elanous 0.2.41 → 0.2.42

package/src/auto-research/research-turn.ts

@@ -3,7 +3,8 @@ import { invokeResearchAxes } from '../research-bridge/index.js';
 import { invokeResearch } from '../research-bridge/invoke.js';
 import type { AxisFanoutResult } from '../research/axis-fanout.js';
 import { routeResearchOutput } from '../research/output-router.js';
-import { runGapFollowup } from '../research/gap-followup.js';
+import { planGapFollowup, runGapFollowup } from '../research/gap-followup.js';
+import type { ResearchGapFollowupInput, ResearchGapFollowupPlan } from '../research/gap-followup.js';
 import { debug, redactSecretText } from '../debug/log.js';
 import {
   buildWorkingBackwardsContractPrompt,
@@ -16,13 +17,16 @@ import { attachResearchEvidence, buildResearchQuestionLadderPrompt,
   type SixLevelResearchQuestionLadder } from './question-ladder-contract.js';
 
 export interface ResearchQuestionLadderRunOptions {
-  generateContract?: (prompt: string) => Promise<string>;
-  generate?: (prompt: string) => Promise<string>;
-  invokeAxes?: (questions: SixLevelResearchQuestionLadder['questions']) => Promise<AxisFanoutResult>;
+  signal?: AbortSignal;
+  generateContract?: (prompt: string, signal?: AbortSignal) => Promise<string>;
+  generate?: (prompt: string, signal?: AbortSignal) => Promise<string>;
+  invokeAxes?: (questions: SixLevelResearchQuestionLadder['questions'], signal?: AbortSignal) => Promise<AxisFanoutResult>;
   routeOutput?: typeof routeResearchOutput;
-  invokeOutputSkill?: (skill: string, args: string) => Promise<{ ok: boolean; output: string; error?: string }>;
+  invokeOutputSkill?: (skill: string, args: string, signal?: AbortSignal) => Promise<{ ok: boolean; output: string; error?: string }>;
   /** Test seam for the bounded Bloom 4/5 evidence-gap follow-up. */
   gapFollowup?: typeof runGapFollowup;
+  /** Test seam for deciding whether production gap-follow-up has any runnable calls. */
+  planGapFollowup?: (input: ResearchGapFollowupInput) => ResearchGapFollowupPlan;
   /** Test/config seam for an explicitly installed private Bloom exemplar pack. */
   bloomExemplarPackPath?: string;
 }
@@ -205,9 +209,10 @@ export async function runResearchQuestionLadder(
   const initial: SixLevelResearchQuestionLadder = { request, questions: [], evidence: [] };
   let contractRaw: string;
   try {
-    contractRaw = await (options.generateContract ?? ((prompt) => defaultGenerate(prompt, 'research-working-backwards-contract')))(
-      buildWorkingBackwardsContractPrompt(request),
+    contractRaw = await (options.generateContract ?? ((prompt, signal) => defaultGenerate(prompt, 'research-working-backwards-contract', signal)))(
+      buildWorkingBackwardsContractPrompt(request), options.signal,
     );
+    if (options.signal?.aborted) return { ok: false, ladder: initial, output: '', error: 'research aborted' };
   } catch (error) {
     const reason = errorText(error);
     debug.log('research.lens', 'working-backwards-contract-failed', { reason }, { level: 'warn' });
@@ -247,9 +252,11 @@ export async function runResearchQuestionLadder(
       domain: exemplar.domain, domainSignal: exemplar.domainSignal,
       source: exemplar.source, count: exemplar.count, levels: exemplar.levels,
     });
-    raw = await (options.generate ?? ((prompt) => defaultGenerate(prompt, 'research-question-ladder')))(
+    raw = await (options.generate ?? ((prompt, signal) => defaultGenerate(prompt, 'research-question-ladder', signal)))(
       `${buildResearchQuestionLadderPrompt(request, sectionIds)}\n\n${exemplar.context ? `${exemplar.context}\n\n` : ''}${ladderContext}${retroGroundingContext(request)}`,
+      options.signal,
     );
+    if (options.signal?.aborted) return { ok: false, ladder: initial, outputContract, output: '', error: 'research aborted' };
     debug.log('research.turn', 'bloom-generated', { questionCount: 6 });
   } catch (error) {
     debug.log('research.turn', 'bloom-failed', { reason: errorText(error) }, { level: 'warn' });
@@ -276,7 +283,8 @@ export async function runResearchQuestionLadder(
   const observedAt = new Date().toISOString();
   let fanout: AxisFanoutResult;
   try {
-    fanout = await (options.invokeAxes ?? invokeResearchAxes)(ladder.questions);
+    fanout = await (options.invokeAxes ?? ((questions, signal) => invokeResearchAxes(questions, { signal })))(ladder.questions, options.signal);
+    if (options.signal?.aborted) return { ok: false, ladder, outputContract, output: '', error: 'research aborted' };
     debug.log('research.turn', 'axis-evidence-collected', {
       axisCount: fanout.axes.length,
       evidenceCount: fanout.axes.reduce((count, axis) => count + axis.evidence.length, 0),
@@ -327,47 +335,92 @@ export async function runResearchQuestionLadder(
       stage: 'axis-fanout', reason: 'every output-contract section slot is filled',
     });
   }
-  if (!finalBoundary.complete) {
-    try {
-      const followup = await (options.gapFollowup ?? runGapFollowup)({
-        request,
-        questions: ladder.questions,
-        evidence: ladder.evidence,
-      });
-      const followupLines = [
-        '## Bounded research gap follow-up',
-        `Rounds: ${followup.rounds} · budget: ${followup.timeBudgetMs}ms`,
-        ...followup.classifications.map((item) => `- ${item.questionId} [${item.action}/${item.reason}]${item.skill ? ` · ${item.skill}` : ''}: ${item.gap}`),
-      ];
-      for (const envelope of followup.followupEvidence) {
-        const attached = attachResearchEvidence(ladder, envelope);
-        if (!attached.ok) {
-          evidenceRejected = true;
