mixdog 0.9.3 → 0.9.4

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "mixdog",
-  "version": "0.9.3",
+  "version": "0.9.4",
   "private": false,
   "type": "module",
   "description": "Standalone mixdog coding-agent CLI/TUI workspace.",
@@ -33,7 +33,7 @@
     "prepack": "npm run build:tui",
     "start": "node src/cli.mjs",
     "smoke": "node scripts/smoke.mjs",
-    "smoke:all": "npm run smoke && npm run smoke:boot && npm run smoke:tools && npm run smoke:output",
+    "smoke:all": "npm run smoke && npm run smoke:boot && npm run smoke:tools && npm run smoke:output && npm run smoke:tui && npm run smoke:freevars",
     "smoke:boot": "node scripts/boot-smoke.mjs",
     "smoke:compact": "node scripts/compact-smoke.mjs",
     "smoke:loop": "node scripts/smoke-loop.mjs",
@@ -41,6 +41,8 @@
     "smoke:loop:report": "node scripts/smoke-loop-report.mjs",
     "smoke:tools": "node scripts/tool-smoke.mjs",
     "smoke:output": "node scripts/output-style-smoke.mjs",
+    "smoke:tui": "node scripts/tui-render-smoke.mjs",
+    "smoke:freevars": "node scripts/freevar-smoke.mjs",
     "test:toolcall": "node --test scripts/toolcall-args-test.mjs",
     "test:providers": "node --test scripts/provider-toolcall-test.mjs",
     "failures": "node scripts/tool-failures.mjs",
@@ -92,6 +94,8 @@
     }
   },
   "devDependencies": {
-    "esbuild": "^0.28.1"
+    "acorn": "^8.17.0",
+    "esbuild": "^0.28.1",
+    "eslint-scope": "^9.1.2"
   }
 }

package/scripts/bench/lead-review-tasks-r3.json

@@ -0,0 +1,20 @@
+[
+  {
+    "id": "lead-review-session-runtime",
+    "agent": "reviewer",
+    "cwd": "C:/Project/mixdog-bench",
+    "prompt": "READ-ONLY review task: you and any delegated agents must NOT modify, create, or delete any files - report findings only. Review the folder src/session-runtime/ for correctness, regression, and security risks. Delegate exploration/reading to your workflow agents where useful and synthesize their findings. Focus on real actionable bugs: broken imports, undefined identifiers, race conditions, wrong fallbacks, data loss paths. Report findings ordered by severity, one line each, with file:line anchors. Skip style nits. End with a one-line verdict."
+  },
+  {
+    "id": "lead-review-channels-lib",
+    "agent": "reviewer",
+    "cwd": "C:/Project/mixdog-bench",
+    "prompt": "READ-ONLY review task: you and any delegated agents must NOT modify, create, or delete any files - report findings only. Review the folder src/runtime/channels/lib/ for correctness, regression, and security risks. Delegate exploration/reading to your workflow agents where useful and synthesize their findings. Focus on real actionable bugs: message loss, double-forwarding, ownership races, watcher leaks, unsafe input handling. Report findings ordered by severity, one line each, with file:line anchors. Skip style nits. End with a one-line verdict."
+  },
+  {
+    "id": "lead-review-compact",
+    "agent": "reviewer",
+    "cwd": "C:/Project/mixdog-bench",
+    "prompt": "READ-ONLY review task: you and any delegated agents must NOT modify, create, or delete any files - report findings only. Review the folder src/runtime/agent/orchestrator/session/compact/ for correctness, regression, and security risks. Delegate exploration/reading to your workflow agents where useful and synthesize their findings. Focus on real actionable bugs: schema validation gaps, data corruption on summarize, lossy round-trips, error swallowing. Report findings ordered by severity, one line each, with file:line anchors. Skip style nits. End with a one-line verdict."
+  }
+]

package/scripts/bench/lead-review-tasks.json

@@ -0,0 +1,20 @@
+[
+  {
+    "id": "lead-review-session-runtime",
+    "agent": "reviewer",
+    "cwd": "C:/Project/mixdog-bench",
+    "prompt": "Review the folder src/session-runtime/ for correctness, regression, and security risks. Delegate exploration/reading to your workflow agents where useful and synthesize their findings. Focus on real actionable bugs: broken imports, undefined identifiers, race conditions, wrong fallbacks, data loss paths. Report findings ordered by severity, one line each, with file:line anchors. Skip style nits. End with a one-line verdict."
