elanous 0.2.0 → 0.2.4

package/src/task-orchestrator/surfaces/pod-command-job.ts

@@ -0,0 +1,348 @@
+// ☸️ 명령 Job — 하니스 없이 스크립트 한 줄을 Pod 에서 돌리고 `~/outbox` 산출을 호스트로 돌려받는다.
+//
+// 자격은 기본 0 이다. `skills` 에 이름이 있을 때만 그 스킬 `.env`(기존 스킬 키 관 · 런 Secret · 0600)를,
+// `llm: 'grok'` 일 때만 grok access 사본(refresh 없음 · `hostGrokCredentials` 검사 재사용)을 싣는다.
+// codex·elanous 자격 파일은 이 Job 에 없다. 격리 관문(호스트 넥서스에 닿으면 종료)은 기존 Job 과 같다.
+// 명령 인자는 셸 문자열로 이어붙이지 않고 `"$@"` 로 따로 인용한다.
+
+import { spawnSync } from 'node:child_process';
+import { debug } from '../../debug/log.js';
+import { effectiveInstanceRoot } from '../../instance/resolve.js';
+import { collectPodArtifacts } from './pod-artifact-return.js';
+import { type PodPoolMember, PodPoolScheduler, resolvePodPoolSpec, parsePodPool, checkPodPool, syncPoolImages, type PoolKubectl } from './pod-pool.js';
+import { readSkillEnvFiles } from './pod-skills.js';
+import { podSourceScript } from './pod-source-receive.js';
+import {
+  POD_CHILD_REQUESTS,
+  POD_JOB_DEADLINE_SECONDS,
+  defaultGhToken,
+  defaultKubectl,
+  hostGrokCredentials,
+  podImageFreshness,
+  podJobName,
+  type Kubectl,
+} from './self-implement-pod.js';
+
+/** CLI `--deadline` 생략 시 — 기존 Pod Job 수명 상한과 같다. */
+export const POD_COMMAND_DEADLINE_SECONDS = POD_JOB_DEADLINE_SECONDS;
+
+const COMMAND_LOG_CHARS = 80;
+
+const GATE = `ok=0
+for i in $(seq 1 60); do
+  if curl -s -m 1 -o /dev/null http://host.orb.internal:31415/health || curl -s -m 1 -o /dev/null http://core.elanous-prod:8080/; then ok=0; else ok=$((ok+1)); fi
+  [ "$ok" -ge 3 ] && { echo "[gate] isolation enforced after \${i} probes"; exit 0; }
+  sleep 0.5
+done
+echo "[gate] ISOLATION NOT ENFORCED within 30s"; exit 1`;
+
+/** 기존 Job 과 같은 `~/outbox` 조각 형식(`ELANOUS_POD_ARTIFACT`). 원장 회수는 하지 않는다. */
+const ARTIFACT_EMIT = `set -o pipefail
+artifact_bytes=0
+if [ -d "$HOME/outbox" ]; then
+  while IFS= read -r -d '' file; do
+    [ -f "$file" ] && [ ! -L "$file" ] || continue
+    relative=\${file#"$HOME/outbox/"}
+    size=$(( $(wc -c < "$file") ))
+    if [ "$size" -gt 5242880 ] || [ $((artifact_bytes + size)) -gt 20971520 ]; then
+      printf 'ELANOUS_POD_ARTIFACT_SKIPPED %s %s\\n' "$relative" "$size"
+      continue
+    fi
+    path_token=$(printf '%s' "$relative" | base64 | tr -d '\\n' | tr '+/' '-_' | tr -d '=')
+    if encoded=$(gzip -c "$file" | base64 | tr -d '\\n'); then
+      artifact_bytes=$((artifact_bytes + size))
+      total=$(( (\${#encoded} + 7999) / 8000 ))
+      for ((n=1; n<=total; n++)); do
+        chunk=\${encoded:$(( (n-1)*8000 )):8000}
+        printf 'ELANOUS_POD_ARTIFACT %s %s/%s %s\\n' "$path_token" "$n" "$total" "$chunk"
+      done
+    else
+      printf 'ELANOUS_POD_ARTIFACT_SKIPPED %s %s\\n' "$relative" "$size"
+    fi
+  done < <(find "$HOME/outbox" -type f -print0)
+fi`;
+
+export interface PodCommandJobInput {
+  name: string;
+  namespace: string;
+  image: string;
+  imagePullPolicy?: 'Never' | 'IfNotPresent';
+  repoUrl: string;
+  command: readonly string[];
+  skills: readonly string[];
+  llm?: 'grok';
+  /** 저장소를 clone 한다 — 비공개 저장소라 GitHub 토큰이 Secret 으로 간다(명시 opt-in). 기본은 clone 없이 `~/work` 에서 이미지의 `elanous` 로 돈다. */
+  clone?: boolean;
+  deadlineSeconds: number;