-          followupLines.push(`- Evidence rejected: ${attached.reason}`);
-          continue;
-        }
-        ladder.evidence = attached.value.evidence;
-        followupLines.push(`- [${envelope.outcome}] ${envelope.sourceKind} · ${envelope.source}: ${envelope.content}`);
+  if (!finalBoundary.complete && !options.signal?.aborted) {
+    const followupStartedAt = performance.now();
+    const followupInput: ResearchGapFollowupInput = {
+      request,
+      questions: ladder.questions,
+      evidence: ladder.evidence,
+      ...(options.signal ? { signal: options.signal } : {}),
+    };
+    let followupPlan: ResearchGapFollowupPlan | undefined;
+    if (options.planGapFollowup || !options.gapFollowup) {
+      try {
+        followupPlan = (options.planGapFollowup ?? planGapFollowup)(followupInput);
+        debug.log('research.turn', 'gap-followup-planned', {
+          rounds: followupPlan.rounds,
+          classifications: followupPlan.classifications.length,
+          calls: followupPlan.followups.length,
+          elapsedMs: Math.round((performance.now() - followupStartedAt) * 10) / 10,
+        });
+      } catch (error) {
+        debug.log('research.turn', 'gap-followup-plan-failed', {
+          reason: errorText(error),
+          elapsedMs: Math.round((performance.now() - followupStartedAt) * 10) / 10,
+        }, { level: 'warn' });
+      }
+    }
+    if (followupPlan && followupPlan.followups.length === 0) {
+      if (followupPlan.classifications.length > 0) {
+        sections.push(
+          '## Bounded research gap follow-up',
+          `Rounds: ${followupPlan.rounds} · budget: ${followupPlan.timeBudgetMs}ms`,
+          ...followupPlan.classifications.map((item) => `- ${item.questionId} [${item.action}/${item.reason}]${item.skill ? ` · ${item.skill}` : ''}: ${item.gap}`),
+        );
       }
-      for (const card of followup.discoveryCards ?? []) followupLines.push(card);
-      if (followup.classifications.length || followup.followupEvidence.length) sections.push(...followupLines);
-      debug.log('research.turn', 'gap-followup-completed', {
-        rounds: followup.rounds,
-        classifications: followup.classifications.length,
-        evidenceCount: followup.followupEvidence.length,
-        evidenceAttached: ladder.evidence.length,
+      debug.log('research.turn', 'gap-followup-skipped', {
+        reason: 'no-runnable-followup-calls',
+        classifications: followupPlan.classifications.length,
+        timeBudgetMs: followupPlan.timeBudgetMs,
+      });
+    } else {
+      const executionStartedAt = performance.now();
+      debug.log('research.turn', 'gap-followup-started', {
+        evidenceCount: ladder.evidence.length,
+        completeEvidenceCount: ladder.evidence.filter((item) => item.outcome === 'complete').length,
       });
-      finalBoundary = planWorkingBackwardsCompletion(outputContract, ladder.questions, ladder.evidence);
-      logCompletionBoundary('gap-followup', finalBoundary);
-      if (finalBoundary.complete) {
-        debug.log('research.turn', 'completion-boundary-stop', {
-          stage: 'gap-followup', reason: 'every output-contract section slot is filled',
+      try {
+        const followup = await (options.gapFollowup ?? runGapFollowup)(followupInput);
+        const followupLines = [
+          '## Bounded research gap follow-up',
+          `Rounds: ${followup.rounds} · budget: ${followup.timeBudgetMs}ms`,
+          ...followup.classifications.map((item) => `- ${item.questionId} [${item.action}/${item.reason}]${item.skill ? ` · ${item.skill}` : ''}: ${item.gap}`),
+        ];
+        for (const envelope of followup.followupEvidence) {
+          const attached = attachResearchEvidence(ladder, envelope);
+          if (!attached.ok) {
+            evidenceRejected = true;
+            followupLines.push(`- Evidence rejected: ${attached.reason}`);
+            continue;
+          }
+          ladder.evidence = attached.value.evidence;
+          followupLines.push(`- [${envelope.outcome}] ${envelope.sourceKind} · ${envelope.source}: ${envelope.content}`);
+        }
+        for (const card of followup.discoveryCards ?? []) followupLines.push(card);
+        if (followup.classifications.length || followup.followupEvidence.length) sections.push(...followupLines);
+        debug.log('research.turn', 'gap-followup-completed', {
+          rounds: followup.rounds,
+          classifications: followup.classifications.length,
+          evidenceCount: followup.followupEvidence.length,
+          evidenceAttached: ladder.evidence.length,
+          elapsedMs: Math.round((performance.now() - executionStartedAt) * 10) / 10,
         });
+        finalBoundary = planWorkingBackwardsCompletion(outputContract, ladder.questions, ladder.evidence);
+        logCompletionBoundary('gap-followup', finalBoundary);
+        if (finalBoundary.complete) {
+          debug.log('research.turn', 'completion-boundary-stop', {
+            stage: 'gap-followup', reason: 'every output-contract section slot is filled',
+          });
+        }
+      } catch (error) {
+        debug.log('research.turn', 'gap-followup-failed', {
+          reason: errorText(error),
+          elapsedMs: Math.round((performance.now() - executionStartedAt) * 10) / 10,
+        }, { level: 'warn' });
       }
-    } catch (error) {
-      debug.log('research.turn', 'gap-followup-failed', { reason: errorText(error) }, { level: 'warn' });
     }
   }
+  if (options.signal?.aborted) return { ok: false, ladder, outputContract, output: '', error: 'research aborted' };
 