+  },
+  {
+    "id": "lead-review-channels-lib",
+    "agent": "reviewer",
+    "cwd": "C:/Project/mixdog-bench",
+    "prompt": "Review the folder src/runtime/channels/lib/ for correctness, regression, and security risks. Delegate exploration/reading to your workflow agents where useful and synthesize their findings. Focus on real actionable bugs: message loss, double-forwarding, ownership races, watcher leaks, unsafe input handling. Report findings ordered by severity, one line each, with file:line anchors. Skip style nits. End with a one-line verdict."
+  },
+  {
+    "id": "lead-review-compact",
+    "agent": "reviewer",
+    "cwd": "C:/Project/mixdog-bench",
+    "prompt": "Review the folder src/runtime/agent/orchestrator/session/compact/ for correctness, regression, and security risks. Delegate exploration/reading to your workflow agents where useful and synthesize their findings. Focus on real actionable bugs: schema validation gaps, data corruption on summarize, lossy round-trips, error swallowing. Report findings ordered by severity, one line each, with file:line anchors. Skip style nits. End with a one-line verdict."
+  }
+]

package/scripts/bench/r4-mixed-tasks.json

@@ -0,0 +1,20 @@
+[
+  {
+    "id": "r4-debug-forwarding-gap",
+    "agent": "reviewer",
+    "cwd": "C:/Project/mixdog-bench",
+    "prompt": "Investigate a real reliability question in this repo: when the Discord channel bridge activates (bridge-state.json written) there can be a window where assistant output is not yet forwarded to the channel - transcript binding, fs.watch on a not-yet-created transcript file, and ownership probes are involved. Trace the full path from bridge activation to first forwarded message in src/runtime/channels/ (index.mjs, lib/output-forwarder.mjs), identify every gap where output produced during the window would be silently dropped or delayed, and state for each gap whether backlog catch-up recovers it. Report findings ordered by impact with file:line anchors. End with a one-line verdict."
+  },
+  {
+    "id": "r4-impl-usage-summary",
+    "agent": "worker",
+    "cwd": "C:/Project/mixdog-bench",
+    "prompt": "Implement a small feature: add a `usage:today` npm script backed by a new scripts/usage-today.mjs that reads ~/.mixdog/data/history/agent-trace.jsonl (respect MIXDOG_DATA_DIR), filters usage rows from local midnight, and prints per-model calls, input/cached/output tokens, and costUsd totals plus a grand total line. Follow the style of scripts/session-bench.mjs arg parsing. Verify with node --check and run it once. Report the diff summary."
+  },
+  {
+    "id": "r4-explain-compact-engine",
+    "agent": "explore",
+    "cwd": "C:/Project/mixdog-bench",
+    "prompt": "Map the compaction subsystem: starting from src/runtime/agent/orchestrator/session/compact/engine.mjs, produce a concise architecture brief - entry points and callers, the decision flow for when compaction fires, how the preserved tail is chosen, where redaction hooks in, and the failure/fallback paths. Anchor every claim with file:line. Keep it under 40 lines."
+  }
+]

package/scripts/bench/review-tasks.json

@@ -0,0 +1,20 @@
+[
+  {
+    "id": "review-session-runtime",
+    "agent": "reviewer",
+    "cwd": "C:/Project/mixdog",
+    "prompt": "Review the folder src/session-runtime/ for correctness, regression, and security risks. Focus on real actionable bugs: broken imports, undefined identifiers, race conditions, wrong fallbacks, data loss paths. Report findings ordered by severity, one line each, with file:line anchors. Skip style nits. End with a one-line verdict."
+  },
+  {
+    "id": "review-channels-lib",
+    "agent": "reviewer",
+    "cwd": "C:/Project/mixdog",
+    "prompt": "Review the folder src/runtime/channels/lib/ for correctness, regression, and security risks. Focus on real actionable bugs: message loss, double-forwarding, ownership races, watcher leaks, unsafe input handling. Report findings ordered by severity, one line each, with file:line anchors. Skip style nits. End with a one-line verdict."
+  },
+  {
+    "id": "review-compact",
+    "agent": "reviewer",
+    "cwd": "C:/Project/mixdog",
+    "prompt": "Review the folder src/runtime/agent/orchestrator/session/compact/ for correctness, regression, and security risks. Focus on real actionable bugs: schema validation gaps, data corruption on summarize, lossy round-trips, error swallowing. Report findings ordered by severity, one line each, with file:line anchors. Skip style nits. End with a one-line verdict."
