elanous 0.2.23 → 0.2.25

package/catalog/models/openrouter/glm-5.3-flash.yaml

@@ -0,0 +1,35 @@
+id: openrouter/z-ai/glm-5.3-flash
+provider: openrouter
+displayName: GLM 5.3 Flash (OpenRouter)
+family: glm-5
+# ⭐ 2026-10-11 MK(CATALOG-DEEPSEEK-FLASH) — 손 레코드(근거는 deepseek-v4-pro-0813.yaml 머리말과 같다).
+#   ⚠️ 이 id(`openrouter/z-ai/…`)는 사다리의 `glm` provider id(`glm-5.3-flash`)와 «다른 id» 다 — 그 구멍(catalog-drift 시험의 glm 줄)은 닫지 않는다.
+# 📏 수치 출처 — OpenRouter 공개 API 실호출 2026-10-10T19:30Z (인증 없음):
+#   ⑴ GET /api/v1/models/z-ai/glm-5.3-flash/endpoints — «Z.AI» 1st-party 공급자 엔드포인트(glm-5.3-flash-20260826 · fp8):
+#      context_length 1048576 · max_completion_tokens 131072 · prompt $0.15 · completion $0.50 · input_cache_read $0.03 /MTok · discount 0
+#   ⑵ GET /api/v1/models — context_length 1048576 · top_provider.max_completion_tokens 943717 · 같은 단가
+#      ⚠️ ⑵ 의 출력은 «다른 공급자» 값이다. deepseek 팔과 같은 원칙(제조사 자체 엔드포인트)으로 ⑴ 을 쓴다.
+#      공급자를 고정하지 않으면 출력 상한은 공급자마다 다르다(131072 ~ 943718).
+# 💲 판촉 할인(discount) 0 · 시간대 가산 없음.
+contextSize: 1048576
+outputMaxTokens: 131072
+# 입력 모달리티 text+image+video (architecture.input_modalities) — 칸이 하나라 images 로 적는다.
+vision: images
+# 기본 effort — OR supported_parameters 에 reasoning_effort 가 있다. 관례(flash 계열 = medium)를 따른다.
+reasoning: medium
+toolCalling: native-openai
+streamingProtocol: sse
+pricing:
+  inputPerMTok: 0.15
+  outputPerMTok: 0.50
+  cachedInputPerMTok: 0.03
+deprecated: null
+tokenizer: unknown
+kind: chat
+# 출시일 — 엔드포인트 이름의 날짜(glm-5.3-flash-20260826) · /api/v1/models `created` 도 2026-08-26(UTC).
+releaseDate: 2026-08-26
+discoveryMeta:
+  source: manual
+  lastSeen: 2026-10-11
+  autoFilled: false
+  confidence: high

package/catalog/providers/openrouter.yaml

@@ -2,6 +2,8 @@
 # ⛔ 모델 YAML 은 «손으로 두지 않는다» — `catalogFromDiscovery: true` 라 발견 스냅숏
 #    (`openrouter` 소스 · /api/v1/models)에서 `openrouter/<vendor>/<model>` 로 접힌다.
 #    근거·결정 = 내부 문서 `RESEARCH-kimi-qwen-glm-and-openrouter-provider-implementation-2026-09-23`
+#  ⚠️ 2026-10-11 예외(CATALOG-DEEPSEEK-FLASH): 창·출력·단가·기본 effort 를 «공급자 엔드포인트»로 실측해 못 박아야 하는
+#    모델만 `catalog/models/openrouter/` 에 손 레코드를 둔다 — YAML 이 폴드를 이긴다(loader.ts foldDiscoveredModels).
 id: openrouter
 displayName: OpenRouter
 aliases: []

package/graphs/implement-loop.yaml

@@ -32,6 +32,9 @@
 #   ⛔⭐ 그리고 ***「거른 뒤의 가장 이른 행」을 「원장의 바닥」으로 읽지 마라*** — 그 둘은 다른 값이다.
 #      바닥을 재려면 그 «필터를 빼고» 같은 이벤트의 가장 이른 행을 본다.
 
