@zhixuan92/multi-model-agent-core 3.10.7 → 3.11.0

package/dist/run-tasks/reviewed-lifecycle.js

@@ -1,14 +1,15 @@
 import { execFile } from 'node:child_process';
 import { promisify } from 'node:util';
-import { computeCostUSD, computeSavedCostUSD } from '../types.js';
+import { computeCostUSD } from '../types.js';
 import { createProvider } from '../provider.js';
 import { delegateWithEscalation } from '../delegate-with-escalation.js';
 import { pickEscalation, pickReviewer, maxRowsFor, } from '../escalation/policy.js';
 import { runWithFallback, makeSyntheticRunResult, TRANSPORT_FAILURES, isReviewTransportFailure, } from '../escalation/fallback.js';
 import { findModelCapabilities, findModelProfile } from '../routing/model-profiles.js';
+import { canonicalIdentity } from '../routing/canonical-model-identity.js';
 import { HeartbeatTimer } from '../heartbeat.js';
 import { newStageIdleTracker, snapshotIdle } from './stage-idle-tracker.js';
-import { DEFAULT_TASK_TIMEOUT_MS, DEFAULT_STALL_TIMEOUT_MS } from '../config/schema.js';
+import { DEFAULT_TASK_TIMEOUT_MS, DEFAULT_STALL_TIMEOUT_MS, MAX_TIME_PRESTOP_RATIO } from '../config/schema.js';
 import { runSpecReview } from '../review/spec-reviewer.js';
 import { makeSkippedReviewResult } from '../review/skipped-result.js';
 import { runQualityReview } from '../review/quality-reviewer.js';
@@ -201,6 +202,17 @@ export async function executeReviewedLifecycle(task, resolved, config, taskIndex
     function providerFor(tier) {
         return providers[tier];
     }
+    // Compute the implementer's canonical identity for reviewer separation (R3).
+    // Used as forbiddenIdentities on reviewer fallback calls so the reviewer
+    // never lands on the same effective backend as the implementer.
+    const implementerIdentity = (() => {
+        try {
+            return canonicalIdentity(resolved.provider.config);
+        }
+        catch {
+            return undefined;
+        }
+    })();
     // Partition filePaths into output targets before the worker runs.
     // Output targets are paths that do not yet exist on disk.
     const { outputTargets } = partitionFilePaths(task.filePaths, task.cwd ?? process.cwd());
@@ -216,17 +228,54 @@ export async function executeReviewedLifecycle(task, resolved, config, taskIndex
         'task_done_summary',
         'fallback', 'fallback_unavailable',
         'escalation', 'escalation_unavailable',
-        'stall_abort', 'cost_check',
+        'stall_abort', 'cost_check', 'time_check',
     ]);
     const shortBatchEarly = verboseBatchIdEarly ? verboseBatchIdEarly.slice(0, 8) : '????????';
     const emitTaskEvent = (event, fields) => {
         if (bus && verboseBatchIdEarly !== undefined) {
+            const schemaEvent = event === 'heartbeat_timer' ? 'task_started' : event;
             const cleaned = {};
             for (const [key, value] of Object.entries(fields)) {
                 if (value !== undefined)
                     cleaned[key] = value;
             }
-            bus.emit({ event, ts: new Date().toISOString(), batchId: verboseBatchIdEarly, taskIndex, ...cleaned });
+            // Keep verbose-line field names stable while emitting schema-declared
+            // telemetry envelopes in their authoritative persisted shape. EventSchemas
+            // validate the full envelope at EventBus.emit in dev/test, so production
+            // emission paths must construct schema-shaped keys before persistence.