+  }
+]

package/scripts/bench/round-codex.json

@@ -0,0 +1,114 @@
+{
+  "round": "1",
+  "runner": "codex",
+  "opts": {
+    "provider": "openai-oauth",
+    "model": "gpt-5.5",
+    "effort": "high",
+    "fast": false,
+    "env": {}
+  },
+  "tasks": 3,
+  "completed": 3,
+  "completion_rate": 100,
+  "sessions": [
+    "019f24e5-98ee-7cf0-afb0-a1a80ac81702",
+    "019f24e5-98f3-7f12-9638-7a899028c7e3",
+    "019f24e5-98f3-7f22-902c-934f0f7cbfb0"
+  ],
+  "results": [
+    {
+      "id": "review-session-runtime",
+      "agent": "reviewer",
+      "ok": true,
+      "ms": 300087,
+      "sessionId": "019f24e5-98ee-7cf0-afb0-a1a80ac81702",
+      "card": {
+        "session": "019f24e5-98ee-7cf0-afb0-a1a80ac81702",
+        "wall_ms": 300087,
+        "uncached_in": 157635,
+        "cached_in": 1071360,
+        "out_tokens": 8581,
+        "cache_ratio": 0.872,
+        "completed": 1
+      },
+      "output": "- High: `src/session-runtime/plugin-mcp.mjs:40` project `.mcp.json` stdio entries can set `cwd` to `..` or an absolute path outside the project, bypassing the stricter project-root guard in `normalizeMcpServerInput` and letting repo-local MCP config run with access scoped outside the opened repo.\n- Medium: `src/session-runtime/workflow.mjs:276` invalid persisted `effort` values throw from read paths like `summarizeWorkflowRoutes`/`agentRouteFromConfig` instead of being ignored, so one bad config entry can break onboarding status or agent listing.\n- Low: `src/session-runtime/provider-models.mjs:42` cached `{id, provider}` fallback metadata is returned before honoring `allowFetch`, so a prior non-fetch lookup prevents a later forced lookup from ever retrieving real model capabilities.\n\nVerdict: no broken imports found, but fix the project MCP cwd escape before trusting repo-local MCP configs."
+    },
+    {
+      "id": "review-channels-lib",
+      "agent": "reviewer",
+      "ok": true,
+      "ms": 172813,
+      "sessionId": "019f24e5-98f3-7f12-9638-7a899028c7e3",
+      "card": {
+        "session": "019f24e5-98f3-7f12-9638-7a899028c7e3",
+        "wall_ms": 172813,
+        "uncached_in": 136264,
+        "cached_in": 1237376,
+        "out_tokens": 8502,
+        "cache_ratio": 0.901,
+        "completed": 1
+      },
+      "output": "- High `src/runtime/channels/lib/webhook/deliveries.mjs:255`: delivery rows are appended asynchronously but the in-memory dedup state advances immediately, so crashes, append failures, out-of-order status writes, or compaction can lose/reorder `received/pending/done` rows and allow duplicate webhook dispatch or permanent dedup blockage.\n- High `src/runtime/channels/lib/event-queue.mjs:239`: queued events are moved to processed at `:241` after a fire-and-forget `injectFn`; if parent IPC drops/fails, the claimed event is acknowledged anyway and the message is lost.\n- Medium `src/runtime/channels/lib/output-forwarder.mjs:316`: actual backend send does not re-check `_isOwner()` at send time, so an ownership handoff during an in-flight drain can let the old owner send while the new owner tails the same transcript, causing double-forwarding.\n- Medium `src/runtime/channels/lib/output-forwarder.mjs:690`: `closeWatcher()` clears `sendRetryTimer` but leaves queued failed items in `finalLane/streamLane`, so transcript rebind/watch error during backoff can strand unsent messages until unrelated new activity kicks the drain.\n- Medium `src/runtime/channels/lib/webhook.mjs:530`: `startNgrok()` schedules `waitAndStart` timers that `stop()` at `:585` never cancels, so reload/stop during the delay can spawn a stale ngrok process after shutdown.\n- Low `src/runtime/channels/lib/inbound-routing.mjs:35`: stale dedup marker files are treated as non-duplicates but never refreshed/replaced, so cross-process duplicate inbound messages after TTL can bypass dedup repeatedly.\n\nVerdict: Not clean yet; I found real message-loss, double-forwarding, and lifecycle leak risks worth fixing before shipping."