+# L0 여정 단계 목록(journey_stage 닫힌 집합) — 노드별 선언 커버리지의 분모.
+# 입구 · 저작 · 배치 · 실행 · 착지 · 실패 갈래 · 종료 (RFC ARG §L0).
+journey_stages: [entry, author, dispatch, run, land, recover, end]
 # ⛔⭐ `graph_id` 는 «원장의 조인 키»다 — 바꾸면 옛 표본과 «비교 불가»가 된다.
 #   📏 2026-09-08 실측: 원장에 graphId="self-implement" 로 15건이 이미 쌓였다.
 #   📌 RFC §4.2 는 이 템플릿을 `implement-loop` 이라 불렀지만, ***코드·원장이 쓰는 이름이 canonical***이다.
@@ -72,6 +75,7 @@ runner_inputs: [ask_text, goal, worktree, branch, review-intent, head-sha, base-
 nodes:
   - node_id: author
     group: author
+    journey_stage: author
     kind: agent
     recipe: goal-author
     max_visits: 2
@@ -84,6 +88,7 @@ nodes:
 
   - node_id: plan
     group: author
+    journey_stage: author
     kind: agent
     recipe: harness-plan-rfc
     max_visits: 2
@@ -96,6 +101,7 @@ nodes:
 
   - node_id: decompose
     group: author
+    journey_stage: author
     kind: judge
     recipe: fabric-decompose-adapter
     max_visits: 12
@@ -108,6 +114,7 @@ nodes:
 
   - node_id: implement
     group: build
+    journey_stage: run
     kind: agent
     recipe: headless-goal-loop        # seams.implement → 자식 PTY
     max_visits: 12
@@ -122,6 +129,7 @@ nodes:
 
   - node_id: gate
     group: verify
+    journey_stage: run
     kind: gate
     recipe: gate-baseline              # 변경 파일 «범위» · 새 회귀만
     max_visits: 11
@@ -136,6 +144,7 @@ nodes:
 
   - node_id: review
     group: verify
+    journey_stage: run
     kind: judge
     recipe: pr-reviewer                # agent-substrate/pr-reviewer
     max_visits: 11
@@ -150,6 +159,7 @@ nodes:
 
   - node_id: rework
     group: verify
+    journey_stage: run
     kind: agent
     recipe: headless-goal-loop         # 같은 루프의 다음 라운드
     max_visits: 10
@@ -164,6 +174,7 @@ nodes:
 
   - node_id: main-sync
     group: land
+    journey_stage: land
     kind: git
     recipe: seams-git-merge-main       # LLM 충돌해결 포함
     max_visits: 2
@@ -176,6 +187,7 @@ nodes:
 
   - node_id: regate
     group: land
+    journey_stage: land
     kind: gate
     recipe: gate-baseline              # 충돌해결 «뒤» 재-gate
     max_visits: 2
@@ -188,6 +200,7 @@ nodes:
 
   - node_id: open-pr
     group: land
+    journey_stage: land
     kind: git
     recipe: seams-gh-open-pr
     # 📏 실측 2026-09-08 (원장 30일 · 전 우주 · 걸음 191개 소급 계산):
@@ -204,6 +217,7 @@ nodes:
 
   - node_id: merge
     group: land
+    journey_stage: land
     kind: git
     recipe: seams-gh-merge-squash
     # 📏 실측 2026-09-08 (원장 30일 · 전 우주 · 걸음 191개 소급 계산):
@@ -222,6 +236,7 @@ nodes:
   # ───────── 여정 앞단: 입구 → 대기열 → 발사 관문 → 배치 ─────────
   - node_id: intake
     group: entry
+    journey_stage: entry
     kind: observe
     recipe: harness-entrance            # src/harness/harness-cli-command.ts (say · ask · TUI /harness)
     max_visits: 1
@@ -233,6 +248,7 @@ nodes:
 
   - node_id: queue
     group: entry
+    journey_stage: entry
     kind: gate
     recipe: harness-queue               # src/harness/harness-queue.ts (directSay 플래그 · 기본 끔 → 건너뜀)
     max_visits: 1
@@ -244,6 +260,7 @@ nodes:
 
   - node_id: launch-gate
     group: entry
+    journey_stage: entry
     kind: gate
     recipe: pre-launch-gate             # harness-cli-command.ts preLaunchGate (예산 · 같은 골 활성 런)
     max_visits: 1
@@ -255,6 +272,7 @@ nodes:
 
