mixdog 0.9.3 → 0.9.5

package/scripts/bench/round-r17-wirebytes.json

@@ -0,0 +1,456 @@
+{
+  "round": "17",
+  "runner": "lead",
+  "opts": {
+    "provider": "openai-oauth",
+    "model": "gpt-5.5",
+    "effort": null,
+    "fast": false,
+    "env": {
+      "MIXDOG_DATA_DIR": "C:\\Users\\tempe\\.mixdog\\bench-data"
+    }
+  },
+  "tasks": 6,
+  "completed": 6,
+  "completion_rate": 100,
+  "sessions": [
+    "sess_55884_1783059815805_1783059816044_1d290d1d86ad152644020350714e0b6a",
+    "sess_48508_1783059815820_1783059816056_cebabb220978834570309408a4932033",
+    "sess_12172_1783059815886_1783059816130_c1422313e74ff647937bfb7ccc713170",
+    "sess_44152_1783059834394_1783059834618_56b4c9c66902e5730af3a1b5d05d85c4",
+    "sess_25912_1783059899801_1783059900029_f989eab921374da513570e09d7d39c55",
+    "sess_60436_1783059931841_1783059932082_99b61f2d2925940ee5554fa92042346c"
+  ],
+  "results": [
+    {
+      "id": "r4-debug-forwarding-gap#1",
+      "agent": "reviewer",
+      "ok": true,
+      "ms": 116038,
+      "sessionId": "sess_55884_1783059815805_1783059816044_1d290d1d86ad152644020350714e0b6a",
+      "output": "sessionId: sess_55884_1783059815805_1783059816044_1d290d1d86ad152644020350714e0b6a\n관련 경로를 먼저 확인해 전체 흐름과 줄 번호를 맞추겠습니다.조사 결과, not-yet-created transcript는 대부분 복구되지만 “이미 존재하는 transcript를 처음 EOF로 묶는” 경로에는 실제 무손실 보장이 없습니다.\n\n- **최대 영향: 기존 transcript를 EOF로 바인딩하면 선행 출력 드롭** — `setContext()`가 기본적으로 `currentSize`로 커서를 잡고(`src/runtime/channels/lib/output-forwarder.mjs:111-129`), `applyTranscriptBinding()` 호출은 즉시 watch만 겁니다(`src/runtime/channels/index.mjs:317-321`). backlog catch-up: **아니오**, `replayFromStart` 또는 같은 transcript의 persisted cursor가 없으면 이미 쓴 assistant 출력은 읽지 않습니다.\n\n- **not-yet-created transcript는 지연 후 대체로 복구** — `fs.watch`는 없는 파일에서 실패하지만(`src/runtime/channels/lib/output-forwarder.mjs:647-670`), re-arm poll이 생성 감지 후 watch 설치와 `forwardNewText()`를 실행합니다(`src/runtime/channels/index.mjs:357-389`, `486-498`). backlog catch-up: **예**, 단 60초 timeout 이후에는 다음 rebind 없이는 stranded입니다.\n\n- **활성화 직후 ownership/backend 연결 전 출력은 지연** — `bridge-state.json`은 먼저 쓰이고(`src/runtime/channels/index.mjs:1735-1744`), owner gate는 `bridgeRuntimeConnected && currentOwnerState().owned`가 되기 전 `forwardNewText()`를 no-op 처리합니다(`src/runtime/channels/index.mjs:316`, `src/runtime/channels/lib/output-forwarder.mjs:351-354`). backlog catch-up: **조건부 예**, 커서가 앞서 EOF로 잡히지 않았을 때 post-connect/rearm flush가 회수합니다(`src/runtime/channels/index.mjs:801-809`).\n\n- **watch 이벤트 중 ownership probe 실패는 지연 가능** — `_isOwner()`는 probe 예외/false를 fail-closed로 처리하고(`src/runtime/channels/lib/output-forwarder.mjs:36-41`), watch flush는 실패 시 별도 retry를 예약하지 않습니다(`src/runtime/channels/lib/output-forwarder.mjs:703-731`). backlog catch-up: **부분적 예**, 커서는 안 움직이므로 다음 write/turn-end/rebind가 있으면 회수되지만 마지막 출력이면 지연될 수 있습니다.\n\n- **stat→watch 설치 사이의 짧은 레이스는 커버됨** — watch 설치 직후 `scheduleWatchFlush()`가 실행되어 그 사이 쓰인 bytes를 읽습니다(`src/runtime/channels/lib/output-forwarder.mjs:653-667`). backlog catch-up: **예**.\n\n판정: bridge activation 이후 pending transcript 경로는 복구 설계가 있지만, 기존 transcript를 처음 바인딩하는 EOF cursor 경로는 assistant 출력 유실 가능성이 남아 있습니다.",
+      "card": {
+        "session": "sess_55884_1783059815805_1783059816044_1d290d1d86ad152644020350714e0b6a",
+        "provider": "openai-oauth",
+        "model": "gpt-5.5",
+        "wall_ms": 105545,
+        "turns": 5,
+        "tool_calls": 7,
+        "tools_per_turn": 1.4,
+        "total_tool_ms": 439,
+        "llm_stream_ms": 104528,
+        "cache_ratio": 0.23191906990789674,
+        "cached_tokens": 23040,
+        "prompt_tokens": 99345,
+        "uncached_tokens": 76305,
+        "output_tokens": 11249,
+        "cache_weighted_input_10": 78609,
+        "cache_weighted_input_25": 82065,
+        "prompt_growth": 38284,
+        "antipatterns": 0,
+        "issues": 10,
+        "issue_types": [