+  runId?: string;
+}
+
+function bashSingleQuote(value: string): string {
+  return `'${value.replace(/'/g, `'\\''`)}'`;
+}
+
+/** 명령 Job 매니페스트. `skills`·`llm` 을 안 주면 Secret 키를 하나도 참조하지 않는다. */
+export function podCommandJobManifest(o: PodCommandJobInput): Record<string, unknown> {
+  const skillNames = [...new Set(o.skills.map((s) => s.trim()).filter(Boolean))];
+  for (const name of skillNames) {
+    if (!/^[A-Za-z0-9][A-Za-z0-9._-]*$/u.test(name)) throw new Error(`pod command skill name rejected: ${name}`);
+  }
+  const grok = o.llm === 'grok';
+  const quotedArgs = o.command.map(bashSingleQuote).join(' ');
+  const credLines = [
+    ...(grok ? ['mkdir -p ~/.grok && install -m 600 /creds/grok-auth.json ~/.grok/auth.json'] : []),
+    ...(skillNames.length
+      ? [`for n in ${skillNames.join(' ')}; do [ -d ~/.claude/skills/$n ] && install -m 600 /creds/skillenv-$n ~/.claude/skills/$n/.env; done`]
+      : []),
+  ];
+  const script = [
+    'set -u',
+    ...credLines,
+    'curl -s -m 3 -o /dev/null http://host.orb.internal:31415/health && { echo "[pod] ISOLATION FAIL"; exit 3; }',
+    ...(o.clone
+      ? ['export GH_TOKEN="$(cat /creds/gh-token)" && gh auth setup-git', podSourceScript({ kind: 'default' }, o.repoUrl)]
+      : ['mkdir -p ~/work && cd ~/work']),
+    `set -- ${quotedArgs}`,
+    '"$@"',
+    'rc=$?',
+    ARTIFACT_EMIT,
+    'exit $rc',
+  ].join('\n');
+  const secretKeys = grok || skillNames.length || o.clone === true;
+  return {
+    apiVersion: 'batch/v1',
+    kind: 'Job',
+    metadata: { name: o.name, namespace: o.namespace, labels: { 'elanous.substrate': 'pod', 'elanous.job': o.name, 'elanous.kind': 'command' } },
+    spec: {
+      backoffLimit: 0,
+      ttlSecondsAfterFinished: 7200,
+      activeDeadlineSeconds: o.deadlineSeconds,
+      template: {
+        metadata: { labels: { 'elanous.job': o.name, 'elanous.kind': 'command' } },
+        spec: {
+          restartPolicy: 'Never',
+          securityContext: { runAsUser: 1000, fsGroup: 1000 },
+          initContainers: [{ name: 'isolation-gate', image: o.image, imagePullPolicy: o.imagePullPolicy ?? 'Never', command: ['bash', '-c'], args: [GATE] }],
+          containers: [{
+            name: 'child', image: o.image, imagePullPolicy: o.imagePullPolicy ?? 'Never',
+            resources: { requests: { ...POD_CHILD_REQUESTS }, limits: { memory: '16Gi', cpu: '4' } },
+            command: ['bash', '-c'], args: [script],
+            env: [
+              ...(o.runId ? [{ name: 'ELANOUS_RUN_ID', value: o.runId }] : []),
+              { name: 'ELANOUS_SUBSTRATE', value: 'pod' },
+              { name: 'ELANOUS_POD_NAME', valueFrom: { fieldRef: { fieldPath: 'metadata.name' } } },
+            ],
+            ...(secretKeys ? { volumeMounts: [{ name: 'creds', mountPath: '/creds', readOnly: true }] } : {}),
+          }],
+          ...(secretKeys ? { volumes: [{ name: 'creds', secret: { secretName: `${o.name}-creds`, defaultMode: 0o400 } }] } : {}),
+        },
+      },
+    },
+  };
+}
+
+export function commandJobSecret(o: {
+  name: string;
+  namespace: string;
+  skills: readonly string[];
+  skillEnvText: Readonly<Record<string, string>>;
