elanous 0.2.43 → 0.2.44

package/src/task-orchestrator/surfaces/pod-warm-worker-runtime.ts

@@ -39,6 +39,18 @@ function readInstalledLockHash(repoDir: string): string | null {
   } catch { return null; }
 }
 
+function pathIsAbsent(path: string): boolean {
+  try { lstatSync(path); return false; }
+  catch (error) { return (error as NodeJS.ErrnoException).code === 'ENOENT'; }
+}
+
+function isRealDirectory(path: string): boolean {
+  try {
+    const stat = lstatSync(path);
+    return stat.isDirectory() && !stat.isSymbolicLink();
+  } catch { return false; }
+}
+
 /** Record the image's installed dependency lock once source verification ties it to the worker key. */
 function rememberInstalledLockHash(repoDir: string, lockHash: string): void {
   try {
@@ -261,7 +273,35 @@ export class WarmWorkerRuntime {
       }, { level: 'warn' }); } catch { /* observation cannot revive an unsafe worker */ }
       return { reply: null, quarantineWorker: true, reason: 'cleanup-failed', artifactsDir };
     }
-    mkdirSync(this.options.scratchDir, { recursive: true });
+    try {
+      (this.options.removeScratch ?? ((path: string) => rmSync(path, { recursive: true, force: true })))(this.options.scratchDir);
+      // A successful rm call is not proof of cleanup: a stub can be a no-op, and an
+      // attacker can replace the path with a symlink before the next mkdir.
+      if (!pathIsAbsent(this.options.scratchDir)) throw new Error('scratch path remains after removal');
+    } catch {
+      this.quarantined = true;
+      try { debug.log('pod.reuse', 'clean-failed', {
+        runId: command.runId, attemptId: command.attemptId, workerId: command.workerId,
+        // The worker outcome stays cleanup-failed for callers; this event reason identifies the scratch phase.
+        // Include only a path that lstat can still observe; a failed rm may also have removed it.
+        reason: 'scratch-removal-failed',
+        remainingPaths: pathIsAbsent(this.options.scratchDir) ? [] : [this.options.scratchDir],
+      }, { level: 'warn' }); } catch { /* observation cannot revive an unsafe worker */ }
+      return { reply: null, quarantineWorker: true, reason: 'cleanup-failed', artifactsDir };
+    }
+    try {
+      // Non-recursive mkdir refuses a path that was recreated (for example as a
+      // symlink) after the absence check. Validate the resulting path type as well.
+      mkdirSync(this.options.scratchDir);
+      if (!isRealDirectory(this.options.scratchDir)) throw new Error('scratch path is not a real directory');
+    } catch {
+      this.quarantined = true;
+      try { debug.log('pod.reuse', 'clean-failed', {
+        runId: command.runId, attemptId: command.attemptId, workerId: command.workerId,
+        reason: 'scratch-directory-unsafe', remainingPaths: [this.options.scratchDir],
+      }, { level: 'warn' }); } catch { /* observation cannot revive an unsafe worker */ }
+      return { reply: null, quarantineWorker: true, reason: 'cleanup-failed', artifactsDir };
+    }
     mkdirSync(artifactsRoot, { recursive: true });
     mkdirSync(artifactsDir, { recursive: true, mode: 0o700 });
     const startedAt = now();
@@ -304,8 +344,10 @@ export class WarmWorkerRuntime {
         reason = 'execution-error';
         break;
       } finally {
-        try { (this.options.removeScratch ?? ((path: string) => rmSync(path, { recursive: true } )))(sandbox); }
-        catch { this.quarantined = true; reason = 'cleanup-failed'; }
+        try {
+          (this.options.removeScratch ?? ((path: string) => rmSync(path, { recursive: true, force: true })))(sandbox);
+          if (!pathIsAbsent(sandbox)) throw new Error('request sandbox remains after removal');
+        } catch { this.quarantined = true; reason = 'cleanup-failed'; }
       }
       if (now() >= command.deadlineAt) reason = 'execution-error';
       if (reason !== 'complete') break;