+          "cache",
+          "cache_break",
+          "cache_miss",
+          "broad_tool_result",
+          "service_tier"
+        ],
+        "read_fragmentation_paths": [],
+        "grep_sweep_count": 0
+      }
+    },
+    {
+      "id": "r4-impl-usage-summary#1",
+      "agent": "worker",
+      "ok": true,
+      "ms": 18575,
+      "sessionId": "sess_48508_1783059815820_1783059816056_cebabb220978834570309408a4932033",
+      "output": "sessionId: sess_48508_1783059815820_1783059816056_cebabb220978834570309408a4932033\n재영님, 진행 계획은 아래처럼 잡겠습니다. 승인해 주시면 바로 구현·검증하겠습니다.\n\n- **스크립트 추가**: `scripts/usage-today.mjs`에서 `MIXDOG_DATA_DIR` 또는 `~/.mixdog/data`의 `history/agent-trace.jsonl`을 읽고 오늘 00:00 이후 usage row만 집계\n- **출력**: 모델별 calls, input/cached/output tokens, costUsd와 grand total 출력\n- **연결/검증**: `package.json`에 `usage:today` 추가 후 `node --check`, 실제 1회 실행까지 확인\n\n진행해도 될까요?",
+      "card": {
+        "session": "sess_48508_1783059815820_1783059816056_cebabb220978834570309408a4932033",
+        "provider": "openai-oauth",
+        "model": "gpt-5.5",
+        "wall_ms": 8082,
+        "turns": 1,
+        "tool_calls": 0,
+        "tools_per_turn": 0,
+        "total_tool_ms": 0,
+        "llm_stream_ms": 7179,
+        "cache_ratio": 0,
+        "cached_tokens": 0,
+        "prompt_tokens": 4748,
+        "uncached_tokens": 4748,
+        "output_tokens": 466,
+        "cache_weighted_input_10": 4748,
+        "cache_weighted_input_25": 4748,
+        "prompt_growth": 0,
+        "antipatterns": 0,
+        "issues": 3,
+        "issue_types": [
+          "cache",
+          "cache_break",
+          "service_tier"
+        ],
+        "read_fragmentation_paths": [],
+        "grep_sweep_count": 0
+      }
+    },
+    {
+      "id": "r4-explain-compact-engine#1",
+      "agent": "explore",
+      "ok": true,
+      "ms": 83971,
+      "sessionId": "sess_12172_1783059815886_1783059816130_c1422313e74ff647937bfb7ccc713170",
+      "output": "sessionId: sess_12172_1783059815886_1783059816130_c1422313e74ff647937bfb7ccc713170\n완료했습니다 — compaction subsystem architecture brief입니다.\n\n```text\n1. Public surface re-exports `semanticCompactMessages` and `recallFastTrackCompactMessages`; existing callers import through `session/compact.mjs`. src/runtime/agent/orchestrator/session/compact.mjs:52\n2. Main auto entry is pre-send in `agentLoop`: drain queued messages, resolve policy, estimate tokens, decide, then choose recall-fasttrack or semantic. src/runtime/agent/orchestrator/session/loop.mjs:324\n3. Manual/API entry is `compact()` → `compactSessionMessages()` → `runSessionCompaction(... force:true)`. src/mixdog-session-runtime.mjs:2452, src/runtime/agent/orchestrator/session/manager.mjs:1950\n4. Auto-clear can pre-compact before clearing when a compact type is requested. src/runtime/agent/orchestrator/session/manager.mjs:1834\n5. Policy is enabled unless `compaction.auto === false` / env disables it, and requires a boundary from explicit boundary/context/auto limit. src/runtime/agent/orchestrator/session/loop/compact-policy.mjs:89\n6. Trigger tokens are explicit sub-boundary auto limit or boundary minus configured/default buffer. src/runtime/agent/orchestrator/session/loop/compact-policy.mjs:107\n7. Compaction fires when reactive overflow is pending, token estimate is unknown, or `messages+reserve >= triggerTokens`. src/runtime/agent/orchestrator/session/loop/compact-policy.mjs:170\n8. Target budget is the compact target plus reserve, clamped to the boundary. src/runtime/agent/orchestrator/session/loop/compact-policy.mjs:163\n9. Auto loop tags reactive retries separately from proactive auto compaction. src/runtime/agent/orchestrator/session/loop.mjs:351\n10. Recall-fasttrack path runs first when policy type is recall-fasttrack; semantic path is the default semantic