mixdog 0.9.3 → 0.9.5

package/scripts/bench/round-r15-fulldefault.json

@@ -0,0 +1,462 @@
+{
+  "round": "15",
+  "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_20052_1783058612106_1783058612356_363946685eccb22a98de415cf55e2b6b",
+    "sess_6636_1783058612119_1783058612368_878b7998c6a201833b0d90d7db428756",
+    "sess_32108_1783058612133_1783058612379_16813f439160d71cffa545f046f32444",
+    "sess_60200_1783058729488_1783058729715_0addf90688066bfe101c1e912ae350eb",
+    "sess_54744_1783058755944_1783058756177_365b216b85254be92613100c34872249",
+    "sess_54388_1783058842867_1783058843114_396cad8719025805d423bfec1e2af48c"
+  ],
+  "results": [
+    {
+      "id": "r4-debug-forwarding-gap#1",
+      "agent": "reviewer",
+      "ok": true,
+      "ms": 143838,
+      "sessionId": "sess_20052_1783058612106_1783058612356_363946685eccb22a98de415cf55e2b6b",
+      "output": "sessionId: sess_20052_1783058612106_1783058612356_363946685eccb22a98de415cf55e2b6b\nFindings ordered by impact:\n\n- **Permanent skip on first bind to an existing, non-persisted transcript** — `OutputForwarder.setContext()` seeds `readFileSize` to current EOF when `catchUpFromPersisted` has no same-transcript cursor, so assistant bytes already written before binding are treated as already read; backlog catch-up does **not** recover them. Anchors: `src/runtime/channels/lib/output-forwarder.mjs:117`, `src/runtime/channels/lib/output-forwarder.mjs:122`, `src/runtime/channels/lib/output-forwarder.mjs:126`, `src/runtime/channels/index.mjs:625`, `src/runtime/channels/index.mjs:441`.\n\n- **Bridge active before runtime ownership/connection is ready** — activation writes `bridge-state.json` before `refreshBridgeOwnership()` completes, while `_isOwner()` stays false until `bridgeRuntimeConnected`; output during this interval is delayed, not read/advanced, and is recovered by the post-connect `forwardNewText()` path if the cursor was not already moved to EOF by the bind case above. Anchors: `src/runtime/channels/index.mjs:1735`, `src/runtime/channels/index.mjs:1744`, `src/runtime/channels/index.mjs:776`, `src/runtime/channels/lib/output-forwarder.mjs:351`, `src/runtime/channels/index.mjs:809`.\n\n- **Transcript path known but file not created yet** — `fs.watch()` fails on the missing file, so creation would be missed without the pending re-arm poll; backlog is recovered when the file appears within the 60s poll window, from offset 0 for a new transcript or from the persisted cursor for the same transcript. Anchors: `src/runtime/channels/lib/output-forwarder.mjs:647`, `src/runtime/channels/lib/output-forwarder.mjs:667`, `src/runtime/channels/index.mjs:357`, `src/runtime/channels/index.mjs:376`, `src/runtime/channels/index.mjs:485`.\n\n- **Pre-connect re-arm tick can fire while ownership is still false** — the poll exits when `bridgeRuntimeConnected` is false, so first output is delayed until the post-connect repair re-arms or forwards; recovered after connect, but if the file is still absent and appears after the 60s re-arm deadline, it remains delayed until another rebind/trigger. Anchors: `src/runtime/channels/index.mjs:372`, `src/runtime/channels/index.mjs:384`, `src/runtime/channels/index.mjs:803`, `src/runtime/channels/index.mjs:807`.\n\n- **Watch flush before ownership is established** — `startWatch()` schedules a flush immediately, but `forwardNewText()` returns before reading while `_isOwner()` is false; backlog is recovered by the explicit post-connect `forwardNewText()` and no cursor is advanced during the failed probe. Anchors: `src/runtime/channels/lib/output-forwarder.mjs:665`, `src/runtime/channels/lib/output-forwarder.mjs:724`, `src/runtime/channels/lib/output-forwarder.mjs:351`, `src/runtime/channels/index.mjs:809`.\n\nVerdict: the bridge mostly delays and catches up, but an existing current transcript with no same persisted cursor can still silently drop pre-bind assistant output by binding at EOF.",