+  grokAuth?: string;
+  ghToken?: string;
+}): Record<string, unknown> | null {
+  const stringData: Record<string, string> = {};
+  if (o.grokAuth !== undefined) stringData['grok-auth.json'] = o.grokAuth;
+  if (o.ghToken !== undefined) stringData['gh-token'] = o.ghToken;
+  for (const name of o.skills) {
+    const text = o.skillEnvText[name];
+    if (text !== undefined) stringData[`skillenv-${name}`] = text;
+  }
+  if (Object.keys(stringData).length === 0) return null;
+  return {
+    apiVersion: 'v1', kind: 'Secret', type: 'Opaque',
+    metadata: { name: `${o.name}-creds`, namespace: o.namespace, labels: { 'elanous.job': o.name } },
+    stringData,
+  };
+}
+
+export function podCommandScript(manifest: Record<string, unknown>): string {
+  const spec = manifest.spec as { template: { spec: { containers: Array<{ args: string[] }> } } };
+  return spec.template.spec.containers[0]!.args[0]!;
+}
+
+export function podCommandSecretKeys(secret: Record<string, unknown> | null): string[] {
+  if (!secret) return [];
+  return Object.keys((secret.stringData as Record<string, string> | undefined) ?? {});
+}
+
+export interface PodCommandResult {
+  exitCode: number;
+  artifactsDir: string;
+  job: string;
+  /** 이 런이 실제로 쓴 이미지(레지스트리 판 태그 또는 로컬 이름). */
+  image?: string;
+}
+
+export interface RunPodCommandOptions {
+  command: readonly string[];
+  pool?: string;
+  skills?: readonly string[];
+  llm?: 'grok';
+  clone?: boolean;
+  ghToken?: () => string;
+  deadlineSeconds?: number;
+  namespace?: string;
+  image?: string;
+  repoUrl?: string;
+  runId?: string;
+  name?: string;
+  pollMs?: number;
+  kubectl?: Kubectl;
+  poolScheduler?: PodPoolScheduler;
+  checkPool?: typeof checkPodPool;
+  syncImages?: typeof syncPoolImages;
+  imageCommit?: string | null;
+  readSkillEnv?: (skills: readonly string[]) => Record<string, string>;
+  grokCredentials?: () => { grokAuth?: string; grokApiKey?: string; ghToken: string };
+  artifactsRoot?: string;
+  sleep?: (ms: number) => Promise<void>;
+  log?: (category: string, event: string, data: Record<string, unknown>) => void;
+  env?: NodeJS.ProcessEnv;
+}
+
+function jobFinished(types: string): 'complete' | 'failed' | null {
+  if (/Complete|SuccessCriteriaMet/.test(types)) return 'complete';
+  if (/Failed|FailureTarget/.test(types)) return 'failed';
+  return null;
+}
+
+function exitCodeFromPod(kubectl: Kubectl, namespace: string, job: string, state: 'complete' | 'failed'): number {
+  const got = kubectl(['-n', namespace, 'get', 'pods', '-l', `job-name=${job}`, '-o', 'jsonpath={.items[0].status.containerStatuses[?(@.name=="child")].state.terminated.exitCode}']);
+  const code = Number(got.stdout.trim());
+  if (got.status === 0 && Number.isInteger(code)) return code;
+  return state === 'complete' ? 0 : 1;
+}
+
+/** 명령 Job 의 목표 이미지 판. 원격 노드면 발사 트리의 HEAD(없으면 로컬 이미지 판) · 이 기계면 로컬 이미지 판. */
+export function podCommandTargetCommit(o: { remote: boolean; gitHead: () => string | null; localImageCommit: () => string | null }): string | null {
+  if (o.remote) return o.gitHead() ?? o.localImageCommit();
+  return o.localImageCommit();
+}
+
+/** 풀 자리 · 이미지 판 · Job 적용 · 완료 대기 · 산출 회수 · Secret 정리. */