   const outputContractSectionIndex = sections.findIndex((section) => section.startsWith('## Output contract\n'));
   if (outputContractSectionIndex < 0) throw new Error('output contract section is missing from research output');
@@ -405,24 +458,43 @@ export async function runResearchQuestionLadder(
     fallbackReason: route?.fallbackReason ?? (route ? undefined : 'router-failed'),
   });
   let routedOutput = '';
-  if (route?.skillCallPlan) {
+  const hasCompleteEvidence = ladder.evidence.some((item) => item.outcome === 'complete');
+  if (route?.skillCallPlan && !hasCompleteEvidence) {
+    debug.log('research.turn', 'output-skill-skipped', {
+      skill: route.skillCallPlan.skill,
+      reason: 'no-complete-evidence',
+      evidenceCount: ladder.evidence.length,
+    });
+  } else if (route?.skillCallPlan) {
+    const outputSkillStartedAt = performance.now();
+    debug.log('research.turn', 'output-skill-started', {
+      skill: route.skillCallPlan.skill,
+      evidenceCount: ladder.evidence.length,
+      completeEvidenceCount: ladder.evidence.filter((item) => item.outcome === 'complete').length,
+    });
     try {
       const formatted = await (options.invokeOutputSkill ?? defaultInvokeOutputSkill)(
-        route.skillCallPlan.skill, route.skillCallPlan.args,
+        route.skillCallPlan.skill, route.skillCallPlan.args, options.signal,
       );
       if (formatted.ok && formatted.output.trim()) {
         routedOutput = `\n\n## Routed ${route.format} output (${route.skillCallPlan.skill})\n\n${formatted.output.trim()}`;
-        debug.log('research.turn', 'output-skill-completed', { skill: route.skillCallPlan.skill, outputChars: formatted.output.length });
+        debug.log('research.turn', 'output-skill-completed', {
+          skill: route.skillCallPlan.skill,
+          outputChars: formatted.output.length,
+          elapsedMs: Math.round((performance.now() - outputSkillStartedAt) * 10) / 10,
+        });
       } else {
         debug.log('research.turn', 'output-skill-fallback', {
           skill: route.skillCallPlan.skill,
           reason: formatted.error ?? 'empty output',
+          elapsedMs: Math.round((performance.now() - outputSkillStartedAt) * 10) / 10,
         }, { level: 'warn' });
       }
     } catch (error) {
       debug.log('research.turn', 'output-skill-fallback', {
         skill: route.skillCallPlan.skill,
         reason: errorText(error),
+        elapsedMs: Math.round((performance.now() - outputSkillStartedAt) * 10) / 10,
       }, { level: 'warn' });
     }
   }
@@ -436,8 +508,8 @@ export async function runResearchQuestionLadder(
   };
 }
 
-async function defaultInvokeOutputSkill(skill: string, args: string) {
-  const result = await invokeResearch(args, { skill });
+async function defaultInvokeOutputSkill(skill: string, args: string, signal?: AbortSignal) {
+  const result = await invokeResearch(args, { skill, signal });
   return { ok: result.ok, output: result.output, ...(result.error ? { error: result.error } : {}) };
 }
 
@@ -451,10 +523,11 @@ export function researchTurnForGrounding(result: ResearchQuestionLadderRunResult
   };
 }
 
-async function defaultGenerate(prompt: string, usageSite: string): Promise<string> {
+async function defaultGenerate(prompt: string, usageSite: string, signal?: AbortSignal): Promise<string> {
   return streamLLM([{ role: 'user', content: prompt }], () => {}, {
     usageSite,
     usagePhase: 'research',
+    ...(signal ? { signal } : {}),
   });
 }
 

