elanous 0.2.20 → 0.2.21

package/graphs/implement-loop.yaml

@@ -38,8 +38,12 @@
 #      (파일 이름은 RFC 어휘를 남겨 둔다 — 사람이 찾는 이름과 조인 키가 다를 수 있다.)
 graph_id: self-implement
 version: 1
-entry_node: author
-terminal_nodes: [merge]
+# ⭐ HARNESS-FULL-GRAPH (0.2.21 · 대표 10-08) — 입구부터 회수까지 «여정 전체»가 노드다.
+#   📄 내부 문서 `RFC-harness-full-journey-as-graph-2026-10-08` · 코드 = src/self-dev/graph-journey-nodes.ts
+#   ⛔ 여정 노드(intake … stopped)의 id 는 원장·라이브 그래프 화면의 조인 키다 — 바꾸지 마라.
+#   🔀 분기는 «데이터»다(`on:` 상태 키 ⊕ `map:` 값→노드) — 기본 = 풀세트, 환경(substrate·freeze·queue)이 일부를 건너뛴다.
+entry_node: intake
+terminal_nodes: [cleanup, stopped]
 
 # Runner-provided inputs are declared here so contract-flow can distinguish them from undeclared gaps.
 runner_inputs: [ask_text, goal, worktree, branch, review-intent, head-sha, base-branch]
@@ -61,8 +65,13 @@ runner_inputs: [ask_text, goal, worktree, branch, review-intent, head-sha, base-
 #   🔑 ⇒ ***여기 적은 도구 권한은 「선언」이고 「제한」이 아니다.*** 자식의 실제 권한은 코드가 정한다.
 #   ⛔ 그러니 이 칸을 좁혀 「자식이 그 도구를 못 쓴다」고 읽지 마라 — 지금은 안 그렇다.
 #   🔎 반증: rg -n "readonly contract" src/self-implement/graph-templates.ts   (⇒ 0건이면 여전히 그렇다)
+# 🧩 `group:` = 나중에 `graphs/units/*.yaml` 로 들어 올릴 «한 노드로 부르는 부분 그래프» 단위(대표 10-08).
+#   entry · dispatch · author · dispatch-pod · build · verify · land · recover · end
+#   ⛔ 지금 파서는 이 칸을 «읽지 않는다»(무시) — 단위 호출은 0.2.22(RFC §6). 묶음은 «위치»가 아니라 이 태그로 정한다
+#      (오버레이가 /nodes/<번호> 로 가리켜 노드를 재배열할 수 없다).
 nodes:
   - node_id: author
+    group: author
     kind: agent
     recipe: goal-author
     max_visits: 2
@@ -74,6 +83,7 @@ nodes:
       outputs: [goal_path, goal_id, goal_type, grounding_context]
 
   - node_id: plan
+    group: author
     kind: agent
     recipe: harness-plan-rfc
     max_visits: 2
@@ -85,6 +95,7 @@ nodes:
       outputs: [rfc_path, markdown, arcs, open_questions]
 
   - node_id: decompose
+    group: author
     kind: judge
     recipe: fabric-decompose-adapter
     max_visits: 12
@@ -96,6 +107,7 @@ nodes:
       outputs: [pieces, ordering, depends_on_edges, hot_path_edges, omitted_goal_count]
 
   - node_id: implement
+    group: build
     kind: agent
     recipe: headless-goal-loop        # seams.implement → 자식 PTY
     max_visits: 12
@@ -109,6 +121,7 @@ nodes:
       outputs: [diff, summary, changed-files]
 
   - node_id: gate
+    group: verify
     kind: gate
     recipe: gate-baseline              # 변경 파일 «범위» · 새 회귀만
     max_visits: 11
@@ -122,6 +135,7 @@ nodes:
       outputs: [passed, gate-log, reflect-gate-facts]
 
   - node_id: review
+    group: verify
     kind: judge
     recipe: pr-reviewer                # agent-substrate/pr-reviewer
     max_visits: 11
@@ -135,6 +149,7 @@ nodes:
       outputs: [verdict, must-fix, should-fix, review-summary]
 
   - node_id: rework
+    group: verify
     kind: agent
     recipe: headless-goal-loop         # 같은 루프의 다음 라운드
     max_visits: 10
@@ -148,6 +163,7 @@ nodes:
       outputs: [diff, summary]
 
   - node_id: main-sync
+    group: land
     kind: git
     recipe: seams-git-merge-main       # LLM 충돌해결 포함
     max_visits: 2
@@ -159,6 +175,7 @@ nodes:
       outputs: [synced, conflicts]
 