+export async function runPodCommand(options: RunPodCommandOptions): Promise<PodCommandResult> {
+  const command = [...options.command];
+  if (command.length === 0) throw new Error('pod command: 명령이 비었다');
+  const skills = [...new Set((options.skills ?? []).map((s) => s.trim()).filter(Boolean))];
+  const llm = options.llm;
+  const log = options.log ?? ((c, e, d) => debug.log(c, e, d));
+  const baseKubectl = options.kubectl ?? defaultKubectl;
+  const sleep = options.sleep ?? ((ms: number) => new Promise<void>((r) => setTimeout(r, ms)));
+  const namespace = options.namespace ?? 'elanous-test';
+  const image = options.image ?? 'elanous-harness:local';
+  const repoUrl = options.repoUrl ?? 'https://github.com/ElanvitalAI/elanous';
+  const deadlineSeconds = options.deadlineSeconds ?? POD_COMMAND_DEADLINE_SECONDS;
+  const name = options.name ?? podJobName(`cmd-${Date.now().toString(36)}`);
+  const env = options.env ?? process.env;
+
+  let member: PodPoolMember | null = null;
+  let pool = options.poolScheduler;
+  if (!pool && options.pool) {
+    const spec = resolvePodPoolSpec(options.pool, env);
+    if (!spec) throw new Error('pod command: 풀 스펙이 비었다');
+    const members = parsePodPool(spec);
+    const checked = (options.checkPool ?? checkPodPool)(members, baseKubectl as PoolKubectl);
+    if (!checked.ok) throw new Error('pod command: 풀의 노드가 하나도 준비되지 않았다');
+    pool = new PodPoolScheduler(checked.ready);
+  }
+  if (pool) {
+    for (;;) {
+      member = pool.tryAcquire();
+      if (member) break;
+      await sleep(options.pollMs ?? 15_000);
+    }
+  }
+  const currentContext = member || options.kubectl ? null : baseKubectl(['config', 'current-context']);
+  const context = member?.context ?? (currentContext?.status === 0 ? currentContext.stdout.trim() : '');
+  const kubectl: Kubectl = (args, stdin) => baseKubectl(context ? ['--context', context, ...args] : [...args], stdin);
+  const cleanupSecret = () => { kubectl(['-n', namespace, 'delete', 'secret', `${name}-creds`, '--ignore-not-found']); };
+  const artifactsDir = `${options.artifactsRoot ?? `${effectiveInstanceRoot()}/pod-artifacts`}/${name}`;
+  try {
+    if (!context && !options.kubectl) throw new Error('pod command: kubectl context 를 확인할 수 없다');
+    let grokAuth: string | undefined;
+    if (llm === 'grok') {
+      const creds = (options.grokCredentials ?? (() => hostGrokCredentials({ env, apiKeyOptIn: false })))();
+      if (!creds.grokAuth) throw new Error('grok: 구독 자격 없음 · API 키 opt-in 꺼짐 또는 키 없음');
+      let parsed: unknown;
+      try { parsed = JSON.parse(creds.grokAuth); } catch { throw new Error('grok: Pod 구독 자격 JSON 이 아니다'); }
+      if (!parsed || typeof parsed !== 'object' || Array.isArray(parsed) || !Object.values(parsed).some((scope) => scope && typeof scope === 'object' && typeof (scope as { key?: unknown }).key === 'string') || /"[^"]*refresh[^"]*"\s*:/i.test(creds.grokAuth)) {
+        throw new Error('grok: Pod 구독 자격에 access 토큰이 없거나 refresh 필드가 있다');
+      }
+      grokAuth = creds.grokAuth;
+    }
+    const skillEnvText = skills.length ? (options.readSkillEnv ?? readSkillEnvFiles)(skills) : {};
+    // ☸️ 원격 노드면 목표 판 = 발사 트리의 HEAD(하니스 발사와 같은 규칙) — 이 기계의 로컬 이미지 판을 따르지 않는다.