   - node_id: dispatch
     group: dispatch
+    journey_stage: dispatch
     kind: judge
     recipe: harness-substrate-select    # resolveLaunchSubstrate → substrate = local | pod
     max_visits: 2
@@ -267,6 +285,7 @@ nodes:
   # ───────── Pod 배치 (substrate == pod 일 때만 — local 은 건너뛴다) ─────────
   - node_id: pool-admit
     group: dispatch-pod
+    journey_stage: dispatch
     kind: gate
     recipe: pod-pool-admission          # src/task-orchestrator/surfaces/self-implement-pod.ts (입장 · 멤버 슬롯 · 메모리 등급)
     max_visits: 4
@@ -278,6 +297,7 @@ nodes:
 
   - node_id: image-sync
     group: dispatch-pod
+    journey_stage: dispatch
     kind: gate
     recipe: pod-pool-image-sync         # src/task-orchestrator/surfaces/pod-pool.ts syncPoolImages (멤버별 · pack 한 번)
     max_visits: 4
@@ -289,6 +309,7 @@ nodes:
 
   - node_id: credential-refresh
     group: dispatch-pod
+    journey_stage: dispatch
     kind: gate
     recipe: pod-credential-prerefresh   # self-implement-pod.ts podCodexCredentialWithPrerefresh
     max_visits: 4
@@ -301,6 +322,7 @@ nodes:
   # ───────── 착지: 동결 검사 → (보류) → 병합 → 회수 ─────────
   - node_id: freeze-check
     group: land
+    journey_stage: land
     kind: gate
     recipe: landing-freeze-admission    # src/self-implement/frozen-merges.ts admitLandingMerge (병합 직전 한 곳)
     max_visits: 4
@@ -312,6 +334,7 @@ nodes:
 
   - node_id: hold
     group: land
+    journey_stage: land
     kind: hitl
     recipe: frozen-merge-queue          # frozen-merges.ts queueFrozenMerge → sweepFrozenMerges 가 해제 뒤 재개
     max_visits: 4
@@ -323,6 +346,7 @@ nodes:
 
   - node_id: cleanup
     group: land
+    journey_stage: land
     kind: git
     recipe: post-merge-cleanup          # src/self-implement/orchestrator.ts (워크트리 ⊕ 브랜치 회수)
     max_visits: 1
@@ -335,6 +359,7 @@ nodes:
   # ───────── 실패 쪽: 결과 해석 → 수확 → heal → 다음 한 수 ─────────
   - node_id: result-parse
     group: recover
+    journey_stage: recover
     kind: observe
     recipe: pod-failure-reason          # src/task-orchestrator/surfaces/pod-failure-reason.ts
     max_visits: 4
@@ -346,6 +371,7 @@ nodes:
 
   - node_id: salvage
     group: recover
+    journey_stage: recover
     kind: git
     recipe: salvage-branch              # salvage/si-task-<job>/… (Pod Job 이 push · salvage-branches.ts 가 읽는다)
     max_visits: 4
@@ -355,8 +381,11 @@ nodes:
       tools: git-write
       outputs: [salvage_ref]
 
+  # 🩹 멈춤 갈래(stop_class)별 처방 노드는 «여기 없다» — 런타임 오버레이 `graphs/overlays/stop-heal-ladder.yaml` 이
+  #   `heal` 에서 `on: stop_class` 간선으로 갈라지는 노드(id = 처방 이름)를 얹는다. 읽는 쪽 = `stop-autoheal.ts`.
   - node_id: heal
     group: recover
+    journey_stage: recover
     kind: judge
     recipe: heal-intake                 # src/self-implement/heal-intake.ts (호스트에서만)
     max_visits: 4
@@ -368,6 +397,7 @@ nodes:
 
   - node_id: next-move
     group: recover
+    journey_stage: recover
     kind: judge
     recipe: task-agent-judge            # src/task-agent/judge.ts (판단부 → 실행부)
     max_visits: 4
@@ -379,6 +409,7 @@ nodes:
 