package/src/task-orchestrator/surfaces/self-implement-pod.ts

@@ -22,7 +22,7 @@ import { isVerifiedOpenGatewayModel } from '../../llm/opengateway-models.js';
 import { configuredOpenGatewayKeys } from '../../llm/provider-credentials.js';
 import { collectPodLedgers, createPodLedgerFollower } from './pod-ledger-collect.js';
 import { captureFailedPodJobs } from './pod-ledger-events.js';
-import { collectPodArtifacts, podNoResultDiagnostic, POD_SALVAGE_BUNDLE_FILE_LIMIT, POD_SALVAGE_BUNDLE_TOTAL_LIMIT } from './pod-artifact-return.js';
+import { collectPodArtifacts, podNoResultChildError, podNoResultDiagnostic, POD_SALVAGE_BUNDLE_FILE_LIMIT, POD_SALVAGE_BUNDLE_TOTAL_LIMIT } from './pod-artifact-return.js';
 import { readPodRepairArtifact } from './pod-repair-commit.js';
 import { isValidAccountName, resolveCodexAccount } from '../../oauth/codex-account.js';
 import { createCodexAccessMirror } from '../../oauth/codex-access-mirror.js';
@@ -32,6 +32,7 @@ import { finishPodFragment, writePodFragment } from '../../harness/self-send-tar
 import { mintRunId, normalizeRunId } from '../../harness/harness-space.js';
 import { TA_LAND_ENV } from '../../task-agent/task-hand.js';
 import { appendRunLedgerEntry, loadRunLedger, runLedgerDir, runLedgerPath, type RunLedgerEntry } from '../../self-implement/run-ledger.js';
+import { childFirstLlmFact, childFirstLlmReportFact } from '../../llm/child-first-llm-observation.js';
 import { recordRunStop } from '../../self-implement/run-stop.js';
 import { gateFailureReceipt, gateFailureReceiptLines } from '../../self-implement/gate-failure-signature.js';
 import { runSuffixKey } from '../../self-implement/salvage-branches.js';
@@ -148,6 +149,19 @@ import { findGitDir } from '../../git-fs/locate.js';
 import { hostLeaseBaseDir } from '../../pod-lease/host-lease.js';
 import { acquireLockAsync } from '../../storage/file-lock.js';
 import { debug, redactSecretText } from '../../debug/log.js';
+
+function readPodChildFirstLlmFact(runId: string, ledgerDir: string, report = false) {
+  try {
+    const entries = loadRunLedger(runId, ledgerDir) ?? [];
+    return report ? childFirstLlmReportFact(entries) : childFirstLlmFact(entries);
+  } catch (error) {
+    debug.log('self-implement.pod', 'child-first-llm-ledger-unavailable', {
+      childRunId: runId,
+      reason: error instanceof Error ? error.message : String(error),
+    }, { level: 'warn' });
+    return undefined;
+  }
+}
 import { enterJourneyNode, exitJourneyNode, passJourneyNode } from '../../self-dev/graph-journey-nodes.js';
 import { routeJourneyEdge } from '../../self-dev/graph-journey-route.js';
 import { selectLiveDetail } from '../../live/detail-switch.js';
@@ -1003,9 +1017,11 @@ function podCodexAccountScript(accounts: readonly string[]): string {
   return [
     'mkdir -p "$HOME/.elanous" "$HOME/.codex"',
     // Names travel as JSON data, never as shell syntax. Reconstruct the account-store keys that elanous resolves at runtime.
-    `POD_CODEX_ACCOUNTS='${JSON.stringify(accounts)}' bun -e 'const fs=require("node:fs"),path=require("node:path");const names=JSON.parse(process.env.POD_CODEX_ACCOUNTS),home=process.env.HOME;const providers={};let version;for(let i=0;i<names.length;i++){const payload=JSON.parse(fs.readFileSync("/creds/codex-"+i+".json","utf8"));const row=JSON.parse(payload.elanousAuth);if(i===0)version=row.version;const entry=row.providers["openai-codex"];if(!entry)throw Error("missing codex account credential "+i);const codexHome=path.join(home,".elanous/codex-accounts",String(i));fs.mkdirSync(codexHome,{recursive:true});fs.writeFileSync(path.join(codexHome,"auth.json"),payload.codexAuth,{mode:0o600});fs.chmodSync(path.join(codexHome,"auth.json"),0o600);providers[names[i]==="default"?"openai-codex":"openai-codex:"+names[i]]={...entry,codexHome}}if(!providers["openai-codex"])providers["openai-codex"]=providers["openai-codex:"+names[0]];fs.writeFileSync(path.join(home,".elanous/auth.json"),JSON.stringify({version,providers}),{mode:0o600});fs.chmodSync(path.join(home,".elanous/auth.json"),0o600);fs.copyFileSync(path.join(home,".elanous/codex-accounts/0/auth.json"),path.join(home,".codex/auth.json"));fs.chmodSync(path.join(home,".codex/auth.json"),0o600)' || exit 5`,
+    `POD_CODEX_ACCOUNTS='${JSON.stringify(accounts)}' bun -e 'const fs=require("node:fs"),path=require("node:path");const names=JSON.parse(process.env.POD_CODEX_ACCOUNTS),home=process.env.HOME;const providers={},mirrorProviders={};let version;for(let i=0;i<names.length;i++){const payload=JSON.parse(fs.readFileSync("/creds/codex-"+i+".json","utf8"));const row=JSON.parse(payload.elanousAuth);if(i===0)version=row.version;const entry=row.providers["openai-codex"];if(!entry||!entry.tokens||typeof entry.tokens.accessToken!=="string"||!entry.tokens.accessToken||typeof entry.tokens.expiresAt!=="number"||!Number.isFinite(entry.tokens.expiresAt))throw Error("missing Codex access snapshot "+i);const accountKey=names[i]==="default"?"openai-codex":"openai-codex:"+names[i];const codexHome=path.join(home,".elanous/codex-accounts",String(i));fs.mkdirSync(codexHome,{recursive:true});fs.writeFileSync(path.join(codexHome,"auth.json"),payload.codexAuth,{mode:0o600});fs.chmodSync(path.join(codexHome,"auth.json"),0o600);providers[accountKey]={...entry,codexHome};mirrorProviders[accountKey]={authMode:"chatgpt",tokens:{accessToken:entry.tokens.accessToken,expiresAt:entry.tokens.expiresAt,refreshToken:"",refresh_token:""}}}if(!providers["openai-codex"])providers["openai-codex"]=providers["openai-codex:"+names[0]];if(!mirrorProviders["openai-codex"])mirrorProviders["openai-codex"]=mirrorProviders["openai-codex:"+names[0]];fs.writeFileSync(path.join(home,".elanous/auth.json"),JSON.stringify({version,providers}),{mode:0o600});fs.chmodSync(path.join(home,".elanous/auth.json"),0o600);const mirrorDir=path.join(home,".elanous/remote-parent-credentials/cli-mirror");fs.mkdirSync(mirrorDir,{recursive:true,mode:0o700});fs.chmodSync(mirrorDir,0o700);const mirrorPath=path.join(mirrorDir,"auth.json");fs.writeFileSync(mirrorPath,JSON.stringify({version,providers:mirrorProviders}),{mode:0o600});fs.chmodSync(mirrorPath,0o600);fs.copyFileSync(path.join(home,".elanous/codex-accounts/0/auth.json"),path.join(home,".codex/auth.json"));fs.chmodSync(path.join(home,".codex/auth.json"),0o600)' || exit 5`,