branch. src/runtime/agent/orchestrator/session/loop.mjs:414, src/runtime/agent/orchestrator/session/loop.mjs:439\n11. Agent-owned sessions are forced to semantic to avoid injecting unrelated main-session recall. src/runtime/agent/orchestrator/session/loop/compact-policy.mjs:39\n12. Semantic compaction sanitizes tool pairs/dedups, no-ops if already within budget and not forced, otherwise redacts before windowing. src/runtime/agent/orchestrator/session/compact/engine.mjs:284\n13. Redaction scrubs sensitive tool-call argument values while preserving message/tool-call shape and order. src/runtime/agent/orchestrator/session/compact/text-utils.mjs:219\n14. Redaction is applied before semantic window selection so preserved system/tail are measured and emitted redacted. src/runtime/agent/orchestrator/session/compact/engine.mjs:306\n15. Redaction is also applied before recall-fasttrack splitting. src/runtime/agent/orchestrator/session/compact/engine.mjs:644\n16. Semantic windowing separates protected prefix, prior summary, and live non-summary messages. src/runtime/agent/orchestrator/session/compact/engine.mjs:100\n17. Semantic tail keeps recent user-anchored turns when user messages exist, bounded by preserved-recent budget and `tailTurns`. src/runtime/agent/orchestrator/session/compact/engine.mjs:134\n18. If no user turns exist, semantic tail starts at a valid non-tool cut point so tool results never become the first preserved message. src/runtime/agent/orchestrator/session/compact/engine.mjs:121\n19. Semantic head is everything before `tailStartIdx`; tail is re-sanitized to repair pair/dedup artifacts. src/runtime/agent/orchestrator/session/compact/engine.mjs:244\n20. Semantic refuses only an empty live window or no compactable prior history without previous summary. src/runtime/agent/orchestrator/session/compact/engine.mjs:254, src/runtime/agent/orchestrator/session/compact/engine.mjs:315\n21. Semantic raises budget to fit mandatory preserved context plus minimum summary room instead of refusing mandatory overflow. src/runtime/agent/orchestrator/session/compact/engine.mjs:319\n22. Semantic prompt is fitted to 85% headroom, sent with low/fast settings and a timeout signal. src/runtime/agent/orchestrator/session/compact/engine.mjs:333\n23. Semantic malformed summaries are schema-repaired before becoming the summary message. src/runtime/agent/orchestrator/session/compact/engine.mjs:371\n24. Recall-fasttrack preserves newest structured turns, default max two user turns and 8k cap. src/runtime/agent/orchestrator/session/compact/engine.mjs:449\n25. Recall tail drops oldest kept turns first, then truncates the oldest surviving content while keeping newest content whole. src/runtime/agent/orchestrator/session/compact/engine.mjs:601\n26. Recall no-user tails are valid and preserve recent assistant/tool structure. src/runtime/agent/orchestrator/session/compact/engine.mjs:613\n27. Recall summary input combines session id, cycle-1 hydrated chunks/raw fallback, and recall text. src/runtime/agent/orchestrator/session/loop/recall-fasttrack.mjs:148\n28. Recall injection is capped to 5% of context window with a floor when context window is known. src/runtime/agent/orchestrator/session/loop/recall-fasttrack.mjs:151\n29. Recall-fasttrack throws on no compactable prior plus empty recall/prior unless `allowEmptyRecall` is true. src/runtime/agent/orchestrator/session/compact/engine.mjs:653\n30. Recall-fasttrack raises budget for mandatory preserved context the same way semantic does. src/runtime/agent/orchestrator/session/compact/engine.mjs:658\n31. Recall-fasttrack then fits a recall summary message and validates final tokens against budget. src/runtime/agent/orchestrator/session/compact/engine.mjs:684\n32. In loop auto path, recall-fasttrack failure is logged and rethrown; semantic fallback is not attempted there. src/runtime/agent/orchestrator/session/loop.mjs:429\n33. In