   - node_id: stopped
     group: end
+    journey_stage: end
     kind: observe
     recipe: journey-stopped             # 거절·포기 — 종결
     max_visits: 1

package/graphs/ops/harness-gc.yaml

@@ -0,0 +1,17 @@
+graph_id: harness-gc
+loop:
+  title: Harness GC
+  description: Reclaim only ladder-safe harness worktrees and merged branch-only refs in bounded ticks
+  owner: OP
+  trigger:
+    cron: '*/10 * * * *'
+    events: [manual]
+version: 1
+entry_node: tick
+terminal_nodes: [done, failed]
+nodes:
+  - { node_id: tick, kind: agent, recipe: 'cmd:harness-gc-tick', max_visits: 1 }
+  - { node_id: done, kind: gate, max_visits: 1 }
+  - { node_id: failed, kind: gate, max_visits: 1 }
+edges:
+  - { from: tick, on: outcome, map: { ok: done, skipped: done, fail: failed, error: failed } }

package/graphs/ops/recipes.yaml

@@ -12,3 +12,6 @@ secret-exposure-scan:
 guardian-shadow:
   command: 'bun "$ELANOUS_GRAPH_DIR/../../bin/elanous.mjs" guardian tick'
   timeout_ms: 120000
+harness-gc-tick:
+  command: 'bun "$ELANOUS_GRAPH_DIR/../../src/harness/harness-gc.ts" --json'
+  timeout_ms: 300000