     `export ELANOUS_CODEX_ACCOUNT='${accounts[0]!.replace(/'/g, `'\\''`)}'`,
     'export ELANOUS_CODEX_ACCOUNT_HOME="$HOME/.elanous/codex-accounts/0" CODEX_HOME="$HOME/.elanous/codex-accounts/0"',
+    'export ELANOUS_REMOTE_CODEX_ACCESS_ONLY=1',
+    'export ELANOUS_REMOTE_CODEX_MIRROR_PATH="$HOME/.elanous/remote-parent-credentials/cli-mirror/auth.json"',
     podQuotaPolicyExport(),
   ].join('\n');
 }
@@ -1504,7 +1520,9 @@ if [ "$found" -eq 0 ]; then echo ELANOUS_RUN_LEDGER_NONE; fi`,
               // ⭐ 런 계약(graph-run-contract.ts) — 이 칸이 «실제로» Pod 이고, 부모가 정한 계약도 Pod 다. 안의 그래프가 노드마다 대조한다.
               { name: ACTUAL_SUBSTRATE_ENV, value: 'pod' }, { name: RUN_CONTRACT_ENV, value: carryRunContract({ substrate: 'pod' }) },
               ...Object.entries(o.armEnv ?? {}).map(([name, value]) => ({ name, value })),
-              ...o.passEnv.map((key) => ({ name: key, valueFrom: { secretKeyRef: { name: `${o.name}-creds`, key: `env-${key}` } } })),
+              // Growth admission is host-validated control state, never a caller-provided Secret env.
+              ...o.passEnv.filter((key) => key !== 'ELANOUS_GRAPH_GROWTH_SCOPE')
+                .map((key) => ({ name: key, valueFrom: { secretKeyRef: { name: `${o.name}-creds`, key: `env-${key}` } } })),
               ...(o.bedrock ? bedrockPodEnv(o.bedrock) : []),
               ...(o.zai || o.zaiFallback ? zaiPodEnv(o.zai ?? o.zaiFallback!) : []),
             ],
@@ -2631,7 +2649,7 @@ export function podSelfImplementSpawn(options: PodSpawnOptions = {}): SelfImplem
           }
         }
         // Grok 자격은 전용 Secret 키 하나로만 보낸다 — passEnv 의 API 키 중복 전달은 막는다.
-        const passKeys = [...new Set([...(options.passEnv ?? []), ...(openrouterCredentialNeeded && openrouterKey ? ['OPENROUTER_API_KEY'] : []), ...(openGateway ? ['OPENGATEWAY_API_KEY'] : []), ...(anthropic ? ['ANTHROPIC_API_KEY'] : [])])].filter((key) => key !== 'ELANOUS_LIVE_DETAIL_UNTIL'
+        const passKeys = [...new Set([...(options.passEnv ?? []), ...(openrouterCredentialNeeded && openrouterKey ? ['OPENROUTER_API_KEY'] : []), ...(openGateway ? ['OPENGATEWAY_API_KEY'] : []), ...(anthropic ? ['ANTHROPIC_API_KEY'] : [])])].filter((key) => key !== 'ELANOUS_GRAPH_GROWTH_SCOPE' && key !== 'ELANOUS_LIVE_DETAIL_UNTIL'
           && key !== POD_GITHUB_CREDENTIAL_TOKEN_ENV && key !== POD_GITHUB_CREDENTIAL_URL_ENV
           && (!grok || !['XAI_API_KEY', 'GROK_API_KEY', 'GROK_CODE_XAI_API_KEY'].includes(key)));
         const skillEnvs: Record<string, string> = options.skillEnv
@@ -2811,7 +2829,7 @@ export function podSelfImplementSpawn(options: PodSpawnOptions = {}): SelfImplem
         const hostKey = (k: string): string | undefined => k === 'OPENROUTER_API_KEY' && openrouterCredentialNeeded ? openrouterKey : k === 'OPENGATEWAY_API_KEY' && openGateway ? openGatewayKey : k === 'ANTHROPIC_API_KEY' && anthropic ? anthropicKey : env[k] ?? (options.readKeyCache ?? defaultReadKeyCache)(k);
         const hostFallbackNames = hostConfig.llm.fallbackChain;
         const zaiOverage = zaiLaunchOverage ?? zaiSettingsFromConfig(hostConfig.llm).overage;
-        const normalizedFallback = normalizeFallbackChain(hostFallbackNames);
+        const normalizedFallback = normalizeFallbackChain(hostFallbackNames, hostConfig.llm.provider);
         let fallbackSteps = normalizedFallback.chain;
         let zaiFallbackSpec: ZaiPodSpec | undefined;
         const codexChild = options.provider === undefined || options.provider === 'openai-codex';
@@ -2996,10 +3014,11 @@ export function podSelfImplementSpawn(options: PodSpawnOptions = {}): SelfImplem
           ...(credentialRelay ? [POD_CREDENTIAL_TOKEN_ENV] : []),
           ...(githubRelay ? [POD_GITHUB_CREDENTIAL_TOKEN_ENV] : []),
         ];
+        const podGrowthScope = env.ELANOUS_GRAPH_GROWTH_SCOPE === 'harness' ? 'harness' : undefined;
         const detail = selectLiveDetail(options.liveDetailFile ? { path: options.liveDetailFile } : {}).state;
         const detailUntil = detail && (detail.scope === 'all' || detail.scope === parentRunId || detail.scope === launchRunId) && detail.until > Date.now()
           ? String(detail.until) : undefined;
-        const armEnv = Object.fromEntries(Object.entries(options.armEnv ?? {}).filter(([key]) => key !== 'ELANOUS_LIVE_DETAIL_UNTIL' && !key.startsWith('ELANOUS_PROMPT_ENHANCE_') && key !== POD_GITHUB_CREDENTIAL_TOKEN_ENV && key !== POD_GITHUB_CREDENTIAL_URL_ENV && (!hostRefresh || !['GH_TOKEN', 'GITHUB_TOKEN', 'GH_CONFIG_DIR'].includes(key))));
+        const armEnv = Object.fromEntries(Object.entries(options.armEnv ?? {}).filter(([key]) => key !== 'ELANOUS_GRAPH_GROWTH_SCOPE' && key !== 'ELANOUS_LIVE_DETAIL_UNTIL' && !key.startsWith('ELANOUS_PROMPT_ENHANCE_') && key !== POD_GITHUB_CREDENTIAL_TOKEN_ENV && key !== POD_GITHUB_CREDENTIAL_URL_ENV && (!hostRefresh || !['GH_TOKEN', 'GITHUB_TOKEN', 'GH_CONFIG_DIR'].includes(key))));
         const seat = env.ELANOUS_HARNESS_SEAT;
         const authorEnv = authorChoice ? {
           ...(authorChoice.modelSource !== 'default' || inferProviderFromModel(authorChoice.model) === 'openai-codex' ? {