+    //   🐞 2026-09-27(🅞 실측): 로컬 `elanous-harness:local` 이 옛 판(30f95cf1)이라 레지스트리의 옛 태그를 골라 새 스킬이 없는 이미지로 돌았다.
+    const imageCommit = options.imageCommit !== undefined
+      ? options.imageCommit
+      : options.kubectl ? null
+        : podCommandTargetCommit({
+          remote: Boolean(member?.sshHost),
+          gitHead: () => spawnSync('git', ['rev-parse', 'HEAD'], { encoding: 'utf8' }).stdout?.trim() || null,
+          localImageCommit: () => podImageFreshness({ image }).imageCommit,
+        });
+    if (pool && member) {
+      const { podSkillsDigest, resolvePodSkills } = await import('./pod-skills.js');
+      const localSkillsDigest = options.kubectl ? null : podSkillsDigest(resolvePodSkills(env).skills).digest;
+      const syncs = await (options.syncImages ?? syncPoolImages)([member], image, imageCommit, { localSkillsDigest });
+      const synced = syncs.get(member.context);
+      if (synced && !synced.ok) throw new Error(`pod command: 이미지 판을 못 맞췄다: ${synced.detail}`);
+      if (synced?.imageRef) member = { ...member, imageRef: synced.imageRef };
+    }
+    const ghToken = options.clone ? (options.ghToken ?? defaultGhToken)() : undefined;
+    const secret = commandJobSecret({ name, namespace, skills, skillEnvText, ...(grokAuth !== undefined ? { grokAuth } : {}), ...(ghToken !== undefined ? { ghToken } : {}) });
+    if (secret) {
+      const applied = kubectl(['apply', '-f', '-'], JSON.stringify(secret));
+      if (applied.status !== 0) throw new Error(applied.stderr.trim() || 'pod command: Secret 적용 실패');
+    }
+    const manifest = podCommandJobManifest({
+      name, namespace, image: member?.imageRef ?? image,
+      ...(member?.imageRef ? { imagePullPolicy: 'IfNotPresent' as const } : {}),
+      repoUrl, command, skills, ...(llm ? { llm } : {}), ...(options.clone ? { clone: true } : {}), deadlineSeconds,
+      ...(options.runId ? { runId: options.runId } : {}),
+    });
+    kubectl(['-n', namespace, 'delete', 'job', name, '--ignore-not-found']);
+    const applied = kubectl(['apply', '-f', '-'], JSON.stringify(manifest));
+    if (applied.status !== 0) {
+      cleanupSecret();
+      throw new Error(applied.stderr.trim() || 'pod command: Job 적용 실패');
+    }
+    let state: 'complete' | 'failed' = 'failed';
+    for (;;) {
+      const got = kubectl(['-n', namespace, 'get', 'job', name, '-o', 'jsonpath={.status.conditions[*].type}']);
+      const finished = jobFinished(got.stdout);
+      if (finished) { state = finished; break; }
+      await sleep(options.pollMs ?? 15_000);
+    }
+    const logs = kubectl(['-n', namespace, 'logs', `job/${name}`, '-c', 'child']).stdout;
+    collectPodArtifacts(logs, {
+      dir: options.artifactsRoot ?? `${effectiveInstanceRoot()}/pod-artifacts`,
+      job: name,
+      log: (c, e, d) => log(c, e, d),
+    });
+    const exitCode = exitCodeFromPod(kubectl, namespace, name, state);