+      "card": {
+        "session": "sess_20052_1783058612106_1783058612356_363946685eccb22a98de415cf55e2b6b",
+        "provider": "openai-oauth",
+        "model": "gpt-5.5",
+        "wall_ms": 133321,
+        "turns": 7,
+        "tool_calls": 9,
+        "tools_per_turn": 1.3,
+        "total_tool_ms": 776,
+        "llm_stream_ms": 131748,
+        "cache_ratio": 0.44157442640570693,
+        "cached_tokens": 101888,
+        "prompt_tokens": 230738,
+        "uncached_tokens": 128850,
+        "output_tokens": 9586,
+        "cache_weighted_input_10": 139038.8,
+        "cache_weighted_input_25": 154322,
+        "prompt_growth": 44194,
+        "antipatterns": 0,
+        "issues": 11,
+        "issue_types": [
+          "cache",
+          "cache_break",
+          "cache_miss",
+          "service_tier"
+        ],
+        "read_fragmentation_paths": [],
+        "grep_sweep_count": 0
+      }
+    },
+    {
+      "id": "r4-impl-usage-summary#1",
+      "agent": "worker",
+      "ok": true,
+      "ms": 263872,
+      "sessionId": "sess_6636_1783058612119_1783058612368_878b7998c6a201833b0d90d7db428756",
+      "output": "sessionId: sess_6636_1783058612119_1783058612368_878b7998c6a201833b0d90d7db428756\n요청 범위가 명확해 바로 구현 후 검증까지 진행하겠습니다.이미 같은 이름의 스크립트가 있어 요구사항과 맞는지 보정한 뒤 검증하겠습니다.현재 구현이 이미 잡혀 있어 검증과 실제 diff 확인으로 마무리하겠습니다.검증은 통과했고, 추적 상태만 확인해 최종 보고하겠습니다.재영님, 완료했습니다 — `usage:today` 스크립트와 `scripts/usage-today.mjs` 구현이 준비됐고 검증도 통과했습니다.\n\n- **Diff 요약**: `package.json`에 `\"usage:today\": \"node scripts/usage-today.mjs\"` 추가, `scripts/usage-today.mjs`는 오늘 자정 이후 usage rows를 모델별/총합으로 집계합니다.\n\n- **검증**: `node --check scripts/usage-today.mjs` 통과, `npm run usage:today` 실행 완료.\n\n- **실행 결과**: 현재 `MIXDOG_DATA_DIR` 기준 trace에서 `GRAND TOTAL`은 `1371 calls`, `51.03M input`, `49.76M cached`, `1.62M output`, `$89.6468`였습니다.\n\n- **작업 상태**: `package.json`은 수정됨, `scripts/usage-today.mjs`는 새 untracked 파일입니다.",
+      "card": {
+        "session": "sess_6636_1783058612119_1783058612368_878b7998c6a201833b0d90d7db428756",
+        "provider": "openai-oauth",
+        "model": "gpt-5.5",
+        "wall_ms": 253360,
+        "turns": 9,
+        "tool_calls": 11,
+        "tools_per_turn": 1.2,
+        "total_tool_ms": 2515,
+        "llm_stream_ms": 65504,
+        "cache_ratio": 0.6846480771224215,
+        "cached_tokens": 132096,
+        "prompt_tokens": 192940,
+        "uncached_tokens": 60844,
+        "output_tokens": 5124,
+        "cache_weighted_input_10": 74053.6,
+        "cache_weighted_input_25": 93868,
+        "prompt_growth": 27537,
+        "antipatterns": 1,
+        "issues": 14,
+        "issue_types": [
+          "cache",
+          "cache_break",
+          "cache_miss",
+          "service_tier",
+          "missed_parallelism",
+          "tool_churn_cluster"
+        ],
+        "read_fragmentation_paths": [],
+        "grep_sweep_count": 0
+      }
+    },
+    {
+      "id": "r4-explain-compact-engine#1",
