elanous 0.2.27 → 0.2.28

package/src/self-implement/rework-policy.ts

@@ -628,17 +628,74 @@ function sameProviderStep(child: ExplicitChildForEscalation): EscalateTarget | n
   return null;
 }
 
+/** escalateChild 매핑 해석 근원 — `env` 가 `config` 를 이기고, 둘 다 없으면 같은 provider 사다리(`ladder`). */
+export type EscalateTargetSource = 'config' | 'env' | 'ladder';
+
+/** `self-dev.rework`/`escalate-target` 관측 한 건. 관측 전용 — 판정·분기는 이 값을 읽지 않는다. */
+export interface EscalateTargetObservation {
+  readonly category: 'self-dev.rework';
+  readonly event: 'escalate-target';
+  readonly data: {
+    readonly source: EscalateTargetSource;
+    readonly from: string;
+    readonly to: string;
+    readonly applied: boolean;
+    /** `applied=false` 의 이유 — provider 경계 또는 불완전한 모델 ID. */
+    readonly reason?: 'cross-provider' | 'invalid-target';
+  };
+}
+
+/** 명시 자식 승급 해석 옵션. `escalateChild`(`tools.selfImplement.escalateChild` from→to 맵)는 호출부가
+ *  config 에서 읽어 전달한다 — ⚠️ Pod 안엔 운영 config 가 없다(POD-CONFIG-SNAPSHOT 전) → Pod 자식은
+ *  `ELANOUS_SELFDEV_ESCALATE_CHILD` env 로만 받는다. 관측 싱크도 호출부가 주입한다(정책은 순수 유지). */
+export interface ExplicitChildEscalateOptions {
+  readonly escalateChild?: Readonly<Record<string, string>>;
+  readonly observe?: (observation: EscalateTargetObservation) => void;
+}
+
+/** 모델 id(`openrouter/z-ai/glm-5.3` 등)의 provider 세그먼트 — 첫 `/` 앞. */
+function escalateModelProvider(model: string): string {
+  return (model.split('/')[0] ?? '').trim();
+}
+
+/** 명시 매핑은 provider/model, OpenRouter 는 openrouter/vendor/model 형태의 완전한 ID 여야 한다. */
+function validEscalateModelId(model: string): boolean {
+  const segments = model.split('/');
+  const required = segments[0] === 'openrouter' ? 3 : 2;
+  return segments.length === required && segments.every((segment) => segment.length > 0 && !/\s/.test(segment));
+}
+
+/** `ELANOUS_SELFDEV_ESCALATE_CHILD=<from>=<to>` 해석 — 첫 `=` 기준. 빈 조각·구문 파손은 매핑 없음. */
+function parseEscalateChildEnv(raw: string | undefined): { from: string; to: string } | undefined {
+  const value = raw?.trim();
+  if (!value) return undefined;
+  const separator = value.indexOf('=');
+  if (separator <= 0) return undefined;
+  const from = value.slice(0, separator).trim();
+  const to = value.slice(separator + 1).trim();
+  return from && to ? { from, to } : undefined;
+}
+
 /** 명시 자식의 승급 타깃 — 같은 provider 안에서만(위 규칙). `ELANOUS_SELFDEV_SOL_*` 는 운영자의 명시
  *  덮어쓰기라 그대로 우선한다. ⛔ `effort` 가 빈 문자열이면 «안 보낸다»는 뜻이다(OpenRouter 는 노력을
- *  보내면 덜 생각했다 — 2026-09-23 실측). */
+ *  보내면 덜 생각했다 — 2026-09-23 실측).
+ *  escalateChild 짝지기(env 가 config 를 이긴다)는 사다리(`sameProviderStep`)보다 «먼저» —
+ *  같은 provider 타깃이면 그 타깃으로, provider 가 다르면 무시 ⊕ 관측하고 사다리로 돌아간다.
+ *  둘 다 없으면 종전 그대로(바이트 불변). */
 export function resolveExplicitChildEscalateTarget(
   tier: EscalateTier,
   env: Record<string, string | undefined> = process.env,
   child?: ExplicitChildForEscalation,
+  options?: ExplicitChildEscalateOptions,
 ): EscalateTarget | null {
   if (tier === 'none') return null;
+  const sink = options?.observe;
+  const emit = sink
+    ? (data: EscalateTargetObservation['data']): void => sink({ category: 'self-dev.rework', event: 'escalate-target', data })
+    : undefined;
   const overrideModel = env.ELANOUS_SELFDEV_SOL_MODEL?.trim();
   if (overrideModel) {
+    if (child) emit?.({ source: 'env', from: child.model, to: overrideModel, applied: true });
     return {
       model: overrideModel,
       provider: env.ELANOUS_SELFDEV_SOL_PROVIDER?.trim() || child?.provider || 'openai-codex',
@@ -646,9 +703,30 @@ export function resolveExplicitChildEscalateTarget(
     };
   }
   if (child === undefined) return null;
+  const effortOverride = env.ELANOUS_SELFDEV_SOL_EFFORT?.trim();
+  // env → config 순서로 이 자식 모델에 대한 매핑을 모은다(없는 층은 건너뛴다).
+  const pairings: Array<{ source: 'env' | 'config'; to: string }> = [];
+  const envPairing = parseEscalateChildEnv(env.ELANOUS_SELFDEV_ESCALATE_CHILD);
+  if (envPairing && envPairing.from === child.model) pairings.push({ source: 'env', to: envPairing.to });
+  const configMap = options?.escalateChild;
+  const configTo = configMap && Object.hasOwn(configMap, child.model) ? configMap[child.model]?.trim() : undefined;
+  if (configTo) pairings.push({ source: 'config', to: configTo });
+  for (const pairing of pairings) {
+    if (!validEscalateModelId(pairing.to)) {
+      emit?.({ source: pairing.source, from: child.model, to: pairing.to, applied: false, reason: 'invalid-target' });
+      continue;
+    }
+    if (escalateModelProvider(pairing.to) === child.provider) {
+      emit?.({ source: pairing.source, from: child.model, to: pairing.to, applied: true });
+      return effortOverride
+        ? { model: pairing.to, provider: child.provider, effort: effortOverride }
+        : { model: pairing.to, provider: child.provider, effort: '' };
+    }
+    emit?.({ source: pairing.source, from: child.model, to: pairing.to, applied: false, reason: 'cross-provider' });
+  }
   const step = sameProviderStep(child);
   if (step === null) return null;
-  const effortOverride = env.ELANOUS_SELFDEV_SOL_EFFORT?.trim();
+  emit?.({ source: 'ladder', from: child.model, to: step.model, applied: true });
   return effortOverride ? { ...step, effort: effortOverride } : step;
 }
 