+            if (schemaEvent === 'task_started') {
+                cleaned.route = routeKey || 'delegate';
+                cleaned.cwd = task.cwd ?? process.cwd();
+                for (const key of ['state', 'stage_count', 'tick_ms', 'reason'])
+                    delete cleaned[key];
+            }
+            if (event === 'verify_step') {
+                if ('exit_code' in cleaned) {
+                    cleaned.exitCode = cleaned.exit_code;
+                    delete cleaned.exit_code;
+                }
+                if ('duration_ms' in cleaned) {
+                    cleaned.durationMs = cleaned.duration_ms;
+                    delete cleaned.duration_ms;
+                }
+                if ('error_message' in cleaned) {
+                    cleaned.errorMessage = cleaned.error_message;
+                    delete cleaned.error_message;
+                }
+            }
+            if (event === 'task_completed') {
+                if ('stages_json' in cleaned) {
+                    cleaned.stages = cleaned.stages_json;
+                    delete cleaned.stages_json;
+                }
+                if (!('cachedTokens' in cleaned))
+                    cleaned.cachedTokens = null;
+                if (!('reasoningTokens' in cleaned))
+                    cleaned.reasoningTokens = null;
+                if (!('stages' in cleaned))
+                    cleaned.stages = JSON.stringify(stats);
+            }
+            bus.emit({ event: schemaEvent, ts: new Date().toISOString(), batchId: verboseBatchIdEarly, taskIndex, ...cleaned });
         }
         if (verboseStreamRaw && (verbose || DEFAULT_MODE_EVENTS.has(event))) {
             verboseStreamRaw(composeVerboseLine({ event, ts: new Date().toISOString(), batch: shortBatchEarly, task: taskIndex, ...toVerboseFields(fields) }));
@@ -398,9 +447,11 @@ export async function executeReviewedLifecycle(task, resolved, config, taskIndex
         }
         if (event.kind === 'turn_complete') {
             heartbeat?.markEvent('llm');
-            const costUSD = computeCostUSD(event.cumulativeInputTokens, event.cumulativeOutputTokens, resolved.provider.config);
-            const savedCostUSD = computeSavedCostUSD(costUSD, event.cumulativeInputTokens, event.cumulativeOutputTokens, task.parentModel);
-            heartbeat?.updateCost(costUSD, savedCostUSD);
+            const providerConfig = _activeRunnerProviderConfig ?? resolved.provider.config;
+            const costUSD = computeCostUSD(event.cumulativeInputTokens, event.cumulativeOutputTokens, providerConfig);
+            _currentRunnerCostUSD = costUSD ?? 0;
+            const cumulativeCostUSD = (_completedRunnerCostUSD ?? 0) + _currentRunnerCostUSD;
+            heartbeat?.updateCost(cumulativeCostUSD, null);
             const nowTurn = Date.now();
             const turnDurMs = nowTurn - prevEventAtMs;
             prevEventAtMs = nowTurn;
@@ -410,7 +461,7 @@ export async function executeReviewedLifecycle(task, resolved, config, taskIndex
                     output_tokens: event.cumulativeOutputTokens,
                     cost: costUSD,
                     duration_ms: turnDurMs,
-                    provider: resolved.provider.config.model,
+                    provider: providerConfig.model,
                 });
             }
         }
@@ -422,7 +473,7 @@ export async function executeReviewedLifecycle(task, resolved, config, taskIndex
     // any in-flight call gets a per-call timeoutMs clamped to remaining
     // budget so it returns its salvage promptly. The user gets *something*
     // back instead of an open-ended retry storm.
-    const taskTimeoutMs = task.timeoutMs ?? config.defaults.timeoutMs ?? DEFAULT_TASK_TIMEOUT_MS;
+    const taskTimeoutMs = task.timeoutMs ?? config.defaults?.timeoutMs ?? DEFAULT_TASK_TIMEOUT_MS;
     const taskDeadlineMs = taskStartMs + taskTimeoutMs;
     // Stall watchdog: when no LLM / tool / text event has fired for this
     // many ms, the in-flight runner is force-aborted via `stallController`.
@@ -484,7 +535,48 @@ export async function executeReviewedLifecycle(task, resolved, config, taskIndex
         const model = provider?.config.model ?? config.agents[tier]?.model ?? resolvedModel;
         return { tier, family: modelFamily(model), model };
     };
-    const runningCostUSD = () => taskCostUSD();
+    // §3.9: runningCostUSD must be cumulative and monotonic across explicit
+    // runner boundaries. Runner progress reports per-runner cumulative token
+    // counts, so lifecycle cost is completed runners + current runner partial.
+    // Boundaries are closed from actual RunResult.usage.costUSD values rather
+    // than inferred from drops; this handles reviewer costs greater than the
+    // implementer and preserves reviewer-provider pricing.
+    let _completedRunnerCostUSD = null;
+    let _currentRunnerCostUSD = 0;
+    let _activeRunnerProviderConfig = null;
+    let _prevRunningCost = null;
+    const runningCostUSD = () => {