@@ -3018,9 +3037,10 @@ export function podSelfImplementSpawn(options: PodSpawnOptions = {}): SelfImplem
         const reviewDeferredToHost = input.autoMerge === true
           && resolvePodHostAcpReview(options.hostAcpReview, env, options.kubectl ? () => undefined : getUserConfig).enabled;
         if (reviewDeferredToHost) debug.log('self-implement.pod', 'review-deferred-armed', { job: name, env: REVIEW_DEFERRED_TO_HOST_ENV, ...(reviewerPin ? { reviewerPinUnused: reviewerPin } : {}) });
-        const jobArmEnv = grounding || credentialRelay || githubRelay || env.ELANOUS_DISPATCH_RECORDED === '1' || options.armEnv || detailUntil || seat || env.ELANOUS_POD_AUTHOR_ON_POD === '1' || reviewerPin || reviewDeferredToHost || budgetSnapshot
+        const jobArmEnv = grounding || credentialRelay || githubRelay || podGrowthScope || env.ELANOUS_DISPATCH_RECORDED === '1' || options.armEnv || detailUntil || seat || env.ELANOUS_POD_AUTHOR_ON_POD === '1' || reviewerPin || reviewDeferredToHost || budgetSnapshot
           ? {
               ...armEnv,
+              ...(podGrowthScope ? { ELANOUS_GRAPH_GROWTH_SCOPE: podGrowthScope } : {}),
               ...(budgetSnapshot ? { [BUDGET_SNAPSHOT_ENV]: encodeBudgetSnapshot(budgetSnapshot) } : {}),
               ...authorEnv,
               ...(reviewerPin ? { [roleLlmEnvName('review')]: reviewerPin } : {}),
@@ -3667,7 +3687,7 @@ export function podSelfImplementSpawn(options: PodSpawnOptions = {}): SelfImplem
           }
           }
         }
-        const childFailure = state === 'failed' && !oomKilled && failedReason !== 'DeadlineExceeded'
+        let childFailure = state === 'failed' && !oomKilled && failedReason !== 'DeadlineExceeded'
           ? lastPodChildFailure(logs) : null;
         let childPrFailure: ReturnType<typeof podPrOpenFailure> = null;
         if (childFailure) {
@@ -3692,6 +3712,7 @@ export function podSelfImplementSpawn(options: PodSpawnOptions = {}): SelfImplem
           };
         }
         let noResultDiagnostic: string | null = null;
+        let noResultLedgerText: string | null = null;
         let noResultLedgerObservation: { available: boolean; bytes: number; rows: number; lastEvent: string | null } = {
           available: false, bytes: 0, rows: 0, lastEvent: null,
         };
@@ -3701,6 +3722,7 @@ export function podSelfImplementSpawn(options: PodSpawnOptions = {}): SelfImplem
         if (noResultCandidate) {
           try {
             const ledgerText = readFileSync(runLedgerPath(liveChildRunId, ledgerDir), 'utf8');
+            noResultLedgerText = ledgerText;
             const ledgerRows = ledgerText.split('\n').filter(Boolean);
             let lastEvent: string | null = null;
             try {
@@ -3713,6 +3735,10 @@ export function podSelfImplementSpawn(options: PodSpawnOptions = {}): SelfImplem
           catch { /* ledger not returned: do not invent a stage */ }
           noResultDiagnostic ??= 'child terminal result missing; last ledger stage=unknown; round=unknown; mustFix=unknown';
         }
+        if (!childFailure && state === 'failed' && noResultCandidate && noResultLedgerText) {
+          const childError = podNoResultChildError(noResultLedgerText);
+          if (childError) childFailure = { stage: 'pr-open-failed', error: childError, rawError: childError };
+        }
         // POD-FAIL-EVIDENCE-CAPTURE: failed Pods are collected/deleted immediately after terminal handling.
         //   Preserve a bounded log tail for every non-OOM/non-deadline failure, including known child stages:
         //   the terminal result line can be followed by the useful stderr that explains the actual exit.
@@ -3792,6 +3818,7 @@ export function podSelfImplementSpawn(options: PodSpawnOptions = {}): SelfImplem
         const noResultIncident = state === 'failed' && failedReason !== 'DeadlineExceeded' && !disposition;
         if (reportDeficit || noResultIncident) {
           const evidence = disposition?.childEvidence;
+          const firstLlm = readPodChildFirstLlmFact(liveChildRunId, ledgerDir, true);
           const diagnosticLines = [
             ...(Array.isArray(evidence?.podLogTail) ? evidence.podLogTail : []),
             ...(typeof evidence?.detail === 'string' ? evidence.detail.split('\n') : []),
@@ -3812,6 +3839,7 @@ export function podSelfImplementSpawn(options: PodSpawnOptions = {}): SelfImplem
             reason: failedReason || containerReason || 'unknown',
             terminationReason: evidence?.terminationReason ?? null,
             providerErrors: disposition?.providerErrors ?? null,
+            firstLlm: firstLlm ?? null,
             oomKilled, podTermination: podTermination ?? null,
             detailTail: evidence?.detailTail ?? null,
             lastLines: diagnosticLines,
@@ -3926,6 +3954,7 @@ export function podSelfImplementSpawn(options: PodSpawnOptions = {}): SelfImplem
         const childProviderFailure = podChildProviderFailureFact(logs, {
           providerHint: options.provider ?? podChildProviderOf(podDelegatedChildLlm(jobArgs)?.provider),
         });
+        const childFirstLlm = readPodChildFirstLlmFact(liveChildRunId, ledgerDir);
         const childEvidenceRow = podChildEvidenceLedgerRow({
           job: name,
           childRunId: liveChildRunId,
@@ -3935,6 +3964,7 @@ export function podSelfImplementSpawn(options: PodSpawnOptions = {}): SelfImplem
           noResultLogEvidence,
           podTermination,
           providerFailure: childProviderFailure,
+          firstLlm: childFirstLlm,
         });
         if (childEvidenceRow) {
           // Every terminal attempt writes a marker, including one with no parseable child facts.