   - node_id: regate
+    group: land
     kind: gate
     recipe: gate-baseline              # 충돌해결 «뒤» 재-gate
     max_visits: 2
@@ -170,6 +187,7 @@ nodes:
       outputs: [passed, gate-log]
 
   - node_id: open-pr
+    group: land
     kind: git
     recipe: seams-gh-open-pr
     # 📏 실측 2026-09-08 (원장 30일 · 전 우주 · 걸음 191개 소급 계산):
@@ -185,6 +203,7 @@ nodes:
       outputs: [pr-number, pr-url]
 
   - node_id: merge
+    group: land
     kind: git
     recipe: seams-gh-merge-squash
     # 📏 실측 2026-09-08 (원장 30일 · 전 우주 · 걸음 191개 소급 계산):
@@ -199,6 +218,176 @@ nodes:
       tools: gh-pr-merge
       outputs: [merged]
 
+  # ⛔ 여정 노드는 «목록 끝»에 둔다 — 오버레이가 JSON 포인터(/nodes/<번호>)로 노드를 가리킨다(rework-patient). 앞에 끼우면 번호가 밀린다.
+  # ───────── 여정 앞단: 입구 → 대기열 → 발사 관문 → 배치 ─────────
+  - node_id: intake
+    group: entry
+    kind: observe
+    recipe: harness-entrance            # src/harness/harness-cli-command.ts (say · ask · TUI /harness)
+    max_visits: 1
+    terminal_stages: []
+    contract:
+      inputs: [ask_text]
+      tools: none
+      outputs: [entrance, journey_key]
+
+  - node_id: queue
+    group: entry
+    kind: gate
+    recipe: harness-queue               # src/harness/harness-queue.ts (directSay 플래그 · 기본 끔 → 건너뜀)
+    max_visits: 1
+    terminal_stages: []
+    contract:
+      inputs: [entrance, seat]
+      tools: queue-write
+      outputs: [queue_item, queue_outcome]
+
+  - node_id: launch-gate
+    group: entry
+    kind: gate
+    recipe: pre-launch-gate             # harness-cli-command.ts preLaunchGate (예산 · 같은 골 활성 런)
+    max_visits: 1
+    terminal_stages: []
+    contract:
+      inputs: [goal_id, budget]
+      tools: read-only
+      outputs: [launch_decision]
+
+  - node_id: dispatch
+    group: dispatch
+    kind: judge
+    recipe: harness-substrate-select    # resolveLaunchSubstrate → substrate = local | pod
+    max_visits: 2
+    terminal_stages: []
+    contract:
+      inputs: [launch_decision]
+      tools: read-only
+      outputs: [substrate]
+
+  # ───────── Pod 배치 (substrate == pod 일 때만 — local 은 건너뛴다) ─────────
+  - node_id: pool-admit
+    group: dispatch-pod
+    kind: gate
+    recipe: pod-pool-admission          # src/task-orchestrator/surfaces/self-implement-pod.ts (입장 · 멤버 슬롯 · 메모리 등급)
+    max_visits: 4
+    terminal_stages: []
+    contract:
+      inputs: [substrate, memory_tier]
+      tools: pod-lease
+      outputs: [pool_member]
+
+  - node_id: image-sync
+    group: dispatch-pod
+    kind: gate
+    recipe: pod-pool-image-sync         # src/task-orchestrator/surfaces/pod-pool.ts syncPoolImages (멤버별 · pack 한 번)
+    max_visits: 4
+    terminal_stages: []
+    contract:
+      inputs: [pool_member]
+      tools: image-ship
+      outputs: [image_current]
+
+  - node_id: credential-refresh
+    group: dispatch-pod
+    kind: gate
+    recipe: pod-credential-prerefresh   # self-implement-pod.ts podCodexCredentialWithPrerefresh
+    max_visits: 4
+    terminal_stages: []
+    contract:
+      inputs: [pool_member]
+      tools: credential-refresh
+      outputs: [usable_accounts]
+
+  # ───────── 착지: 동결 검사 → (보류) → 병합 → 회수 ─────────
+  - node_id: freeze-check
+    group: land
+    kind: gate
+    recipe: landing-freeze-admission    # src/self-implement/frozen-merges.ts admitLandingMerge (병합 직전 한 곳)