+    },
+    {
+      "id": "review-compact",
+      "agent": "reviewer",
+      "ok": true,
+      "ms": 140945,
+      "sessionId": "019f24e5-98f3-7f22-902c-934f0f7cbfb0",
+      "card": {
+        "session": "019f24e5-98f3-7f22-902c-934f0f7cbfb0",
+        "wall_ms": 140945,
+        "uncached_in": 102061,
+        "cached_in": 416512,
+        "out_tokens": 7120,
+        "cache_ratio": 0.803,
+        "completed": 1
+      },
+      "output": "- High `src/runtime/agent/orchestrator/session/compact/engine.mjs:588` - `recallTailStartIndex()` relies on object identity, so when `fitSingleRecallTurnToCap()` clones/truncates the first tail user message, `indexOf()` fails and the fallback can misplace the tail, duplicating summarized user content and dropping/reordering recent messages.\n- High `src/runtime/agent/orchestrator/session/compact/text-utils.mjs:269` - `redactMessageToolCallSecrets()` returns unchanged unless `m.toolCalls` exists, leaving Anthropic-style `content` / `assistantBlocks` `tool_use` inputs unredacted in preserved tails.\n- High `src/runtime/agent/orchestrator/session/compact/text-utils.mjs:12` - the sensitive-key regex misses common camelCase token keys like `accessToken`, `idToken`, and `sessionToken`, so those secrets survive both prompt summaries and verbatim tail redaction.\n- Medium `src/runtime/agent/orchestrator/session/compact/budget.mjs:79` - `drainSessionCycle1()` swallows `cycle1` errors and returns a “successful” partial drain, so callers can proceed with raw pending rows while diagnostics lose the actual failure.\n- Medium `src/runtime/agent/orchestrator/session/compact/summary-schema.mjs:68` - semantic schema validation only checks top-level `##` anchors, while `truncateSummaryBySections()` can drop required `### Done/In Progress/Blocked` progress structure and still treat the summary as valid.\n\nVerdict: Needs fixes before compaction can be considered secure and loss-safe."
+    }
+  ],
+  "task_complete": true,
+  "task_errors": [],
+  "score_complete": true,
+  "score_errors": [],
+  "group": {
+    "n": 3,
+    "wall_ms": 204615,
+    "uncached_in": 131986.7,
+    "cached_in": 908416,
+    "out_tokens": 8067.7,
+    "cache_ratio": 0.9,
+    "completed": 1
+  },
+  "cards": [
+    {
+      "session": "019f24e5-98ee-7cf0-afb0-a1a80ac81702",
+      "wall_ms": 300087,
+      "uncached_in": 157635,
+      "cached_in": 1071360,
+      "out_tokens": 8581,
+      "cache_ratio": 0.872,
+      "completed": 1
+    },
+    {
+      "session": "019f24e5-98f3-7f12-9638-7a899028c7e3",
+      "wall_ms": 172813,
+      "uncached_in": 136264,
+      "cached_in": 1237376,
+      "out_tokens": 8502,
+      "cache_ratio": 0.901,
+      "completed": 1
+    },
+    {
+      "session": "019f24e5-98f3-7f22-902c-934f0f7cbfb0",
+      "wall_ms": 140945,
+      "uncached_in": 102061,
+      "cached_in": 416512,
+      "out_tokens": 7120,
+      "cache_ratio": 0.803,
+      "completed": 1
+    }
+  ]
+}

package/scripts/bench/round-mixdog-lead-r3.json

@@ -0,0 +1,269 @@
+{
+  "round": "3",
+  "runner": "lead",
+  "opts": {
+    "provider": null,
+    "model": null,
+    "effort": null,
+    "fast": false,
+    "env": {
+      "MIXDOG_DATA_DIR": "C:\\Users\\tempe\\.mixdog\\bench-data"
+    }
+  },
+  "tasks": 3,
+  "completed": 3,
