advantage ?? undefined,
-                    });
+            else if (diagnosis) {
+                // (shouldSkipEvolution returns true for null, so diagnosis is non-null here)
+                const advantage = diagnosis.advantage;
+                const badAdvantage = advantage !== null && advantage < threshold;
+                // Crash-resume dedup: a 'rollback' ledger entry whose episodeId is THIS
+                // episode means runEpisode already applied the rollback before the host
+                // crashed (the rollback head is monotonic — re-calling rollbackPolicyVersion
+                // would stack a SECOND, duplicate rollback version). When present, reuse its
+                // recorded version axis and SKIP the re-rollback; only ensure the
+                // reject-buffer entry + the 'rolled-back' stage advance complete.
+                const ledger = await readPolicyLedger(repoRoot, targetId);
+                const priorRollback = ledger.find((e) => e.action === 'rollback' && e.episodeId === episodeId);
+                if (priorRollback) {
+                    // The prior rollback already advanced to `priorRollback.version`, rolling
+                    // FORWARD to the content of the version immediately before the rejected
+                    // edit's head. Reconstruct the reject-buffer axis from that entry:
+                    //   toVersion   = the (rolled-back) version the rejected edit reached
+                    //               = priorRollback.version - 1 (the head before the rollback)
+                    //   fromVersion = the prior good policy restored (one before that head)
+                    const toVersion = priorRollback.version - 1;
+                    const fromVersion = resolveRollbackTarget(ep.policyVersionBaseline, toVersion);
                     await ensureRejectBufferEntry(repoRoot, {
                         episodeId,
                         targetId,
-                        fromVersion: rollbackTarget,
-                        toVersion: headBeforeRollback,
+                        fromVersion: fromVersion ?? toVersion,
+                        toVersion,
                         advantage,
                         diagnosis,
                     });
                     await advanceEpisodeStage({ repoRoot, episodeId, stage: 'rolled-back' });
                 }
                 else {
-                    await advanceEpisodeStage({ repoRoot, episodeId, stage: 'kept' });
+                    const headBeforeRollback = await currentPolicyVersion(repoRoot, targetId);
+                    const rollbackTarget = resolveRollbackTarget(ep.policyVersionBaseline, headBeforeRollback);
+                    if (badAdvantage && rollbackTarget !== null && headBeforeRollback !== null) {
+                        await rollbackPolicyVersion({
+                            repoRoot,
+                            targetId,
+                            episodeId,
+                            toVersion: rollbackTarget,
+                            advantage: advantage ?? undefined,
+                        });
+                        await ensureRejectBufferEntry(repoRoot, {
+                            episodeId,
+                            targetId,
+                            fromVersion: rollbackTarget,
+                            toVersion: headBeforeRollback,
+                            advantage,
+                            diagnosis,
+                        });
+                        await advanceEpisodeStage({ repoRoot, episodeId, stage: 'rolled-back' });
+                    }
+                    else {
+                        await advanceEpisodeStage({ repoRoot, episodeId, stage: 'kept' });
+                    }
+                }
+                // 预测校准 (idempotent if the original run already settled it) + 步长 schedule
+                // + calibration note, mirroring runEpisode's (g) step.
+                try {
+                    await reconcilePrediction({ repoRoot, targetId, episodeId });
                 }
+                catch {
+                    // best-effort: advisory only
+                }
+                const afterDecision = await readEpisode(repoRoot, episodeId);
+                const scheduledBudget = afterDecision.stage === 'rolled-back'
+                    ? scheduledEditBudget(await readPolicyLedger(repoRoot, targetId), editBudget)
+                    : editBudget;
+                const calibrationNote = await summarizeCalibration(repoRoot, targetId);
+                evolution = await runEvolvingAgent({
+                    repoRoot,
+                    episodeId,
+                    targetId,
+                    editBudget: scheduledBudget,
+                    ...(calibrationNote ? { calibrationNote } : {}),
+                    spawn: opts.spawn,
+                    ...(opts.agentTimeoutMs !== undefined ? { timeoutMs: opts.agentTimeoutMs } : {}),
+                    ...(opts.harness ? { harness: opts.harness } : {}),
+                });
             }
-            // 预测校准 (idempotent if the original run already settled it) + 步长 schedule
-            // + calibration note, mirroring runEpisode's (g) step.