package/src/autopilot/pr-manager.ts

@@ -488,12 +488,15 @@ function editExistingPr(run: CmdRunner, input: UpsertPrInput, url: string): Upse
 /** Same-card rework may have a new run branch. Merge the bound PR head into that work,
  * then update the bound branch with a plain fast-forward push. Every identity read is
  * checked again after the inventory snapshot; an ambiguous or moved PR stays refused. */
-function reuseSameCellPr(run: CmdRunner, input: UpsertPrInput, target: NonNullable<GuardVerdict['sameCellReuse']>, recheck: () => GuardVerdict): UpsertPrOutcome {
+function reuseSameCellPr(run: CmdRunner, input: UpsertPrInput, target: NonNullable<GuardVerdict['sameCellReuse']>,
+  initialInventory: GuardVerdict, recheck: () => GuardVerdict): UpsertPrOutcome {
   const cwd = input.worktreePath;
   const pushEvidence = { ref: `refs/heads/${target.branch}`, remoteHead: target.headSha, cardId: target.cardId };
-  const fail = (detail: string): UpsertPrOutcome => {
+  const fail = (detail: string, inventoryEvidence?: { inventoryBefore: unknown; inventoryAfter: unknown; inventoryDiff: unknown }): UpsertPrOutcome => {
     const safeDetail = sanitizeCredentialText(detail).slice(0, 300);
-    debug.log('autopilot.pr-manager', 'same-cell-push-decision', { ...pushEvidence, outcome: 'refused', reason: safeDetail }, { level: 'warn' });
+    debug.log('autopilot.pr-manager', 'same-cell-push-decision', {
+      ...pushEvidence, outcome: 'refused', reason: safeDetail, ...inventoryEvidence,
+    }, { level: 'warn' });
     return { ok: false, reason: 'guard', detail: `same-cell reuse: ${safeDetail}` };
   };
   if (!/^[0-9a-f]{40}$/i.test(target.headSha) || !Number.isSafeInteger(target.number) || target.number < 1
@@ -617,11 +620,39 @@ function reuseSameCellPr(run: CmdRunner, input: UpsertPrInput, target: NonNullab
     || finalPr.url !== pr.url) return fail('PR changed before push');
   const finalRemote = run('git', ['-C', cwd, 'ls-remote', '--exit-code', 'origin', remoteRef]);
   if (!finalRemote.ok || finalRemote.out.trim() !== `${target.headSha}\t${remoteRef}`) return fail('remote head changed before push');
-  const latest = recheck().sameCellReuse;
+  const latestInventory = recheck();
+  const latest = latestInventory.sameCellReuse;
   if (!latest || latest.number !== target.number || latest.branch !== target.branch
     || latest.headSha !== target.headSha || latest.cellId !== target.cellId
-    || latest.cardId !== target.cardId || Boolean(latest.legacyBound) !== Boolean(target.legacyBound))
-    return fail('OPEN PR inventory changed before push');
+    || latest.cardId !== target.cardId || Boolean(latest.legacyBound) !== Boolean(target.legacyBound)) {
+    const snapshot = (verdict: GuardVerdict) => ({
+      conflicts: [...(verdict.sameCellConflicts ?? [])].map((row) => ({ pr: row.pr,
+        branch: row.branch === null ? null : sanitizeCredentialText(row.branch).slice(0, 180),
+        headSha: /^[0-9a-f]{40}$/i.test(row.headSha) ? row.headSha.toLowerCase() : 'invalid',
+        cardId: row.cardId === null ? null : sanitizeCredentialText(row.cardId).slice(0, 128), slice: row.slice ?? null,
+      })).sort((a, b) => a.pr - b.pr),
+      reusable: verdict.sameCellReuse ? { number: verdict.sameCellReuse.number,
+        branch: sanitizeCredentialText(verdict.sameCellReuse.branch).slice(0, 180),
+        headSha: /^[0-9a-f]{40}$/i.test(verdict.sameCellReuse.headSha) ? verdict.sameCellReuse.headSha.toLowerCase() : 'invalid',
+        cellId: sanitizeCredentialText(verdict.sameCellReuse.cellId).slice(0, 128),
+        cardId: sanitizeCredentialText(verdict.sameCellReuse.cardId).slice(0, 128),
+        legacyBound: Boolean(verdict.sameCellReuse.legacyBound) } : null,
+    });
+    const before = snapshot(initialInventory);
+    const after = snapshot(latestInventory);
+    const beforeByPr = new Map(before.conflicts.map((row) => [row.pr, row]));
+    const afterByPr = new Map(after.conflicts.map((row) => [row.pr, row]));
+    const inventoryDiff = {
+      added: after.conflicts.filter((row) => !beforeByPr.has(row.pr)),
+      removed: before.conflicts.filter((row) => !afterByPr.has(row.pr)),
+      changed: after.conflicts.flatMap((row) => {
+        const old = beforeByPr.get(row.pr);
+        return old && JSON.stringify(old) !== JSON.stringify(row) ? [{ pr: row.pr, before: old, after: row }] : [];
+      }),
+      reusableChanged: JSON.stringify(before.reusable) !== JSON.stringify(after.reusable),
+    };
+    return fail('OPEN PR inventory changed before push', { inventoryBefore: before, inventoryAfter: after, inventoryDiff });
+  }
   // The result descends from the exact fetched head. The explicit lease binds the
   // remote update to that head, so a delete or move between reads cannot be revived.
   // The Pod hook receives these values for this one push only. It checks the Pod card,
@@ -857,7 +888,7 @@ export function makePrManager(
       };
       const earlyVerdict = earlyGuard ? siblingSliceAllowed(earlyGuard) : undefined;
       const freshGuard = () => checkCreate!({ cwd, base: comparisonBase, seat: input.seat ?? seatFromEnvironment(), body: input.body }, run);
-      if (earlyVerdict?.sameCellReuse) return reuseSameCellPr(run, input, earlyVerdict.sameCellReuse, freshGuard);
+      if (earlyVerdict?.sameCellReuse) return reuseSameCellPr(run, input, earlyVerdict.sameCellReuse, earlyVerdict, freshGuard);
       if (earlyVerdict?.status === 'refuse') return { ok: false, reason: 'guard', detail: sameCellGuardDetail(earlyVerdict, candidateCardId) };
       const earlyInventoryUnknown = !!candidateCardId && (!checkCreate || (earlyVerdict?.status === 'unknown'
         && earlyVerdict.reasons.length === 1 && earlyVerdict.reasons[0]?.code === 'inventory-unknown'));

package/src/autopilot/pty-control-loop.ts

@@ -73,6 +73,8 @@ export interface ControlObservation {
   readonly subjectPtyId?: string;
   /** 이 판단이 속한 harness run. harness 밖 감독자는 생략한다. */
   readonly runId?: string;
+  /** Recent same-run chart cells pulled immediately before this supervision judgment. */
+  readonly runChart?: readonly { device: string; stopKey: string; observed: string; verdict: string; action: string; handoff: string | null; why: string }[];
   /** Last structured action that could not be performed; brain sees the reason next round. */
   readonly unavailableAction?: string;
 }