+        const current = _completedRunnerCostUSD !== null || _currentRunnerCostUSD !== 0
+            ? (_completedRunnerCostUSD ?? 0) + _currentRunnerCostUSD
+            : null;
+        if (process.env.NODE_ENV === 'test' || process.env.NODE_ENV === 'development') {
+            if (_prevRunningCost !== null && current !== null && current < _prevRunningCost) {
+                throw new Error(`runningCostUSD non-monotonic: prev=${_prevRunningCost} now=${current}`);
+            }
+            _prevRunningCost = current;
+        }
+        return current;
+    };
+    const runAccounted = async (provider, call) => {
+        if (_activeRunnerProviderConfig !== null) {
+            throw new Error('lifecycle cost accounting runner overlap');
+        }
+        _activeRunnerProviderConfig = provider.config;
+        _currentRunnerCostUSD = 0;
+        try {
+            const result = await call();
+            const actualCost = result?.usage?.costUSD
+                ?? result?.metrics?.costUSD
+                ?? _currentRunnerCostUSD;
+            _completedRunnerCostUSD = (_completedRunnerCostUSD ?? 0) + actualCost;
+            _currentRunnerCostUSD = 0;
+            heartbeat?.updateCost(_completedRunnerCostUSD, null);
+            return result;
+        }
+        finally {
+            _activeRunnerProviderConfig = null;
+        }
+    };
     const policyEscalated = { spec: false, quality: false, diff: false };
     const emitFallback = (p) => {
         emitTaskEvent('fallback', p);
@@ -541,11 +633,21 @@ export async function executeReviewedLifecycle(task, resolved, config, taskIndex
             ...(fallbackOverrides.length > 0 ? { fallbackOverrides } : {}),
         };
     };
-    const abortReviewLoop = (base, terminationReason, message, aborting) => ({
+    const abortReviewLoop = (base, terminationReason, message, aborting, wallClockMs) => ({
         ...base,
         status: 'incomplete',
         workerStatus: 'review_loop_aborted',
-        terminationReason,
+        terminationReason: terminationReason === 'round_cap'
+            ? 'round_cap'
+            : {
+                cause: terminationReason === 'cost_ceiling' ? 'cost_exceeded' : 'time_ceiling',
+                turnsUsed: base.turns,
+                hasFileArtifacts: (base.filesWritten ?? []).length > 0,
+                usedShell: (base.toolCalls ?? []).some(c => c.startsWith('shell') || c.startsWith('runShell')),
+                workerSelfAssessment: 'review_loop_aborted',
+                wasPromoted: false,
+                ...(wallClockMs !== undefined ? { wallClockMs } : {}),
+            },
         reviewRounds: reviewRounds(),
         error: message,
         specReviewStatus: aborting === 'spec' ? 'changes_required' : (base.specReviewStatus ?? 'approved'),
@@ -562,7 +664,7 @@ export async function executeReviewedLifecycle(task, resolved, config, taskIndex
         const verification = await runVerifyStage({
             cwd,
             verifyCommand: task.verifyCommand,
-            taskTimeoutMs: task.timeoutMs ?? config.defaults.timeoutMs ?? DEFAULT_TASK_TIMEOUT_MS,
+            taskTimeoutMs: task.timeoutMs ?? config.defaults?.timeoutMs ?? DEFAULT_TASK_TIMEOUT_MS,
             taskStartMs,
         });
         latestVerification = verification;
@@ -589,7 +691,7 @@ export async function executeReviewedLifecycle(task, resolved, config, taskIndex
         const cause = typeof result.terminationReason === 'object' ? result.terminationReason.cause : result.terminationReason;
         const capExhausted = result.capExhausted
             ?? (result.status === 'cost_exceeded' || cause === 'cost_exceeded' || cause === 'cost_ceiling' ? 'cost'
-                : result.status === 'timeout' || cause === 'timeout' ? 'wall_clock'
+                : result.status === 'timeout' || cause === 'timeout' || cause === 'time_ceiling' ? 'wall_clock'
                     : result.status === 'incomplete' && result.turns > 1 ? 'turn'
                         : undefined);
         const lifecycleClarificationRequested = result.lifecycleClarificationRequested
@@ -605,7 +707,7 @@ export async function executeReviewedLifecycle(task, resolved, config, taskIndex
         return signalize({
             output: '',
             status: 'error',
-            usage: { inputTokens: 0, outputTokens: 0, totalTokens: 0, costUSD: null },
+            usage: { inputTokens: 0, outputTokens: 0, totalTokens: 0, costUSD: null, costDeltaVsParentUSD: null, cachedTokens: null, reasoningTokens: null },
             turns: 0,
             filesRead: [],
             filesWritten: [],
@@ -831,7 +933,7 @@ export async function executeReviewedLifecycle(task, resolved, config, taskIndex
                 return withVerification({
                     output: `Sub-agent error: task.cwd ${cwd} had pre-existing modifications`,
                     status: 'error',
-                    usage: { inputTokens: 0, outputTokens: 0, totalTokens: 0, costUSD: null },