+    max_visits: 4
+    terminal_stages: []
+    contract:
+      inputs: [pr-number]
+      tools: read-only
+      outputs: [freeze]
+
+  - node_id: hold
+    group: land
+    kind: hitl
+    recipe: frozen-merge-queue          # frozen-merges.ts queueFrozenMerge → sweepFrozenMerges 가 해제 뒤 재개
+    max_visits: 4
+    terminal_stages: []
+    contract:
+      inputs: [pr-number, head-sha]
+      tools: queue-write
+      outputs: [held]
+
+  - node_id: cleanup
+    group: land
+    kind: git
+    recipe: post-merge-cleanup          # src/self-implement/orchestrator.ts (워크트리 ⊕ 브랜치 회수)
+    max_visits: 1
+    terminal_stages: []
+    contract:
+      inputs: [worktree, branch]
+      tools: git-write
+      outputs: [cleaned]
+
+  # ───────── 실패 쪽: 결과 해석 → 수확 → heal → 다음 한 수 ─────────
+  - node_id: result-parse
+    group: recover
+    kind: observe
+    recipe: pod-failure-reason          # src/task-orchestrator/surfaces/pod-failure-reason.ts
+    max_visits: 4
+    terminal_stages: []
+    contract:
+      inputs: [pod_log]
+      tools: read-only
+      outputs: [failure_reason]
+
+  - node_id: salvage
+    group: recover
+    kind: git
+    recipe: salvage-branch              # salvage/si-task-<job>/… (Pod Job 이 push · salvage-branches.ts 가 읽는다)
+    max_visits: 4
+    terminal_stages: []
+    contract:
+      inputs: [worktree]
+      tools: git-write
+      outputs: [salvage_ref]
+
+  - node_id: heal
+    group: recover
+    kind: judge
+    recipe: heal-intake                 # src/self-implement/heal-intake.ts (호스트에서만)
+    max_visits: 4
+    terminal_stages: []
+    contract:
+      inputs: [failure_reason, salvage_ref]
+      tools: read-only
+      outputs: [heal_signature]
+
+  - node_id: next-move
+    group: recover
+    kind: judge
+    recipe: task-agent-judge            # src/task-agent/judge.ts (판단부 → 실행부)
+    max_visits: 4
+    terminal_stages: []
+    contract:
+      inputs: [heal_signature]
+      tools: read-only
+      outputs: [next_action]
+
+  - node_id: stopped
+    group: end
+    kind: observe
+    recipe: journey-stopped             # 거절·포기 — 종결
+    max_visits: 1
+    terminal_stages: []
+    contract:
+      inputs: []
+      tools: none
+      outputs: []
+
 # ═══════════════════════════════════════════════════════════════════════
 # 엣지 — ⛔ 출처는 «코드 선언 ⊕ 원장 실측» 둘 다다. 어느 하나만 보면 틀린다.
 #   📏 2026-09-08 실측 (pipeline-traversal-shadow 182 걸음 · 판정 함수가 «실제로 본» 것):
@@ -207,14 +396,83 @@ nodes:
 #         gate→main-sync · gate→open-pr · review→open-pr
 #   🔑 그래서 엣지마다 `observed` 를 «값으로» 적는다 — 「선언했다」와 「돈다」를 가른다.
 edges:
+  # ───── 여정 앞단 — 분기는 «데이터»(on/map). 값은 코드가 낸다(RFC §4.3 ⑵) ─────
+  - from: intake
+    on: queue
+    map:
+      queued: queue                    # harness.queue.directSay 켬 ⊕ 자리 있음
+      direct: launch-gate              # 기본(플래그 끔) — 대기열을 건너뛴다
+
+  - from: queue
+    on: queue_outcome
+    map:
+      launched: launch-gate            # 대기열이 발사한 자식이 다시 intake 를 밟는다(queueLaunchId 로 잇는다)
+      waiting: stopped                 # 이 프로세스는 여기서 끝난다(발사는 대기열 tick 의 몫)
+      refused: stopped                 # 같은 요청이 이미 대기열에 있다
+
+  - from: launch-gate
+    on: outcome
+    map:
+      pass: dispatch
+      refuse: stopped
+
+  - from: dispatch
+    to: author
+
   - from: author
     to: plan
 