+    cleanupSecret();
+    log('pod.command', 'job-finished', {
+      job: name,
+      exitCode,
+      skills,
+      llm: llm ?? null,
+      command: command.join(' ').slice(0, COMMAND_LOG_CHARS),
+    });
+    return { exitCode, artifactsDir, job: name, image: member?.imageRef ?? image };
+  } catch (err) {
+    cleanupSecret();
+    throw err;
+  } finally {
+    if (member && pool) pool.release(member);
+  }
+}

package/src/task-orchestrator/surfaces/pod-ledger-collect.ts

@@ -0,0 +1,177 @@
+import { appendFileSync, existsSync, mkdirSync, readFileSync, writeFileSync } from 'node:fs';
+import { gunzipSync } from 'node:zlib';
+import { debug } from '../../debug/log.js';
+import { runLedgerDir, runLedgerPath } from '../../self-implement/run-ledger.js';
+
+type Ledger = { runId: string; jsonl: string };
+type Incomplete = { runId: string; reason: string };
+type ParseResult = Ledger[] | { error: Incomplete[]; ledgers: Ledger[] };
+
+const NON_PROGRESS_EVENTS = new Set(['progress-delivery-outcome']);
+
+/** Reassemble only complete, unambiguous transfers; retain good runs when another run is incomplete. */
+export function parsePodLedgerChunks(logs: string): ParseResult {
+  const groups = new Map<string, { total: number; chunks: Map<number, string>; reason?: string }>();
+  for (const line of logs.split(/\r?\n/)) {
+    if (!line.startsWith('ELANOUS_RUN_LEDGER ')) continue;
+    const match = /^ELANOUS_RUN_LEDGER (\S+) (\d+)\/(\d+) ([A-Za-z0-9+/=]+)$/.exec(line);
+    if (!match) {
+      const runId = /^ELANOUS_RUN_LEDGER (\S+)/.exec(line)?.[1];
+      if (runId) {
+        const group = groups.get(runId) ?? { total: 0, chunks: new Map<number, string>() };
+        group.reason = 'malformed chunk';
+        groups.set(runId, group);
+      }
+      continue;
+    }
+    const [, runId, partText, totalText, chunk] = match;
+    const part = Number(partText);
+    const total = Number(totalText);
+    const group = groups.get(runId!) ?? { total, chunks: new Map<number, string>() };
+    groups.set(runId!, group);
+    if (!Number.isSafeInteger(part) || !Number.isSafeInteger(total) || part < 1 || total < 1 || part > total || chunk!.length > 8000 || group.total !== total || group.chunks.has(part)) {
+      group.reason = 'invalid or duplicate chunk';
+    } else {
+      group.chunks.set(part, chunk!);
+    }
+  }
+  const ledgers: Ledger[] = [];
+  const error: Incomplete[] = [];
+  for (const [runId, group] of groups) {
+    let reason = group.reason;
+    if (!reason && group.chunks.size !== group.total) reason = 'missing chunk';
+    if (!reason) {
+      for (let i = 1; i <= group.total; i++) {
+        if (!group.chunks.has(i)) { reason = 'missing chunk'; break; }
+      }
+    }
+    if (!reason) {
+      try {
+        runLedgerPath(runId);
+        let encoded = '';
