elanous 0.2.24 → 0.2.25

package/src/self-implement/orchestrator.ts

@@ -18,6 +18,7 @@ import {
   parseGateLogTestRuns,
   type PrEvidenceInput,
 } from './pr-evidence-artifact.js';
+import { REVIEW_DEFERRED_TO_HOST_MERGE_REASON } from './review-deferred-to-host.js';
 import { assessImplementationArtifactCompleteness } from './implementation-artifact-completeness.js';
 import { runContractAtNode, type RunContract } from './graph-run-contract.js';
 export {
@@ -91,6 +92,7 @@ import { classifyError } from '../session-runtime/retry-policy.js';
 import { decomposeSelfDevGoal, inferHotPaths, type SelfDevDecomposeOptions } from '../self-dev/decompose.js';
 import { boundaryGlobs, boundaryViolations } from './goal-boundary.js';
 import { askAtExecution, type ExecutionClarificationResult } from './execution-clarification.js';
+import { classifyOutOfScope } from './scope-adjacency.js';
 import { targetScopedGoalText } from './goal-text-path-scope.js';
 import { dropStaleReverts, type RevertGuardResult } from './revert-guard.js';
 import type { SelfDevGoalType } from '../self-dev/orchestrate.js';
@@ -157,10 +159,11 @@ import { resolveAdaptiveMaxReworkDecision, resolveEscalateTier, resolveEscalateT
 import type { SupervisionReworkSource } from './supervision-vocabulary.js';
 import { classifyReworkBudgetShadow } from './classify-shadow.js';
 import { createReworkBudgetJudgment, REWORK_BUDGET_JUDGMENT, reworkBudgetEvidenceForObservation, reworkBudgetWorkflow } from './rework-budget-judgment.js';
+import { formatGateFailedId, type PremiseReviewCard, type PremiseReviewInput } from './premise-review.js';
 import { runWorkflowToCompletion } from '../workflow-runtime/executor.js';
 import { mapReworkVerdict, supervisionObservationFields } from './supervision-vocabulary.js';
 import { resolveRunOutcome } from './run-outcome.js';
-import { authorLimitationCountFromGoal, harvestEvidenceLines, harvestedEvidenceObservation, MAX_HARVESTED_EVIDENCE_CHARS, parseOffDiffEvidence, requiredEvidenceFromGoal, runRequiredEvidenceChecks, coverRequiredEvidence, formatUncoveredEvidenceFeedback, type OffDiffEvidenceParse } from './off-diff-evidence.js';
+import { authorLimitationCountFromGoal, harvestEvidenceLines, harvestedEvidenceObservation, MAX_HARVESTED_EVIDENCE_CHARS, parseOffDiffEvidence, requiredEvidenceFromGoal, runRequiredEvidenceChecks, coverRequiredEvidence, settleUncoveredEvidence, formatUncoveredEvidenceFeedback, type OffDiffEvidenceParse } from './off-diff-evidence.js';
 import { pressDecisionSignals, type DecisionSignalPressResult } from './decision-signal-press.js';
 import { inspectDecisionObservations, inspectDecisionSignalKinds, inspectDecisionSignalObservations } from '../../scripts/ask-marker-check.js';
 
@@ -580,6 +583,8 @@ export interface SelfImplementReview {
   reviewed: boolean;
   /** 리뷰가 실행되지 않은 실패 사유. 리뷰 엔진이 제공한 경우에만 관측으로 전달한다. */
   failureReason?: string;
+  /** REVIEW-ACP-POD ②-b — 리뷰를 발사 호스트(Claude ACP)로 미뤘다. `reviewed` 는 거짓 그대로다(통과가 아니다). */
+  deferredToHost?: true;
   /** 리뷰 엔진이 실제로 사용한 diff 예산. 없는 값은 미검토/no-diff 등으로, 완전한 diff의 증거가 아니다. */
   diffTruncated?: boolean;
   diffShownChars?: number;
@@ -817,11 +822,20 @@ function declaredPathsFromGoalDocumentText(document: string): { readable: boolea
   if (originalAsk === null) return { readable: false, paths: [] };
   const declared = parseDeclaredTargetPathsFromAsk(originalAsk);
   if (!declared.readable) return declared;
-  const newFiles = originalAsk.split(/\r?\n/).flatMap((line) => {
+  const newFiles = declaredNewFilePaths(originalAsk);
+  return newFiles.length > 0 ? { readable: true, paths: [...new Set([...declared.paths, ...newFiles])] } : declared;
+}
+
+function declaredNewFilePaths(askText: string): readonly string[] {
+  return askText.split(/\r?\n/).flatMap((line) => {
     const match = DECLARED_NEW_FILES_LINE.exec(line);
     return match ? parseDeclaredTargetPathsFromAsk(`대상 경로: ${match[1]}`).paths : [];
   });
-  return newFiles.length > 0 ? { readable: true, paths: [...new Set([...declared.paths, ...newFiles])] } : declared;
+}
+
+/** P3 전제의 경계 — ask 머리줄 «대상 경로:» ⊕ «신설:» 줄의 합집합. 못 읽으면 빈 목록(그때 P3 는 unknown). */
+function declaredPremiseScopePaths(askText: string): string[] {
+  return [...new Set([...parseDeclaredTargetPathsFromAsk(askText).paths, ...declaredNewFilePaths(askText)])];
 }
 