manual/session runner, recall-fasttrack failure can degrade to semantic once if semantic is enabled and provider is available. src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:237\n34. Semantic/manual non-recall failure falls back to anchor-independent tool-output pruning if the salvaged transcript fits the effective budget. src/runtime/agent/orchestrator/session/loop.mjs:501\n35. Session-runner mirrors the same non-recall prune fallback. src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:322\n36. If no fallback succeeds in the loop, telemetry/event records `overflow_failed` and raises `AgentContextOverflowError`. src/runtime/agent/orchestrator/session/loop.mjs:552\n37. `AgentContextOverflowError` reports stage, budget, reserve, context window, estimate, and original cause. src/runtime/agent/orchestrator/session/loop/context-overflow.mjs:4\n38. On success, loop replaces messages, clears provider state, resets iteration count, records telemetry, and emits compact event. src/runtime/agent/orchestrator/session/loop.mjs:631\n```",
+      "card": {
+        "session": "sess_12172_1783059815886_1783059816130_c1422313e74ff647937bfb7ccc713170",
+        "provider": "openai-oauth",
+        "model": "gpt-5.5",
+        "wall_ms": 73427,
+        "turns": 9,
+        "tool_calls": 11,
+        "tools_per_turn": 1.2,
+        "total_tool_ms": 1049,
+        "llm_stream_ms": 72134,
+        "cache_ratio": 0.8001111265453535,
+        "cached_tokens": 207360,
+        "prompt_tokens": 259164,
+        "uncached_tokens": 51804,
+        "output_tokens": 4631,
+        "cache_weighted_input_10": 72540,
+        "cache_weighted_input_25": 103644,
+        "prompt_growth": 45252,
+        "antipatterns": 1,
+        "issues": 14,
+        "issue_types": [
+          "cache",
+          "cache_break",
+          "readonly_stall",
+          "service_tier",
+          "missed_parallelism",
+          "tool_churn_cluster"
+        ],
+        "read_fragmentation_paths": [],
+        "grep_sweep_count": 0
+      }
+    },
+    {
+      "id": "r4-debug-forwarding-gap#2",
+      "agent": "reviewer",
+      "ok": true,
+      "ms": 119286,
+      "sessionId": "sess_44152_1783059834394_1783059834618_56b4c9c66902e5730af3a1b5d05d85c4",
+      "output": "sessionId: sess_44152_1783059834394_1783059834618_56b4c9c66902e5730af3a1b5d05d85c4\n요청 범위의 핵심 파일을 먼저 확인해 활성화부터 첫 전달까지 흐름을 추적하겠습니다.Done — traced activation to first forward; one path can silently skip existing output, while most startup races are delayed and caught up.\n\n- **Highest impact: EOF baseline on first existing transcript** — activation uses `catchUpFromPersisted`, but if no matching persisted cursor exists, `setContext()` starts at `currentSize`, so assistant bytes already in the transcript are marked read and not replayed. Backlog catch-up: **No**. Anchors: `src/runtime/channels/index.mjs:625`, `src/runtime/channels/lib/output-forwarder.mjs:117`.\n\n- **Missing transcript file** — `fs.watch()` cannot attach to a not-yet-created `.jsonl`, so output waits for the re-arm poll; if the file appears within 60s, `forwardNewText()` replays from the saved cursor/start. Backlog catch-up: **Yes within 60s; no automatic recovery after timeout**. Anchors: `src/runtime/channels/lib/output-forwarder.mjs:647`, `src/runtime/channels/index.mjs:357`.\n\n- **Bridge active before backend/ownership ready** — `bridge-state.json` is written before `backend.connect()` and `bridgeRuntimeConnected=true`; early flushes fail owner probe and read nothing. Backlog catch-up: **Yes on normal connect via post-connect `forwardNewText()`/re-arm; no if this process loses ownership**. Anchors: `src/runtime/channels/index.mjs:1735`, `src/runtime/channels/index.mjs:776`, `src/runtime/channels/lib/output-forwarder.mjs:351`.