+  "completion_rate": 100,
+  "sessions": [
+    "sess_44376_1783036165596_1783036165772_a77d5225c11eaee87b9509c9611c40c3",
+    "sess_53612_1783036165611_1783036165786_6799f6b975f8c1531cf1766d5802f647",
+    "sess_40692_1783036165624_1783036165796_a84440b9682c554ba565b3270485c54d"
+  ],
+  "results": [
+    {
+      "id": "lead-review-session-runtime",
+      "agent": "reviewer",
+      "ok": true,
+      "ms": 396471,
+      "sessionId": "sess_44376_1783036165596_1783036165772_a77d5225c11eaee87b9509c9611c40c3",
+      "output": "sessionId: sess_44376_1783036165596_1783036165772_a77d5225c11eaee87b9509c9611c40c3\n4개 리뷰 모두 완료. 보고 전에 핵심 발견들을 직접 교차 검증하겠습니다.리뷰 4건을 모두 회수하고 핵심 주장(특히 High 2건)을 직접 교차 검증했습니다. review-mcp의 High 1건은 오탐으로 판정해 제외했습니다.\n\n**발견 사항 (심각도순)**\n\n- **High** `src/session-runtime/provider-usage.mjs:54` — in-flight `providerSetupPromise`가 `invalidateProviderCaches()` 이후에도 `.then`에서 `providerSetupCache`를 재기록해, 로그인/키 변경 직후 무효화된 provider·auth 스냅샷이 되살아나는 레이스 (provider-models의 `providerModelsLoadSeq` 같은 시퀀스 가드 부재, `mixdog-session-runtime.mjs:1050-1062`에서 확인).\n\n- **Medium** `src/session-runtime/provider-models.mjs:211` — `collectSearchProviderModels` force 경로가 시퀀스 가드 없이 `searchProviderModelsCache`를 덮어써, 무효화 이전에 시작된 새로고침이 stale 검색 모델 목록을 복원할 수 있음 (같은 파일 227행의 가드 패턴과 비대칭).\n\n- **Medium** `src/session-runtime/config-helpers.mjs:49,66` — preset에 `provider`가 없으면 `resolveDefaultProvider(config)` 대신 하드코딩 `DEFAULT_PROVIDER`('anthropic-oauth')로 떨어져, `config.defaultProvider` 설정을 무시하고 잘못된 provider로 라우팅 (79행의 비-preset 경로와 모순).\n\n- **Low** `src/session-runtime/effort.mjs:87` — reasoning 메타데이터가 전혀 없는 모델이 provider 전체 effort 옵션으로 폴백해, 비-reasoning 모델(quick/static 행)에 effort 선택·저장이 노출됨.\n\n**기각한 보고**: `native-search.mjs:207`의 \"openai nativeTools 무시\" High는 오탐 — `openai`는 registry에서 `OpenAIDirectProvider`(openai-ws.mjs)로 라우팅되고, 이는 `buildRequestBody`(openai-oauth.mjs:546-550)를 통해 `nativeTools`를 전달하며, compat 프리셋 provider들은 애초에 `SEARCH_CAPABLE_PROVIDERS`에 없습니다. tool-catalog/workflow/warmup-schedulers/mcp 계열 파일들은 깨진 import·미정의 식별자 없이 깨끗했습니다.\n\n**평결**: 깨진 import나 데이터 손실 경로는 없으며, 실질 리스크는 캐시 무효화 레이스 1건(High)과 폴백 불일치 2건 수준입니다.",
+      "card": {
+        "session": "sess_44376_1783036165596_1783036165772_a77d5225c11eaee87b9509c9611c40c3",
+        "provider": "anthropic-oauth",
+        "model": "claude-fable-5",
+        "wall_ms": 385974,
+        "turns": 72,
+        "tool_calls": 99,
+        "tools_per_turn": 1.4,
+        "total_tool_ms": 5723,
+        "llm_stream_ms": 625711,
+        "cache_ratio": 0.797586432611301,
+        "cached_tokens": 1938608,
+        "prompt_tokens": 2430593,
+        "uncached_tokens": 116796,
+        "output_tokens": 12031,
+        "cache_weighted_input_10": 310656.80000000005,
+        "cache_weighted_input_25": 601448,
+        "prompt_growth": 78094,
+        "antipatterns": 2,
+        "issues": 23,
+        "issue_types": [
+          "readonly_stall",
+          "cache",
+          "cache_break",
+          "tool_error",
+          "broad_tool_result",
+          "service_tier",
+          "missed_parallelism",
+          "tool_churn_cluster"
+        ],
+        "read_fragmentation_paths": [],
+        "grep_sweep_count": 0
+      }
+    },
+    {
+      "id": "lead-review-channels-lib",
+      "agent": "reviewer",
+      "ok": true,