+      "agent": "explore",
+      "ok": true,
+      "ms": 117362,
+      "sessionId": "sess_32108_1783058612133_1783058612379_16813f439160d71cffa545f046f32444",
+      "output": "sessionId: sess_32108_1783058612133_1783058612379_16813f439160d71cffa545f046f32444\n완료했습니다. compaction subsystem 요약은 아래와 같습니다.\n\n- **진입점**: public API는 `semanticCompactMessages()`와 `recallFastTrackCompactMessages()`이며 `compact.mjs`가 재수출합니다. `src/runtime/agent/orchestrator/session/compact/engine.mjs:278`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:445`, `src/runtime/agent/orchestrator/session/compact.mjs:52-55`\n\n- **주요 호출자**: active loop pre-send에서 recall-fasttrack 또는 semantic을 호출하고, manager runner는 manual/auto-clear compaction에서 같은 엔진을 호출합니다. `src/runtime/agent/orchestrator/session/loop.mjs:414-460`, `src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:207-224`, `src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:280-306`\n\n- **세션 경로**: `askSession()`은 `agentLoop()`로 들어가 pre-send compaction을 통과하며, `compactSessionMessages()`와 `clearSessionMessages()`는 `runSessionCompaction()`으로 수동 실행합니다. `src/runtime/agent/orchestrator/session/manager.mjs:1379-1400`, `src/runtime/agent/orchestrator/session/manager.mjs:1849-1856`, `src/runtime/agent/orchestrator/session/manager.mjs:1943-1957`\n\n- **발화 조건**: loop는 steering을 먼저 주입한 뒤 정책을 만들고, `messageTokensEst + reserveTokens >= triggerTokens`이면 compact합니다. `src/runtime/agent/orchestrator/session/loop.mjs:324-354`, `src/runtime/agent/orchestrator/session/loop/compact-policy.mjs:89-148`, `src/runtime/agent/orchestrator/session/loop/compact-policy.mjs:170-174`\n\n- **reactive 경로**: provider context overflow가 나면 한 번만 `reactiveOverflowRetryPending`을 세우고 다음 pre-send compact를 강제합니다. `src/runtime/agent/orchestrator/session/loop.mjs:982-1010`, `src/runtime/agent/orchestrator/session/loop.mjs:351-364`\n\n- **manual/auto-clear 조건**: runner는 closed/disabled/no-boundary를 빠져나가고, manual은 force이며 auto는 `pressureTokens < triggerTokens`면 “below threshold”로 종료합니다. `src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:143-180`\n\n- **semantic tail 선택**: 먼저 system/protected prefix, previous summary, live를 분리하고, user turn이 있으면 최근 `tailTurns`를 `preserveRecentBudget` 안에서 보존합니다. `src/runtime/agent/orchestrator/session/compact/engine.mjs:100-111`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:235-242`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:134-185`\n\n- **no-user tail 선택**: user turn이 없으면 tool result로 시작하지 않는 cut point만 허용하고, 가장 최근부터 예산에 맞는 suffix를 고릅니다. `src/runtime/agent/orchestrator/session/compact/engine.mjs:121-127`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:193-232`\n\n- **semantic window 결과**: 선택된 live prefix는 summary 대상 head가 되고 suffix는 verbatim tail로 남으며, tail은 tool pair sanitize/dedup/reconcile을 거칩니다. `src/runtime/agent/orchestrator/session/compact/engine.mjs:244-267`\n\n- **recall tail 선택**: recall-fasttrack은 live에서 최근 최대 2 user-anchored turns를 보존하고 cap 초과 시 오래된 turn drop 후 단일 turn을 truncate합니다. `src/runtime/agent/orchestrator/session/compact/engine.mjs:467-471`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:593-611`\n\n- **recall no-user tail**: single-turn agent처럼 user가 없는 최신 assistant/tool 구조도 valid tail로 보존합니다. `src/runtime/agent/orchestrator/session/compact/engine.mjs:613-619`\n\n- **redaction hook**: semantic은 no-op이면 redaction 없이 sanitized transcript를 반환하고, 실제 compact 시에는 window selection 전에 tool-call