\n\n- **No channel binding** — if persisted status and main-channel fallback cannot provide a channel, binding returns and `forwardNewText()` no-ops. Backlog catch-up: **No automatic recovery until a later inbound establishes a channel, and existing bytes may then hit the EOF-baseline issue**. Anchors: `src/runtime/channels/index.mjs:610`, `src/runtime/channels/lib/output-forwarder.mjs:353`.\n\n- **Covered micro-gap** — bytes written between `setContext()` stat and `fs.watch()` install are explicitly flushed after watcher install. Backlog catch-up: **Yes**. Anchor: `src/runtime/channels/lib/output-forwarder.mjs:665`.\n\nVerdict: the bridge is mostly self-healing for startup timing races, but activation can still silently drop pre-bind output when an existing transcript is first bound without a matching persisted cursor.",
+      "card": {
+        "session": "sess_44152_1783059834394_1783059834618_56b4c9c66902e5730af3a1b5d05d85c4",
+        "provider": "openai-oauth",
+        "model": "gpt-5.5",
+        "wall_ms": 108812,
+        "turns": 6,
+        "tool_calls": 5,
+        "tools_per_turn": 0.8,
+        "total_tool_ms": 304,
+        "llm_stream_ms": 107642,
+        "cache_ratio": 0.7334463495955811,
+        "cached_tokens": 117248,
+        "prompt_tokens": 159859,
+        "uncached_tokens": 42611,
+        "output_tokens": 10625,
+        "cache_weighted_input_10": 54335.8,
+        "cache_weighted_input_25": 71923,
+        "prompt_growth": 36198,
+        "antipatterns": 1,
+        "issues": 11,
+        "issue_types": [
+          "cache",
+          "cache_break",
+          "broad_tool_result",
+          "service_tier",
+          "missed_parallelism",
+          "tool_churn_cluster"
+        ],
+        "read_fragmentation_paths": [],
+        "grep_sweep_count": 0
+      }
+    },
+    {
+      "id": "r4-impl-usage-summary#2",
+      "agent": "worker",
+      "ok": true,
+      "ms": 58353,
+      "sessionId": "sess_25912_1783059899801_1783059900029_f989eab921374da513570e09d7d39c55",
+      "output": "sessionId: sess_25912_1783059899801_1783059900029_f989eab921374da513570e09d7d39c55\n요청 범위가 명확해 바로 구현하고 검증까지 진행하겠습니다.이미 같은 이름의 스크립트가 있어 현재 구현 상태를 기준으로 누락 여부와 diff를 확인하겠습니다.구현 파일이 이미 작업트리에 있어 요구사항 충족 여부를 검증하고 diff를 정리하겠습니다.완료했습니다. `usage:today` 스크립트와 `scripts/usage-today.mjs`가 추가되어 오늘 사용량을 모델별/총합으로 출력합니다.\n\n- **Diff 요약**: `package.json`에 `usage:today` 추가, `scripts/usage-today.mjs` 신규 추가.\n- **검증**: `node --check scripts/usage-today.mjs` 통과, `npm run usage:today` 실행 완료.\n- **참고**: 작업 전부터 `src/runtime/channels/lib/event-queue.mjs`, `src/runtime/channels/lib/output-forwarder.mjs`에 별도 수정이 있었습니다.",
+      "card": {
+        "session": "sess_25912_1783059899801_1783059900029_f989eab921374da513570e09d7d39c55",
+        "provider": "openai-oauth",
+        "model": "gpt-5.5",
+        "wall_ms": 47887,
+        "turns": 6,
+        "tool_calls": 10,
+        "tools_per_turn": 1.7,
+        "total_tool_ms": 2312,
+        "llm_stream_ms": 45195,
+        "cache_ratio": 0.6885261237663127,
+        "cached_tokens": 58880,
+        "prompt_tokens": 85516,
+        "uncached_tokens": 26636,
+        "output_tokens": 3571,
+        "cache_weighted_input_10": 32524,
+        "cache_weighted_input_25": 41356,
+        "prompt_growth": 16109,
+        "antipatterns": 0,
+        "issues": 8,
+        "issue_types": [
+          "cache",
+          "cache_break",
+          "service_tier"
+        ],
+        "read_fragmentation_paths": [],
+        "grep_sweep_count": 0
+      }
+    },
+    {
+      "id": "r4-explain-compact-engine#2",
+      "agent": "explore",
+      "ok": true,
+      "ms": 60242,
+      "sessionId": "sess_60436_1783059931841_1783059932082_99b61f2d2925940ee5554fa92042346c",