   - from: plan
     to: decompose
 
+  # ⭐ 실행 칸(substrate)이 «Pod 쪽 세 노드»를 켜고 끈다 — local 은 곧장 구현.
   - from: decompose
-    to: implement
+    on: substrate
+    map:
+      local: implement
+      pod: pool-admit
+
+  - from: pool-admit
+    on: outcome
+    map:
+      admitted: image-sync
+      refused: stopped
+
+  - from: image-sync
+    to: credential-refresh
+
+  - from: credential-refresh
+    on: outcome
+    map:
+      usable: implement
+      exhausted: stopped
+
+  # 실패 쪽 — Pod 구현이 rc≠0 이면 결과 해석으로.
+  - from: implement
+    on: pod_result
+    map:
+      failed: result-parse
+
+  - from: result-parse
+    to: salvage
+
+  # ⚠️ heal 은 실패 사건 «접수»다 — 호스트에서만(Pod 안에서는 시작 안 함 · POD-RESULT-PARSE).
+  - from: salvage
+    to: heal
+
+  - from: heal
+    to: next-move
+
+  # 출구 라벨 = task-agent 판단부의 수(TaskMove · src/task-agent/judge.ts). 관측 배선은 0.2.22(판단이 카드 단위라 런 조인 먼저).
+  - from: next-move
+    on: move
+    map:
+      narrow-relaunch: dispatch
+      salvage-relaunch: dispatch
+      alternative: dispatch
+      wait-retry: dispatch
+      decision-card: stopped
+      wait: stopped
 
   - from: implement
     to: gate
@@ -264,6 +522,20 @@ edges:
     to: open-pr
     observed: 80
 
+  # ⭐ HARNESS-FULL-GRAPH — 코드 간선 open-pr→merge 사이에 «동결 검사»가 늘 끼어 있었다(병합 직전 한 곳).
   - from: open-pr
-    to: merge
-    observed: 80
+    to: freeze-check
+    observed: 80                       # 옛 open-pr→merge 관측 — 그 사이에 동결 검사가 늘 있었다
+
+  - from: freeze-check
+    on: freeze
+    map:
+      open: merge
+      frozen: hold
+      taken: stopped                   # 다른 착지자(재개 sweep)가 그 PR 을 쥐었다
+
+  - from: hold
+    to: freeze-check                   # 해제 뒤 재검사
+
+  - from: merge
+    to: cleanup

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "elanous",
   "mcpName": "io.github.ElanvitalAI/elanous",
-  "version": "0.2.20",
+  "version": "0.2.21",
   "description": "Self-healing coding harness where eyes, hands, and memory all turn on one sentence — it authors its own goals, implements them in isolated worktrees, and reviews its own work.",
   "license": "Apache-2.0",
   "author": "Elanvital AI",

package/scripts/coord-channel-watch.sh

@@ -240,15 +240,42 @@ lsof_fd1() {
   return 0
 }
 
+# 🐧 lsof 가 없는 리눅스(게이트 Pod 이미지 · 0.2.20 컷 실측)에서는 /proc/<pid>/fd/1 을 읽어
+#   lsof -Ftn 과 «같은 모양»(t<type>\nn<name>)으로 낸다 — 아래 판정 표를 그대로 쓴다.
+#   ⛔ 못 읽으면(죽은 pid·권한) 1 을 돌려 「모름」으로 끝낸다.
+proc_fd1() {
+  local pid="$1" target
+  target=$(readlink "/proc/$pid/fd/1" 2>/dev/null) || return 1
+  [ -n "$target" ] || return 1
+  case "$target" in
+    pipe:*)   printf 'tPIPE\nn%s\n' "$target" ;;
+    socket:*) printf 'tunix\nn%s\n' "$target" ;;
+    *)
+      if [ -c "$target" ]; then printf 'tCHR\nn%s\n' "$target"
+      elif [ -f "$target" ]; then printf 'tREG\nn%s\n' "$target"
+      else printf 'tunknown\nn%s\n' "$target"
+      fi ;;
+  esac
+  return 0
+}
+
 delivery_of() {
   local pid="$1" out t n
   case "$pid" in
     ''|*[!0-9]*) printf '⚠️ «못 쟀다»(감시자 pid 를 «증명하지 못했다») — 「정상」이 아니라 «모름»이다'; return 0 ;;
   esac
-  command -v lsof >/dev/null 2>&1 || { printf '⚠️ «못 쟀다»(lsof 없음) — 「정상」이 아니라 «모름»이다'; return 0; }
-  if ! out=$(lsof_fd1 "$pid"); then
-    printf '⚠️ «못 쟀다»(lsof 가 %s틱 안에 안 끝나 죽였다) — 「정상」이 아니라 «모름»이다' "$LSOF_WAIT_TICKS"
-    return 0
+  if command -v lsof >/dev/null 2>&1; then
+    if ! out=$(lsof_fd1 "$pid"); then
+      printf '⚠️ «못 쟀다»(lsof 가 %s틱 안에 안 끝나 죽였다) — 「정상」이 아니라 «모름»이다' "$LSOF_WAIT_TICKS"
+      return 0
+    fi
+  elif command -v readlink >/dev/null 2>&1 && [ -d /proc/self/fd ]; then
+    if ! out=$(proc_fd1 "$pid"); then
+      printf '⚠️ «못 쟀다»(/proc/%s/fd/1 을 못 읽었다 · lsof 없음) — 「정상」이 아니라 «모름»이다' "$pid"
+      return 0
+    fi
+  else
+    printf '⚠️ «못 쟀다»(lsof 없음) — 「정상」이 아니라 «모름»이다'; return 0
   fi
   t=$(printf '%s\n' "$out" | sed -n 's/^t//p' | head -1)
   n=$(printf '%s\n' "$out" | sed -n 's/^n//p' | head -1)