 /** Only paths explicitly designated as targets in the ask or piece text protect an intentional revert. */
@@ -6446,7 +6460,21 @@ async function runSelfImplementInner(
         judgmentStarted = true;
         const judgmentCallLLM = s.judgmentCallLLM;
         if (diagnosedDecision?.verdict !== 'CONTRACT-CONFLICT' && !judgmentCallLLM) throw new Error('rework-budget@v1 requires a judgment provider');
-        const judgmentProvider = judgmentCallLLM ? createReworkBudgetJudgment(judgmentCallLLM) : undefined;
+        // FLEX-RFC §A-③ — UNCONVERGEABLE 직전의 전제 되짚기 입력. 이 루프 뒤쪽 «flex-signals» 관측(6895~)과 같은
+        // 실제 런 값으로 조립한다. 이 생산 배선이 없으면 카드는 시험에서만 세워지고 원장에는 남지 않는다.
+        const premiseLatestEntry = roundSummaryEntries[roundSummaryEntries.length - 1];
+        const premiseInput: PremiseReviewInput = {
+          runId,
+          gateFailedIdsByRound: premiseLatestEntry?.kind === 'gate'
+            ? roundSummaryEntries.filter((entry) => entry.kind === 'gate').map((entry) => entry.failedIds)
+            : [],
+          failureAttributions: reworkKind === 'gate' && gate ? gate.baselineFailures?.map(({ attribution }) => attribution) : undefined,
+          reviewRepeat,
+          round,
+          plannedRounds: effectiveMax,
+          targetPaths: declaredPremiseScopePaths(opts.feature),
+        };
+        const judgmentProvider = judgmentCallLLM ? createReworkBudgetJudgment(judgmentCallLLM, { premiseInput }) : undefined;
         try {
           const result = diagnosedDecision?.verdict === 'CONTRACT-CONFLICT'
             ? undefined
@@ -6628,6 +6656,17 @@ async function runSelfImplementInner(
             : { evidence: null };
           observe('rework-budget', { round, verdict: parsedDecision.verdict, reason: parsedDecision.reason.slice(0, 240), effectiveMaxBefore: before, effectiveMaxAfter: effectiveMax, carriedBudgetBefore, carriedBudgetAfter: judgedEffectiveMax, historyRounds: priorHistory.length, kind: reworkKind, applied: budgetDecision.applied, stopped: budgetDecision.stop, exit: budgetDecision.exit, shadowed: budgetDecision.shadowed === true, repeatedBlockingFindingCount: repeatedBlockingIds?.length ?? null, repeatedBlockingFindingIds: repeatedBlockingIds ?? null, normalizedRepeatedReviewFindingCount: recurrence?.normalizedRepeatedReviewFindingCount ?? null, ordinaryRepeatedReviewFindingCount: recurrence?.ordinaryRepeatedReviewFindingCount ?? null, previouslyDismissedRepeatedReviewFindingCount: recurrence?.previouslyDismissedRepeatedReviewFindingCount ?? null, citedReviewSymbolRepeatCount: recurrence?.citedReviewSymbolRepeatCount ?? null, citedReviewSymbolBaseNameRepeatCount: recurrence?.citedReviewSymbolBaseNameRepeatCount ?? null, reviewFindingComparableCount: recurrence?.comparableFindings ?? null, reviewFindingKeyRepeatCount: recurrence?.reviewFindingKeyRepeatCount ?? null, ...recurrenceDisagreementObservation, ...evidenceObservation, judgePredictedBefore: predictionAccuracyForObservation.predicted, judgeFulfilledBefore: predictionAccuracyForObservation.fulfilled, judgeMissedBefore: predictionAccuracyForObservation.missed, judgePendingBefore: predictionAccuracyForObservation.pending, ...(budgetDecision.shadowed ? { wouldExit: budgetDecision.wouldExit } : {}), ...(budgetDecision.contractConflictDisposition ? { contractConflictDisposition: budgetDecision.contractConflictDisposition } : {}), ...(pendingPlanRevision?.relaxation ? { contractConflictRelaxation: { ...pendingPlanRevision.relaxation, reason: pendingPlanRevision.reason, application: pendingPlanRevision.application?.status ?? 'failed' } } : {}), ...supervisionObservation }, { ...(budgetDecision.exit === 'blocked' ? { level: 'warn' as const } : {}), compact: { stringMax: 256, arrayMax: 6, maxDepth: 5 } });
           if (parsedDecision.verdict === 'UNCONVERGEABLE' && budgetDecision.stop && budgetDecision.exit === 'blocked') {
+            // FLEX-RFC §A-③(0.2.25 FLEX-PREMISE-REVIEW) — 확정된 UNCONVERGEABLE 의 전제 되짚기 카드를
+            // «별도 원장 사건»으로 남긴다. 카드는 판정 provider 가 UNCONVERGEABLE 을 «내기 전에» 세웠다
+            // (createReworkBudgetJudgment 가 premiseInput 으로 buildPremiseCard 를 부른다) — 이 자리는 그
+            // 카드를 원장에 옮기기만 한다. 카드가 없으면 사건도 없다.
+            const evidenceRecord: unknown = evidenceObservation.evidence;
+            const premiseReview = typeof evidenceRecord === 'object' && evidenceRecord !== null
+              ? (evidenceRecord as { premiseReview?: PremiseReviewCard }).premiseReview
+              : undefined;
+            if (premiseReview) {
+              observe('premise-review', { round, runId: premiseReview.runId, findings: premiseReview.findings, proposal: premiseReview.proposal ?? null }, { level: 'warn' as const, compact: { stringMax: 256, arrayMax: 6, maxDepth: 5 } });
+            }
             try {
               const shadow = inferDecompositionShadow(opts.feature);
               observe('decomposition-shadow', { round, ...shadow });
@@ -7130,8 +7169,12 @@ async function runSelfImplementInner(
     offDiffEvidence = parseOffDiffEvidence(evidenceText);
     const requiredEvidence = requiredEvidenceFromGoal(opts.feature);
     const limitationCount = authorLimitationCountFromGoal(opts.feature);
-    runRequiredEvidenceChecks(opts.feature, wt.path, runId);
-    const coverage = coverRequiredEvidence(requiredEvidence, offDiffEvidence.items);