package/src/self-implement/speed-claims.ts

@@ -2,7 +2,7 @@ import { spawnSync } from 'node:child_process';
 import { readFileSync, readdirSync } from 'node:fs';
 import { dirname, isAbsolute, join, resolve } from 'node:path';
 import { fileURLToPath } from 'node:url';
-import { prodInstanceRoot } from '../instance/resolve.js';
+import { prodInstanceRoot, releaseLedgerRoot } from '../instance/resolve.js';
 import { seatOfTree } from '../loops/orchestrator/traffic.js';
 import { getUserConfig, ORCHESTRATOR_DEFAULTS } from '../user-config.js';
 import { isRunStatus } from './run-status-mapping.js';
@@ -30,6 +30,7 @@ export interface SpeedClaimsMeasurement {
   rows: SpeedClaimsRow[];
   note: string;
   parent: ParentSpeedClaimsMeasurement;
+  daily: SpeedClaimsDailyMeasurement;
 }
 export interface ParentSpeedClaimsEvent {
   ts: string;
@@ -58,6 +59,32 @@ export interface ParentSpeedClaimsMeasurement {
   note: string;
 }
 
+export interface SpeedClaimsReleaseEvent {
+  ts: string;
+  version: string;
+}
+
+export interface SpeedClaimsDailyRow {
+  date: string;
+  landedPrCount: number | null;
+  releasedCount: number | null;
+}
+
+export interface SpeedClaimsDailySourceStatus {
+  status: 'measured' | 'unmeasured';
+  undatedCount: number;
+}
+
+export interface SpeedClaimsDailyMeasurement {
+  scope: 'last-30-days';
+  windowStart: string;
+  windowEnd: string;
+  landedPrs: SpeedClaimsDailySourceStatus;
+  releases: SpeedClaimsDailySourceStatus;
+  rows: SpeedClaimsDailyRow[];
+  note: string;
+}
+
 export interface SpeedClaimsOptions {
   dir?: string;
   list?: (path: string) => string[];
@@ -65,6 +92,8 @@ export interface SpeedClaimsOptions {
   readSeat?: (runId: string, ledgerDirectory: string, entries: readonly RunLedgerEntry[]) => SpeedClaimsSeat;
   readParentEvents?: () => readonly ParentSpeedClaimsEvent[];
   readParentQueue?: () => unknown;
+  now?: number;
+  readReleaseEvents?: () => readonly SpeedClaimsReleaseEvent[];
 }
 