package/src/tool-hints/gate.ts

@@ -36,7 +36,7 @@ export function evaluateGate(
   const key = cacheKey(catalog, hints, signals);
   const cached = CACHE.get(key);
   if (cached) {
-    recordGateEvaluation(catalog.length, cached.filtered.length, {
+    recordGateEvaluation(catalog, cached, {
       defaultEnabled: notRunStage(),
       probe: notRunStage(),
       intentScope: notRunStage(),
@@ -58,10 +58,12 @@ export function evaluateGate(
   };
   const boost = new Map<string, number>();
   const paramDefaults = new Map<string, Record<string, unknown>>();
+  const excludedReasons = new Map<string, string>();
 
   // ─── Step 1: seed with defaultEnabled entries ──────────────────
   for (const tool of catalog) {
     if (tool.defaultEnabled) enabled.add(tool.id);
+    else excludedReasons.set(tool.id, 'default-disabled');
   }
   stages.defaultEnabled = executedStage(catalog.length, enabled.size);
 
@@ -74,6 +76,7 @@ export function evaluateGate(
     const mode = tool.probe.onFail ?? 'hide';
     if (mode === 'hide') enabled.delete(tool.id);
     else if (mode === 'disable') enabled.delete(tool.id);
+    if (mode === 'hide' || mode === 'disable') excludedReasons.set(tool.id, `probe-${mode}`);
     // 'warn': leave enabled; caller may read hintReasons to surface
   }
   stages.probe = executedStage(beforeProbe, enabled.size);
@@ -93,7 +96,10 @@ export function evaluateGate(
     for (const tool of catalog) {
       const scope: ToolIntentScope = tool.intentScope ?? 'always';
       if (scope === 'always') continue;
-      if (!activeScopes.has(scope)) enabled.delete(tool.id);
+      if (!activeScopes.has(scope)) {
+        enabled.delete(tool.id);
+        excludedReasons.set(tool.id, `intent-scope:${scope}`);
+      }
     }
     stages.intentScope = executedStage(beforeIntentScope, enabled.size);
   }
@@ -107,7 +113,10 @@ export function evaluateGate(
     for (const tool of catalog) {
       if (!tool.minTier) continue;
       const need = tierRank[tool.minTier] ?? 2;
-      if (activeRank < need) enabled.delete(tool.id);
+      if (activeRank < need) {
+        enabled.delete(tool.id);
+        excludedReasons.set(tool.id, `min-tier:${tool.minTier}`);
+      }
     }
     stages.minTier = executedStage(beforeMinTier, enabled.size);
   }
@@ -126,8 +135,14 @@ export function evaluateGate(
       if (seenForKind.has(comboKey)) continue;
       seenForKind.set(comboKey, { toolId, kind: h.kind });
 
-      if (h.kind === 'enable')  enabled.add(toolId);
-      if (h.kind === 'disable') enabled.delete(toolId);
+      if (h.kind === 'enable') {
+        enabled.add(toolId);
+        excludedReasons.delete(toolId);
+      }
+      if (h.kind === 'disable') {
+        enabled.delete(toolId);
+        excludedReasons.set(toolId, 'hint-disabled');
+      }
       if (h.kind === 'prefer') boost.set(toolId, (boost.get(toolId) ?? 0) + 2);
       if (h.kind === 'boost')  boost.set(toolId, (boost.get(toolId) ?? 0) + 1);
       if (h.kind === 'avoid')  boost.set(toolId, (boost.get(toolId) ?? 0) - 1);
@@ -181,11 +196,21 @@ export function evaluateGate(
     const beforeModelFamilyCap = keep.length;
     const cap = signals.researchAgent ? 16 : 8;
     if (keep.length > cap) {
-      keep = keep
+      const requiredResearchTools = signals.intentResearch
+        ? ['web_search', 'web_fetch'].filter(id => enabled.has(id))
+        : [];
+      const required = new Set(requiredResearchTools);
+      const ranked = keep
+        .filter(id => !required.has(id))
         .map(id => ({ id, score: boost.get(id) ?? 0 }))
         .sort((a, b) => b.score - a.score || a.id.localeCompare(b.id))
-        .slice(0, cap)
+        .slice(0, Math.max(0, cap - requiredResearchTools.length))
         .map(x => x.id);
+      keep = [...requiredResearchTools, ...ranked];
+      const selectedBeforeStableSort = new Set(keep);
+      for (const id of enabled) {
+        if (!selectedBeforeStableSort.has(id)) excludedReasons.set(id, 'model-family-cap');
+      }
     }
     stages.modelFamilyCap = executedStage(beforeModelFamilyCap, keep.length);
   }
@@ -204,6 +229,7 @@ export function evaluateGate(
     boost: Object.fromEntries(boost),
     paramDefaults: Object.fromEntries(paramDefaults),
     hintReasons,
+    excludedReasons: Object.fromEntries(excludedReasons),
   };
 
   if (CACHE.size >= CACHE_MAX) {
@@ -212,7 +238,7 @@ export function evaluateGate(
     if (firstKey !== undefined) CACHE.delete(firstKey);
   }
   CACHE.set(key, decision);
-  recordGateEvaluation(catalog.length, decision.filtered.length, stages, false);
+  recordGateEvaluation(catalog, decision, stages, false);
   return decision;
 }
 