+    const evidenceChecks = runRequiredEvidenceChecks(opts.feature, wt.path, runId);
+    const tagCoverage = coverRequiredEvidence(requiredEvidence, offDiffEvidence.items);
+    // EVIDENCE-WIRING-EXEMPT — a requirement whose own `|| bun test …` check passed is covered by what the harness measured.
+    const settledByCheck = settleUncoveredEvidence(tagCoverage.uncovered, { checks: evidenceChecks, required: requiredEvidence });
+    if (settledByCheck.coveredByCheck.length) debug.log('self-implement.evidence', 'covered-by-check', { round, tags: settledByCheck.coveredByCheck });
+    const coverage = { ...tagCoverage, uncovered: settledByCheck.uncovered, uncoveredCount: settledByCheck.uncovered.length };
     evidenceCoverage = requiredEvidence.length || limitationCount
       ? {
         required: requiredEvidence.length,
@@ -7226,21 +7269,35 @@ async function runSelfImplementInner(
       // 🩸 10-09: 범위 타입검사 임시 tsconfig(.elanous-typecheck-scope-*.json · scoped-typecheck-config.ts)가 옛 가지에 커밋된 뒤 지워지면
       //   «추적 파일 삭제» 라 impact=critical → 결정 카드가 «irreversible» 로 대표에게 갔다(D-20·35·43·80). 하니스 임시물은 범위 판정 밖이다.
       const scope = detectDeclaredScopeDiff({ goalDocument, changedFiles: gateRouteFiles?.filter(file => !harnessSeededPaths.includes(file) && !isHarnessTempArtifact(file)), nameCap: gateRouteFiles?.length });
-      if (scope.status === 'known' && scope.outsideCount !== null && scope.outsideCount > 0) {
-        const result: ExecutionClarificationResult = await (s.askExecutionClarification ?? askAtExecution)({
+      // SCOPE-HELD-DELIVER — no declared target means the declaration was unreadable (e.g. paths listed with spaces), not
+      // that every edit is outside it: treat it like an unknown declaration and go on to the gate.
+      if (scope.status === 'known' && scope.declaredCount === 0 && (scope.outsideCount ?? 0) > 0) {
+        observe('execution-scope-undeclared', { outsideCount: scope.outsideCount });
+        progress('implementing', `⚠️ 대상 경로를 하나도 못 읽었다 — 범위 밖 ${scope.outsideCount}개를 묻지 않고 게이트로 간다`);
+      } else if (scope.status === 'known' && scope.outsideCount !== null && scope.outsideCount > 0) {
+        const impact = scope.outsideNames.some(file => {
+          if (existsSync(join(wt.path, file))) return false;
+          const tracked = runGitCommand(wt.path, ['ls-files', '--error-unmatch', '--', file], { encoding: 'utf8', timeout: 20_000 });
+          return tracked.status === 0;
+        }) ? 'critical' : 'high';
+        const declaredPaths = declaredPathsFromGoalDocumentText(goalDocument!).paths;
+        const { adjacent, other } = classifyOutOfScope(scope.outsideNames, declaredPaths);
+        // Only edits next to a declared file, none of them a deletion: the gate judges them — nobody has to answer first.
+        const adjacentOnly = impact !== 'critical' && !scope.nameCapReached && other.length === 0 && adjacent.length === scope.outsideCount;
+        if (adjacentOnly) {
+          observe('execution-scope-adjacent-accepted', { outsideCount: scope.outsideCount, adjacent });
+          progress('implementing', `범위 밖 인접 편집 ${adjacent.length}개(대상과 같은 폴더) — 묻지 않고 게이트로 간다: ${adjacent.join(' · ')}`);
+        }
+        const result: ExecutionClarificationResult = adjacentOnly ? { choice: 'Expand scope', by: 'assumed' } : await (s.askExecutionClarification ?? askAtExecution)({
           id: 'execution_scope', decision: 'Changing files outside the authored target paths',
           prompt: 'Implementation changed files outside the authored target paths. Keep the original scope or expand it?',
           whyNow: 'Continuing to gate or publish these edits would expand the requested scope.',
-          impact: scope.outsideNames.some(file => {
-            if (existsSync(join(wt.path, file))) return false;
-            const tracked = runGitCommand(wt.path, ['ls-files', '--error-unmatch', '--', file], { encoding: 'utf8', timeout: 20_000 });
-            return tracked.status === 0;
-          }) ? 'critical' : 'high',
+          impact,
           options: [
             { label: 'Stay in scope', description: 'Do not accept out-of-scope changes.', recommended: true },
             { label: 'Expand scope', description: 'Accept the changed files outside the target paths.' },
           ],
-        }, { runId, outsideFiles: scope.outsideNames, outsideCount: scope.outsideCount, declaredPaths: declaredPathsFromGoalDocumentText(goalDocument!).paths }, { writeRunLedger: s.writeRunLedger });
+        }, { runId, outsideFiles: scope.outsideNames, outsideCount: scope.outsideCount, declaredPaths }, { writeRunLedger: s.writeRunLedger });
         if ('parked' in result || 'deferred' in result || result.choice !== 'Expand scope') {
           finalizeSupervisorDeliveries('execution-scope-not-approved');
           observe('execution-scope-held', { outsideCount: scope.outsideCount, disposition: 'parked' in result ? 'parked' : 'in-scope-only' });
@@ -7487,7 +7544,7 @@ async function runSelfImplementInner(
       }
       lastGateUnverifiedKey = unverifiedRepeatKey(gate.unverified);
       roundSummaryEntries.push({ round, kind: 'gate', tried: impl.summary.replace(/\s+/g, ' ').trim().slice(0, 200),
-        failedIds: gate.baselineFailures?.map(({ name, file, attribution }) => `${name}@${file ?? '?'} (${attribution})`),