package/graphs/overlays/README.md

@@ -18,7 +18,7 @@
 ```yaml
 overlay_id: patient-research
 target: default-loop
-stage: launch            # launch | runtime
+stage: launch            # launch | runtime | heal(멈춤 갈래 사다리 전용 · stop-autoheal 만 읽는다)
 patch:
   - op: replace
     path: /nodes/2/max_visits

package/graphs/overlays/stop-heal-ladder.yaml

@@ -0,0 +1,105 @@
+# 멈춤 갈래 사다리(STOP-AUTOHEAL)를 «선언»으로 — 0.2.24 GRAPH-HEAL-NODES 첫 조각.
+#
+# 한 문서(`---` 로 나눈다) = 멈춤 갈래 하나 = 처방 노드 하나. 조건은 실행기가 이미 주는 `stop_class` 다.
+#   stop_class == env-unrelated      → clean-workspace-regate
+#   stop_class == main-sync          → rebase-regate
+#   stop_class == pod-died           → resume-from-salvage
+#   stop_class == evidence-uncovered → derive-evidence-rejudge
+# ⛔ 이 넷 밖의 갈래(review-repeat · review-out-of-scope · fabric · launch-failed · unclassified)는 선언이 «없다» —
+#   처방은 `needs-owner` 이고 그것은 코드 기본(`AUTOHEAL_ACTION_BY_CLASS`)이 말한다.
+#
+# 🔑 노드 id 가 곧 처방 이름이다. `src/self-implement/stop-autoheal.ts` 는 그 갈래에 얹힌 오버레이가 «더한 노드 id» 로
+#   처방을 고르고, 선언이 없거나 못 읽으면 코드 표로 간다(원장·관측의 `source` = graph | code-default).
+#   ⇒ 이 파일과 코드 표가 «같은 답»을 내는지는 시험이 문다(stop-autoheal.test.ts).
+# ⚠️ 이 조각은 표를 선언으로 «옮기기만» 한다 — 노드의 recipe 는 주입 seam 이름(`stopAutohealActions`)이다.
+#   새 처방 노드(예: strategy-switch)는 다음 조각이다.
+#
+# 🚨 경로·값은 «파싱된 스펙» 키(camelCase)다 — `nodeId`·`maxVisits`.
+# ⛔ stage 는 `heal` 이다 — `runtime` 이면 실행기가 라운드마다 원장의 stop 줄로 `stop_class` 를 채워 이 노드들을
+#   «실행 그래프»에 얹는다(버전 해시·원장 값이 바뀐다). `heal` 은 `stop-autoheal.ts` 만 고른다.
+# 📌 한 번만 시도한다(원장 가드와 같은 규율) — 그래서 maxVisits 1 이다.
+# 📌 들어오는 간선 = `heal` 이 stop_class 로 가른다 · 나가는 간선 = 처방 결과(healed → 재발사 · needs-owner → 종결).
+---
+overlay_id: stop-heal-ladder-env-unrelated
+target: self-implement
+stage: heal
+applies_when: stop_class == env-unrelated
+patch:
+  - op: add
+    path: /nodes/-
+    value:
+      nodeId: clean-workspace-regate
+      kind: gate
+      recipe: stop-autoheal:clean-workspace-regate
+      maxVisits: 1
+      journeyStage: recover
+      contract: { inputs: [worktree], tools: read-and-run-tests, outputs: [passed, gate-log] }
+  - op: add
+    path: /edges/-
+    value: { from: heal, on: stop_class, map: { env-unrelated: clean-workspace-regate } }
+  - op: add
+    path: /edges/-
+    value: { from: clean-workspace-regate, on: outcome, map: { healed: dispatch, needs-owner: stopped } }
+---
+overlay_id: stop-heal-ladder-main-sync
+target: self-implement
+stage: heal
+applies_when: stop_class == main-sync
+patch:
+  - op: add
+    path: /nodes/-
+    value:
+      nodeId: rebase-regate
+      kind: git
+      recipe: stop-autoheal:rebase-regate
+      maxVisits: 1
+      journeyStage: recover
+      contract: { inputs: [worktree, base-branch], tools: git-write, outputs: [synced, passed] }
+  - op: add
+    path: /edges/-
+    value: { from: heal, on: stop_class, map: { main-sync: rebase-regate } }
+  - op: add
+    path: /edges/-
+    value: { from: rebase-regate, on: outcome, map: { healed: dispatch, needs-owner: stopped } }
+---
+overlay_id: stop-heal-ladder-pod-died
+target: self-implement
+stage: heal
+applies_when: stop_class == pod-died
+patch:
+  - op: add
+    path: /nodes/-
+    value:
+      nodeId: resume-from-salvage
+      kind: git
+      recipe: stop-autoheal:resume-from-salvage
+      maxVisits: 1
+      journeyStage: recover
+      contract: { inputs: [salvage_ref], tools: git-write, outputs: [resumed] }
+  - op: add
+    path: /edges/-
+    value: { from: heal, on: stop_class, map: { pod-died: resume-from-salvage } }
+  - op: add
+    path: /edges/-
+    value: { from: resume-from-salvage, on: outcome, map: { healed: dispatch, needs-owner: stopped } }
+---
+overlay_id: stop-heal-ladder-evidence-uncovered
+target: self-implement
+stage: heal
+applies_when: stop_class == evidence-uncovered
+patch:
+  - op: add
+    path: /nodes/-
+    value:
+      nodeId: derive-evidence-rejudge
+      kind: judge