package/src/capture/tui-self-report.ts

@@ -31,7 +31,7 @@
 import type { SelfReportFrame } from './self-report-frame.js';
 import { resolveInstanceName } from '../instance-identity.js';
 import { getHarnessRunId } from '../harness/harness-space.js';
-import { injectKey, setFrameObserver, splitKeys, termSize as tuiTermSize, type Key } from '../tui.js';
+import { injectKey, routeExternalTuiKey, setFrameObserver, splitKeys, termSize as tuiTermSize, type Key } from '../tui.js';
 import { canTransitionAccessMode, type PtyAccessMode, type PtyTransitionPolicy } from '../pty-shell/pty-ref.js';
 import { resolveWriteDecision, type PtyWriteActor } from '../pty-shell/pty-write-arbiter.js';
 import { registerPtyControlTarget, startPtyControlPoller, type PtyControlTarget } from '../pty-shell/registry.js';
@@ -43,7 +43,7 @@ import {
   updatePtyManifestBuild,
 } from '../pty-shell/pty-manifest.js';
 import { writeHarnessScreen } from '../harness/harness-screen.js';
-import { runningBuild, startTuiRunStateRecorder } from '../pty-shell/tui-run-state.js';
+import { runningBuild, seatCompareDepsFromEnv, startTuiRunStateRecorder } from '../pty-shell/tui-run-state.js';
 import type { TuiSelfRestartDeps } from '../pty-shell/tui-run-state.js';
 import { lastHumanKeyAt } from '../tui.js';
 import { debug } from '../debug/log.js';
@@ -176,9 +176,19 @@ export function createTuiControlTarget(surfaceId: string, deps: TuiControlTarget
     stopTurn: (force: true) => force && stopTurn(),
     canWrite: (actor: PtyWriteActor = 'human') => resolveWriteDecision(accessMode, actor).allow,
     write(chars: string): void {
+      const keys = parse(chars);
+      const remaining: Key[] = [];
+      let routedToModal = false;
+      for (const key of keys) {
+        if (routeExternalTuiKey(key)) routedToModal = true;
+        else remaining.push(key);
+      }
+      if (routedToModal) {
+        if (remaining.length > 0 && !remaining.every(inject)) throw new Error('tui-input-undelivered');
+        return;
+      }
       // ⭐ 이 write 는 «바깥»(제어 IPC)에서만 온다 — 키보드와 섞이기 전에 라우터가 먼저 가른다.
       if (routeExternal(chars) !== 'passthrough') return;
-      const keys = parse(chars);
       if (!keys.length || !keys.every(inject)) throw new Error('tui-input-undelivered');
     },
     resize: () => { throw new Error('tui-resize-unsupported'); },
@@ -444,9 +454,15 @@ export function startTuiSelfReport(deps: TuiSelfReportDeps = {}): () => void {
   const startRunState = deps.startRunState
     ?? (deps.registerRow
       ? () => () => {}
-      : (id: string) => startTuiRunStateRecorder(id, {
-        selfRestart: deps.selfRestart === null ? undefined : (deps.selfRestart ?? productionTuiSelfRestart(readActiveSessionId)),
-      }));
+      : (id: string) => {
+        // TUI-SEAT-PILOT-2H-COMPARE — 판정 신호 한 줄은 env(ELANOUS_SEAT_COMPARE_RUN_ID)가 있을 때만 켜진다.
+        // env 가 없으면 빈 spread → 종전 기록 경로 그대로(보존 계약).
+        const seatCompare = seatCompareDepsFromEnv();
+        return startTuiRunStateRecorder(id, {
+          selfRestart: deps.selfRestart === null ? undefined : (deps.selfRestart ?? productionTuiSelfRestart(readActiveSessionId)),
+          ...(seatCompare ? { seatCompare } : {}),
+        });
+      });
   try { stopRunState = startRunState(surfaceId); } catch { /* fail-soft — recorder logs its own write failures */ }
   // 미룬 우편 깨움(TUI-H4) — 이 TUI 가 턴 중·5분 창이라 못 받은 깨움을 idle 로 «기록»된 뒤 한 번 다시 두드린다.
   // 실 행 등록과 짝일 때만(시험이 행 등록을 갈아 끼우면 0) · 무거운 모듈은 늦게 불러온다.

package/src/chat/index.ts

@@ -523,7 +523,7 @@ export const SLASH_COMMANDS: SlashCommand[] = ([
   { name: 'tablet',    aliases: [],        description: 'Tablet mode — collapse layout to log + input', subcommands: ['on', 'off', 'toggle', 'auto', 'status', 'browser-preview', 'bp'] },
   { name: 'surface',   aliases: [], description: 'Surface control — open a pane/catalog launcher (/surface catalog · browser · preview · log · scratch · obsidian · skill) OR set a preferred LLM surface (/surface coding-agent · research-agent · control-agent · ops-agent · … · clear · status)', subcommands: ['catalog', 'browser', 'preview', 'browser-preview', 'log', 'scratch', 'obsidian', 'skill', 'skill-file', 'agents', 'coding-agent', 'coding-chat', 'research-agent', 'research-chat', 'control-agent', 'ops-ui-agent', 'ops-fleet-agent', 'ops-agent', 'clear', 'status'] },
   { name: 'term',  aliases: ['terminal'], description: 'Interactive terminal modal — new / list / attach / detach / switch / kill / snapshot / resume', subcommands: ['new', 'list', 'attach', 'detach', 'switch', 'kill', 'snapshot', 'resume'] },
-  { name: 'side', aliases: [], description: '사이드탭 기초 — 목록 · 셸 생성 · 번호로 PTY 보기/attach', subcommands: ['list', 'new', 'attach'] },
+  { name: 'side', aliases: [], description: '사이드탭 — 목록 · 셸/Codex/Claude 탭 · attach · 미션 결과 파일', subcommands: ['list', 'new', 'attach', 'open', 'mission'] },
   { name: 'claude', aliases: [], description: 'Spawn claude-code CLI in a terminal modal' },
   { name: 'codex',  aliases: [], description: 'Spawn codex CLI in a terminal modal' },
   { name: 'gemini', aliases: [], description: 'Spawn gemini CLI in a terminal modal' },

package/src/cli/fixtures/review-loop-breaker/opwaiver-32646-1694ccc7a170.json

@@ -0,0 +1 @@
+{"pr": 32646, "headCommit": "1694ccc7a1704d451508008068265aeee45cde24", "reviewed": true, "verdict": "pass", "mustFix": [], "shouldFix": [], "reviewer": "OP-waiver", "model": "none (OP adjudication)", "waiver": {"by": "OP", "approvedBy": "OP 2026-10-29 10:5x «무한 리뷰 중단·웨이버» · r9 초과", "adjudicates": "pr-32646-2dcdb05e4306.json (gpt-6.1-terra r9 fail · mustFix 1: evidencePathPattern 이 @·공백·한글 경로를 잘라 출력)", "headTestRun": {"host": "node-b", "sha": "1694ccc7a1704d451508008068265aeee45cde24", "command": "bun test src/dashboard/seat-boot-preamble.test.ts", "pass": 20, "fail": 0, "expects": 97, "exit": 0, "rawOutput": "~/elanous-hq/seat-state/OPC/evidence/32646-1694ccc7a170-node-b.txt"}, "diffChecked": "2dcdb05e→1694ccc7: seat-boot-preamble.ts +7/-5 · test +16/-8 — 경로 끝 경계 판정 · 공백 경로는 확장자 꼬리 확인 · 시험이 proof@2.md·검증@2.md 전체 보존 ⊕ 공백·# 경로는 잘린 포인터 미출력 확인 — OP 직접 대조", "dispositions": [{"finding": 1, "disposition": "fixed-in-this-head", "evidence": "seat-boot-preamble.test.ts 추가 단언(asciiSpecialPath·unicodeSpecialPath 포함 · spaced/delimiter 잘린 출력 없음) · node-b 20/0"}]}, "summary": "OP waiver — r9 지적 수리 확인."}