+      "ms": 287195,
+      "sessionId": "sess_53612_1783036165611_1783036165786_6799f6b975f8c1531cf1766d5802f647",
+      "output": "sessionId: sess_53612_1783036165611_1783036165786_6799f6b975f8c1531cf1766d5802f647\n핵심 High 항목들을 직접 교차 검증하겠습니다.4개 리뷰 스코프 결과를 받아 High 항목을 코드로 직접 교차 검증했습니다(event-queue 배칭, output-forwarder drain, webhook fail-open, PowerShell 인젝션, cron 덮어쓰기 모두 확인됨).\n\n**High**\n- `event-queue.mjs:198` — 저순위 배칭이 `item.name`만으로 그룹핑 후 `claimedPairs[0].item`(:219)으로 발송해, 같은 이름의 다른 채널/지시 이벤트가 첫 채널로 병합·오배송되고 나머지는 유실됨.\n- `output-forwarder.mjs:575` — `forwardFinalText()`가 finally에서 `sending`만 해제하고 drain을 재기동하지 않아, 전송 중 큐잉된 항목의 `_kickDrain()`이 :436에서 no-op 되어 다음 활동 전까지 좌초됨.\n- `webhook.mjs:338` — `secret`/`parser` 없는 config 등록 엔드포인트는 경고만 남기고 통과(fail-open)해 무서명 POST가 특권 디스패치로 주입 가능.\n- `webhook/deliveries.mjs:255` — 전송 dedup 상태가 durable write 전에 성공 처리되어, flush 전 크래시 시 재시작 후 재배달/이중 디스패치 가능.\n- `runtime-paths.mjs:247` — UI-heartbeat-stale 판정 시 `prev=null`로 처리돼 좀비 워커가 `ui_heartbeat_at` 없이 오너 상태를 즉시 재게시, 테이크오버를 차단할 수 있음.\n- `runtime-paths.mjs:73` — 오너 생존 판정이 PID만 신뢰해 PID 재사용 시 stale `active-instance.json`이 무기한 고정되어 인수 불가.\n- `voice-runtime-fetcher.mjs:432` — PowerShell `-Command`에 JSON 이중따옴표 경로를 삽입해 경로 내 `$()`/백틱이 설치 중 명령 실행으로 이어지는 인젝션.\n- `voice-runtime-fetcher.mjs:174` — 락 해제가 `lockPath`를 무조건 삭제해, age 기반 재클레임 후 구 설치자가 새 오너의 락을 지워 동시 설치/손상 허용.\n- `scheduler.mjs:368` — 중복 스케줄 이름이 `cronJobs` 엔트리를 덮어써 이전 `ScheduledTask`가 누수, `destroyCronJobs()` 후에도 이중 발화.\n\n**Medium**\n- `event-queue.mjs:77` — 시작 시 `in-progress/` 전체를 나이/오너 확인 없이 재큐잉해, 핸드오프 중 구 오너가 처리 중인 파일이 중복 주입될 수 있음.\n- `backend-dispatch.mjs:17` — `dispatchReply()`가 부분 전송 실패 시 `resumeToken`을 버려, 재시도 때 이미 전송된 청크를 재발송(이중 게시).\n- `webhook.mjs:139` — 요청 바디 스트림에 `error`/`aborted` 핸들러가 없어 클라이언트 리셋 시 요청이 미해결로 남거나 스트림 에러가 unhandled로 표면화.\n- `session-control.mjs:15` — 동시 `controlClaudeSession()` 호출이 하나의 컨트롤 파일을 공유해 나중 명령이 미판독 명령을 덮어씀(첫 호출 타임아웃/유실).\n- `status-snapshot.mjs:218` — `stopSnapshotWriter()`가 전역 스냅샷을 무조건 unlink해 핸드오프 중 구 오너가 새 오너의 스냅샷을 삭제.\n- `session-discovery.mjs:26` — parent PID 캐시가 만료 없이 PID만으로 키잉되어 PID 재사용 시 stale 세션/트랜스크립트에 바인딩.\n- `webhook/signature.mjs:32` — Stripe 검증이 첫 `v1=` 서명만 확인해 시크릿 로테이션 중 유효 전송이 거부됨.\n- `whisper-server.m",
+      "card": {
+        "session": "sess_53612_1783036165611_1783036165786_6799f6b975f8c1531cf1766d5802f647",
+        "provider": "anthropic-oauth",
+        "model": "claude-fable-5",
+        "wall_ms": 276579,
+        "turns": 49,
+        "tool_calls": 67,
+        "tools_per_turn": 1.4,
+        "total_tool_ms": 4244,
+        "llm_stream_ms": 672149,
+        "cache_ratio": 0.6616259058956849,
+        "cached_tokens": 1003052,
+        "prompt_tokens": 1516041,
+        "uncached_tokens": 139818,
+        "output_tokens": 14174,
+        "cache_weighted_input_10": 240123.2,
+        "cache_weighted_input_25": 390581,
+        "prompt_growth": 75338,
+        "antipatterns": 1,
+        "issues": 16,
+        "issue_types": [
+          "cache",
+          "cache_break",
+          "broad_tool_result",
+          "readonly_stall",
+          "service_tier",
+          "missed_parallelism",
+          "tool_churn_cluster"
+        ],
+        "read_fragmentation_paths": [],
+        "grep_sweep_count": 0
+      }
+    },
+    {
+      "id": "lead-review-compact",