+        failedIds: gate.baselineFailures?.map(formatGateFailedId),
         filesTouched: gateRouteFiles,
       });
       round++;
@@ -7831,6 +7888,14 @@ async function runSelfImplementInner(
         continue;
       }
     }
+    // EVIDENCE-WIRING-EXEMPT — a commandless [preservation] is covered once the changed-file gate passed.
+    if (gate.passed && evidenceCoverage?.uncoveredCount) {
+      const settledByGate = settleUncoveredEvidence(evidenceCoverage.uncovered, { gatePassed: true, required: requiredEvidenceFromGoal(opts.feature) });
+      if (settledByGate.coveredByGate.length) {
+        debug.log('self-implement.evidence', 'covered-by-gate', { round, tags: settledByGate.coveredByGate });
+        evidenceCoverage = { ...evidenceCoverage, uncovered: settledByGate.uncovered, uncoveredCount: settledByGate.uncovered.length };
+      }
+    }
     if (gate.passed && autoMergeEnabled(opts) && evidenceCoverage?.uncoveredCount && !uncoveredFeedbackUsed && round < effectiveMax) {
       const uncovered = [...evidenceCoverage.uncovered];
       const nextRound = round + 1;
@@ -7869,8 +7934,11 @@ async function runSelfImplementInner(
     ? { completionDisposition }
     : {};
   const reviewReal = !!review?.reviewed;
+  // REVIEW-ACP-POD ②-b — 호스트가 병합하는(mergeByHost) 런에서만 «미룬 리뷰»를 merge-ready 로 받는다. 판정은 호스트 ACP 리뷰가 한다.
+  //   ⛔ 그 밖의 런에선 종전 그대로 no-real-review(HITL) — 리뷰 없이 병합하는 길을 새로 열지 않는다.
+  const reviewDeferredToHost = !reviewReal && review?.deferredToHost === true && opts.mergeByHost === true && review.mustFix.length === 0;
   const reviewedHeadResult = reviewReal && reviewedHeadCommit ? { reviewedHeadCommit } : {};
-  const reviewDiffComplete = review?.diffTruncated === false;
+  const reviewDiffComplete = review?.diffTruncated === false || reviewDeferredToHost;
   const requiredEvidenceMissing = evidenceCoverage !== undefined && evidenceCoverage.uncoveredCount > 0;
   const { latestSignalIncomplete } = signalIncompleteState(roundClassifications);
   const autoMerge = autoMergeEnabled(opts);
@@ -7880,7 +7948,7 @@ async function runSelfImplementInner(
   //   종전 `verdict !== 'fail'` 만 보던 이 자리에서 auto 가 됐다(#25914 · #25934 · #25918 실물).
   //   ⭐ 반사가 must-fix 를 «전부» 기각하면 위에서 `review.mustFix = []` 로 비우므로 그 길은 종전 그대로다.
   const reviewMustFixCount = review?.mustFix?.length ?? 0;
-  let canAuto = !gateTimeoutUnmeasured && autoMerge && reviewReal && review!.verdict !== 'fail' && reviewMustFixCount === 0 && reviewDiffComplete && !requiredEvidenceMissing && !latestSignalIncomplete && !decisionSignalRed;
+  let canAuto = !gateTimeoutUnmeasured && autoMerge && (reviewReal || reviewDeferredToHost) && review!.verdict !== 'fail' && reviewMustFixCount === 0 && reviewDiffComplete && !requiredEvidenceMissing && !latestSignalIncomplete && !decisionSignalRed;
   // ⭐⭐ 통과의 «이유»를 둘로 가른다 — ⛔ 막지는 «않는다»(대표 판단 2026-08-11: ⓐ 문면만).
   //   📏 근거(`JDG-T36` · merge-decision 300건 전수): `verifyByBreaking` 이 실린 8건의 auto 중 ***5건***이
   //     ***ran=true 인데 distinguishes=0***(아무것도 안 가르는 자기검증)이었고, 그 다섯이 «전부»
@@ -7891,7 +7959,8 @@ async function runSelfImplementInner(
   //     그건 `skippedReason` 이 이미 가른다. 그래서 여기 조건에 넣지 않는다.
   const verifyInconclusive = gate.verifyByBreaking?.ran === true && (gate.verifyByBreaking.distinguishes ?? 0) === 0;
   const verifyNothingAtBase = verifyInconclusive && (gate.verifyByBreaking?.missingAtBase ?? 0) > 0;
-  let mergeReason = canAuto ? (verifyNothingAtBase ? 'review-clean-verify-nothing-at-base' : verifyInconclusive ? 'review-clean-verify-inconclusive' : 'review-clean-armed')
+  let mergeReason = canAuto && reviewDeferredToHost ? REVIEW_DEFERRED_TO_HOST_MERGE_REASON
+    : canAuto ? (verifyNothingAtBase ? 'review-clean-verify-nothing-at-base' : verifyInconclusive ? 'review-clean-verify-inconclusive' : 'review-clean-armed')
     : gateTimeoutUnmeasured ? GATE_TIMEOUT_UNMEASURED_MERGE_REASON
     : !autoMerge ? 'no-auto-flag'
       : requiredEvidenceMissing ? 'required-evidence-uncovered'
@@ -8311,7 +8380,7 @@ async function runSelfImplementInner(
     if (opts.mergeByHost) {
       observe('merge-ready', { number: pr.number, checkedHeadCommit });
       progress('pr-opened', `호스트 재게이트 대기 (#${pr.number})`);
-      return { ok: true, stage: 'merge-ready', node: 'open-pr', ...resolveRunOutcome({ termination: 'completed' }), ...decisionSignalResult, ...preservedCompletionResult, ...(completionStatus ? { completionStatus } : {}), sessionId, worktreePath: wt.path, branch: wt.branch, gate, ...(review ? { review } : {}), ...reviewedHeadResult, checkedHeadCommit: checkedHeadCommit!, prUrl: pr.url, prNumber: pr.number };
+      return { ok: true, stage: 'merge-ready', node: 'open-pr', ...resolveRunOutcome({ termination: 'completed' }), ...decisionSignalResult, ...preservedCompletionResult, ...(completionStatus ? { completionStatus } : {}), sessionId, worktreePath: wt.path, branch: wt.branch, gate, ...(review ? { review } : {}), ...reviewedHeadResult, ...(reviewDeferredToHost ? { mergeReason } : {}), checkedHeadCommit: checkedHeadCommit!, prUrl: pr.url, prNumber: pr.number };
     }
     // Landing freeze: the in-flight marker is written before the freeze is read, so a `freeze on` racing this merge waits.
     // The resume sweep merges from the repository that owns the run worktree (the worktree itself may be cleaned up).