+        for (let i = 1; i <= group.total; i++) encoded += group.chunks.get(i)!;
+        const compressed = Buffer.from(encoded, 'base64');
+        if (compressed.toString('base64') !== encoded) throw new Error('invalid base64');
+        const raw = gunzipSync(compressed);
+        const jsonl = raw.toString('utf8');
+        if (!Buffer.from(jsonl, 'utf8').equals(raw)) throw new Error('invalid UTF-8');
+        ledgers.push({ runId, jsonl });
+      } catch (e) { reason = e instanceof Error ? e.message : String(e); }
+    }
+    if (reason) error.push({ runId, reason });
+  }
+  return error.length ? { error, ledgers } : ledgers;
+}
+
+/** Never replace a host ledger, even if another collector races this one —
+ *  except a ledger that THIS run's live follower created (`replace`): that one is a partial copy and the final is complete. */
+export function collectPodLedgers(
+  logs: string,
+  { dir = runLedgerDir(), log = (c, e, d) => debug.log(c, e, d), replace = new Set<string>() }: {
+    dir?: string;
+    log?: (category: string, event: string, data: Record<string, unknown>) => void;
+    replace?: ReadonlySet<string>;
+  } = {},
+): void {
+  const parsed = parsePodLedgerChunks(logs);
+  const ledgers = Array.isArray(parsed) ? parsed : parsed.ledgers;
+  for (const { runId, reason } of Array.isArray(parsed) ? [] : parsed.error) {
+    log('self-implement.pod', 'ledger-collect-incomplete', { runId, reason });
+  }
+  for (const { runId, jsonl } of ledgers) {
+    try {
+      mkdirSync(dir, { recursive: true });
+      writeFileSync(runLedgerPath(runId, dir), jsonl, { flag: replace.has(runId) ? 'w' : 'wx' });
+      log('self-implement.pod', 'ledger-collected', { runId, lines: jsonl.split('\n').filter(Boolean).length, bytes: Buffer.byteLength(jsonl) });
+    } catch (e) {
+      if ((e as NodeJS.ErrnoException).code === 'EEXIST') {
+        // A reattached Job may have already delivered this exact complete snapshot.
+        // An unrelated host ledger (or a partial copy) must still remain untouched.
+        try {
+          const existing = readFileSync(runLedgerPath(runId, dir));
+          log('self-implement.pod', existing.equals(Buffer.from(jsonl)) ? 'ledger-collect-already-complete' : 'ledger-collect-skipped', { runId, reason: 'exists' });
+        } catch (readError) {
+          log('self-implement.pod', 'ledger-collect-incomplete', { runId, reason: readError instanceof Error ? readError.message : String(readError) });
+        }
+      } else log('self-implement.pod', 'ledger-collect-incomplete', { runId, reason: e instanceof Error ? e.message : String(e) });
+    }
+  }
+}
+
+/** ⭐ 런 «도중» 원장 증분 회수(🅣 요청 2026-09-26 · 힐 루프의 runtime 성형이 Pod 판에서 성립하는 전제).
+ *  호스트 폴링(15초)마다 Pod 안 `run-ledger/<runId>.jsonl` 의 «새 바이트»만 가져와 «완성된 줄»만 호스트 원장에 잇는다.
+ *  ⛔ 원장 줄만 — 값·비밀·화면은 안 가져온다. ⛔ 호스트에 «남이 만든» 같은 원장이 있으면 손대지 않는다(종료 회수의 wx 규칙과 같은 뜻).