 const NOTE = 'Child-ledger elapsed minutes from the first child event (or checkpoint creation when unavailable) to the recorded child end, including retries and resumes; this is not the parent launch-to-done interval. Only finished runs enter quartiles (linear interpolation at ranks (n-1)*p). In-progress runs (including parked) and missing, unreadable or invalid timestamps are excluded, not assigned zero. Seats come from the same state root checkpoint, launch queue or recorded launch tree; unresolved seats are counted separately. A completed run without a recorded merge is not a merge.';
@@ -192,6 +221,98 @@ function readParentQueueFromState(): unknown {
   return JSON.parse(readFileSync(join(prodInstanceRoot(), 'harness', 'queue.json'), 'utf8')) as unknown;
 }
 
+const DAILY_WINDOW_DAYS = 30;
+const DAY_MS = 86_400_000;
+const DAILY_NOTE = 'Daily rows cover the last 30 UTC calendar days ending on the measurement day. Landed PRs are child run-ledger merged events (merged=true), identified by PR number with the run id as fallback when no number is recorded, de-duplicated across runs, retries and resumes, and attributed to the earliest recorded landing timestamp. Issued releases use the release ledger publishedAt, never a tag target commit date, and are de-duplicated by version. Identities without any valid timestamp are counted as undated, never as a day. A count of zero means the day was queried and had no events; unmeasured means the source could not be queried, and the JSON value is null. Start-to-land and external-change-to-reflection timings stay with the parent and child measurements above.';
+const VERSION_TAG = /^v?\d+(?:\.\d+)+(?:-[\w.]+)?$/;
+
+interface DailyObservation {
+  key: string;
+  ts: number | null;
+}
+
+function utcDayOf(ts: number): string {
+  return new Date(ts).toISOString().slice(0, 10);
+}
+
+/** Landed-PR identity: the recorded PR number, else the run id (one run lands at most one PR). */
+function landedPrIdentity(data: unknown, runId: string): string {
+  const number = object(data)?.number;
+  if (typeof number === 'number' && Number.isInteger(number) && number > 0) return `pr-${number}`;
+  if (typeof number === 'string' && /^\d+$/.test(number.trim()) && Number(number.trim()) > 0) return `pr-${Number(number.trim())}`;
+  return runId;
+}
+
+/** De-duplicate repeated identities by their earliest valid timestamp; identities without any valid timestamp are reported as undated, never dropped silently. */
+function dailyBuckets(observations: readonly DailyObservation[]): { counts: Map<string, number>; undated: number } {
+  const earliest = new Map<string, number>();
+  const undated = new Set<string>();
+  for (const { key, ts } of observations) {
+    if (key.length === 0) continue;
+    if (ts === null) { undated.add(key); continue; }
+    const known = earliest.get(key);
+    if (known === undefined || ts < known) earliest.set(key, ts);
+  }
+  const counts = new Map<string, number>();
+  for (const ts of earliest.values()) {
+    const day = utcDayOf(ts);
+    counts.set(day, (counts.get(day) ?? 0) + 1);
+  }
+  let undatedCount = 0;
+  for (const key of undated) if (!earliest.has(key)) undatedCount++;
+  return { counts, undated: undatedCount };
+}
+
+/** Read-only daily census of landed PRs and issued releases over the last 30 UTC days. */
+function measureDailySpeedClaims(
+  landed: readonly DailyObservation[],
+  landedMeasured: boolean,
+  releases: readonly SpeedClaimsReleaseEvent[] | null,
+  now: number,
+): SpeedClaimsDailyMeasurement {
+  const windowEnd = utcDayOf(now);
+  const startMs = Date.parse(`${windowEnd}T00:00:00.000Z`) - (DAILY_WINDOW_DAYS - 1) * DAY_MS;
+  const landedBuckets = dailyBuckets(landed);
+  const releaseBuckets = releases === null
+    ? null
+    : dailyBuckets(releases.map((event) => ({ key: event.version.trim().replace(/^v(?=\d)/, ''), ts: parseTime(event.ts) })));
+  const landedComplete = landedMeasured && landedBuckets.undated === 0;
+  const releasesComplete = releaseBuckets !== null && releaseBuckets.undated === 0;
+  return {
+    scope: 'last-30-days', windowStart: utcDayOf(startMs), windowEnd,
+    landedPrs: { status: landedComplete ? 'measured' : 'unmeasured', undatedCount: landedBuckets.undated },
+    releases: { status: releasesComplete ? 'measured' : 'unmeasured', undatedCount: releaseBuckets?.undated ?? 0 },
+    rows: Array.from({ length: DAILY_WINDOW_DAYS }, (_, index) => {
+      const date = utcDayOf(startMs + index * DAY_MS);
+      return {
+        date,
+        landedPrCount: landedComplete ? (landedBuckets.counts.get(date) ?? 0) : null,
+        releasedCount: releasesComplete ? (releaseBuckets!.counts.get(date) ?? 0) : null,
+      };
+    }),
+    note: DAILY_NOTE,
+  };
+}
+
+/** Release manifests record the actual publication time; lightweight tags have no creation timestamp. */
+function readReleaseEventsFromLedger(): SpeedClaimsReleaseEvent[] {
+  const root = join(releaseLedgerRoot(), 'release');
+  const versions = readdirSync(root).filter((version) => VERSION_TAG.test(version));
+  const events: SpeedClaimsReleaseEvent[] = [];
+  for (const version of versions) {
+    // Absence and corruption are both unmeasured; only a readable manifest proves preparation.
+    const value: unknown = JSON.parse(readFileSync(join(root, version, 'release.json'), 'utf8')) as unknown;
+    const record = object(value);
+    if (record?.version !== version) throw new Error(`release manifest version mismatch: ${version}`);
+    if (!Object.hasOwn(record, 'publishedAt')) continue; // Prepared, not yet published.
+    if (typeof record.publishedAt !== 'string' || parseTime(record.publishedAt) === null) throw new Error(`invalid release publication time: ${version}`);
+    events.push({ version, ts: record.publishedAt });
+  }
+  // A checkout without tags or a release history cannot establish a queried zero.
+  if (versions.length === 0) throw new Error('release publication history unavailable');
+  return events;
+}
+
 /** Parent launch-to-done census; no child ledger id is used as a join key. */
 export function measureParentSpeedClaims(
   events: readonly ParentSpeedClaimsEvent[], queue?: unknown,
@@ -262,7 +383,13 @@ export function measureParentSpeedClaims(
 