package/src/self-implement/premise-review.ts

@@ -0,0 +1,159 @@
+/**
+ * 전제 다섯 되짚기 — premise-review 카드 순수 계산(FLEX-RFC §A-③).
+ *
+ * P1 골 · P2 자(판정 시험) · P3 범위(대상 경로) · P4 쪼갬 · P5 환경.
+ * ⛔ LLM 없이 잴 수 있는 둘만 판정한다 — P3(범위) · P2(preexisting 자).
+ * ⛔ unknown 은 «맞다(ok)»가 아니다 — 못 잰 전제는 ok 로 지어내지 않는다.
+ */
+
+import type { FlexSignalsInput } from './flex-signals.js';
+
+/** 못 잰 근거의 표기 — P1·P4·P5 는 이 값으로 unknown 으로만 나간다. */
+const UNMEASURED = '못 잼';
+
+export type PremiseKey = 'P1' | 'P2' | 'P3' | 'P4' | 'P5';
+export type PremiseVerdict = 'ok' | 'wrong' | 'unknown';
+
+export interface PremiseFinding {
+  premise: PremiseKey;
+  verdict: PremiseVerdict;
+  evidence: string;
+}
+
+/** `computeFlexSignals` 와 같은 꼴의 값 객체 + 카드가 가리킬 run 식별자와 P3 판정의 경계. */
+export interface PremiseReviewInput extends FlexSignalsInput {
+  runId: string;
+  /** 이 골이 선언한 대상 경로 — 없으면 P3 는 unknown(경계를 못 잼). */
+  targetPaths?: readonly string[] | undefined;
+}
+
+export interface PremiseReviewCard {
+  runId: string;
+  findings: PremiseFinding[];
+  /** wrong 이 있을 때만 «가장 좁은 것» 하나 — wrong 이 없으면 undefined. */
+  proposal: PremiseFinding | undefined;
+}
+
+/** 최신 라운드의 실패 id 목록 — `gateFailedIdsByRound` 의 마지막 칸이 최신(`computeFlexSignals` 와 같은 순서). */
+function latestFailingIds(input: FlexSignalsInput): readonly string[] | undefined {
+  const latest = input.gateFailedIdsByRound.length === 0
+    ? undefined
+    : input.gateFailedIdsByRound[input.gateFailedIdsByRound.length - 1];
+  return latest !== undefined && latest.length > 0 ? latest : undefined;
+}
+
+/**
+ * 게이트 실패 id — 생산자(orchestrator 의 roundSummaryEntries.failedIds)와 P3 파서가 «같은 꼴»을 쓰도록
+ * 둘 다 이 함수 짝을 거친다. 꼴: `${name}@${file ?? '?'} (${attribution})`.
+ */
+export function formatGateFailedId(failure: { name: string; file?: string | undefined; attribution: string }): string {
+  return `${failure.name}@${failure.file ?? '?'} (${failure.attribution})`;
+}
+
+/**
+ * 실패 id 에서 시험 파일을 뽑는다 — `name@` 접두와 ` (귀속)` 접미를 뗀다. 이름에 `@` 가 섞여도
+ * (`user@example.com`·`@scope/pkg`)·경로에 `@`(`src/@scope/x.ts`)가 있어도 가른다: 경로 안의 `@` 는
+ * 세그먼트 머리(맨 앞 또는 `/` 바로 뒤)에만 온다. 그래서 «공백 없고 · 확장자로 끝나고 · 안의 `@` 가 전부
+ * 세그먼트 머리인» 후보를 처음 주는 `@` 가 구분자다. 그런 `@` 가 없거나 파일이 `?`·빈 값이면 undefined
+ * (못 잼) — fail-closed. 남는 모호함: 이름이 `/` 로 끝나면(`a/@src/x.ts`) 구분자를 경로로 읽는다.
+ */
+export function gateFailedIdFile(id: string): string | undefined {
+  const withoutAttribution = id.replace(/\s*\([^)]*\)\s*$/, '');
+  for (let at = withoutAttribution.indexOf('@'); at >= 0; at = withoutAttribution.indexOf('@', at + 1)) {
+    const file = withoutAttribution.slice(at + 1);
+    if (file === '' || file === '?' || /\s/.test(file)) continue;
+    if (!/\.[A-Za-z0-9]+$/.test(file)) continue;
+    if (/[^/]@/.test(file)) continue;
+    return file;
+  }
+  return undefined;
+}
+
+/**
+ * 시험 파일 → 짝 소스. `dir/x.test.ts` → `dir/x.ts`(같은 폴더). 시험 전용 폴더(`test/`·`__tests__/`)의 시험은
+ * 짝을 추측하지 않는다(undefined · fail-closed). 시험 꼴이 아닌 파일은 그 자신이 소스다.
+ */
+export function pairedSourceOf(file: string): string | undefined {
+  const normalized = file.replace(/^\.\//, '');
+  const testMatch = /^(.*)\.test\.([cm]?[jt]sx?)$/.exec(normalized);
+  if (!testMatch) return normalized;
+  if (/(^|\/)(test|tests|__tests__)\//.test(normalized)) return undefined;
+  return `${testMatch[1]}.${testMatch[2]}`;
+}
+
+function insideTargets(path: string, targets: readonly string[]): boolean {
+  return targets.some((target) => {
+    const t = target.replace(/^\.\//, '').replace(/\/+$/, '');
+    return path === t || path.startsWith(`${t}/`);
+  });
+}
+
+/** P2 자 — 판정 시험이 base 에서도 실패(preexisting)면 wrong. 귀속을 못 잼·하나라도 unknown 이면 unknown. */
+function judgeP2(input: PremiseReviewInput): PremiseFinding {
+  const attributions = input.failureAttributions;
+  if (attributions === undefined || attributions.length === 0) {
+    return { premise: 'P2', verdict: 'unknown', evidence: UNMEASURED };
+  }
+  const preexistingIndex = attributions.indexOf('preexisting');
+  if (preexistingIndex >= 0) {
+    const latest = latestFailingIds(input);
+    const evidence = latest !== undefined && preexistingIndex < latest.length
+      ? latest[preexistingIndex]!
+      : 'preexisting';
+    return { premise: 'P2', verdict: 'wrong', evidence };
+  }
+  if (attributions.includes('unknown')) {
+    return { premise: 'P2', verdict: 'unknown', evidence: UNMEASURED };
+  }
+  return { premise: 'P2', verdict: 'ok', evidence: `preexisting 0/${attributions.length}` };
+}
+
+/**
+ * P3 범위 — 실패 시험 파일을 짝 소스로 옮겨 대상 경로와 견준다. 짝지은 소스(또는 시험 파일 자체)가
+ * 대상 경로 안이면 in · 짝은 지었는데 밖이면 wrong(그 시험 파일을 근거로). 파일을 못 읽거나 짝을 못 지은
+ * 실패가 하나라도 있고 wrong 이 없으면 unknown — 추측으로 ok 를 만들지 않는다.
+ */
+function judgeP3(input: PremiseReviewInput): PremiseFinding {
+  const latest = latestFailingIds(input);
+  const targets = input.targetPaths;
+  if (latest === undefined) return { premise: 'P3', verdict: 'unknown', evidence: UNMEASURED };
+  if (targets === undefined || targets.length === 0) return { premise: 'P3', verdict: 'unknown', evidence: UNMEASURED };
+  let unmeasured: string | undefined;
+  for (const id of latest) {
+    const file = gateFailedIdFile(id);
+    if (file === undefined) { unmeasured ??= `${UNMEASURED}: 실패 파일 없음 (${id})`; continue; }
+    if (insideTargets(file, targets)) continue;
+    const source = pairedSourceOf(file);
+    if (source === undefined) { unmeasured ??= `${UNMEASURED}: 짝 소스 없음 (${file})`; continue; }
+    if (!insideTargets(source, targets)) return { premise: 'P3', verdict: 'wrong', evidence: file };
+  }
+  if (unmeasured !== undefined) return { premise: 'P3', verdict: 'unknown', evidence: unmeasured };
+  return { premise: 'P3', verdict: 'ok', evidence: `failing files all inside target paths (${latest.length})` };
+}
+
+/**
+ * wrong 이 여럿이면 «가장 좁은 것» 하나만 제안한다(RFC §A-③ — F6 카드 → F7 조치). 좁음 = 수정 폭:
+ * P2 자가 가장 좁다(판정 자만 고친다 · 골을 안 건드린다) → P5 환경(재무대) → P4 쪼갬 → P3 범위(경계를 넓힌다)
+ * → P1 골(요구 자체를 바꾼다)이 가장 넓다. findings 배열 순서에 우연히 맡기지 않고 이 rank 로 고른다.
+ */
+const NARROWEST_FIRST: readonly PremiseKey[] = ['P2', 'P5', 'P4', 'P3', 'P1'];
+
+export function narrowestWrong(findings: readonly PremiseFinding[]): PremiseFinding | undefined {
+  for (const premise of NARROWEST_FIRST) {
+    const wrong = findings.find((finding) => finding.premise === premise && finding.verdict === 'wrong');
+    if (wrong !== undefined) return wrong;
+  }
+  return undefined;
+}
+
+/** 전제 다섯을 되짚어 카드를 세운다 — 순수 계산, LLM 을 부르지 않는다. */
+export function buildPremiseCard(input: PremiseReviewInput): PremiseReviewCard {
+  const findings: PremiseFinding[] = [
+    { premise: 'P1', verdict: 'unknown', evidence: UNMEASURED },
+    judgeP2(input),
+    judgeP3(input),
+    { premise: 'P4', verdict: 'unknown', evidence: UNMEASURED },
+    { premise: 'P5', verdict: 'unknown', evidence: UNMEASURED },
+  ];
+  return { runId: input.runId, findings, proposal: narrowestWrong(findings) };
+}