secret 값을 redacts합니다. `src/runtime/agent/orchestrator/session/compact/engine.mjs:286-314`\n\n- **redaction 적용 범위**: recall-fasttrack도 split 전에 redacts하고, 보존된 system/tail은 redacted form으로 측정·출력됩니다. `src/runtime/agent/orchestrator/session/compact/engine.mjs:644-651`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:389-392`\n\n- **redaction 방식**: sensitive key 값과 raw string secret pair만 `[redacted]`로 바꾸며 shape, order, non-sensitive args는 보존합니다. `src/runtime/agent/orchestrator/session/compact/text-utils.mjs:219-229`, `src/runtime/agent/orchestrator/session/compact/text-utils.mjs:291-317`\n\n- **semantic 실패 조건**: provider 없음, empty live, head 없음, prompt fit 실패, empty provider summary, summary fit 실패, final budget 초과가 throw됩니다. `src/runtime/agent/orchestrator/session/compact/engine.mjs:278-280`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:254-256`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:315-336`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:371-397`\n\n- **recall 실패 조건**: compactable prior/recall text가 없거나 summary/final result가 budget에 맞지 않으면 throw됩니다. `src/runtime/agent/orchestrator/session/compact/engine.mjs:653-668`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:684-700`\n\n- **fallback**: loop/runner에서 semantic 실패는 `pruneToolOutputsUnanchored()`로 한 번 복구하지만 recall-fasttrack 실패에는 silent prune을 금지합니다. `src/runtime/agent/orchestrator/session/loop.mjs:500-537`, `src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:322-347`\n\n- **recall degraded fallback**: manager runner의 recall-fasttrack 실패는 semantic이 enabled이고 provider가 있으면 semantic compact를 한 번 시도합니다. `src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:231-279`\n\n- **hard failure 기록**: fallback도 실패하면 loop는 `overflow_failed` telemetry 후 `AgentContextOverflowError`를 던지고, runner는 `manual_failed`/`post_turn_failed`를 session.compaction에 남깁니다. `src/runtime/agent/orchestrator/session/loop.mjs:546-626`, `src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:348-397`",
+      "card": {
+        "session": "sess_32108_1783058612133_1783058612379_16813f439160d71cffa545f046f32444",
+        "provider": "openai-oauth",
+        "model": "gpt-5.5",
+        "wall_ms": 106845,
+        "turns": 8,
+        "tool_calls": 10,
+        "tools_per_turn": 1.3,
+        "total_tool_ms": 1070,
+        "llm_stream_ms": 105884,
+        "cache_ratio": 0.7214462101076363,
+        "cached_tokens": 153088,
+        "prompt_tokens": 212196,
+        "uncached_tokens": 59108,
+        "output_tokens": 4515,
+        "cache_weighted_input_10": 74416.8,
+        "cache_weighted_input_25": 97380,
+        "prompt_growth": 52013,
+        "antipatterns": 0,
+        "issues": 10,
+        "issue_types": [
+          "cache",
+          "cache_break",
+          "service_tier"
+        ],
+        "read_fragmentation_paths": [],
+        "grep_sweep_count": 0
+      }
+    },
+    {
+      "id": "r4-debug-forwarding-gap#2",
+      "agent": "reviewer",
+      "ok": true,
+      "ms": 113378,
+      "sessionId": "sess_60200_1783058729488_1783058729715_0addf90688066bfe101c1e912ae350eb",