+      "output": "sessionId: sess_60436_1783059931841_1783059932082_99b61f2d2925940ee5554fa92042346c\n완료했습니다. compaction subsystem 아키텍처 브리프입니다.\n\n- **Entry points**: public facade exports `semanticCompactMessages`/`recallFastTrackCompactMessages` from `compact.mjs`, backed by `engine.mjs` exports at `src/runtime/agent/orchestrator/session/compact.mjs:52-54`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:278`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:445`.\n\n- **Callers**: pre-send agent loop calls recall-fasttrack or semantic compaction at `src/runtime/agent/orchestrator/session/loop.mjs:414-459`; manual/clear/session compaction uses `runSessionCompaction` via manager at `src/runtime/agent/orchestrator/session/manager.mjs:1849`, `src/runtime/agent/orchestrator/session/manager.mjs:1950`.\n\n- **Fire decision**: loop drains queued steering before measuring, resolves worker policy, estimates transcript tokens, then calls `shouldCompactForSession` at `src/runtime/agent/orchestrator/session/loop.mjs:324-353`.\n\n- **Thresholds**: auto compaction is disabled by `cfg.auto === false` or env flag, otherwise boundary/trigger/reserve/type/tail settings are derived at `src/runtime/agent/orchestrator/session/loop/compact-policy.mjs:90-148`.\n\n- **Trigger rule**: compaction fires on reactive overflow, unknown token estimate, or `messageTokens + reserveTokens >= triggerTokens` at `src/runtime/agent/orchestrator/session/loop/compact-policy.mjs:170-174`.\n\n- **Path choice**: loop prefers recall-fasttrack when policy says so, otherwise semantic, otherwise throws unavailable compact type at `src/runtime/agent/orchestrator/session/loop.mjs:414-498`.\n\n- **Semantic window**: engine sanitizes, splits protected/system prefix, previous summary, and live non-summary messages at `src/runtime/agent/orchestrator/session/compact/engine.mjs:100-111`.\n\n- **Semantic tail**: if user turns exist, it keeps up to configured recent turns within preserved budget; if no user turns, it cuts on non-`tool` boundaries to keep provider-valid pairs at `src/runtime/agent/orchestrator/session/compact/engine.mjs:134-185`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:193-233`.\n\n- **Tail repair**: selected tail is sanitized/deduped/reconciled after the cut so orphan tool pairs are repaired at `src/runtime/agent/orchestrator/session/compact/engine.mjs:244-250`.\n\n- **Recall tail**: recall-fasttrack preserves newest turns, drops oldest turns until under cap, truncates the oldest surviving content if needed, then returns `head` before the tail at `src/runtime/agent/orchestrator/session/compact/engine.mjs:593-619`.\n\n- **Redaction hook**: both semantic and recall paths redact tool-call secrets after no-op budget check and before window/tail selection at `src/runtime/agent/orchestrator/session/compact/engine.mjs:306-314`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:644-651`.\n\n- **Redaction scope**: redaction targets assistant `toolCalls` argument values, with sensitive key/raw-string patterns handled in text utils at `src/runtime/agent/orchestrator/session/compact/text-utils.mjs:12-17`, `src/runtime/agent/orchestrator/session/compact/text-utils.mjs:121-127`, `src/runtime/agent/orchestrator/session/compact/text-utils.mjs:321-323`.\n\n- **Semantic failure checks**: missing provider, no compactable head, prompt fit failure, empty provider summary, summary fit failure, or final budget overflow throw from `src/runtime/agent/orchestrator/session/compact/engine.mjs:279-280`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:315-317`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:335-386`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:394-397`.\n\n- **Recall failure checks**: no compactable prior/recall, empty recall, summary fit failure, or final budget overflow throw from `src/runtime/agent/orchestrator/session/compact/engine.mjs:653-668`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:684-699`.