package/src/self-implement/regate-detail.ts

@@ -0,0 +1,79 @@
+/**
+ * HOST-REGATE-REASON-LOST (2nd · 2026-10-12): a host-regate failure must say «why» on the ledger.
+ *
+ * The first repair (#26215) put `hostRegateFailure` on the pr-terminal row but kept only the first 300 characters,
+ * and the gate's own `run()` cut every git/gh error to its first 300 characters as well. A git merge lists one
+ * `CONFLICT` line per file and gh puts the GraphQL reason at the end, so a head cut loses exactly the cause
+ * (run-caaf045f · PR #26472: the autopsy could not tell a conflict from a gate failure).
+ *
+ * The detail is bounded (ledger rows are append-only and replayed) but keeps both ends.
+ */
+export const HOST_REGATE_DETAIL_MAX = 4096;
+
+/** Keep `text` within `max` characters, preserving its head (an eighth) and its tail (the rest). */
+export function clipRegateDetail(text: string, max = HOST_REGATE_DETAIL_MAX): string {
+  if (text.length <= max) return text;
+  const marker = ` …[${text.length - max} chars elided]… `;
+  const head = Math.floor(max / 8);
+  const tail = Math.max(0, max - head - marker.length);
+  return text.slice(0, head) + marker + text.slice(text.length - tail);
+}
+
+/** Files named by `git merge` CONFLICT lines (`CONFLICT (content): Merge conflict in <path>` and the rename/delete forms). */
+export function parseConflictFiles(text: string): string[] {
+  const files = new Set<string>();
+  for (const raw of text.split('\n')) {
+    const line = /^CONFLICT \(([^)]*)\): (.*)$/.exec(raw);
+    if (!line) continue;
+    const rest = line[2]!;
+    const inPath = /Merge conflict in (.+)$/.exec(rest);
+    // modify/delete · rename/delete · file/directory name the path first; other free text is not a path.
+    const path = inPath ? inPath[1] : /^(?:modify\/delete|rename\/delete)$/.test(line[1]!)
+      ? /^(.+?) (?:deleted|renamed) in /.exec(rest)?.[1]
+      : undefined;
+    if (path) files.add(path);
+  }
+  return [...files].sort();
+}
+
+export type MergeFailureKind = 'base-modified' | 'head-moved' | 'conflict' | 'branch-policy' | 'other';
+
+/** Classify a `gh pr merge` failure so a reader can tell a race from a conflict from a policy block without the raw text. */
+export function classifyMergeFailure(detail: string, expectedHead?: string, actualHead?: string | null): MergeFailureKind {
+  if (/base branch was modified/i.test(detail)) return 'base-modified';
+  if (/head branch was modified/i.test(detail)) return 'head-moved';
+  if (expectedHead && actualHead && expectedHead !== actualHead) return 'head-moved';
+  if (/required status|review is required|protected branch|branch policy|not authorized|merge queue/i.test(detail)) return 'branch-policy';
+  if (/\bconflicts?\b/i.test(detail) && !/\bno conflicts?\b/i.test(detail)) return 'conflict';
+  return 'other';
+}
+
+/** One host-regate failure. The optional facts are filled for the pr-head, merge-conflict and merge steps. */
+export interface HostRegateFailure {
+  step: string;
+  detail: string;
+  expectedHead?: string;
+  actualHead?: string | null;
+  conflictFiles?: string[];
+  /** Number omitted from conflictFiles when the ledger's 200-path bound is reached. */
+  conflictFilesOmitted?: number;
+  mergeFailureKind?: MergeFailureKind;
+}
+
+/** The optional facts of a failure, copied only when well-formed (ledger rows come back from disk). */
+export function regateFailureFacts(failure: unknown): Omit<HostRegateFailure, 'step' | 'detail'> {
+  if (!failure || typeof failure !== 'object') return {};
+  const f = failure as Record<string, unknown>;
+  const sha = (value: unknown): value is string => typeof value === 'string' && /^[0-9a-f]{40}$/i.test(value);
+  const kinds: readonly MergeFailureKind[] = ['base-modified', 'head-moved', 'conflict', 'branch-policy', 'other'];
+  const conflictFiles = Array.isArray(f.conflictFiles) ? f.conflictFiles.filter((p): p is string => typeof p === 'string') : undefined;
+  const priorOmitted = Number.isSafeInteger(f.conflictFilesOmitted) && (f.conflictFilesOmitted as number) > 0 ? f.conflictFilesOmitted as number : 0;
+  const conflictFilesOmitted = Math.max(0, (conflictFiles?.length ?? 0) - 200) + priorOmitted;
+  return {
+    ...(sha(f.expectedHead) ? { expectedHead: f.expectedHead } : {}),
+    ...(sha(f.actualHead) || f.actualHead === null ? { actualHead: f.actualHead as string | null } : {}),
+    ...(conflictFiles ? { conflictFiles: conflictFiles.slice(0, 200) } : {}),
+    ...(conflictFilesOmitted > 0 ? { conflictFilesOmitted } : {}),
+    ...(typeof f.mergeFailureKind === 'string' && kinds.includes(f.mergeFailureKind as MergeFailureKind) ? { mergeFailureKind: f.mergeFailureKind as MergeFailureKind } : {}),
+  };
+}