+            await closeEpisodeBestEffort(repoRoot, episodeId);
+        }
+        else if (stage === 'rolled-back' || stage === 'kept') {
+            // The decision already ran (and the original episode settled the prediction);
+            // re-settle idempotently for the crash window, then schedule + calibrate.
             try {
                 await reconcilePrediction({ repoRoot, targetId, episodeId });
             }
             catch {
                 // best-effort: advisory only
             }
-            const afterDecision = await readEpisode(repoRoot, episodeId);
-            const scheduledBudget = afterDecision.stage === 'rolled-back'
+            const scheduledBudget = stage === 'rolled-back'
                 ? scheduledEditBudget(await readPolicyLedger(repoRoot, targetId), editBudget)
                 : editBudget;
             const calibrationNote = await summarizeCalibration(repoRoot, targetId);
@@ -598,39 +752,34 @@ export async function resumeEpisode(opts) {
                 editBudget: scheduledBudget,
                 ...(calibrationNote ? { calibrationNote } : {}),
                 spawn: opts.spawn,
+                ...(opts.agentTimeoutMs !== undefined ? { timeoutMs: opts.agentTimeoutMs } : {}),
+                ...(opts.harness ? { harness: opts.harness } : {}),
             });
+            await closeEpisodeBestEffort(repoRoot, episodeId);
         }
-        await closeEpisodeBestEffort(repoRoot, episodeId);
-    }
-    else if (ep.stage === 'rolled-back' || ep.stage === 'kept') {
-        // The decision already ran (and the original episode settled the prediction);
-        // re-settle idempotently for the crash window, then schedule + calibrate.
-        try {
-            await reconcilePrediction({ repoRoot, targetId, episodeId });
-        }
-        catch {
-            // best-effort: advisory only
+        else if (stage === 'evolved' ||
+            stage === 'evolution-refused' ||
+            stage === 'abstained') {
+            await closeEpisodeBestEffort(repoRoot, episodeId);
         }
-        const scheduledBudget = ep.stage === 'rolled-back'
-            ? scheduledEditBudget(await readPolicyLedger(repoRoot, targetId), editBudget)
-            : editBudget;
-        const calibrationNote = await summarizeCalibration(repoRoot, targetId);
-        evolution = await runEvolvingAgent({
+        // earlier stages (and a non-auto-resumable 'errored'): not auto-resumable here
+        // — reported as-is.
+    }
+    catch (err) {
+        // A thrown decision/evolving step records a DURABLE terminal 'errored' stage so
+        // the resumed episode is never left stuck at 'scored'/'rolled-back'/'kept'
+        // (indistinguishable from a still-running episode). Mirrors
+        // runEpisodeAfterCreate's catch — including the `timeout:` marker so a timed-out
+        // resume is distinguishable from a hard crash. Best-effort: a failed record must
+        // not mask the original throw, which still propagates to the caller.
+        await advanceEpisodeStage({
             repoRoot,
             episodeId,
-            targetId,
-            editBudget: scheduledBudget,
-            ...(calibrationNote ? { calibrationNote } : {}),
-            spawn: opts.spawn,
-        });
-        await closeEpisodeBestEffort(repoRoot, episodeId);
-    }
-    else if (ep.stage === 'evolved' ||
-        ep.stage === 'evolution-refused' ||
-        ep.stage === 'abstained') {
-        await closeEpisodeBestEffort(repoRoot, episodeId);
+            stage: 'errored',
+            patch: { terminalError: terminalErrorLabel(err) },
+        }).catch(() => { });
+        throw err;
     }
-    // earlier stages: not auto-resumable here — reported as-is.
     const after = await readEpisode(repoRoot, episodeId);
     return { episodeId, resumedFrom, stage: after.stage, evolution };
 }