\n\n- **Loop fallback**: semantic failures may recover with non-LLM unanchored tool-output prune; recall failures skip that fallback and escalate to overflow at `src/runtime/agent/orchestrator/session/loop.mjs:501-538`, `src/runtime/agent/orchestrator/session/loop.mjs:546-626`.\n\n- **Manager fallback**: session/manual compaction degrades recall-fasttrack to semantic once, then non-recall failures may prune; unrecovered failures are recorded and returned without mutating messages at `src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:237-270`, `src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:322-397`.\n\n- **Successful mutation**: loop replaces in-memory messages and clears provider state; manager writes `session.messages`, clears `session.providerState`, and records compact metadata at `src/runtime/agent/orchestrator/session/loop.mjs:630-647`, `src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:410-445`.",
+      "card": {
+        "session": "sess_60436_1783059931841_1783059932082_99b61f2d2925940ee5554fa92042346c",
+        "provider": "openai-oauth",
+        "model": "gpt-5.5",
+        "wall_ms": 49755,
+        "turns": 6,
+        "tool_calls": 7,
+        "tools_per_turn": 1.2,
+        "total_tool_ms": 656,
+        "llm_stream_ms": 48561,
+        "cache_ratio": 0.5869960043588812,
+        "cached_tokens": 51712,
+        "prompt_tokens": 88096,
+        "uncached_tokens": 36384,
+        "output_tokens": 3580,
+        "cache_weighted_input_10": 41555.2,
+        "cache_weighted_input_25": 49312,
+        "prompt_growth": 30432,
+        "antipatterns": 1,
+        "issues": 10,
+        "issue_types": [
+          "cache",
+          "cache_break",
+          "broad_tool_result",
+          "service_tier",
+          "tool_churn_cluster"
+        ],
+        "read_fragmentation_paths": [],
+        "grep_sweep_count": 0
+      }
+    }
+  ],
+  "task_complete": true,
+  "task_errors": [],
+  "score_complete": true,
+  "score_errors": [],
+  "group": {
+    "n": 6,
+    "wall_ms": 65584.7,
+    "turns": 5.5,
+    "tool_calls": 6.7,
+    "tools_per_turn": 1.1,
+    "total_tool_ms": 793.3,
+    "llm_stream_ms": 64206.5,
+    "cache_ratio": 0.5,
+    "cached_tokens": 76373.3,
+    "prompt_tokens": 116121.3,
+    "uncached_tokens": 39748,
+    "output_tokens": 5687,
+    "cache_weighted_input_10": 47385.3,
+    "cache_weighted_input_25": 58841.3,
+    "prompt_growth": 27712.5,
+    "antipatterns": 0.5,
+    "issues": 9.3,
+    "grep_sweep_count": 0
+  },
+  "cards": [
+    {
+      "session": "sess_55884_1783059815805_1783059816044_1d290d1d86ad152644020350714e0b6a",
+      "provider": "openai-oauth",
+      "model": "gpt-5.5",
+      "wall_ms": 105545,
+      "turns": 5,
+      "tool_calls": 7,
+      "tools_per_turn": 1.4,
+      "total_tool_ms": 439,
+      "llm_stream_ms": 104528,
+      "cache_ratio": 0.23191906990789674,
+      "cached_tokens": 23040,
+      "prompt_tokens": 99345,
+      "uncached_tokens": 76305,
+      "output_tokens": 11249,
+      "cache_weighted_input_10": 78609,
+      "cache_weighted_input_25": 82065,
+      "prompt_growth": 38284,
+      "antipatterns": 0,
+      "issues": 10,
+      "issue_types": [
+        "cache",
+        "cache_break",
+        "cache_miss",
+        "broad_tool_result",
+        "service_tier"
+      ],
+      "read_fragmentation_paths": [],
+      "grep_sweep_count": 0
+    },
+    {
+      "session": "sess_48508_1783059815820_1783059816056_cebabb220978834570309408a4932033",
+      "provider": "openai-oauth",
+      "model": "gpt-5.5",
+      "wall_ms": 8082,
+      "turns": 1,
+      "tool_calls": 0,
+      "tools_per_turn": 0,
+      "total_tool_ms": 0,
+      "llm_stream_ms": 7179,
+      "cache_ratio": 0,
+      "cached_tokens": 0,
+      "prompt_tokens": 4748,
+      "uncached_tokens": 4748,
+      "output_tokens": 466,