package/src/self-implement/review-deferred-to-host.ts

@@ -0,0 +1,21 @@
+// REVIEW-ACP-POD ②-b — Pod 안 리뷰를 «호스트 리뷰로 미룸».
+//
+// S  ②-a(#26457)로 호스트가 Pod merge-ready PR 을 Claude ACP 로 리뷰한다. 그런데 Pod 안 하니스 리뷰는 그대로 api 로 돌아
+//    켜면 리뷰가 «둘»(돈 두 번)이다.
+// C  오케스트레이터는 «실제 리뷰»(`reviewed=true`)가 있어야 merge-ready 를 낸다 — 그냥 끄면 `no-real-review` 로 멎는다.
+// Q  Pod 안 리뷰를 어떻게 «정직하게» 비우나?
+// A  발사 호스트가 ACP 리뷰를 맡을 때만 자식 env 에 표지를 싣는다. 자식은 api 리뷰를 부르지 않고
+//    `reviewed=false` ⊕ `failureReason=deferred-to-host` 를 낸다(⛔ verdict 를 clean 으로 꾸미지 않는다).
+//    오케스트레이터는 «호스트 병합(mergeByHost)» 런에서만 그 상태를 merge-ready 로 받고 `mergeReason` 에 적어 호스트로 돌려보낸다.
+//    호스트는 그 표지를 보면 ACP 리뷰를 «반드시» 하고, 못 하면(refused) 병합하지 않는다(fail-closed).
+
+/** 발사 호스트 → Pod 자식: 「리뷰는 호스트가 Claude ACP 로 한다 — Pod 안 api 리뷰를 부르지 마라」. 값 `1` 만 켠다. */
+export const REVIEW_DEFERRED_TO_HOST_ENV = 'ELANOUS_REVIEW_DEFERRED_TO_HOST';
+/** 미룬 리뷰의 `failureReason`(`reviewed=false` 의 «왜»). */
+export const REVIEW_DEFERRED_TO_HOST_FAILURE = 'deferred-to-host';
+/** 자식 merge-ready 결과의 `mergeReason` — 호스트가 이 값으로 «Pod 리뷰 없음»을 안다. */
+export const REVIEW_DEFERRED_TO_HOST_MERGE_REASON = 'review-deferred-to-host';
+
+export function reviewDeferredToHostFromEnv(env: Readonly<Record<string, string | undefined>>): boolean {
+  return env[REVIEW_DEFERRED_TO_HOST_ENV]?.trim() === '1';
+}

package/src/self-implement/rework-budget-judgment.ts

@@ -1,4 +1,5 @@
 import { executeClassifyNode } from '../workflow-runtime/nodes/classify.js';
+import { buildPremiseCard, type PremiseReviewCard, type PremiseReviewInput } from './premise-review.js';
 import { SUPERVISION_REWORK_SOURCES, type SupervisionReworkSource } from './supervision-vocabulary.js';
 import type { ClassifyNode, JudgmentContext, JudgmentResult, NodeExecContext, WorkflowDefinition, WorkflowDeps } from '../workflow-runtime/types.js';
 
@@ -22,6 +23,8 @@ interface ObservedValue {
 interface ReworkBudgetEvidence {
   observations: Record<'goal' | 'outcome' | 'history' | 'kind', ObservedValue>;
   modelReasoning: string;
+  /** UNCONVERGEABLE 직전에 세운 premise-review 카드(FLEX-RFC §A-③). 카드를 안 만든 판정에는 키가 없다. */
+  premiseReview?: PremiseReviewCard | undefined;
 }
 
 type EvidencedJudgmentResult = JudgmentResult & { evidence: ReworkBudgetEvidence };
@@ -108,7 +111,13 @@ export function reworkBudgetEvidenceForObservation(readEvidence: () => unknown):
   }
 }
 