+      recipe: stop-autoheal:derive-evidence-rejudge
+      maxVisits: 1
+      journeyStage: recover
+      contract: { inputs: [diff, review-summary], tools: read-only, outputs: [verdict] }
+  - op: add
+    path: /edges/-
+    value: { from: heal, on: stop_class, map: { evidence-uncovered: derive-evidence-rejudge } }
+  - op: add
+    path: /edges/-
+    value: { from: derive-evidence-rejudge, on: outcome, map: { healed: dispatch, needs-owner: stopped } }

package/graphs/release/release-loop.yaml

@@ -16,11 +16,11 @@ nodes:
   - { node_id: checklist-gate, kind: gate, recipe: 'cmd:checklist-gate', max_visits: 1 }
   # cmd:docs dispatches by graph node ID: bounded prefetch before gate, TUI join after upgrade, docs write after both.
   - { node_id: prefetch, kind: agent, recipe: 'cmd:docs', max_visits: 1 }
-  - { node_id: gate, kind: gate, recipe: 'cmd:gate', max_visits: 1, notify: start }
+  - { node_id: gate, kind: gate, recipe: 'cmd:gate', max_visits: 1, notify: { on: [fail] } }
   - { node_id: mac-smoke, kind: gate, recipe: 'cmd:mac-smoke', max_visits: 1 }
   - { node_id: export-check, kind: gate, recipe: 'cmd:export-check', max_visits: 1 }
-  - { node_id: pwa, kind: gate, recipe: 'cmd:pwa', max_visits: 1, notify: start }
-  - { node_id: prepare, kind: agent, recipe: 'cmd:prepare', max_visits: 1, notify: start }
+  - { node_id: pwa, kind: gate, recipe: 'cmd:pwa', max_visits: 1, notify: { on: [fail] } }
+  - { node_id: prepare, kind: agent, recipe: 'cmd:prepare', max_visits: 1, notify: { on: [fail] } }
   - { node_id: upgrade, kind: gate, recipe: 'cmd:upgrade', max_visits: 1 }
   - { node_id: tui, kind: gate, recipe: 'cmd:docs', max_visits: 1 }
   - { node_id: docs, kind: agent, recipe: 'cmd:docs', max_visits: 1 }
@@ -28,10 +28,10 @@ nodes:
   - { node_id: notes-check, kind: gate, recipe: 'cmd:notes-check', max_visits: 1 }
   - { node_id: auto-approve, kind: gate, recipe: 'cmd:auto-approve', max_visits: 1 }
   - { node_id: approve-publish, kind: hitl, recipe: 'approval:publish', max_visits: 1 }
-  - { node_id: publish, kind: git, recipe: 'cmd:publish-command', max_visits: 1 }
+  - { node_id: publish, kind: git, recipe: 'cmd:publish-command', max_visits: 1, notify: { on: [ok, fail] } }
   - { node_id: npm-publish, kind: git, recipe: 'cmd:npm-publish', max_visits: 1 }
   - { node_id: docs-land, kind: git, recipe: 'cmd:docs-land', max_visits: 1 }
-  - { node_id: verify, kind: gate, recipe: 'cmd:verify', max_visits: 1 }
+  - { node_id: verify, kind: gate, recipe: 'cmd:verify', max_visits: 1, notify: { on: [ok, fail] } }
   - { node_id: vault-note, kind: gate, recipe: 'cmd:vault-note', max_visits: 1 }
   - { node_id: release-story, kind: gate, recipe: 'cmd:release-story', max_visits: 1 }
   - { node_id: ops-upgrade, kind: agent, recipe: 'cmd:ops-upgrade', max_visits: 1 }

package/graphs/steward/recipes.yaml

@@ -1,3 +1,7 @@
+steward-rejudge:
+  command: 'bun "$ELANOUS_GRAPH_DIR/../../src/steward/rejudge-tick.ts"'
+  # The node kills its own `tasks rejudge` child at 600 s; this leaves room for the log line.
+  timeout_ms: 660000
 steward-sync:
   command: 'bun "$ELANOUS_GRAPH_DIR/../../src/steward/triage.ts" sync'
   timeout_ms: 120000

package/graphs/steward/steward.yaml

@@ -5,9 +5,15 @@ loop:
     cron: '*/15 8-23 * * *'
     events: [directive, linear-webhook]
 version: 1
-entry_node: sync
+entry_node: rejudge
 terminal_nodes: [done, failed]
 nodes:
+  # TA-REJUDGE-PERIODIC — `tasks rejudge --all` on a cadence (off unless taskAgent.rejudgeEveryMinutes > 0).
+  # First so a Linear outage in sync cannot starve it; either outcome continues to sync.
+  - node_id: rejudge
+    kind: agent
+    recipe: cmd:steward-rejudge
+    max_visits: 1
   - node_id: sync
     kind: agent
     recipe: cmd:steward-sync
@@ -31,6 +37,9 @@ nodes:
     kind: gate
     max_visits: 1
 edges:
+  - from: rejudge
+    on: outcome
+    map: { ok: sync, fail: sync }
   - from: sync