+      "cache_weighted_input_10": 4748,
+      "cache_weighted_input_25": 4748,
+      "prompt_growth": 0,
+      "antipatterns": 0,
+      "issues": 3,
+      "issue_types": [
+        "cache",
+        "cache_break",
+        "service_tier"
+      ],
+      "read_fragmentation_paths": [],
+      "grep_sweep_count": 0
+    },
+    {
+      "session": "sess_12172_1783059815886_1783059816130_c1422313e74ff647937bfb7ccc713170",
+      "provider": "openai-oauth",
+      "model": "gpt-5.5",
+      "wall_ms": 73427,
+      "turns": 9,
+      "tool_calls": 11,
+      "tools_per_turn": 1.2,
+      "total_tool_ms": 1049,
+      "llm_stream_ms": 72134,
+      "cache_ratio": 0.8001111265453535,
+      "cached_tokens": 207360,
+      "prompt_tokens": 259164,
+      "uncached_tokens": 51804,
+      "output_tokens": 4631,
+      "cache_weighted_input_10": 72540,
+      "cache_weighted_input_25": 103644,
+      "prompt_growth": 45252,
+      "antipatterns": 1,
+      "issues": 14,
+      "issue_types": [
+        "cache",
+        "cache_break",
+        "readonly_stall",
+        "service_tier",
+        "missed_parallelism",
+        "tool_churn_cluster"
+      ],
+      "read_fragmentation_paths": [],
+      "grep_sweep_count": 0
+    },
+    {
+      "session": "sess_44152_1783059834394_1783059834618_56b4c9c66902e5730af3a1b5d05d85c4",
+      "provider": "openai-oauth",
+      "model": "gpt-5.5",
+      "wall_ms": 108812,
+      "turns": 6,
+      "tool_calls": 5,
+      "tools_per_turn": 0.8,
+      "total_tool_ms": 304,
+      "llm_stream_ms": 107642,
+      "cache_ratio": 0.7334463495955811,
+      "cached_tokens": 117248,
+      "prompt_tokens": 159859,
+      "uncached_tokens": 42611,
+      "output_tokens": 10625,
+      "cache_weighted_input_10": 54335.8,
+      "cache_weighted_input_25": 71923,
+      "prompt_growth": 36198,
+      "antipatterns": 1,
+      "issues": 11,
+      "issue_types": [
+        "cache",
+        "cache_break",
+        "broad_tool_result",
+        "service_tier",
+        "missed_parallelism",
+        "tool_churn_cluster"
+      ],
+      "read_fragmentation_paths": [],
+      "grep_sweep_count": 0
+    },
+    {
+      "session": "sess_25912_1783059899801_1783059900029_f989eab921374da513570e09d7d39c55",
+      "provider": "openai-oauth",
+      "model": "gpt-5.5",
+      "wall_ms": 47887,
+      "turns": 6,
+      "tool_calls": 10,
+      "tools_per_turn": 1.7,
+      "total_tool_ms": 2312,
+      "llm_stream_ms": 45195,
+      "cache_ratio": 0.6885261237663127,
+      "cached_tokens": 58880,
+      "prompt_tokens": 85516,
+      "uncached_tokens": 26636,
+      "output_tokens": 3571,
+      "cache_weighted_input_10": 32524,
+      "cache_weighted_input_25": 41356,
+      "prompt_growth": 16109,
+      "antipatterns": 0,
+      "issues": 8,
+      "issue_types": [
+        "cache",
+        "cache_break",
+        "service_tier"
+      ],
+      "read_fragmentation_paths": [],
+      "grep_sweep_count": 0
+    },
+    {
+      "session": "sess_60436_1783059931841_1783059932082_99b61f2d2925940ee5554fa92042346c",
+      "provider": "openai-oauth",
+      "model": "gpt-5.5",
+      "wall_ms": 49755,
+      "turns": 6,
+      "tool_calls": 7,
+      "tools_per_turn": 1.2,
+      "total_tool_ms": 656,
+      "llm_stream_ms": 48561,
+      "cache_ratio": 0.5869960043588812,
+      "cached_tokens": 51712,
+      "prompt_tokens": 88096,
+      "uncached_tokens": 36384,
+      "output_tokens": 3580,
+      "cache_weighted_input_10": 41555.2,
+      "cache_weighted_input_25": 49312,
+      "prompt_growth": 30432,
+      "antipatterns": 1,
+      "issues": 10,
+      "issue_types": [
+        "cache",
+        "cache_break",
+        "broad_tool_result",
+        "service_tier",
+        "tool_churn_cluster"
+      ],
+      "read_fragmentation_paths": [],
+      "grep_sweep_count": 0
+    }
+  ]
+}