package/scripts/docs-lint.ts

@@ -13,6 +13,7 @@ import {
 } from '../src/autopilot/discovery/doc-lint.js';
 import { parseIndexLinks } from '../src/autopilot/discovery/doc-inventory.js';
 import { execFileSync } from 'node:child_process';
+import { runGitCommand } from '../src/git-fs/runner.js';
 import { existsSync, readFileSync, readdirSync, statSync } from 'node:fs';
 import { basename, join, relative } from 'node:path';
 import { sourceIdentifierInventory, inlineCodeIdentifiers } from '../src/autopilot/discovery/source-identifiers.js';
@@ -80,7 +81,9 @@ const readText = (relPath: string): string | null => { try { return readFileSync
 //   전부 「못 읽음」이 되어 이 규칙이 통째로 잡음이 된다. 그래서 파일명 → repo 경로를 «한 번» 만든다.
 //   ⛔ 같은 파일명이 둘 이상이면 «해석하지 않는다** — 틀린 파일을 대면 「어긋남」이 거짓으로 난다.
 const sourcePathsByBasename = new Map<string, string | null>();
-for (const relPath of execFileSync('git', ['ls-files', '*.ts', '*.tsx'], { cwd: repoRoot, encoding: 'utf-8' }).split('\n')) {
+const lsFiles = runGitCommand(repoRoot, ['ls-files', '*.ts', '*.tsx'], { encoding: 'utf-8', maxBuffer: 64 * 1024 * 1024 });
+if (lsFiles.status !== 0) throw new Error(`git ls-files failed: ${lsFiles.stderr.trim() || `rc=${lsFiles.status}`}`);
+for (const relPath of lsFiles.stdout.split('\n')) {
   if (relPath === '') continue;
   const base = basename(relPath);
   sourcePathsByBasename.set(base, sourcePathsByBasename.has(base) ? null : relPath);

package/scripts/marketing/generate-ax-brief.ts

@@ -1,10 +1,11 @@
 import { readFileSync, writeFileSync } from 'node:fs';
-import { resolve } from 'node:path';
+import { basename, resolve } from 'node:path';
 import { FEATURE_MATURITY } from '../../src/maturity/feature-maturity.js';
 import type { ChecklistItem } from '../../src/release-loop/checklist.js';
 import type { ReleaseSchedule } from '../../src/release-loop/release-schedule.js';
 import type { MergedRunLedgerQuery } from '../../src/self-implement/run-ledger.js';
 import { generateFeatureMap } from './generate-feature-map.js';
+import { draftReleaseStory } from '../release-story/draft.js';
 
 const MAP = resolve(import.meta.dir, '../../docs/marketing/MAP-value-props-to-features.md');
 const AUTONOMY = new Set(['ORCH1', 'ORCH2', 'LOOP-LIVE1', 'AUTOQ-LOOP', 'AUTHOR-PAR', 'RUN-HELPER', 'FINISH-RATE']);
@@ -110,6 +111,22 @@ export function generateAxBrief(version: string, sources: AxBriefSources): strin
   ].join('\n');
 }
 