@@ -237,18 +263,23 @@ function executedStage(before: number, after: number): GateStage {
 }
 
 function recordGateEvaluation(
-  catalogCount: number,
-  filteredCount: number,
+  catalog: NativeToolCatalogEntry[],
+  decision: GateDecision,
   stages: GateStages,
   cacheHit: boolean,
 ): void {
-  // This is a per-call summary: names stay out so high-cardinality catalogs do not flood observability.
+  const active = new Set(decision.filtered);
   debug.log('capability.resolve', 'tool-gate-evaluated', {
     cacheHit,
-    catalogCount,
-    filteredCount,
-    removedCount: catalogCount - filteredCount,
+    catalogCount: catalog.length,
+    filteredCount: decision.filtered.length,
+    removedCount: catalog.length - decision.filtered.length,
     stages,
+    activeTools: decision.filtered.map((id) => catalog.find((tool) => tool.id === id)?.displayName ?? id),
+    excludedTools: catalog.filter((tool) => !active.has(tool.id)).map((tool) => ({
+      name: tool.displayName,
+      reason: decision.excludedReasons?.[tool.id] ?? 'unclassified-filter',
+    })),
   });
 }
 

package/src/tool-hints/signals.ts

@@ -120,7 +120,7 @@ const RESEARCH_ASCII_RE = /\b(research|find out|look up|investigate|latest (on|n
 const RESEARCH_KO_RE = /(최신|뉴스|찾아(?:줘|봐)|알아봐|조사|리서치)/;
 const RESEARCH_SKILL_RE = /\b(?:omni[ -]market|omni[ -]crawl|social[ -]pulse)\b/i;
 const RESEARCH_EXPLANATION_RE = /(?:왜|이유|원인|전망|분석|비교|조사|\bwhy\b|\bcause\b|\boutlook\b|\banalysis\b|\bcompare\b)/i;
-const RESEARCH_TOPIC_RE = /(?:금리|채권|시장|경제|물가|환율|주가|기업|산업|정책|선거|역사|기후|질병|의학|과학|\bbond\b|\byield\b|\bmarket\b|\beconomy\b|\binflation\b|\bstock\b|\bpolicy\b|\belection\b|\bhistory\b|\bclimate\b|\bhealth\b|\bscience\b)/i;
+const RESEARCH_TOPIC_RE = /(?:금리|채권|시장|경제|물가|환율|주가|기업|산업|정책|선거|역사|기후|질병|의학|과학|반도체|HBM\d?|메모리|공급망|양산|인증|마이크론|삼성전자|SK하이닉스|\bbond\b|\byield\b|\bmarket\b|\beconomy\b|\binflation\b|\bstock\b|\bpolicy\b|\belection\b|\bhistory\b|\bclimate\b|\bhealth\b|\bscience\b)/i;
 const RESEARCH_CODING_RE = /(?:코드|함수|버그|테스트|빌드|커밋|하니스|설정|오류|구현|수리|고쳐|수정|저장소|\brepo\b|\bcode\b|\bfunction\b|\bbug\b|\btest\b|\bbuild\b|\bcommit\b|\bconfig\b|\bimplement\b|\bfix\b)/i;
 const DEEP_RESEARCH_INTENT_TERMS = [
   'research', 'deep research', 'triangulate', 'compare sources',

package/src/tool-hints/types.ts

@@ -102,4 +102,6 @@ export interface GateDecision {
   paramDefaults: Record<string, Record<string, unknown>>;
   /** Reasons surfaced by current hints — for prompt injection. */
   hintReasons: string[];
+  /** Why a catalog entry was not selected, keyed by stable catalog id. */
+  excludedReasons?: Record<string, string>;
 }

package/src/tool-runtime/browser-hand-runtime.ts

@@ -14,6 +14,7 @@ import { requestConfirmation } from '../hitl/confirm.js';
 import type { ToolRuntime, ToolRuntimeContext } from './types.js';
 import { coordinateBrowserRoundtrip } from '../agent-mission/browser-roundtrip-coordinator.js';
 import { appendRunLedgerEntry } from '../self-implement/run-ledger.js';
+import { getHarnessRunId } from '../harness/harness-space.js';
 
 const UUID = /^[0-9a-f]{8}-[0-9a-f]{4}-[1-8][0-9a-f]{3}-[89ab][0-9a-f]{3}-[0-9a-f]{12}$/i;
 const SELECTION_COMMENT = '이것도 반영';
@@ -194,6 +195,15 @@ async function loadPersona(id: string) {
 