 /** Read-only elapsed-time census of canonical local run ledgers, including unsuccessful runs. */
 export function measureSpeedClaims(options: SpeedClaimsOptions = {}): SpeedClaimsMeasurement {
+  const now = options.now ?? Date.now();
   const ledgerDirectory = resolve(options.dir ?? runLedgerDir());
+  let releaseEvents: readonly SpeedClaimsReleaseEvent[] | null = null;
+  try {
+    const events = (options.readReleaseEvents ?? readReleaseEventsFromLedger)();
+    releaseEvents = Array.isArray(events) ? events : null;
+  } catch { /* An unreadable release source marks releases unmeasured; it does not erase landed-PR observations. */ }
   let parent: ParentSpeedClaimsMeasurement;
   try {
     const events = (options.readParentEvents ?? readParentEventsFromLogs)();
@@ -272,21 +399,27 @@ export function measureSpeedClaims(options: SpeedClaimsOptions = {}): SpeedClaim
   } catch {
     parent = { ...measureParentSpeedClaims([]), status: 'cannot-measure' };
   }
-  const result: SpeedClaimsMeasurement = {
+  const result: Omit<SpeedClaimsMeasurement, 'daily'> = {
     scope: 'self-implement-run-ledger', ledgerDirectory, status: 'measured',
     completedCount: 0, inProgressCount: 0, cannotMeasureCount: 0, unknownSeatCount: 0, rows: [], note: NOTE, parent,
   };
   let files: string[];
   try { files = (options.list ?? readdirSync)(ledgerDirectory); }
-  catch { return { ...result, status: 'cannot-measure' }; }
+  catch { return { ...result, status: 'cannot-measure', daily: measureDailySpeedClaims([], false, releaseEvents, now) }; }
 