-export function createReworkBudgetJudgment(callLLM: WorkflowDeps['callLLM']) {
+/** `createReworkBudgetJudgment` 의 선택 인자 — 주지 않으면 카드 없이 지금 동작 그대로. */
+export interface ReworkBudgetJudgmentOpts {
+  /** UNCONVERGEABLE 판정 직전에 세울 premise-review 카드의 입력(값 객체). */
+  premiseInput?: PremiseReviewInput | undefined;
+}
+
+export function createReworkBudgetJudgment(callLLM: WorkflowDeps['callLLM'], opts: ReworkBudgetJudgmentOpts = {}) {
   let latestEvidence: ReworkBudgetEvidence | undefined;
   const provider = async (contractId: string, ctx: JudgmentContext): Promise<JudgmentResult> => {
     if (contractId !== REWORK_BUDGET_JUDGMENT) {
@@ -145,6 +154,7 @@ export function createReworkBudgetJudgment(callLLM: WorkflowDeps['callLLM']) {
       toolPolicy: {},
     };
     let modelReasoning = '';
+    let premiseReview: PremiseReviewCard | undefined;
     const result = await executeClassifyNode(node, execContext, {
       callLLM: async (args) => {
         modelReasoning = await callLLM(args);
@@ -152,7 +162,11 @@ export function createReworkBudgetJudgment(callLLM: WorkflowDeps['callLLM']) {
       },
       runBash: async () => ({ stdout: '', stderr: '', exitCode: 0 }),
     });
-    const evidence = (): ReworkBudgetEvidence => ({ observations, modelReasoning });
+    const evidence = (): ReworkBudgetEvidence => ({
+      observations,
+      modelReasoning,
+      ...(premiseReview === undefined ? {} : { premiseReview }),
+    });
     const evidenced = (judgment: JudgmentResult): EvidencedJudgmentResult => {
       const resultWithEvidence = withEvidence(judgment, evidence());
       latestEvidence = resultWithEvidence.evidence;
@@ -172,7 +186,13 @@ export function createReworkBudgetJudgment(callLLM: WorkflowDeps['callLLM']) {
           + ` expected one of EXTEND|SUFFICIENT|UNCONVERGEABLE`,
       });
     }
-    return evidenced({ ok: true, output: picked, verdict: mapReworkBudgetVerdict(picked, kind) });
+    const verdict = mapReworkBudgetVerdict(picked, kind);
+    // FLEX-RFC §A-③ — UNCONVERGEABLE 을 «내기 전에» 전제를 되짚는다. 카드는 기존 증거 관측에 실려
+    // 원장('rework-budget' 사건)으로 나간다 — 사건 형식은 그대로, 카드는 증거 안의 추가 칸이다.
+    if (picked === 'UNCONVERGEABLE' && opts.premiseInput !== undefined) {
+      premiseReview = buildPremiseCard(opts.premiseInput);
+    }
+    return evidenced({ ok: true, output: picked, verdict });
   };
   return Object.assign(provider, { evidenceForObservation: () => latestEvidence });
 }

package/src/self-implement/running-runs.ts

@@ -566,7 +566,7 @@ export function queryRunningRuns(options: { includeTest?: boolean; runIds?: read
   }
 }
 
-function isPartialRunningRuns(result: RunningRunsResult): boolean {
+export function isPartialRunningRuns(result: RunningRunsResult): boolean {
   return (result.count !== undefined && typeof result.count !== 'number' && result.count.count === null)
     || result.completeness === 'partial' || hasUnreadableLedgerObservation(result.ledger) || (result.unreadable?.length ?? 0) > 0;
 }
@@ -624,7 +624,7 @@ function formatElapsed(ms: number): string {
 export function renderRunProgressTable(result: RunningRunsResult, nowMs: number = Date.now()): string {
   const partial = isPartialRunningRuns(result);
   const header = partial
-    ? `부분 관측 — 최소 ${runningRunsLowerBound(result)} (읽지 못한 원장: ${runningRunsUnreadable(result).map((item) => `${item.dir}:${item.reason}`).join(',') || '없음'})`
+    ? `부분 관측 — 최소 ${runningRunsLowerBound(result)} (읽지 못한 원장: ${runningRunsUnreadable(result).map((item) => `${item.dir}:${item.reason}`).join(',') || '목록 미관측'})`
     : `도는 런 ${result.counts.running + result.counts['probable-running']} (확인 ${result.counts.running} · 추정 ${result.counts['probable-running']})`;
   const rows = buildRunProgressRows(result, nowMs).map((row) => {
     const phaseUnreadable = row.phaseObservation === 'unreadable' || row.phaseObservation === 'partially-unreadable';

package/src/self-implement/scope-adjacency.ts

@@ -0,0 +1,23 @@
+// SCOPE-HELD-DELIVER (0.2.25) — which out-of-scope edits sit right next to a declared target.
+// 10-11~12: 34 scope-held runs auto-answered «Stay in scope» after 30 unanswered minutes and threw away finished work;
+// the outside files were a declared file's same-folder neighbours (its other tests, a sibling helper, the folder's
+// plugin.json). Those are adjacent: the run goes on to its gate instead of parking on a question nobody answers.
+
+const normalize = (path: string): string => path.replace(/\\/g, '/').replace(/^\.\//, '').replace(/\/+$/, '');
+const dirOf = (path: string): string => { const at = path.lastIndexOf('/'); return at < 0 ? '' : path.slice(0, at); };
+
+/** Split out-of-scope files into those adjacent to a declared target and the rest. A file is adjacent when it sits in
+ *  the same folder as a declared file. A declared folder adds nothing (its contents are already in scope), and a
+ *  declared repo-root file adds nothing (the root holds the whole repo's config). */
+export function classifyOutOfScope(outsideFiles: readonly string[], declaredPaths: readonly string[]): { adjacent: string[]; other: string[] } {
+  const folders = new Set<string>();
+  for (const declared of declaredPaths.map(normalize)) {
+    const dir = dirOf(declared);
+    // A path with an extension names a file; without one it names a folder.
+    if (dir && /\.[A-Za-z0-9]+$/.test(declared.slice(dir.length + 1))) folders.add(dir);
+  }
+  const adjacent: string[] = [];
+  const other: string[] = [];
+  for (const file of outsideFiles) (folders.has(dirOf(normalize(file))) ? adjacent : other).push(file);
+  return { adjacent, other };
+}