 /** Attached-CDP browser action stays behind harness browser-act's armed, host and reversibility checks. */
 export async function dispatchBrowserHand(args: BrowserHandArgs, _ctx?: ToolRuntimeContext, deps: BrowserHandDeps = {}): Promise<BrowserHandResult> {
+  const runId = _ctx?.runId || getHarnessRunId() || undefined;
+  const browserActDeps = {
+    ...deps.browserActDeps,
+    ...(_ctx?.runId
+      ? { getRunId: () => _ctx.runId! }
+      : deps.browserActDeps?.getRunId
+        ? {}
+        : { getRunId: () => runId ?? null }),
+  };
   const url = args.url.trim();
   const target = args.target.trim();
   const issueId = args.linearSubtaskId.trim();
@@ -240,7 +250,7 @@ export async function dispatchBrowserHand(args: BrowserHandArgs, _ctx?: ToolRunt
     };
     return (deps.browserAct ?? performBrowserAction)({
       url: actionUrl, target, armed: true, entryPoint: 'tool-runtime/browser-hand', ...(persona ? { persona } : {}),
-    }, { ...deps.browserActDeps, observe, reversibilityPolicy: 'navigation-only' });
+    }, { ...browserActDeps, observe, reversibilityPolicy: 'navigation-only' });
   };
   try { sourceText = await inspect(url, pagePort); } catch { /* unavailable observation stays unknown */ }
   let block = blockedKind(sourceText);

package/src/tool-runtime/browser-runtime.ts

@@ -50,9 +50,9 @@ export interface BrowserRuntimeDeps {
   observe?: BrowserActionObserver;
 }
 
-function observeBrowserAction(data: Record<string, unknown>): void {
+function observeBrowserAction(data: Record<string, unknown>, explicitRunId?: string): void {
   try {
-    const runId = getHarnessRunId() || null;
+    const runId = explicitRunId || getHarnessRunId() || null;
     (deps.observe ?? ((event, detail) => debug.log('harness.browser-action', event, detail)))('executed', {
       runId,
       sessionId: 'browser-runtime',
@@ -95,17 +95,20 @@ async function defaultCloseClient(): Promise<void> {
   singletonPromise = null;
 }
 
-let deps: BrowserRuntimeDeps = {
+const DEFAULT_DEPS: BrowserRuntimeDeps = {
   getClient: defaultGetClient,
   closeClient: defaultCloseClient,
 };
+let deps: BrowserRuntimeDeps = DEFAULT_DEPS;
 let attachedControlSignalBus: ReturnType<typeof defaultControlSignalBus> | null = null;
+let detachBrowserControlSignalConsumer: (() => void) | null = null;
 
 function ensureBrowserControlSignalConsumer(): void {
   const bus = defaultControlSignalBus();
   if (attachedControlSignalBus === bus) return;
+  detachBrowserControlSignalConsumer?.();
   attachedControlSignalBus = bus;
-  bus.subscribe(
+  detachBrowserControlSignalConsumer = bus.subscribe(
     { kind: 'browser-cdp-stop', minUrgency: 'quick-pass', surface: 'browser' },
     () => {
       void deps.closeClient?.();
@@ -117,6 +120,14 @@ export function setBrowserRuntimeDeps(next: Partial<BrowserRuntimeDeps>): void {
   deps = { ...deps, ...next };
 }
 
+/** Reset dependency injection after a run-scoped observability test. */
+export function _resetBrowserRuntimeDepsForTest(): void {
+  detachBrowserControlSignalConsumer?.();
+  detachBrowserControlSignalConsumer = null;
+  deps = DEFAULT_DEPS;
+  attachedControlSignalBus = null;
+}
+
 export async function closeBrowserRuntime(): Promise<void> {
   if (deps.closeClient) await deps.closeClient();
 }
@@ -189,6 +200,7 @@ const MAX_NAV_TIMEOUT_MS = 60_000;
 
 export async function dispatchBrowserNavigate(args: BrowserNavigateArgs, ctx?: ToolRuntimeContext): Promise<BrowserNavigateResult> {
   ensureBrowserControlSignalConsumer();
+  const runId = ctx?.runId || getHarnessRunId() || undefined;
   const url = String(args.url ?? '').trim();
   const input = args.input;
   // Input is never an implicit side effect of navigation. Approval is scoped
@@ -215,7 +227,7 @@ export async function dispatchBrowserNavigate(args: BrowserNavigateArgs, ctx?: T
     if (!approved) return { output: '# BrowserNavigate: input not approved', finalUrl: '', title: '', loadMs: 0 };
   }
   if (!url) {
-    observeBrowserAction({ action: 'navigate', url: '', target: 'url', ok: false, error: 'url is required' });
+    observeBrowserAction({ action: 'navigate', url: '', target: 'url', ok: false, error: 'url is required' }, runId);
     return { output: '# BrowserNavigate: error\nurl is required', finalUrl: '', title: '', loadMs: 0 };
   }
   const waitForLoad = args.waitForLoad !== false;
@@ -226,12 +238,12 @@ export async function dispatchBrowserNavigate(args: BrowserNavigateArgs, ctx?: T
     client = await deps.getClient();
   } catch (err) {
     const error = err instanceof Error ? err.message : String(err);
-    observeBrowserAction({ action: 'navigate', url, target: url, ok: false, error });
+    observeBrowserAction({ action: 'navigate', url, target: url, ok: false, error }, runId);
     throw err;
   }
   if (!client) {
     const availability = getBrowserCdpAvailability();
-    observeBrowserAction({ action: 'navigate', url, target: url, ok: false, error: 'chrome-unavailable' });
+    observeBrowserAction({ action: 'navigate', url, target: url, ok: false, error: 'chrome-unavailable' }, runId);
     return {
       output: `# BrowserNavigate: Chrome unavailable\n${availability.note}`,
       finalUrl: '',
@@ -264,7 +276,7 @@ export async function dispatchBrowserNavigate(args: BrowserNavigateArgs, ctx?: T
     await client.navigate(url);
   } catch (err) {
     const error = err instanceof Error ? err.message : String(err);
-    observeBrowserAction({ action: 'navigate', url, target: url, ok: false, error, loadMs: Date.now() - started });
+    observeBrowserAction({ action: 'navigate', url, target: url, ok: false, error, loadMs: Date.now() - started }, runId);
     return {
       output: `# BrowserNavigate: error\n${error}`,
       finalUrl: url,
@@ -333,7 +345,7 @@ export async function dispatchBrowserNavigate(args: BrowserNavigateArgs, ctx?: T
     timedOut ? 'warning: load event timed out' : '',
   ].filter(Boolean).join('\n');
 