+  const landed: DailyObservation[] = [];
+  let ledgerUnreadable = false;
   const groups = new Map<string, { seat: SpeedClaimsSeat; outcome: SpeedClaimsOutcome; values: number[] }>();
   for (const file of files.sort()) {
     const runId = file.endsWith('.jsonl') ? file.slice(0, -'.jsonl'.length) : '';
     if (!CANONICAL_RUN_ID.test(runId)) continue;
     try {
       const entries = loadRunLedger(runId, ledgerDirectory, options.read);
-      if (!entries || entries.length === 0) { result.cannotMeasureCount++; continue; }
+      if (!entries || entries.length === 0) { ledgerUnreadable = true; result.cannotMeasureCount++; continue; }
+      const merges = entries.filter((entry) => entry.event === 'merged' && entry.data.merged === true);
+      const numbered = new Set(merges.map((entry) => landedPrIdentity(entry.data, runId)).filter((key) => key !== runId));
+      if (numbered.size > 1) ledgerUnreadable = true; // Conflicting PR identities cannot establish a daily count.
+      else for (const entry of merges) landed.push({ key: numbered.size === 1 ? [...numbered][0]! : runId, ts: parseTime(entry.timestamp) });
       const end = endOfRun(entries);
       const first = time(entries[0]);
       const createdAt = checkpointFor(runId, ledgerDirectory)?.createdAt;
@@ -306,14 +439,14 @@ export function measureSpeedClaims(options: SpeedClaimsOptions = {}): SpeedClaim
       group.values.push((endTime! - start) / 60_000);
       groups.set(key, group);
       result.completedCount++;
-    } catch { result.cannotMeasureCount++; }
+    } catch { result.cannotMeasureCount++; ledgerUnreadable = true; }
   }
   result.rows = [...groups.values()].map(({ seat, outcome, values }) => {
     values.sort((a, b) => a - b);
     return { seat, outcome, count: values.length, underOneMinuteCount: values.filter((minutes) => minutes < 1).length, medianMinutes: percentile(values, 0.5), q1Minutes: percentile(values, 0.25), q3Minutes: percentile(values, 0.75) };
   }).sort((a, b) => SEATS.indexOf(a.seat) - SEATS.indexOf(b.seat) || a.outcome.localeCompare(b.outcome));
   if (result.cannotMeasureCount > 0) result.status = 'cannot-measure';
-  return result;
+  return { ...result, daily: measureDailySpeedClaims(landed, !ledgerUnreadable, releaseEvents, now) };
 }
 
 export function renderSpeedClaims(result: SpeedClaimsMeasurement): string {
@@ -334,5 +467,11 @@ export function renderSpeedClaims(result: SpeedClaimsMeasurement): string {
     `parent cannot measure (excluded): ${result.parent.cannotMeasureCount}`,
     ...result.parent.rows.map((row) => `seat=${row.seat} stage=${row.stage} merged=${row.merged} count=${row.count} median=${row.medianMinutes}m q1=${row.q1Minutes}m q3=${row.q3Minutes}m`),
     `parent note: ${result.parent.note}`,
+    'daily landed PRs / issued releases (last 30 days, UTC):',
+    `daily window: ${result.daily.windowStart}..${result.daily.windowEnd}`,
+    `landed PRs status: ${result.daily.landedPrs.status} (undated: ${result.daily.landedPrs.undatedCount})`,
+    `issued releases status: ${result.daily.releases.status} (undated: ${result.daily.releases.undatedCount})`,
+    ...result.daily.rows.map((row) => `date=${row.date} landed=${row.landedPrCount ?? 'unmeasured'} released=${row.releasedCount ?? 'unmeasured'}`),
+    `daily note: ${result.daily.note}`,
   ].join('\n');
 }

package/src/skills/tools/persistent-grounding.ts

@@ -212,6 +212,26 @@ function groundingMessages(goal: string, terms: readonly string[]): LLMMessage[]
   }];
 }
 