+/** Derive the next release's internal brief from its ledgers and the RELEASE-STORY version-note draft.
+ * The release loop supplies its version, ledger sources and isolated story paths; the standalone CLI stays unchanged. */
+export function generateNextVersionAxBrief(
+  nextVersion: string,
+  sources: AxBriefSources,
+  storyOptions: Omit<Parameters<typeof draftReleaseStory>[0], 'version'> = {},
+): string {
+  const brief = generateAxBrief(nextVersion, sources);
+  const story = draftReleaseStory({ ...storyOptions, version: nextVersion });
+  const announcementFile = story.files.find((file) => basename(file) === 'announcement.md');
+  const announcement = story.status === 'drafted' && announcementFile
+    ? readFileSync(announcementFile, 'utf8').replace(/^# [^\r\n]*\r?\n/, '').trim()
+    : '사용자 대상 변경 없음 — RELEASE-STORY 초안을 건너뜀';
+  return `${brief}\n## 5. 다음 판 변화·근거 (RELEASE-STORY 초안)\n\n> 공개 전 CMO 확인 필요 · 원천: release/next.md, 해당 판의 green 칸, 연결된 검증 소구점 원장\n\n${announcement}\n`;
+}
+
 if (import.meta.main) {
   const args = process.argv.slice(2);
   const option = (name: string) => {

package/scripts/measure-nl-routing-corpus.ts

@@ -3,7 +3,7 @@ import { readFileSync, writeFileSync } from 'node:fs';
 import { resolve } from 'node:path';
 import { buildEvalPromptToolSurface, EVAL_PROMPT_TOOL_SURFACES, isEvalPromptToolSurface, resolveReproModelId, runEvalPrompt, type EvalPromptToolSurface } from '../src/eval-prompt-cli.js';
 import { getModelFamily } from '../src/models/prompts.js';
-import { resolveDefaultProvider } from '../src/llm.js';
+import { NoLlmProviderAvailableError, resolveDefaultProvider } from '../src/llm.js';
 import { filterToolsByDeny } from '../src/tool-runtime/tool-deny.js';
 import { getUserConfig } from '../src/user-config.js';
 import { resolveObserveOnlyDecision, type ObserveOnlyDecision } from '../src/self-implement/observe-only.js';
@@ -80,8 +80,16 @@ function formatMachineLoad(load: MachineLoad): string {
 
 function authoritativeSurfaceToolNames(surface: EvalPromptToolSurface): string[] {
   const cfg = getUserConfig();
-  const provider = resolveDefaultProvider();
-  const modelFamily = getModelFamily(resolveReproModelId(undefined, provider.defaultModel));
+  // ⛔ 안전 관문(노출된 변이 툴)은 «provider 해석보다 먼저» 닫혀야 한다 — 자격이 없는 기계(게이트 Pod)에서
+  //   provider 해석이 먼저 던지면 관문이 안내 대신 「No LLM provider」로 끝났다(0.2.20 컷 pod-7·pod-22).
+  //   provider 가 없으면 러너의 마지막 기본 모델로 서피스를 재서 관문을 그대로 판정한다(런은 어차피 못 돈다).
+  let providerDefaultModel: string | undefined;
+  try {
+    providerDefaultModel = resolveDefaultProvider().defaultModel;
+  } catch (error) {
+    if (!(error instanceof NoLlmProviderAvailableError)) throw error;
+  }
+  const modelFamily = getModelFamily(resolveReproModelId(undefined, providerDefaultModel));
   const { specs } = buildEvalPromptToolSurface(surface, modelFamily, cfg);
   const tools = filterToolsByDeny(specs, cfg.chat.toolDeny) ?? specs;
   return tools.map((tool) => tool.name);

package/scripts/release-loop/gate-node.ts

@@ -20,6 +20,7 @@ import { runPodCommand, parsePodCpu, resolvePodCpu, type RunPodCommandOptions, t
 import { POD_BUN_CACHE_HOST_PATH, parseInstallSeconds, podBunCacheVolume } from '../../src/task-orchestrator/surfaces/pod-bun-cache.js';
 import { PodPoolScheduler, parsePodPool, checkPodPool } from '../../src/task-orchestrator/surfaces/pod-pool.js';
 import { POD_INSTALL_SLOTS_DEFAULT, POD_INSTALL_SLOTS_HOST_PATH, installSlotScript } from '../../src/task-orchestrator/surfaces/pod-install-slots.js';
+import { gateDepsInstallScript, parseGateDeps } from '../../src/task-orchestrator/surfaces/pod-gate-deps.js';
 import { gateShardsPath, writeGateShards, type GateShard } from '../../src/release-loop/gate-shards.js';
 import { defaultKubectl } from '../../src/task-orchestrator/surfaces/self-implement-pod.js';
 