+                    usage: { inputTokens: 0, outputTokens: 0, totalTokens: 0, costUSD: null, costDeltaVsParentUSD: null, cachedTokens: null, reasoningTokens: null },
                     turns: 0,
                     filesRead: [],
                     filesWritten: [],
@@ -858,7 +960,7 @@ export async function executeReviewedLifecycle(task, resolved, config, taskIndex
             isTransportFailure: (r) => TRANSPORT_FAILURES.has(r.status) && r.capExhausted === undefined,
             getStatus: (r) => r.status,
             makeSyntheticFailure: (assigned) => makeSyntheticRunResult(assigned, 'all_tiers_unavailable'),
-            call: (provider) => delegateWithEscalation(withDoneCondition(task), [provider], { explicitlyPinned: false, onProgress: wrappedOnProgress, taskDeadlineMs, abortSignal: stallController.signal, assignedTier: initialDecision.impl }),
+            call: (provider) => runAccounted(provider, () => delegateWithEscalation(withDoneCondition(task), [provider], { explicitlyPinned: false, onProgress: wrappedOnProgress, taskDeadlineMs, abortSignal: stallController.signal, assignedTier: initialDecision.impl })),
         });
         if (initialImpl.fallbackFired || initialImpl.bothUnavailable) {
             fallbackOverrides.push({
@@ -1047,10 +1149,11 @@ export async function executeReviewedLifecycle(task, resolved, config, taskIndex
                 isTransportFailure: (r) => isReviewTransportFailure(r),
                 getStatus: (r) => r.status,
                 makeSyntheticFailure: () => makeSkippedReviewResult('all_tiers_unavailable'),
-                call: (provider) => runDiffReview({ cwd, diff: evidence.fullDiff, diffTruncated: evidence.diffTruncated, verification, worker: { call: (prompt, opts) => provider.run(prompt, { abortSignal: opts?.abortSignal, timeoutMs: opts?.timeoutMs }) }, taskDeadlineMs, abortSignal: stallController.signal }),
+                forbiddenIdentities: implementerIdentity ? [implementerIdentity] : undefined,
+                call: (provider) => runAccounted(provider, () => runDiffReview({ cwd, diff: evidence.fullDiff, diffTruncated: evidence.diffTruncated, verification, worker: { call: (prompt, opts) => provider.run(prompt, { cwd: opts?.cwd ?? cwd, abortSignal: opts?.abortSignal, timeoutMs: opts?.timeoutMs }) }, taskDeadlineMs, abortSignal: stallController.signal })),
             });
             if (diffCall.fallbackFired) {
-                emitFallback({ batchId: heartbeatWiring?.batchId ?? '', taskIndex, loop: 'diff', attempt: 0, role: 'diffReviewer', assignedTier: diffReviewerTier, usedTier: diffCall.usedTier, reason: diffCall.fallbackReason, triggeringStatus: diffCall.fallbackTriggeringStatus, violatesSeparation: diffCall.usedTier === implementerHistory[implementerHistory.length - 1] });
+                emitFallback({ batchId: heartbeatWiring?.batchId ?? '', taskIndex, loop: 'diff', attempt: 0, role: 'diffReviewer', assignedTier: diffReviewerTier, usedTier: diffCall.usedTier, reason: diffCall.fallbackReason, triggeringStatus: diffCall.fallbackTriggeringStatus, violatesSeparation: diffCall.usedTier === implementerHistory[implementerHistory.length - 1], fallbackSeparationRespected: diffCall.fallbackSeparationRespected, assignedIdentity: diffCall.assignedIdentity ?? null, usedIdentity: diffCall.usedIdentity ?? null });
                 fallbackOverrides.push({ role: 'diffReviewer', loop: 'diff', attempt: 0, assigned: diffReviewerTier, used: diffCall.usedTier, reason: diffCall.fallbackReason, triggeringStatus: diffCall.fallbackTriggeringStatus, bothUnavailable: diffCall.bothUnavailable });
             }
             if (diffCall.bothUnavailable) {
@@ -1118,7 +1221,8 @@ export async function executeReviewedLifecycle(task, resolved, config, taskIndex
                 isTransportFailure: (r) => isReviewTransportFailure(r),
                 getStatus: (r) => r.status,
                 makeSyntheticFailure: () => makeSkippedReviewResult('all_tiers_unavailable'),
-                call: (provider) => runSpecReview(provider, packet, effectiveImplReport, fileContents, implResult.toolCalls, task.planContext, evidence.block, taskDeadlineMs, stallController.signal, wrappedOnProgress),
+                forbiddenIdentities: implementerIdentity ? [implementerIdentity] : undefined,
+                call: (provider) => runAccounted(provider, () => runSpecReview(provider, packet, effectiveImplReport, fileContents, implResult.toolCalls, task.planContext, evidence.block, taskDeadlineMs, stallController.signal, wrappedOnProgress, cwd)),
             });
             specReviewDurationMs += Date.now() - initialSpecReviewIterStart;
             if (initialSpecReview.bothUnavailable) {
@@ -1129,7 +1233,7 @@ export async function executeReviewedLifecycle(task, resolved, config, taskIndex
             else {
                 specReviewerHistory.push(initialSpecReview.usedTier);
                 if (initialSpecReview.fallbackFired) {