+function unreadEvidencePaths(evidence: string, readFiles: ReadonlySet<string>): string[] {
+  return Array.from(new Set(evidence.split(/\r?\n/)
+    .flatMap(repositoryRelativeCodePaths)
+    .filter((path) => isCodeRootPath(path) && !readFiles.has(path)))).slice(0, 8);
+}
+
+function rejectedEvidenceFeedback(evidence: string, readFiles: ReadonlySet<string>): LLMMessage {
+  const unread = unreadEvidencePaths(evidence, readFiles);
+  const read = Array.from(readFiles).slice(0, 8);
+  return {
+    role: 'user',
+    content: [
+      'Your completion evidence was rejected because it cited a repository path that Read did not verify.',
+      `Unverified paths: ${unread.length ? unread.join(', ') : '(none in allowed code roots)'}`,
+      `Paths actually Read: ${read.length ? read.join(', ') : '(none)'}`,
+      'Try once more. Read each relevant unverified path before citing it, or remove it from the evidence. Cite only repository-relative code paths that this grounding run actually Read, each with a concrete reason. If no path is relevant, do not invent evidence.',
+    ].join('\n'),
+  };
+}
+
 /** Bounded, shared, read-only grounding loop. Failure or no candidates returns null. */
 export async function groundPersistently(goal: string, cwd: string, deps: PersistentGroundingDeps = {}): Promise<PersistentGroundingResult | null> {
   const startedAt = performance.now();
@@ -250,20 +270,53 @@ export async function groundPersistently(goal: string, cwd: string, deps: Persis
       },
     };
     const options: GoalLoopOptions = { maxIterations, objective: `Persistent repository grounding: ${goal.slice(0, 200)}` };
-    const result = await loop(ctx, options);
-    const verified = result.goalComplete && result.stopReason === 'goal_complete'
+    const first = await loop(ctx, options);
+    const firstVerified = first.goalComplete && first.stopReason === 'goal_complete'
       ? verifiedEvidenceFromCompletion(completionEvidence, readFiles)
       : {
         files: [],
         evidence: [],
         rejection: { noPath: 0, outsideRoot: 0, notRead: 0, noMeaningfulText: 0, samples: [] },
       };
+    let result = first;
+    let verified = firstVerified;
+    let feedbackRetry = false;
+    let retryIterations = 0;
+    if (firstVerified.rejection.notRead > 0) {
+      feedbackRetry = true;
+      const feedback = rejectedEvidenceFeedback(completionEvidence, readFiles);
+      completionEvidence = '';
+      try {
+        const retry = await loop({ ...ctx, sessionId: `${sessionId}-evidence-retry`, messages: [...ctx.messages, feedback] },
+          { ...options, maxIterations: 2 });
+        retryIterations = retry.iterations;
+        const retryVerified = retry.goalComplete && retry.stopReason === 'goal_complete'
+          ? verifiedEvidenceFromCompletion(completionEvidence, readFiles)
+          : null;
+        if (retryVerified) {
+          verified = {
+            files: Array.from(new Set([...firstVerified.files, ...retryVerified.files])),
+            evidence: Array.from(new Set([...firstVerified.evidence, ...retryVerified.evidence])),
+            rejection: retryVerified.rejection,
+          };
+          result = retry;
+        }
+      } catch (error) {
+        try {
+          debug.log('grounding.persistent', 'evidence-retry-failed', { sessionId,
+            reason: error instanceof Error ? error.message.slice(0, 200) : String(error).slice(0, 200) }, { level: 'warn' });
+        } catch {
+          // A failed retry and its telemetry cannot erase evidence already verified by Read.
+        }
+      }
+    }
+    const totalIterations = first.iterations + retryIterations;
     try {
       debug.log('grounding.persistent', 'finished', {
         sessionId,
         maxIterations,
         durationMs: durationMs(),
-      iterations: result.iterations,
+      iterations: totalIterations,
       stopReason: result.stopReason,
       // ⛔ 「끝났다」가 두 뜻이다 — ⓐ 스스로 목표를 끝냈다 ⓑ 턴이 «그냥» 끝났다.
       //   📏 2026-08-11 72차 전수(400건): 후보 0인 finished 가 5건이고 ***다섯 다 end_turn*** 이었다.
@@ -273,6 +326,8 @@ export async function groundPersistently(goal: string, cwd: string, deps: Persis
       candidateCount: verified.files.length,
       evidenceCount: verified.evidence.length,
       evidenceRejection: verified.rejection,
+      feedbackRetry,
+      ...(feedbackRetry ? { firstEvidenceRejection: firstVerified.rejection } : {}),
         toolCalls,
       });
     } catch {
@@ -281,7 +336,7 @@ export async function groundPersistently(goal: string, cwd: string, deps: Persis
     return {
       files: verified.files,
       evidence: verified.evidence,
-      iterations: result.iterations,
+      iterations: totalIterations,
       stopReason: result.stopReason,
     };
   } catch (error) {