package/src/cli/fixtures/review-loop-breaker/opwaiver-32853-bd8f73595c83.json

@@ -0,0 +1 @@
+{"pr": 32853, "headCommit": "bd8f73595c83e3bee478faa3b765d4082c4c5aa6", "reviewed": true, "verdict": "pass", "mustFix": [], "shouldFix": [], "reviewer": "OP-waiver", "model": "none (OP adjudication)", "waiver": {"by": "OP", "approvedBy": "OP 2026-10-29 10:5x «b OP 판정 예외 승인 · 무한 리뷰 중단·웨이버»", "adjudicates": "pr-32853-1c0f33c905e0.json (gpt-6.1-terra r11 fail mustFix 5) — 이후 수정 1c0f33c9→f3d60e98→bd8f7359 은 driver.test.ts 만(import 1 · 동적 경로 · 구독 fake · fail-closed 회귀 추가) · OP 가 diff 직접 대조", "headTestRun": {"host": "node-b", "sha": "bd8f73595c83e3bee478faa3b765d4082c4c5aa6", "command": "bun test src/agent-mission/driver.test.ts", "pass": 140, "fail": 0, "expects": 586, "exit": 0, "reportedBy": "OPC #51223"}, "dispositions": [{"finding": 1, "disposition": "already-fixed", "evidence": "driver.test.ts «$name accepts only its observed final-input suggestion after marker, answer, and stable idle»"}, {"finding": 2, "disposition": "already-fixed", "evidence": "driver.test.ts «completion record preserves a final answer longer than 100 lines…» · «long final answer is preserved while output logging remains bounded» · driver.ts slice(-3500) 은 LLM 입력·reviewSummary 뿐"}, {"finding": 3, "disposition": "out-of-scope-followup", "evidence": "실 CLI 자동 시험 → AGENT-MISSION-BENCH-READY-FU · 실측 run-ab227633 양팔 통과"}, {"finding": 4, "disposition": "already-fixed", "evidence": "driver.test.ts «empty input prompt never completes even after a stable interval»"}, {"finding": 5, "disposition": "fixed-in-this-head", "evidence": "readdirSync import 추가(f3d60e98) · fail-closed 쓰기 실패 회귀 추가(bd8f7359)"}]}, "summary": "OP waiver(재작성) — 정확 head 실측 140/0 근거."}

package/src/cli/fixtures/review-loop-breaker/opwaiver-33097-1cc85900413c.json

@@ -0,0 +1 @@
+{"pr": 33097, "headCommit": "1cc85900413cfcb2fdc427d4ca85c51fc0eb8358", "reviewed": true, "verdict": "pass", "mustFix": [], "shouldFix": [], "reviewer": "OP-waiver", "model": "none (OP adjudication)", "waiver": {"by": "OP", "approvedBy": "OP 2026-10-29 10:5x «무한 리뷰 중단·웨이버» 일반 승인 · r7 상한 도달", "adjudicates": "pr-33097-1cc85900413c.json (gpt-6.1-terra r7 fail mustFix 2 — 리뷰 입력이 PR 본문·실행 로그를 못 봐 재개)", "headTestRun": {"host": "node-b", "sha": "1cc85900413cfcb2fdc427d4ca85c51fc0eb8358", "command": "bun test src/cli/pr-land-heavy-remote.test.ts src/task-orchestrator/surfaces/pod-warm-switch.test.ts src/task-orchestrator/surfaces/warm-reuse-evidence.test.ts src/cli/pr-land-pod-shadow-job.test.ts src/task-orchestrator/surfaces/pod-warm-worker-provider.test.ts", "pass": 130, "fail": 0, "expects": 536, "exit": 0, "rawOutput": "~/elanous-hq/seat-state/TC/evidence/33097-1cc85900-node-b.txt"}, "dispositions": [{"finding": 1, "disposition": "evidence-provided", "evidence": "원문 파일에 명령·EXIT_CODE=0·130 pass · (pass) «heavy land gate decides warm reuse once per landing — no evidence=off · 3 matching=live · 1 mismatch=recent-mismatch» · «persisted writer rows flow through the shared reader…» — OP 가 원문·head 시험 존재 대조"}, {"finding": 2, "disposition": "already-fixed", "evidence": "warm-reuse-evidence.test.ts:56 «expired or corrupt rows cannot revive older warm matches» · :69 «corruption after recent passing rows invalidates the entire evidence window» · :83 «a corrupt recent mismatch row cannot restore the preceding three green samples» — 원문에서 corrupt (pass) 3"}]}, "summary": "OP waiver — r7 stale reopen · 정확 head node-b 원문 130/0 근거."}

package/src/cli/fixtures/review-loop-breaker/opwaiver-33106-972f31e4a3ca.json

@@ -0,0 +1 @@
+{"pr":33106,"headCommit":"972f31e4a3cab0dc62873f915f0c71a2882b6044","reviewed":true,"verdict":"pass","mustFix":[],"shouldFix":[],"reviewer":"OP-waiver","model":"none (OP adjudication)","waiver":{"by":"OP","approvedBy":"OP 2026-10-29 12:2x «0.2.41 재개 최적 경로» · 10:5x 웨이버 승인","adjudicates":"gpt-6.1-terra r6 on cc59ca7b (fail · mustFix 2 · shouldFix 1)","headTestRun":{"host":"node-b","sha":"972f31e4a3cab0dc62873f915f0c71a2882b6044","scope":"resume-release.test.ts 전체(apps/pwa/node_modules 부재 전·후 모두 · afterAll rmdirSync 경로 실제 통과)","pass":40,"fail":0,"reportedBy":"MK #51574"},"parentTestRun":{"host":"node-b","sha":"cc59ca7b1e264f4f23504d5523a32efbecf300b9","scope":"resume-release·release-branch·gate-autoheal·gate-shard-receipts 4 suite","pass":287,"fail":0,"expects":1677,"reportedBy":"MK #51494"},"diffChecked":"cc59ca7b→32646bb4: resume-release.ts +13/-5 · resume-release.test.ts +45/-? — 저장된 cutoff baseline 과 receipt base40 대조 · 부재/불일치 시 full(fail-closed) · OP 직접 대조 · 32646bb4→30001d64→972f31e4: 시험 파일만 — Pod checkout 의 apps/pwa/node_modules 부재(기존 28·새 3 실패 공통 원인 · 급행 게이트 123723) 를 빈 디렉터리 fixture 로 · 정리는 rmdirSync","mergeGate":"급행 열차 gate 가 착지 전 전체 반경 재측정","dispositions":[{"finding":1,"disposition":"fixed-in-this-head","evidence":"expectedBase40 = cutoff 노드 baseline.sha · 불일치/부재 → fail(full gate)"},{"finding":2,"disposition":"out-of-scope-followup","evidence":"Goodhart(합성 gate 출력 대신 실 graph 상태) → 0.2.42 GATE-MOCK-EVENT-SCENARIOS"},{"finding":3,"disposition":"fixed-in-this-head","evidence":"SIGTERM 시험이 완료 receipt 확인 뒤 중단"}],"expressGateFinding":"batch-minim5-express-123723 introduced 3 = 환경 원인(verifyResumePreflight 기본 commandRoot 의 apps/pwa/node_modules 부재) — 같은 파일 preexisting 28 과 동일 원인"},"summary":"OP waiver — 0.2.41 필수 GATE-RESUME-STALLED-ONLY."}