-                    emitFallback({ batchId: heartbeatWiring?.batchId ?? '', taskIndex, loop: 'spec', attempt: 0, role: 'specReviewer', assignedTier: initialReviewerTier, usedTier: initialSpecReview.usedTier, reason: initialSpecReview.fallbackReason, triggeringStatus: initialSpecReview.fallbackTriggeringStatus, violatesSeparation: initialSpecReview.usedTier === implementerHistory[implementerHistory.length - 1] });
+                    emitFallback({ batchId: heartbeatWiring?.batchId ?? '', taskIndex, loop: 'spec', attempt: 0, role: 'specReviewer', assignedTier: initialReviewerTier, usedTier: initialSpecReview.usedTier, reason: initialSpecReview.fallbackReason, triggeringStatus: initialSpecReview.fallbackTriggeringStatus, violatesSeparation: initialSpecReview.usedTier === implementerHistory[implementerHistory.length - 1], fallbackSeparationRespected: initialSpecReview.fallbackSeparationRespected, assignedIdentity: initialSpecReview.assignedIdentity ?? null, usedIdentity: initialSpecReview.usedIdentity ?? null });
                     fallbackOverrides.push({ role: 'specReviewer', loop: 'spec', attempt: 0, assigned: initialReviewerTier, used: initialSpecReview.usedTier, reason: initialSpecReview.fallbackReason, triggeringStatus: initialSpecReview.fallbackTriggeringStatus, bothUnavailable: false });
                 }
             }
@@ -1149,6 +1253,11 @@ export async function executeReviewedLifecycle(task, resolved, config, taskIndex
                     emitTaskEvent('cost_check', { stage: 'spec_rework', tripped: true, cost_used_usd: currentCostUSD, cost_cap_usd: maxCostUSD, cost_available: true });
                     return abortReviewLoop(finalImplResult, 'cost_ceiling', 'cost ceiling reached before spec rework', 'spec');
                 }
+                const wallClock = Date.now() - taskStartMs;
+                if (wallClock >= MAX_TIME_PRESTOP_RATIO * taskTimeoutMs) {
+                    emitTaskEvent('time_check', { stage: 'spec_rework', tripped: true, wallClockMs: wallClock, timeoutMs: taskTimeoutMs });
+                    return abortReviewLoop(finalImplResult, 'time_ceiling', `time ceiling reached before spec rework (${wallClock}ms >= 0.8 × ${taskTimeoutMs}ms)`, 'spec', wallClock);
+                }
                 const decision = pickEscalation({ loop: 'spec', attemptIndex: specAttemptIndex, baseTier: resolved.slot });
                 if (decision.isEscalated)
                     emitEscalationEvent('spec', specAttemptIndex, decision);
@@ -1156,7 +1265,7 @@ export async function executeReviewedLifecycle(task, resolved, config, taskIndex
                 transitionStage('spec_review', 'spec_rework', { stage: 'spec_rework', stageIndex: 3, reviewRound: specAttemptIndex, attemptCap: maxSpecRows }, { attempt: specAttemptIndex, attemptCap: maxSpecRows, implTier: decision.impl, reviewerTier: decision.reviewer, escalated: decision.isEscalated });
                 const feedback = specResult.findings.length > 0 ? `\n\n## Spec Review Feedback (round ${specAttemptIndex}):\n${specResult.findings.map(f => `- ${f}`).join('\n')}` : '';
                 const reworkTask = withDoneCondition({ ...task, prompt: `${task.prompt}${feedback}` });
-                const reworkCall = await runWithFallback({ assigned: decision.impl, providerFor, unavailableTiers: specUnavailable, isTransportFailure: (r) => TRANSPORT_FAILURES.has(r.status) && r.capExhausted === undefined, getStatus: (r) => r.status, makeSyntheticFailure: (assigned) => makeSyntheticRunResult(assigned, 'all_tiers_unavailable'), call: (provider) => delegateWithEscalation(reworkTask, [provider], { explicitlyPinned: true, onProgress: wrappedOnProgress, taskDeadlineMs, abortSignal: stallController.signal, assignedTier: decision.impl }) });
+                const reworkCall = await runWithFallback({ assigned: decision.impl, providerFor, unavailableTiers: specUnavailable, isTransportFailure: (r) => TRANSPORT_FAILURES.has(r.status) && r.capExhausted === undefined, getStatus: (r) => r.status, makeSyntheticFailure: (assigned) => makeSyntheticRunResult(assigned, 'all_tiers_unavailable'), call: (provider) => runAccounted(provider, () => delegateWithEscalation(reworkTask, [provider], { explicitlyPinned: true, onProgress: wrappedOnProgress, taskDeadlineMs, abortSignal: stallController.signal, assignedTier: decision.impl })) });
                 if (reworkCall.fallbackFired || reworkCall.bothUnavailable)
                     fallbackOverrides.push({ role: 'implementer', loop: 'spec', attempt: specAttemptIndex, assigned: decision.impl, used: reworkCall.usedTier, reason: (reworkCall.fallbackReason ?? reworkCall.unavailableReason), triggeringStatus: reworkCall.fallbackTriggeringStatus, bothUnavailable: reworkCall.bothUnavailable });
                 if (reworkCall.fallbackFired) {
@@ -1180,7 +1289,7 @@ export async function executeReviewedLifecycle(task, resolved, config, taskIndex
                 commitReworkStage(stats, 'spec_rework', specReworkAcc, implementerAgentInfo);
                 transitionStage('spec_rework', 'spec_review', { stage: 'spec_review', stageIndex: 2, reviewRound: specAttemptIndex + 1, attemptCap: maxSpecRows }, null);
                 const reReviewIterStart = Date.now();