package/src/skills/tools/shell.ts

@@ -31,9 +31,13 @@ import {
   runShell as runnerDispatch,
   getShellRunnerDeps,
 } from '../../shell-runner/dispatch.js';
-import type { ShellResult as RunnerResult } from '../../shell-runner/types.js';
+import type { ShellHandle, ShellResult as RunnerResult } from '../../shell-runner/types.js';
 import { debug } from '../../debug/log.js';
 import { getSessionCwd } from '../../session/working-dir.js';
+import {
+  globalTaskNotificationQueue,
+  type TaskNotificationQueue,
+} from '../../agent/task-notification.js';
 
 /** LLMToolSpec — the shape surfaced to the model. Kept narrow so
  *  accidental misuse is hard: no env override, no cwd outside the
@@ -178,6 +182,62 @@ function logShellDispatch(event: string, data: Record<string, unknown>): void {
   debug.log('shell.dispatch', event, data);
 }
 
+// ── bg completion → turn notification queue (TUI-BG-SHELL) ───────────
+// When a bg-mode shell handle settles, its result is no longer awaited
+// by anyone — the spawn already returned the `outcome=started` stub.
+// To surface the eventual result we enqueue one task-notification into
+// the existing turn notification queue (injected at turn-start by
+// chat/skill-runner). One notice per handle; opening an idle turn from
+// the wake signal is deliberately NOT wired here (TUI-EVENT-TURN-QUEUE).
+
+export interface ShellCompletionNoticeInput {
+  handle: ShellHandle;
+  ok: boolean;
+  exitCode: number | null;
+  durationMs: number;
+  /** Set when ok === false — the runner's failure reason. */
+  errorMessage?: string;
+  /** Caller-supplied free-form label (RunShell's `description` arg). */
+  label?: string;
+}
+
+/** Per-handle dedup guard — module-local, keyed by the handle OBJECT.
+ *  A registry replacement may mint a NEW handle that reuses an id; the
+ *  shell layer must not swallow that fresh completion as a duplicate
+ *  settle (an id-keyed record here would). */
+const notifiedHandles = new WeakSet<ShellHandle>();
+let nextShellCompletionNoticeId = 0;
+
+/** Enqueue one bg shell-handle completion notice. Returns true when this
+ *  handle had not notified before (the notice was handed to `queue`),
+ *  false on a duplicate settle of the same handle. The queue's own
+ *  The shared queue deduplicates by taskId even after drain, so each
+ *  completed handle gets its own notification id. The original handle id
+ *  remains in the output for ShellPoll and human follow-up.
+ *  `queue` defaults to the global turn notification queue; injectable
+ *  so tests can isolate. */
+export function enqueueShellCompletionNotice(
+  input: ShellCompletionNoticeInput,
+  queue: TaskNotificationQueue = globalTaskNotificationQueue,
+): boolean {
+  if (notifiedHandles.has(input.handle)) return false;
+  notifiedHandles.add(input.handle);
+  const header = `RunShell bg handle=${input.handle.id} ${input.ok ? 'completed' : 'failed'}`;
+  const output = `${header}${input.ok || input.exitCode !== null ? ` exitCode=${input.exitCode ?? 'null'}` : ''} elapsed=${input.durationMs}ms${input.errorMessage ? ` error=${input.errorMessage}` : ''}`;
+  queue.enqueue({
+    taskId: `shell:${input.handle.id}:${++nextShellCompletionNoticeId}`,
+    agentName: 'RunShell',
+    ...(input.label ? { label: input.label } : {}),
+    state: input.ok ? 'done' : 'error',
+    output,
+    truncated: false,
+    ...(input.errorMessage ? { errorMessage: input.errorMessage } : {}),
+    durationMs: input.durationMs,
+    finishedAt: Date.now(),
+  });
+  return true;
+}
+
 export async function dispatchRunShell(
   rawArgs: Record<string, unknown>,
   opts: { signal?: AbortSignal } = {},
@@ -269,6 +329,16 @@ async function dispatchViaRunner(args: {
           exitCode: result.exitCode ?? null,
           durationMs: result.durationMs,
         });
+        const ok = result.outcome === 'exit' && result.exitCode === 0;
+        enqueueShellCompletionNotice({
+          handle,
+          ok,
+          exitCode: result.exitCode ?? null,
+          durationMs: result.durationMs,
+          ...(!ok ? { errorMessage: result.exitCode !== undefined
+            ? `exitCode=${result.exitCode}` : `outcome=${result.outcome}` } : {}),
+          ...(args.description !== undefined ? { label: args.description } : {}),
+        });
       },
       (err: unknown) => {
         debug.log('skill-tool.shell', 'bg-finished', {
@@ -277,6 +347,14 @@ async function dispatchViaRunner(args: {
           durationMs: Date.now() - bgStartedAt,
           error: err instanceof Error ? err.message : String(err),
         });
+        enqueueShellCompletionNotice({
+          handle,
+          ok: false,
+          exitCode: null,
+          durationMs: Date.now() - bgStartedAt,
+          errorMessage: err instanceof Error ? err.message : String(err),
+          ...(args.description !== undefined ? { label: args.description } : {}),
+        });
       },
     );
     return {