package/src/cli/fixtures/review-loop-breaker/opwaiver-33113-3ab51f05968f.json

@@ -0,0 +1 @@
+{"pr": 33113, "headCommit": "3ab51f05968f926047a670d11c2afcca698df028", "reviewed": true, "verdict": "pass", "mustFix": [], "shouldFix": [], "reviewer": "OP-waiver", "model": "none (OP adjudication)", "waiver": {"by": "OP", "approvedBy": "OP 2026-10-29 10:5x «무한 리뷰 중단·웨이버» · identity 지적 r1·r2 반복", "adjudicates": "ux-coord-entrance-fold-terra-r2-33113-f5db480b56d5.json (fail · mustFix 2)", "headTestRun": {"host": "node-b", "sha": "3ab51f05968f926047a670d11c2afcca698df028", "tests": {"task-hand": "1/0", "coordinator-shadow": "9/0", "tick": "5/0"}, "exit": 0, "rawOutput": "~/elanous-hq/seat-state/UX/evidence/33113-3ab51f05968f-node-b.txt", "fileHashesMatchHead": 7}, "dispositions": [{"finding": 1, "disposition": "reviewer-misread-contract", "evidence": "칸 COORD-ENTRANCE-FOLD «조율자 그림자가 ④ builder 로 같은 칸·같은 지시서 3일 연속» — ④ builder 는 지시서를 «만드는» caller 단계(OP #51545·#51569 정의 · PR 본문 기록). recipient(--seat)는 받는 자리로 증거 행에만 남긴다. 리뷰어는 builder 4 를 «받는 쪽»으로 읽었다."}, {"finding": 2, "disposition": "fixed-in-this-head", "evidence": "coordinator-shadow.ts:120 그룹 키 = agent·cell·directive(recipient 제외) · :116-124 recipient 는 그룹 속성으로만 · recipient 없는 옛 행도 같은 그룹 — OP 가 head 코드 직접 대조"}]}, "summary": "OP waiver — 계약 오독 1 · 수리 확인 1."}

package/src/cli/fixtures/review-loop-breaker/pr-33097-1cc85900413c.json

@@ -0,0 +1 @@
+{"pr":"33097","model":"gpt-6.1-terra","verdict":"fail","mustFix":["통합 head의 필수 수용 증거가 미완료다: 골 문서가 요구한 `bun test src/cli/pr-land-heavy-remote.test.ts src/task-orchestrator/surfaces/pod-warm-switch.test.ts`와 생산 호출 경로를 지나는 mock 무대의 명령·종료 코드·출력 원문이 없다. 장애 환경을 피한 판단은 타당하지만, node-b에서 실행해 일치 3건의 live, mismatch의 recent-mismatch, 착지당 판정 1줄과 기존 시험 통과를 확인해야 한다.","`readWarmReuseEvidence()`는 손상되거나 유효하지 않은 행을 조용히 제거한다. 최근 mismatch 행이 손상되면 그 앞의 성공 행들이 연속 표본으로 복원될 수 있어 fail-closed가 아니다. 손상된 창은 미측정으로 처리하고 회귀 시험을 추가하라."],"shouldFix":[],"requirements":["DELTA-FINDING 1: OPEN — 이번 delta에 필수 실행 결과와 출력 증거가 없다.","DELTA-FINDING 2: OPEN — 회귀 시험은 추가됐지만 구현 변경이나 시험 통과 증거는 없다.","관찰: 추가된 시험은 임시 디렉터리의 파일만 쓰고 정리한다. 실제 fail-closed 동작은 시험 실행 결과로 확인해야 한다."],"reviewed":true,"diffBudget":{"truncated":false,"shownChars":1645,"totalChars":1645,"omittedFiles":0},"referencedFilesOpened":false,"referencedFilesRead":0,"reviewInputChars":2488,"reviewDurationMs":6440,"basedOn":"10cd6c80fe8d036f7530fd373fcfffc1aa9be044","basedOnReview":"warm-land-evidence-terra-0241-r5-33097-10cd6c80fe8d.json","deltaFindings":[{"finding":"통합 head의 필수 수용 증거가 미완료다: 골 문서가 요구한 `bun test src/cli/pr-land-heavy-remote.test.ts src/task-orchestrator/surfaces/pod-warm-switch.test.ts`와 생산 호출 경로를 지나는 mock 무대의 명령·종료 코드·출력 원문이 없다. 장애 환경을 피한 판단은 타당하지만, node-b에서 실행해 일치 3건의 live, mismatch의 recent-mismatch, 착지당 판정 1줄과 기존 시험 통과를 확인해야 한다.","status":"open"},{"finding":"`readWarmReuseEvidence()`는 손상되거나 유효하지 않은 행을 조용히 제거한다. 최근 mismatch 행이 손상되면 그 앞의 성공 행들이 연속 표본으로 복원될 수 있어 fail-closed가 아니다. 손상된 창은 미측정으로 처리하고 회귀 시험을 추가하라.","status":"open"}],"excludedNewFindings":2,"deltaInputChars":2488,"deltaDurationMs":6440,"deltaEfficiency":{"inputRatio":0.8632893823733518,"timeRatio":0.9649385675756668,"maxRatio":0.3333333333333333,"withinTarget":false},"deltaOutcome":"reviewed","reviewRoute":"api","fallbackAttempted":false,"requestedModel":"gpt-6.1-terra","fallbackNotAttemptedReason":"primary-review-succeeded","skippedAcpFallbackBackends":[{"backend":"gemini","reason":"Confirmed live 2026-08-19 (first observed 2026-08-12): gemini-cli is rejected with IneligibleTierError (free-tier UNSUPPORTED_CLIENT). Antigravity CLI (agy) replaces it but does NOT speak ACP, so this ACP entry has no successor yet — use the PTY backend (--backend gemini) instead."}],"respondedProvider":"openai-codex","respondedModel":"gpt-6.1-terra","reviewerSource":"cli","headCommit":"1cc85900413cfcb2fdc427d4ca85c51fc0eb8358","currentHeadCommit":"1cc85900413cfcb2fdc427d4ca85c51fc0eb8358","stale":false,"refetched":false,"diffSource":"sha-pinned","diffBaseCommit":"10cd6c80fe8d036f7530fd373fcfffc1aa9be044","localHeadComparison":"unavailable","localHead":"59018b9bd0d74574b50bdf4880373756335758a8","prHead":"1cc85900413cfcb2fdc427d4ca85c51fc0eb8358","localHeadComparisonReason":"detached-head","reviewerContextLoaded":0,"reviewerContextFailed":0,"reviewerContextFailures":[],"reviewerContextSource":"pr-diff-files","reviewerContextSourcePaths":[],"decompositionContextStatus":"absent","findingPathHeadChecks":[],"pathWarnings":[],"evidence":"saved-review-delta","prDiffFiles":1}