-                const reviewCall = await runWithFallback({ assigned: decision.reviewer, providerFor, unavailableTiers: specUnavailable, isTransportFailure: (r) => isReviewTransportFailure(r), getStatus: (r) => r.status, makeSyntheticFailure: () => makeSkippedReviewResult('all_tiers_unavailable'), call: (provider) => runSpecReview(provider, packet, finalImplReport, fileContents, finalImplResult.toolCalls, task.planContext, evidence.block, taskDeadlineMs, stallController.signal, wrappedOnProgress) });
+                const reviewCall = await runWithFallback({ assigned: decision.reviewer, providerFor, unavailableTiers: specUnavailable, isTransportFailure: (r) => isReviewTransportFailure(r), getStatus: (r) => r.status, makeSyntheticFailure: () => makeSkippedReviewResult('all_tiers_unavailable'), forbiddenIdentities: implementerIdentity ? [implementerIdentity] : undefined, call: (provider) => runAccounted(provider, () => runSpecReview(provider, packet, finalImplReport, fileContents, finalImplResult.toolCalls, task.planContext, evidence.block, taskDeadlineMs, stallController.signal, wrappedOnProgress, cwd)) });
                 specReviewDurationMs += Date.now() - reReviewIterStart;
                 if (reviewCall.bothUnavailable) {
                     emitFallbackUnavailable({ batchId: heartbeatWiring?.batchId ?? '', taskIndex, loop: 'spec', attempt: specAttemptIndex, role: 'specReviewer', assignedTier: decision.reviewer, reason: reviewCall.unavailableReason });
@@ -1190,7 +1299,7 @@ export async function executeReviewedLifecycle(task, resolved, config, taskIndex
                 else {
                     specReviewerHistory.push(reviewCall.usedTier);
                     if (reviewCall.fallbackFired) {
-                        emitFallback({ batchId: heartbeatWiring?.batchId ?? '', taskIndex, loop: 'spec', attempt: specAttemptIndex, role: 'specReviewer', assignedTier: decision.reviewer, usedTier: reviewCall.usedTier, reason: reviewCall.fallbackReason, triggeringStatus: reviewCall.fallbackTriggeringStatus, violatesSeparation: reviewCall.usedTier === implementerHistory[implementerHistory.length - 1] });
+                        emitFallback({ batchId: heartbeatWiring?.batchId ?? '', taskIndex, loop: 'spec', attempt: specAttemptIndex, role: 'specReviewer', assignedTier: decision.reviewer, usedTier: reviewCall.usedTier, reason: reviewCall.fallbackReason, triggeringStatus: reviewCall.fallbackTriggeringStatus, violatesSeparation: reviewCall.usedTier === implementerHistory[implementerHistory.length - 1], fallbackSeparationRespected: reviewCall.fallbackSeparationRespected, assignedIdentity: reviewCall.assignedIdentity ?? null, usedIdentity: reviewCall.usedIdentity ?? null });
                         fallbackOverrides.push({ role: 'specReviewer', loop: 'spec', attempt: specAttemptIndex, assigned: decision.reviewer, used: reviewCall.usedTier, reason: reviewCall.fallbackReason, triggeringStatus: reviewCall.fallbackTriggeringStatus, bothUnavailable: false });
                     }
                 }
@@ -1234,7 +1343,7 @@ export async function executeReviewedLifecycle(task, resolved, config, taskIndex
             qualityReviewT0 = Date.now();
             qualityReviewC0 = runningCostUSD();
             const initialQualityIterStart = Date.now();
-            const initialQuality = await runWithFallback({ assigned: qualityReviewerTier, providerFor, unavailableTiers: qualityUnavailable, isTransportFailure: (r) => isReviewTransportFailure(r), getStatus: (r) => r.status, makeSyntheticFailure: () => makeSkippedReviewResult('all_tiers_unavailable'), call: (provider) => runQualityReview(provider, packet, specReport ?? finalImplReport, fileContents, finalImplResult.toolCalls, finalImplResult.filesWritten, evidence.block, qualityReviewPromptBuilder, finalImplResult.output, taskDeadlineMs, stallController.signal, wrappedOnProgress) });
+            const initialQuality = await runWithFallback({ assigned: qualityReviewerTier, providerFor, unavailableTiers: qualityUnavailable, isTransportFailure: (r) => isReviewTransportFailure(r), getStatus: (r) => r.status, makeSyntheticFailure: () => makeSkippedReviewResult('all_tiers_unavailable'), forbiddenIdentities: implementerIdentity ? [implementerIdentity] : undefined, call: (provider) => runAccounted(provider, () => runQualityReview(provider, packet, specReport ?? finalImplReport, fileContents, finalImplResult.toolCalls, finalImplResult.filesWritten, evidence.block, qualityReviewPromptBuilder, finalImplResult.output, taskDeadlineMs, stallController.signal, wrappedOnProgress, cwd)) });
             qualityReviewDurationMs += Date.now() - initialQualityIterStart;
             if (initialQuality.bothUnavailable) {
                 emitFallbackUnavailable({ batchId: heartbeatWiring?.batchId ?? '', taskIndex, loop: 'quality', attempt: 0, role: 'qualityReviewer', assignedTier: qualityReviewerTier, reason: initialQuality.unavailableReason });
@@ -1244,7 +1353,7 @@ export async function executeReviewedLifecycle(task, resolved, config, taskIndex