package/src/steward/rejudge-tick.ts

@@ -19,6 +19,7 @@ import { debug } from '../debug/log.js';
 import { effectiveInstanceRoot } from '../instance/resolve.js';
 import { spawnElanous } from '../instance/spawn-elanous.js';
 import { getUserConfig } from '../user-config.js';
+import { resyncPendingPath } from './resync-tick.js';
 
 export const REJUDGE_TICK_EVENT = 'rejudge-tick';
 export const REJUDGE_DEFAULT_MAX_LOAD = 60;
@@ -49,6 +50,7 @@ export interface RejudgeTickDeps {
   config?: () => { rejudgeEveryMinutes?: number; rejudgeMaxLoad?: number } | undefined;
   load?: () => number;
   stampPath?: string;
+  pendingPath?: string;
   run?: () => Promise<RejudgeRun>;
   log?: (event: string, data: Record<string, unknown>) => void;
 }
@@ -110,7 +112,9 @@ export async function rejudgeTick(deps: RejudgeTickDeps = {}): Promise<RejudgeTi
     const now = (deps.now ?? (() => new Date()))();
     const stampPath = deps.stampPath ?? defaultStampPath();
     const lastStartedAt = readStamp(stampPath);
-    if (lastStartedAt && now.getTime() - Date.parse(lastStartedAt) < everyMinutes * 60_000 - REJUDGE_JITTER_MS) {
+    const pendingAt = readPending(deps.pendingPath ?? resyncPendingPath());
+    const savedSinceLastRun = pendingAt !== null && (!lastStartedAt || Date.parse(pendingAt) > Date.parse(lastStartedAt));
+    if (!savedSinceLastRun && lastStartedAt && now.getTime() - Date.parse(lastStartedAt) < everyMinutes * 60_000 - REJUDGE_JITTER_MS) {
       return emit({ decision: 'not-due', everyMinutes, lastStartedAt });
     }
     const maxLoad = positive(cfg?.rejudgeMaxLoad) ?? REJUDGE_DEFAULT_MAX_LOAD;
@@ -133,6 +137,13 @@ export async function rejudgeTick(deps: RejudgeTickDeps = {}): Promise<RejudgeTi
   }
 }
 
+function readPending(path: string): string | null {
+  try {
+    const value = (JSON.parse(readFileSync(path, 'utf8')) as { savedAt?: unknown }).savedAt;
+    return typeof value === 'string' && Number.isFinite(Date.parse(value)) ? value : null;
+  } catch { return null; }
+}
+
 if (import.meta.main) {
   try { await (await import('../domains/standalone-log-sink.js')).registerStandaloneLogSink('steward'); }
   catch (error) { console.error(`steward rejudge: log sink registration failed — ${error instanceof Error ? error.message : String(error)}`); }