package/src/cli/pr-land-heavy-remote.ts

@@ -6,6 +6,8 @@ import { effectiveInstanceRoot } from '../instance/resolve.js';
 import { getUserConfig } from '../user-config.js';
 import { debug, redactSecretText } from '../debug/log.js';
 import { gateRemoteHostCap, realGateRemoteRunner, resolveGateRemoteSettings, runOnRemote, type GateRemoteRunner, type RemoteRunOutcome } from '../self-implement/gate-remote.js';
+import { decideWarmReuse, type WarmReuseEvidenceByConsumer, type WarmReuseSwitchConfig } from '../task-orchestrator/surfaces/pod-warm-switch.js';
+import { buildLandWarmEvidence, type LandShadowSampleWindow } from './pr-land-pod-shadow-job.js';
 
 const SHA = /^[0-9a-f]{40}$/i;
 // This is the stable fallback order when settings.host is unavailable; an explicitly configured allowlisted host
@@ -66,6 +68,10 @@ export interface HeavyRemoteDeps {
   /** Test seam — same cadence as self gate (`dispatchHeavyCheck`): sleep at most 15s while a host stays busy. */
   sleep?: (ms: number) => void;
   observe?: (event: string, data: Record<string, unknown>) => void;
+  /** WARM-REUSE-LAND-CONSUMER-WIRE — land warm 재사용 config seam(기본은 `getUserConfig().pod?.warmReuse`). */
+  warmReuseConfig?: () => WarmReuseSwitchConfig | undefined;
+  /** WARM-REUSE-LAND-CONSUMER-WIRE — land shadow Job 결과 창 seam(기본 없음 = 증거 없음 = off). */
+  landShadowSamples?: () => LandShadowSampleWindow | undefined;
 }
 
 function evidencePath(input: HeavyRemoteInput): string {
@@ -209,6 +215,32 @@ export function ensurePrLandHeavyRemote(input: HeavyRemoteInput, deps: HeavyRemo
   if (!Number.isSafeInteger(input.pr) || input.pr <= 0 || !SHA.test(input.head) || !SHA.test(input.main)) {
     return { ok: false, source: 'unmeasured', reason: 'invalid-pr-head-or-main', elapsedMs: 0, evidencePath: path };
   }
+  // WARM-REUSE-LAND-CONSUMER-WIRE — 무거운 착지 판정마다 land 소비자 판정을 1줄 남긴다(pod.warm/switch-decision).
+  // 판정은 그림자 전용이다: live 여부와 무관하게 원격 게이트는 기존(fresh) 경로를 그대로 간다.
+  // 증거는 land shadow Job 결과 창(주입 seam · 기본 없음=증거 없음=off)에서 만들고, 판정 실패는 착지를 깨지 않는다.
+  let warmConfig: WarmReuseSwitchConfig | undefined;
+  let warmEvidence: WarmReuseEvidenceByConsumer | undefined;
+  let warmStage: 'config' | 'evidence' = 'config';
+  try {
+    warmConfig = deps.warmReuseConfig?.() ?? getUserConfig().pod?.warmReuse;
+  } catch (error) {
+    debug.log('pod.warm', 'land-shadow-unavailable', {
+      pr: input.pr, stage: warmStage,
+      reason: `warm-${warmStage}-read-failed`,
+      detail: error instanceof Error ? error.message : String(error), execution: 'fresh-gate',
+    }, { level: 'warn' });
+  }
+  try {
+    warmStage = 'evidence';
+    warmEvidence = buildLandWarmEvidence(deps.landShadowSamples?.(), { head: input.head, main: input.main });
+  } catch (error) {
+    debug.log('pod.warm', 'land-shadow-unavailable', {
+      pr: input.pr, stage: warmStage, reason: 'warm-evidence-read-failed',
+      detail: error instanceof Error ? error.message : String(error), execution: 'fresh-gate',
+    }, { level: 'warn' });
+  }
+  // Read failures leave config/evidence absent, but every valid landing still emits one fail-closed decision.
+  decideWarmReuse('land', warmConfig, warmEvidence);
   // A cached success is still a landing claim; accept only a measured host in the landing allowlist.
   const settings = deps.settings ?? resolveGateRemoteSettings(getUserConfig().gateRemote);
   if (!allowedHost(settings.host)) {