+      "agent": "reviewer",
+      "ok": true,
+      "ms": 426500,
+      "sessionId": "sess_40692_1783036165624_1783036165796_a84440b9682c554ba565b3270485c54d",
+      "output": "sessionId: sess_40692_1783036165624_1783036165796_a84440b9682c554ba565b3270485c54d\n세 결과를 받았으니 핵심 앵커들을 직접 확인해 검증하겠습니다.주요 주장들을 앵커에서 직접 검증하겠습니다.리뷰 완료 — 에이전트 결과를 앵커에서 직접 검증했고, 아래는 확인된 이슈만 심각도순으로 정리한 것입니다.\n\n- **High** `text-utils.mjs:295` — `redactMessageToolCallSecrets`가 `m.toolCalls`만 스크럽하고 content 배열의 `tool_use` 블록 `input`은 건너뜀(해당 형태 존재는 `context-utils.mjs:551` 주석이 확인) → 보존 tail에 시크릿이 미리댁션 상태로 잔존.\n\n- **High** `engine.mjs:93` — `splitTurnStartIndexForBudget`가 tool-role 인덱스를 컷 지점으로 반환 가능(비교: `findValidCutIndices`는 tool 제외) → 고아 tool result가 head에서도 tail에서도 빠져 무음 손실.\n\n- **High** `text-utils.mjs:29` — camel 정규식의 `s?$`가 주석(\"`maxOutputTokens`는 매치 안 됨\")과 모순, `maxOutputTokens`/`budgetTokens`/`completionTokens`가 `[redacted]`로 치환돼 보존 tool-call 인자가 손상됨.\n\n- **Medium** `summary.mjs:427` (동일 패턴 `:577`) — prior 요약을 remainingTokens 전체로 먼저 이진탐색 최대화한 뒤 recall을 끼워 넣어, 토큰 압박 시 신규 recall 블록이 전부 탈락하고 stale prior만 남음.\n\n- **Medium** `budget.mjs:203` — `pruneToolOutputText`가 maxChars 무시하고 고정 1000/600 슬라이스 사용; `maxToolOutputChars < ~1700`이면 프루닝 결과가 원본보다 커져 예산 회복 실패(head/tail 중복 포함).\n\n- **Medium** `engine.mjs:588` — `recallTailStartIndex`가 객체 동일성(`indexOf`) 의존, fitting이 첫 tail 메시지를 클론하면 `live.length - tail.length` 폴백이 잘못된 경계를 선택해 head/tail 분할이 어긋남.\n\n- **Low** `summary-schema.mjs:71` — 검증이 최상위 `##` 앵커만 확인해 `### Done/In Progress/Blocked` 누락이 통과하고, `:309`의 섹션 중간절단이 서브헤딩을 잘라도 여전히 유효 판정됨.\n\n- **Low** `summary.mjs:400` — `stripNestedSummaryHeaderLines`가 `:318`의 `semantic=true provider=... model=...` 라인을 안 지워 반복 컴팩션 시 합성 메타데이터가 prior 본문에 누적.\n\n- **Low** `text-utils.mjs:118` — `truncateMiddle`이 UTF-16 코드유닛 단위로 잘라 서로게이트 쌍 분리 가능(같은 패턴 `budget.mjs:203-204`).\n\n핵심 컴팩션 경로 자체는 견고하나, 시크릿 리댁션 우회와 턴 분할 시 tool result 무음 손실 두 건은 실사용 전 수정이 필요합니다.",
+      "card": {
+        "session": "sess_40692_1783036165624_1783036165796_a84440b9682c554ba565b3270485c54d",
+        "provider": "anthropic-oauth",
+        "model": "claude-fable-5",
+        "wall_ms": 416022,
+        "turns": 33,
+        "tool_calls": 45,
+        "tools_per_turn": 1.4,
+        "total_tool_ms": 1888,
+        "llm_stream_ms": 569465,
+        "cache_ratio": 0.767026954989235,
+        "cached_tokens": 667631,
+        "prompt_tokens": 870414,
+        "uncached_tokens": 53129,
+        "output_tokens": 10398,
+        "cache_weighted_input_10": 119892.1,
+        "cache_weighted_input_25": 220036.75,
+        "prompt_growth": 49784,
+        "antipatterns": 2,
+        "issues": 12,
+        "issue_types": [
+          "cache",
+          "cache_break",
+          "broad_tool_result",
+          "readonly_stall",
+          "service_tier",
+          "duplicate_tool",
+          "missed_parallelism",
+          "tool_churn_cluster"
+        ],
+        "read_fragmentation_paths": [],
+        "grep_sweep_count": 0
+      }
+    }
+  ],
+  "task_complete": true,
+  "task_errors": [],
+  "score_complete": true,
+  "score_errors": [],
+  "group": {
+    "n": 3,
+    "wall_ms": 359525,
+    "turns": 51.3,
+    "tool_calls": 70.3,
+    "tools_per_turn": 1.4,
+    "total_tool_ms": 3951.7,
+    "llm_stream_ms": 622441.7,