@@ -253,7 +254,7 @@ export function shardParts(paths: readonly string[], seconds: (file: string) =>
  * leaving `install.log`/`install.rc`, `part-<i>.log`/`.rc`/`.junit.xml`, and `shard.timeout` (the part index) when the
  * deadline cut the shard. A part without `.rc` did not finish. The shell itself always exits 0 (as before).
  */
-export function gateShardShell(o: { parts: readonly (readonly string[])[]; cdpPatterns: readonly string[]; softSeconds: number; cachePrefix: string; installSlots: number }): string {
+export function gateShardShell(o: { parts: readonly (readonly string[])[]; cdpPatterns: readonly string[]; softSeconds: number; cachePrefix: string; installSlots: number; depsDir?: string }): string {
   const run = o.parts.map((part, index) => {
     const ignores = o.cdpPatterns.filter((pattern) => part.includes(pattern)).flatMap((pattern) => ['--path-ignore-patterns', pattern]);
     return `run_part ${index} ${['bun', 'run', 'test:deterministic', ...ignores, ...part.map(asPath)].map(quote).join(' ')}`;
@@ -276,9 +277,8 @@ export function gateShardShell(o: { parts: readonly (readonly string[])[]; cdpPa
     '  echo "$prc" > "$O/part-$i.rc"',
     '}',
     'if cd .. && cd repo; then',
-    ...(o.installSlots > 0 ? ['  install_slot_acquire >> "$O/install.log" 2>&1'] : []),
-    '  { bun install && (cd apps/pwa && bun install); } >> "$O/install.log" 2>&1; irc=$?',
-    ...(o.installSlots > 0 ? ['  install_slot_release'] : []),
+    // GATE-IMAGE-DEPS: link the image's baked node_modules when the lockfiles match; only a miss installs (behind a slot).
+    ...gateDepsInstallScript({ log: '"$O/install.log"', installSlots: o.installSlots, ...(o.depsDir ? { depsDir: o.depsDir } : {}) }).split('\n').map((line) => `  ${line}`),
     '  echo "$irc" > "$O/install.rc"',
     `  if [ "$irc" -eq 0 ]; then ${run || 'true'}; fi`,
     'else echo 5 > "$O/install.rc"; fi',
@@ -800,7 +800,7 @@ export function createGateRunner(repo: string, remote?: string, commandOverride?
       }
       const installSeconds = parseInstallSeconds(output ?? '');
       // GATE-TIMEOUT-HEAL ③: a cut shard is «reason: timeout» (with the files that still finished), not a bare rc=null.
-      if (!reused && !descent) debug.log('release-loop.gate', 'pod-shard', { shard, files: paths, durationMs, rc: rc ?? null, attempt: depth + 1, installSeconds, deadlineSeconds: shardDeadline, parts: parts.length,
+      if (!reused && !descent) debug.log('release-loop.gate', 'pod-shard', { shard, files: paths, durationMs, rc: rc ?? null, attempt: depth + 1, installSeconds, gateDeps: parseGateDeps(output ?? ''), deadlineSeconds: shardDeadline, parts: parts.length,
         ...(timedOut ? { reason: 'timeout', finishedFiles: finished.length } : {}) });
       if (!reused && !descent) markShard(shard, { ...(installSeconds !== null ? { installSec: installSeconds } : {}), ...(rc !== undefined ? { rc } : {}), ...(timedOut ? { state: 'timeout' as const } : {}) });
       let clean = output?.replace(/\x1b\[[0-9;]*[A-Za-z]/g, '');

package/scripts/release-loop/shard-plan.ts

@@ -1,15 +1,60 @@
-import { readdirSync, readFileSync } from 'node:fs';
-import { join } from 'node:path';
+import { readdirSync, readFileSync, statSync } from 'node:fs';
+import { basename, dirname, join, resolve } from 'node:path';
 import { debug } from '../../src/debug/log.js';
 