     on: outcome
     map: { ok: triage, fail: failed }

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "elanous",
   "mcpName": "io.github.ElanvitalAI/elanous",
-  "version": "0.2.23",
+  "version": "0.2.25",
   "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/add-api-key.sh

@@ -121,7 +121,7 @@ REFRESH="$(dirname "$LOADER")/refresh_api_key_cache.sh"
 if [ -f "$REFRESH" ] && ! grep -q "\"$ENV_NAME\"" "$REFRESH"; then
   echo ""
   echo "   🔲 ⓷ 다른 기계에도 퍼뜨리려면 — ⛔ «자동으로 안 넣었다»"
-  echo "      먼저 AWS Secrets Manager 에 비밀 «$ENV_NAME» 을 만들고, 그 «뒤»에:"
+  echo "      먼저 AWS Secrets Manager 에 비밀 «${ENV_NAME}» 을 만들고, 그 «뒤»에:"
   echo "        fetch_secret \"$ENV_NAME\" \"\${CACHE_DIR}/$FILE_NAME\""
   echo "      ⚠️ 비밀이 «없는» 상태로 넣으면 그 스크립트가 set -e 라 다른 키 갱신까지 멈춘다."
 fi

package/scripts/bench-report.ts

@@ -13,6 +13,7 @@
 import { spawnSync } from 'node:child_process';
 import { readFileSync } from 'node:fs';
 import { dirname, join } from 'node:path';
+import { dedupePodRollupRows } from './usage-rollup.js';
 
 export interface LogRow { category?: string; event?: string; data?: Record<string, unknown> | null }
 export interface ArmDecl { id: string; provider: string; model: string | null }
@@ -40,6 +41,40 @@ export interface ArmReport {
   reviewMustFixCount: number | null;
 }
 
+export interface ProbeJudgment {
+  model: string;
+  runs: number;
+  passed: number;
+  failed: number;
+  unmeasured: number;
+  verdict: 'pass' | 'fail' | 'unmeasured';
+}
+
+export function judgeProbe(
+  arms: readonly ArmReport[],
+  opts?: { passNumerator?: number; passDenominator?: number },
+): ProbeJudgment[] {
+  const passNumerator = opts?.passNumerator ?? 8;
+  const passDenominator = opts?.passDenominator ?? 9;
+  const groups = new Map<string, { runs: number; passed: number; failed: number; unmeasured: number }>();
+  for (const arm of arms) {
+    const model = arm.declared?.model ?? '못 이음';
+    const group = groups.get(model) ?? { runs: 0, passed: 0, failed: 0, unmeasured: 0 };
+    group.runs++;
+    if (arm.gatePassed === null || arm.leakCalls === null) group.unmeasured++;
+    else if (arm.gatePassed === true && arm.leakCalls === 0) group.passed++;
+    else group.failed++;
+    groups.set(model, group);
+  }
+  return [...groups].map(([model, counts]) => ({
+    model,
+    ...counts,
+    verdict: counts.unmeasured > 0
+      ? 'unmeasured'
+      : counts.passed * passDenominator >= counts.runs * passNumerator ? 'pass' : 'fail',
+  }));
+}
+
 const num = (v: unknown): number => (typeof v === 'number' && Number.isFinite(v) ? v : 0);
 const str = (v: unknown): string | null => (typeof v === 'string' && v.length > 0 ? v : null);
 const armKey = (id: string): string => (id.startsWith('pod/') ? id : `pod/${id}`);
@@ -99,7 +134,8 @@ export function buildBenchReport(rows: readonly LogRow[], runId: string): { runI
     a.reviewMustFixCount = typeof r.data!.reviewMustFixCount === 'number' ? r.data!.reviewMustFixCount : null;
   }
   const leak = new Map<string, number>();
-  for (const r of mine) {
+  // ⭐ METER-POD-ROLLUP-GAP-2 — 같은 Pod 의 재방출(부분 ⊕ 최종 · abort ⊕ resume)은 podRollupKey 로 한 번만 센다(팔 사용량 이중계상 방지).
+  for (const r of dedupePodRollupRows(mine, (row) => (row.event === 'llm-usage' ? row.data ?? {} : {}))) {
     if (r.event !== 'llm-usage') continue;
     const site = str(r.data!.site);
     const id = str(r.data!.armId);
@@ -203,6 +239,15 @@ export function renderBenchReport(rep: ReturnType<typeof buildBenchReport>): str
   lines.push(leaked.length
     ? `⛔ 팔 밖 모델로 간 호출이 있다(${leaked.join(', ')}) — 이 판으로 팔을 비교하지 않는다.`
     : '⛔ 한 판(팔당 N=1)이다 — 결론이 아니라 관측이다. 비교는 팔마다 표본이 쌓인 뒤에.');
+  const probe = judgeProbe(rep.arms);
+  lines.push(
+    '',
+    '## 프로브 판정',
+    '',
+    '| 모델 | 통과/런 | 판정 |',
+    '|---|---:|---|',
+    ...probe.map((row) => `| ${row.model} | ${row.passed}/${row.runs} | ${row.verdict} |`),
+  );
   return lines.join('\n');
 }