+    "cache_ratio": 0.7,
+    "cached_tokens": 1203097,
+    "prompt_tokens": 1605682.7,
+    "uncached_tokens": 103247.7,
+    "output_tokens": 12201,
+    "cache_weighted_input_10": 223557.4,
+    "cache_weighted_input_25": 404021.9,
+    "prompt_growth": 67738.7,
+    "antipatterns": 1.7,
+    "issues": 17,
+    "grep_sweep_count": 0
+  },
+  "cards": [
+    {
+      "session": "sess_44376_1783036165596_1783036165772_a77d5225c11eaee87b9509c9611c40c3",
+      "provider": "anthropic-oauth",
+      "model": "claude-fable-5",
+      "wall_ms": 385974,
+      "turns": 72,
+      "tool_calls": 99,
+      "tools_per_turn": 1.4,
+      "total_tool_ms": 5723,
+      "llm_stream_ms": 625711,
+      "cache_ratio": 0.797586432611301,
+      "cached_tokens": 1938608,
+      "prompt_tokens": 2430593,
+      "uncached_tokens": 116796,
+      "output_tokens": 12031,
+      "cache_weighted_input_10": 310656.80000000005,
+      "cache_weighted_input_25": 601448,
+      "prompt_growth": 78094,
+      "antipatterns": 2,
+      "issues": 23,
+      "issue_types": [
+        "readonly_stall",
+        "cache",
+        "cache_break",
+        "tool_error",
+        "broad_tool_result",
+        "service_tier",
+        "missed_parallelism",
+        "tool_churn_cluster"
+      ],
+      "read_fragmentation_paths": [],
+      "grep_sweep_count": 0
+    },
+    {
+      "session": "sess_53612_1783036165611_1783036165786_6799f6b975f8c1531cf1766d5802f647",
+      "provider": "anthropic-oauth",
+      "model": "claude-fable-5",
+      "wall_ms": 276579,
+      "turns": 49,
+      "tool_calls": 67,
+      "tools_per_turn": 1.4,
+      "total_tool_ms": 4244,
+      "llm_stream_ms": 672149,
+      "cache_ratio": 0.6616259058956849,
+      "cached_tokens": 1003052,
+      "prompt_tokens": 1516041,
+      "uncached_tokens": 139818,
+      "output_tokens": 14174,
+      "cache_weighted_input_10": 240123.2,
+      "cache_weighted_input_25": 390581,
+      "prompt_growth": 75338,
+      "antipatterns": 1,
+      "issues": 16,
+      "issue_types": [
+        "cache",
+        "cache_break",
+        "broad_tool_result",
+        "readonly_stall",
+        "service_tier",
+        "missed_parallelism",
+        "tool_churn_cluster"
+      ],
+      "read_fragmentation_paths": [],
+      "grep_sweep_count": 0
+    },
+    {
+      "session": "sess_40692_1783036165624_1783036165796_a84440b9682c554ba565b3270485c54d",
+      "provider": "anthropic-oauth",
+      "model": "claude-fable-5",
+      "wall_ms": 416022,
+      "turns": 33,
+      "tool_calls": 45,
+      "tools_per_turn": 1.4,
+      "total_tool_ms": 1888,
+      "llm_stream_ms": 569465,
+      "cache_ratio": 0.767026954989235,
+      "cached_tokens": 667631,
+      "prompt_tokens": 870414,
+      "uncached_tokens": 53129,
+      "output_tokens": 10398,
+      "cache_weighted_input_10": 119892.1,
+      "cache_weighted_input_25": 220036.75,
+      "prompt_growth": 49784,
+      "antipatterns": 2,
+      "issues": 12,
+      "issue_types": [
+        "cache",
+        "cache_break",
+        "broad_tool_result",
+        "readonly_stall",
+        "service_tier",
+        "duplicate_tool",
+        "missed_parallelism",
+        "tool_churn_cluster"
+      ],
+      "read_fragmentation_paths": [],
+      "grep_sweep_count": 0
+    }
+  ]
+}