 /** Plan weights for a Pod sweep: wall-scaled when the previous cut recorded shard wall times, junit seconds otherwise. */
-export function planDurations(durationSource: string | undefined): { durations: Map<string, number>; wall?: { shards: number; scaledFiles: number } } {
+export function planDurations(durationSource: string | undefined): { durations: Map<string, number>; wall?: { shards: number; scaledFiles: number }; fallback?: string } {
   if (!durationSource) return { durations: new Map() };
-  const junit = readFileDurations(durationSource);
-  const wall = readFileWallWeights(durationSource, junit);
+  const primary = planDurationsFrom(durationSource);
+  if (primary.durations.size) return primary;
+  // GATE-PLAN-TIMINGS: the baseline cut left no reports (0.2.20: baseline 0.2.19 was gated from another universe, so
+  // all 5,585 files were «unknown», planned 3.9 min, measured 13–20 min). Use the most recent sibling cut instead.
+  const fallback = latestSiblingCutDir(durationSource);
+  if (!fallback) {
+    debug.log('release-loop.gate', 'plan-duration-fallback', { source: durationSource, fallback: null });
+    return primary;
+  }
+  const result = planDurationsFrom(fallback);
+  debug.log('release-loop.gate', 'plan-duration-fallback', { source: durationSource, fallback, files: result.durations.size, wall: !!result.wall });
+  return result.durations.size ? { ...result, fallback } : primary;
+}
+
+function planDurationsFrom(dir: string): { durations: Map<string, number>; wall?: { shards: number; scaledFiles: number } } {
+  const junit = readFileDurations(dir);
+  const wall = readFileWallWeights(dir, junit);
   return wall ? { durations: wall.weights, wall: { shards: wall.shards, scaledFiles: wall.scaledFiles } } : { durations: junit };
 }
 
+/**
+ * For a `<root>/<version>/<a>/<b>` duration source (the gate passes `<ledger>/release/<baseline>/gate-logs/cut`), the
+ * sibling `<root>/<other>/<a>/<b>` directory whose newest `*.junit.xml` is the most recent — «the latest gate». The
+ * source itself is skipped. Undefined when the path is too shallow or no sibling holds a junit report.
+ */
+export function latestSiblingCutDir(durationSource: string): string | undefined {
+  const source = resolve(durationSource);
+  const b = basename(source);
+  const a = basename(dirname(source));
+  const root = dirname(dirname(dirname(source)));
+  if (!a || !b || root === dirname(dirname(source))) return undefined;
+  let versions: string[];
+  try { versions = readdirSync(root, { withFileTypes: true }).filter((entry) => entry.isDirectory()).map((entry) => entry.name); }
+  catch { return undefined; }
+  let best: { dir: string; mtimeMs: number } | undefined;
+  for (const version of versions.sort()) {
+    const dir = join(root, version, a, b);
+    if (dir === source) continue;
+    let newest = -1;
+    try {
+      for (const name of readdirSync(dir)) {
+        if (!name.endsWith('.junit.xml')) continue;
+        try { newest = Math.max(newest, statSync(join(dir, name)).mtimeMs); } catch { /* vanished mid-scan */ }
+      }
+    } catch { continue; }
+    if (newest >= 0 && (!best || newest > best.mtimeMs)) best = { dir, mtimeMs: newest };
+  }
+  return best?.dir;
+}
+
 /** Observation for GATE-PLAN-WALLTIME, emitted once the shards are planned so `maxPlannedMin` is the heaviest shard. */
 export function logPlanWalltime(wall: { shards: number; scaledFiles: number } | undefined, shards: ReadonlyArray<{ plannedSeconds: number }>): void {
   if (!wall) return;

package/scripts/rhythm/daily.ts

@@ -587,8 +587,9 @@ export async function main(args: string[] = process.argv.slice(2), deps: DailyDe
   for (const arg of args) if (!['--dry-run', '--no-news', '--json'].includes(arg)) throw new Error(`알 수 없는 옵션: ${arg}`);
   const result = await runDaily({ dryRun: args.includes('--dry-run'), noNews: args.includes('--no-news'), stage }, deps);
   (deps.print ?? console.log)(args.includes('--json') ? JSON.stringify(result) : `${result.markdown}\n파일: ${result.file} · ${result.deliveryLine}`);
-  // 발송 설정 OFF 는 degraded 로 기록하고 done 으로 진행한다. 실제 발송 실패·볼트 사본 실패는 그래프를 멈춘다.
+  // 그래프 deliver 는 미발송도 출력·tick 에 기록한 뒤 실패로 종결한다.
   if ((stage === 'compose' || !stage) && result.vaultFatal) throw new Error(`볼트 사본 실패 · ${result.vaultError}`);
+  if (stage === 'deliver' && result.deliveryState === 'send-disabled') throw new Error('발송 실패 · 발송 비활성');
   if (stage === 'deliver' && result.sendError) throw new Error(`발송 실패 · ${result.sendError}`);
   return result;
 }