+ *  ⭐ 이 추종기가 만든 파일은 종료 때 완본으로 교체된다(`owned` → collectPodLedgers 의 replace). */
+export function createPodLedgerFollower(opts: {
+  runId: string;
+  /** Pod 안에서 셸 한 줄을 실행해 stdout 을 돌려준다(kubectl exec). */
+  exec: (script: string) => { status: number | null; stdout: string; stderr: string };
+  dir?: string;
+  log?: (category: string, event: string, data: Record<string, unknown>) => void;
+  stallMinutes?: number;
+  now?: () => number;
+  onStall?: (message: string) => void;
+}): { poll(): void; readonly owned: boolean } {
+  const dir = opts.dir ?? runLedgerDir();
+  const log = opts.log ?? ((c, e, d) => debug.log(c, e, d));
+  const now = opts.now ?? Date.now;
+  const stallMinutes = opts.stallMinutes ?? 30;
+  if (!Number.isFinite(stallMinutes) || stallMinutes <= 0) throw new RangeError('stallMinutes must be positive');
+  const onStall = opts.onStall ?? ((message: string) => console.error(message));
+  const path = runLedgerPath(opts.runId, dir);   // runId 검증(경로·셸에 안전한 문자만)도 여기서 된다
+  let offset = 0;
+  let pending = '';
+  let owned = false;
+  let disabled = false;
+  let lastProgressAt = now();
+  let lastProgressEvent: string | null = null;
+  let nextStallMinute = stallMinutes;
+  let stalled = false;
+  function checkStall() {
+    const idleMinutes = Math.max(0, (now() - lastProgressAt) / 60_000);
+    if (idleMinutes >= nextStallMinute) {
+      log('self-implement.pod', 'stalled', { runId: opts.runId, lastProgressEvent, idleMinutes });
+      onStall(`[pod] 진행 없음 ${Math.floor(idleMinutes)}분 — 마지막 진행 ${lastProgressEvent ?? '없음'}`);
+      stalled = true;
+      nextStallMinute = (Math.floor(idleMinutes / stallMinutes) + 1) * stallMinutes;
+    }
+  }
+  return {
+    get owned() { return owned; },
+    poll() {
+      if (disabled) return;
+      if (!owned && existsSync(path)) {
+        disabled = true;
+        log('self-implement.pod', 'ledger-live-skipped', { runId: opts.runId, reason: 'host-ledger-exists' });
+        return;
+      }
+      const r = opts.exec(`f="\${ELANOUS_STATE_DIR:-$HOME/.elanous}/run-ledger/${opts.runId}.jsonl"; [ -f "$f" ] || exit 0; tail -c +${offset + 1} "$f"`);
+      if (r.status !== 0) { log('self-implement.pod', 'ledger-live-unavailable', { runId: opts.runId, status: r.status, stderr: r.stderr.trim().slice(0, 200) }); checkStall(); return; }
+      if (!r.stdout) { checkStall(); return; }
+      offset += Buffer.byteLength(r.stdout);
+      const text = pending + r.stdout;
+      const cut = text.lastIndexOf('\n');
+      if (cut < 0) { pending = text; checkStall(); return; }
+      const complete = text.slice(0, cut + 1);
+      pending = text.slice(cut + 1);
+      mkdirSync(dir, { recursive: true });
+      appendFileSync(path, complete);
+      owned = true;
+      log('self-implement.pod', 'ledger-live-appended', { runId: opts.runId, lines: complete.split('\n').filter(Boolean).length, bytes: Buffer.byteLength(complete), offset });
+      for (const line of complete.split('\n')) {
+        if (!line) continue;
+        let entry: unknown;
+        try { entry = JSON.parse(line); } catch { continue; }
+        const event = (entry && typeof entry === 'object' && 'event' in entry) ? entry.event : undefined;
+        if (typeof event !== 'string' || !event || NON_PROGRESS_EVENTS.has(event)) continue;
+        const idleMinutes = Math.max(0, (now() - lastProgressAt) / 60_000);
+        if (stalled) log('self-implement.pod', 'stall-cleared', { runId: opts.runId, idleMinutes });
+        stalled = false;
+        lastProgressAt = now();
+        lastProgressEvent = event;
+        nextStallMinute = stallMinutes;
+      }
+      checkStall();
+    },
+  };
+}