-  observeBrowserAction({ action: 'navigate', url: finalUrl, target: url, ok: true, titleLength: title.length, loadMs, timedOut });
+  observeBrowserAction({ action: 'navigate', url: finalUrl, target: url, ok: true, titleLength: title.length, loadMs, timedOut }, runId);
   return { output, finalUrl, title, loadMs, timedOut: timedOut || undefined };
 }
 
@@ -405,8 +417,9 @@ function escapeForJs(s: string): string {
   return s.replace(/\\/g, '\\\\').replace(/'/g, "\\'");
 }
 
-export async function dispatchBrowserRead(args: BrowserReadArgs): Promise<BrowserReadResult> {
+export async function dispatchBrowserRead(args: BrowserReadArgs, ctx?: ToolRuntimeContext): Promise<BrowserReadResult> {
   ensureBrowserControlSignalConsumer();
+  const runId = ctx?.runId || getHarnessRunId() || undefined;
   const mode: BrowserReadMode = args.mode ?? 'text';
   const selector = typeof args.selector === 'string' ? args.selector.trim() : '';
   const maxChars = Math.min(
@@ -419,12 +432,12 @@ export async function dispatchBrowserRead(args: BrowserReadArgs): Promise<Browse
     client = await deps.getClient();
   } catch (err) {
     const error = err instanceof Error ? err.message : String(err);
-    observeBrowserAction({ action: 'read', target: selector || 'document.body', mode, ok: false, error, attachmentRef: null, captureOutcome: mode === 'screenshot' ? 'error' : 'not-requested' });
+    observeBrowserAction({ action: 'read', target: selector || 'document.body', mode, ok: false, error, attachmentRef: null, captureOutcome: mode === 'screenshot' ? 'error' : 'not-requested' }, runId);
     throw err;
   }
   if (!client) {
     const availability = getBrowserCdpAvailability();
-    observeBrowserAction({ action: 'read', target: selector || 'document.body', mode, ok: false, error: 'chrome-unavailable', attachmentRef: null, captureOutcome: mode === 'screenshot' ? 'unavailable' : 'not-requested' });
+    observeBrowserAction({ action: 'read', target: selector || 'document.body', mode, ok: false, error: 'chrome-unavailable', attachmentRef: null, captureOutcome: mode === 'screenshot' ? 'unavailable' : 'not-requested' }, runId);
     return {
       output: `# BrowserRead: Chrome unavailable\n${availability.note}`,
       mode,
@@ -435,7 +448,7 @@ export async function dispatchBrowserRead(args: BrowserReadArgs): Promise<Browse
     try {
       const png = await client.screenshot({ format: 'png' });
       const b64 = png.toString('base64');
-      observeBrowserAction({ action: 'read', target: 'screenshot', mode, ok: true, bytes: png.byteLength, base64Length: b64.length, attachmentRef: null, captureOutcome: 'inline-base64' });
+      observeBrowserAction({ action: 'read', target: 'screenshot', mode, ok: true, bytes: png.byteLength, base64Length: b64.length, attachmentRef: null, captureOutcome: 'inline-base64' }, runId);
       return {
         output: `# BrowserRead: screenshot (${png.byteLength} bytes PNG, ${b64.length} chars base64)`,
         mode,
@@ -443,7 +456,7 @@ export async function dispatchBrowserRead(args: BrowserReadArgs): Promise<Browse
       };
     } catch (err) {
       const error = err instanceof Error ? err.message : String(err);
-      observeBrowserAction({ action: 'read', target: 'screenshot', mode, ok: false, error, attachmentRef: null, captureOutcome: 'error' });
+      observeBrowserAction({ action: 'read', target: 'screenshot', mode, ok: false, error, attachmentRef: null, captureOutcome: 'error' }, runId);
       return {
         output: `# BrowserRead: screenshot error\n${error}`,
         mode,
@@ -462,7 +475,7 @@ export async function dispatchBrowserRead(args: BrowserReadArgs): Promise<Browse
     value = v == null ? '' : String(v);
   } catch (err) {
     const error = err instanceof Error ? err.message : String(err);
-    observeBrowserAction({ action: 'read', target: selector || 'document.body', mode, ok: false, error, attachmentRef: null, captureOutcome: 'not-requested' });
+    observeBrowserAction({ action: 'read', target: selector || 'document.body', mode, ok: false, error, attachmentRef: null, captureOutcome: 'not-requested' }, runId);
     return {
       output: `# BrowserRead: evaluate error\n${error}`,
       mode,
@@ -472,7 +485,7 @@ export async function dispatchBrowserRead(args: BrowserReadArgs): Promise<Browse
   const truncated = value.length > maxChars;
   const out = truncated ? value.slice(0, maxChars) : value;
 
-  observeBrowserAction({ action: 'read', target: selector || 'document.body', mode, ok: true, charLength: value.length, returnedChars: out.length, truncated, attachmentRef: null, captureOutcome: 'not-requested' });
+  observeBrowserAction({ action: 'read', target: selector || 'document.body', mode, ok: true, charLength: value.length, returnedChars: out.length, truncated, attachmentRef: null, captureOutcome: 'not-requested' }, runId);
   return {
     output: [
       `# BrowserRead (${mode}${selector ? ` · selector=${selector}` : ''})`,
@@ -490,8 +503,8 @@ export async function dispatchBrowserRead(args: BrowserReadArgs): Promise<Browse
 export const browserReadRuntime: ToolRuntime<BrowserReadArgs, BrowserReadResult> = {
   id: 'browser_read',
   spec: buildBrowserReadTool(),
-  async run(req, _ctx: ToolRuntimeContext) {
-    return dispatchBrowserRead(req);
+  async run(req, ctx: ToolRuntimeContext) {
+    return dispatchBrowserRead(req, ctx);
   },
 };