             else {
                 qualityReviewerHistory.push(initialQuality.usedTier);
                 if (initialQuality.fallbackFired) {
-                    emitFallback({ batchId: heartbeatWiring?.batchId ?? '', taskIndex, loop: 'quality', attempt: 0, role: 'qualityReviewer', assignedTier: qualityReviewerTier, usedTier: initialQuality.usedTier, reason: initialQuality.fallbackReason, triggeringStatus: initialQuality.fallbackTriggeringStatus, violatesSeparation: initialQuality.usedTier === implementerHistory[implementerHistory.length - 1] });
+                    emitFallback({ batchId: heartbeatWiring?.batchId ?? '', taskIndex, loop: 'quality', attempt: 0, role: 'qualityReviewer', assignedTier: qualityReviewerTier, usedTier: initialQuality.usedTier, reason: initialQuality.fallbackReason, triggeringStatus: initialQuality.fallbackTriggeringStatus, violatesSeparation: initialQuality.usedTier === implementerHistory[implementerHistory.length - 1], fallbackSeparationRespected: initialQuality.fallbackSeparationRespected, assignedIdentity: initialQuality.assignedIdentity ?? null, usedIdentity: initialQuality.usedIdentity ?? null });
                     fallbackOverrides.push({ role: 'qualityReviewer', loop: 'quality', attempt: 0, assigned: qualityReviewerTier, used: initialQuality.usedTier, reason: initialQuality.fallbackReason, triggeringStatus: initialQuality.fallbackTriggeringStatus, bothUnavailable: false });
                 }
             }
@@ -1288,8 +1397,6 @@ export async function executeReviewedLifecycle(task, resolved, config, taskIndex
                             : 'error',
                     iterationIndex: 1,
                     findingsReviewed: annotated.length,
-                    findingsFlagged: 0, // legacy field — severity correction tracked elsewhere now
-                    severityCorrections: 0, // reviewerSeverity field removed in 3.10.5
                     meanConfidence,
                     durationMs: Date.now() - qualityReviewT0,
                     costUSD: runningCostUSD() !== null && qualityReviewC0 !== null ? runningCostUSD() - qualityReviewC0 : null,
@@ -1307,6 +1414,11 @@ export async function executeReviewedLifecycle(task, resolved, config, taskIndex
                         emitTaskEvent('cost_check', { stage: 'quality_rework', tripped: true, cost_used_usd: currentCostUSD, cost_cap_usd: maxCostUSD, cost_available: true });
                         return abortReviewLoop(finalImplResult, 'cost_ceiling', 'cost ceiling reached before quality rework', 'quality');
                     }
+                    const wallClock = Date.now() - taskStartMs;
+                    if (wallClock >= MAX_TIME_PRESTOP_RATIO * taskTimeoutMs) {
+                        emitTaskEvent('time_check', { stage: 'quality_rework', tripped: true, wallClockMs: wallClock, timeoutMs: taskTimeoutMs });
+                        return abortReviewLoop(finalImplResult, 'time_ceiling', `time ceiling reached before quality rework (${wallClock}ms >= 0.8 × ${taskTimeoutMs}ms)`, 'quality', wallClock);
+                    }
                     const decision = pickEscalation({ loop: 'quality', attemptIndex: qualityAttemptIndex, baseTier: resolved.slot });
                     if (decision.isEscalated)
                         emitEscalationEvent('quality', qualityAttemptIndex, decision);
@@ -1314,7 +1426,7 @@ export async function executeReviewedLifecycle(task, resolved, config, taskIndex
                     transitionStage('quality_review', 'quality_rework', { stage: 'quality_rework', stageIndex: 5, reviewRound: qualityAttemptIndex, attemptCap: maxQualityRows }, { attempt: qualityAttemptIndex, attemptCap: maxQualityRows, implTier: decision.impl, reviewerTier: decision.reviewer, escalated: decision.isEscalated });
                     const feedback = qualityResult.findings.length > 0 ? `\n\n## Quality Review Feedback (round ${qualityAttemptIndex}):\n${qualityResult.findings.map(f => `- ${f}`).join('\n')}` : '';
                     const reworkTask = withDoneCondition({ ...task, prompt: `${task.prompt}${feedback}` });
-                    const reworkCall = await runWithFallback({ assigned: decision.impl, providerFor, unavailableTiers: qualityUnavailable, isTransportFailure: (r) => TRANSPORT_FAILURES.has(r.status) && r.capExhausted === undefined, getStatus: (r) => r.status, makeSyntheticFailure: (assigned) => makeSyntheticRunResult(assigned, 'all_tiers_unavailable'), call: (provider) => delegateWithEscalation(reworkTask, [provider], { explicitlyPinned: true, onProgress: wrappedOnProgress, taskDeadlineMs, abortSignal: stallController.signal, assignedTier: decision.impl }) });
+                    const reworkCall = await runWithFallback({ assigned: decision.impl, providerFor, unavailableTiers: qualityUnavailable, isTransportFailure: (r) => TRANSPORT_FAILURES.has(r.status) && r.capExhausted === undefined, getStatus: (r) => r.status, makeSyntheticFailure: (assigned) => makeSyntheticRunResult(assigned, 'all_tiers_unavailable'), call: (provider) => runAccounted(provider, () => delegateWithEscalation(reworkTask, [provider], { explicitlyPinned: true, onProgress: wrappedOnProgress, taskDeadlineMs, abortSignal: stallController.signal, assignedTier: decision.impl })) });
                     if (reworkCall.fallbackFired || reworkCall.bothUnavailable)
                         fallbackOverrides.push({ role: 'implementer', loop: 'quality', attempt: qualityAttemptIndex, assigned: decision.impl, used: reworkCall.usedTier, reason: (reworkCall.fallbackReason ?? reworkCall.unavailableReason), triggeringStatus: reworkCall.fallbackTriggeringStatus, bothUnavailable: reworkCall.bothUnavailable });
                     if (reworkCall.fallbackFired)
@@ -1335,7 +1447,7 @@ export async function executeReviewedLifecycle(task, resolved, config, taskIndex
                     commitReworkStage(stats, 'quality_rework', qualityReworkAcc, implementerAgentInfo);
                     transitionStage('quality_rework', 'quality_review', { stage: 'quality_review', stageIndex: 4, reviewRound: qualityAttemptIndex + 1, attemptCap: maxQualityRows }, null);
                     const qReReviewIterStart = Date.now();
-                    const reviewCall = await runWithFallback({ assigned: decision.reviewer, providerFor, unavailableTiers: qualityUnavailable, isTransportFailure: (r) => isReviewTransportFailure(r), getStatus: (r) => r.status, makeSyntheticFailure: () => makeSkippedReviewResult('all_tiers_unavailable'), call: (provider) => runQualityReview(provider, packet, finalImplReport, fileContents, finalImplResult.toolCalls, finalImplResult.filesWritten, evidence.block, qualityReviewPromptBuilder, finalImplResult.output, taskDeadlineMs, stallController.signal, wrappedOnProgress) });
+                    const reviewCall = await runWithFallback({ assigned: decision.reviewer, providerFor, unavailableTiers: qualityUnavailable, isTransportFailure: (r) => isReviewTransportFailure(r), getStatus: (r) => r.status, makeSyntheticFailure: () => makeSkippedReviewResult('all_tiers_unavailable'), forbiddenIdentities: implementerIdentity ? [implementerIdentity] : undefined, call: (provider) => runAccounted(provider, () => runQualityReview(provider, packet, finalImplReport, fileContents, finalImplResult.toolCalls, finalImplResult.filesWritten, evidence.block, qualityReviewPromptBuilder, finalImplResult.output, taskDeadlineMs, stallController.signal, wrappedOnProgress, cwd)) });
                     qualityReviewDurationMs += Date.now() - qReReviewIterStart;
                     if (reviewCall.bothUnavailable) {
                         emitFallbackUnavailable({ batchId: heartbeatWiring?.batchId ?? '', taskIndex, loop: 'quality', attempt: qualityAttemptIndex, role: 'qualityReviewer', assignedTier: decision.reviewer, reason: reviewCall.unavailableReason });
@@ -1345,7 +1457,7 @@ export async function executeReviewedLifecycle(task, resolved, config, taskIndex
                     else {
                         qualityReviewerHistory.push(reviewCall.usedTier);
                         if (reviewCall.fallbackFired) {
-                            emitFallback({ batchId: heartbeatWiring?.batchId ?? '', taskIndex, loop: 'quality', attempt: qualityAttemptIndex, role: 'qualityReviewer', assignedTier: decision.reviewer, usedTier: reviewCall.usedTier, reason: reviewCall.fallbackReason, triggeringStatus: reviewCall.fallbackTriggeringStatus, violatesSeparation: reviewCall.usedTier === implementerHistory[implementerHistory.length - 1] });
+                            emitFallback({ batchId: heartbeatWiring?.batchId ?? '', taskIndex, loop: 'quality', attempt: qualityAttemptIndex, role: 'qualityReviewer', assignedTier: decision.reviewer, usedTier: reviewCall.usedTier, reason: reviewCall.fallbackReason, triggeringStatus: reviewCall.fallbackTriggeringStatus, violatesSeparation: reviewCall.usedTier === implementerHistory[implementerHistory.length - 1], fallbackSeparationRespected: reviewCall.fallbackSeparationRespected, assignedIdentity: reviewCall.assignedIdentity ?? null, usedIdentity: reviewCall.usedIdentity ?? null });
                             fallbackOverrides.push({ role: 'qualityReviewer', loop: 'quality', attempt: qualityAttemptIndex, assigned: decision.reviewer, used: reviewCall.usedTier, reason: reviewCall.fallbackReason, triggeringStatus: reviewCall.fallbackTriggeringStatus, bothUnavailable: false });
                         }
                     }
@@ -1515,6 +1627,8 @@ export async function executeReviewedLifecycle(task, resolved, config, taskIndex
                     toolCalls: r.toolCalls?.length ?? 0,
                     inputTokens: r.usage.inputTokens,
                     outputTokens: r.usage.outputTokens,
+                    cachedTokens: r.usage.cachedTokens ?? null,
+                    reasoningTokens: r.usage.reasoningTokens ?? null,
                     costUSD: r.usage.costUSD,
                     taskMaxIdleMs: r.taskMaxIdleMs ?? null,
                     stallTriggered: r.stallTriggered ?? false,