+      "output": "sessionId: sess_60200_1783058729488_1783058729715_0addf90688066bfe101c1e912ae350eb\n조사 결과, 첫 전달 전 창에는 “기존 파일을 EOF로 바인딩”하는 영구 드롭 경로가 있고, 나머지는 대체로 지연 후 catch-up됩니다.\n\n1. **영구 드롭 가능성 — 기존 transcript가 새로 발견되면 EOF부터 시작합니다.** `rebindTranscriptContext()`가 활성화 복구에서 `catchUpFromPersisted`로 바인딩하고, persisted path가 다르면 `setContext()`가 `currentSize`로 커서를 둡니다. 이미 써진 출력은 읽지 않습니다. Catch-up: **복구 안 됨**. `src/runtime/channels/index.mjs:625`, `src/runtime/channels/index.mjs:438`, `src/runtime/channels/lib/output-forwarder.mjs:117`\n\n2. **not-yet-created transcript — `fs.watch` 설치 실패 창은 보정됩니다.** `startWatch()`는 없는 파일에서 catch/close되고, `schedulePendingTranscriptRearm()`가 파일 생성 후 watch를 다시 설치하고 `forwardNewText()`로 backlog를 읽습니다. Catch-up: **60초 내 생성이면 복구, timeout 후 자동 복구 없음**. `src/runtime/channels/index.mjs:339`, `src/runtime/channels/index.mjs:357`, `src/runtime/channels/index.mjs:376`, `src/runtime/channels/index.mjs:384`, `src/runtime/channels/lib/output-forwarder.mjs:647`\n\n3. **bridge-state 이후 backend connect 전 — ownership probe가 fail-closed라 전달이 지연됩니다.** 활성화는 `bridge-state.json`을 먼저 쓰고, binding task는 `backend.connect()` 전 시작되지만 `_isOwner()`는 `bridgeRuntimeConnected` 전까지 false입니다. 커서는 전진하지 않으며 post-connect `forwardNewText()`/rearm가 다시 읽습니다. Catch-up: **복구됨, 단 1번 EOF 바인딩이면 이미 누락**. `src/runtime/channels/index.mjs:1735`, `src/runtime/channels/index.mjs:758`, `src/runtime/channels/index.mjs:776`, `src/runtime/channels/index.mjs:801`, `src/runtime/channels/lib/output-forwarder.mjs:36`\n\n4. **stat 이후 watch 설치 전 쓰기 — 즉시 flush로 보정됩니다.** `setContext()`가 커서를 잡은 뒤 `startWatch()`가 설치되고, 설치 성공 시 `scheduleWatchFlush()`를 호출해 그 사이 추가된 bytes를 읽습니다. Catch-up: **복구됨**. `src/runtime/channels/lib/output-forwarder.mjs:112`, `src/runtime/channels/lib/output-forwarder.mjs:665`, `src/runtime/channels/lib/output-forwarder.mjs:703`, `src/runtime/channels/lib/output-forwarder.mjs:724`\n\n5. **channelId 미해결 — 전달 대상이 없으면 forwarding 자체가 시작되지 않습니다.** 활성 상태에서 persisted channel 또는 main channel fallback이 없으면 bind가 반환되고, `forwardNewText()`도 channel 없이 false입니다. Catch-up: **나중에 채널이 잡혀도 1번 커서 조건에 따라 조건부**. `src/runtime/channels/index.mjs:610`, `src/runtime/channels/index.mjs:624`, `src/runtime/channels/lib/output-forwarder.mjs:351`\n\nVerdict: **가장 위험한 실제 누락 경로는 “이미 존재하지만 persisted cursor가 없는 transcript를 EOF로 바인딩”하는 경우이며, missing-file/watch/ownership 창은 대부분 bounded catch-up으로 복구됩니다.**",
+      "card": {
+        "session": "sess_60200_1783058729488_1783058729715_0addf90688066bfe101c1e912ae350eb",
+        "provider": "openai-oauth",
+        "model": "gpt-5.5",
+        "wall_ms": 102888,
+        "turns": 8,
+        "tool_calls": 9,
+        "tools_per_turn": 1.1,
+        "total_tool_ms": 820,
+        "llm_stream_ms": 102098,
+        "cache_ratio": 0.7833074378747296,
+        "cached_tokens": 164352,
+        "prompt_tokens": 209818,
+        "uncached_tokens": 45466,
+        "output_tokens": 10858,
+        "cache_weighted_input_10": 61901.2,
+        "cache_weighted_input_25": 86554,
+        "prompt_growth": 38295,
+        "antipatterns": 1,
+        "issues": 12,
+        "issue_types": [
+          "cache",
+          "cache_break",
+          "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": 108658,
+      "sessionId": "sess_54744_1783058755944_1783058756177_365b216b85254be92613100c34872249",
+      "output": "sessionId: sess_54744_1783058755944_1783058756177_365b216b85254be92613100c34872249\n요청 범위가 명확해서 바로 구현하고 검증까지 진행하겠습니다.이미 같은 이름의 스크립트가 있어 요구사항 충족 여부를 확인한 뒤 필요한 보정만 하겠습니다.구현 내용은 이미 작업 트리에 반영되어 있어 문법 검사와 실제 실행으로 확인하겠습니다.검증은 통과했고, 마지막으로 실제 변경 diff만 확인하겠습니다.재영님, 완료되었습니다: `usage:today` 스크립트와 `scripts/usage-today.mjs`가 추가되어 오늘 사용량을 모델별/총합으로 출력합니다.\n\n- **Diff 요약**: `package.json`에 `\"usage:today\": \"node scripts/usage-today.mjs\"` 추가, 새 스크립트는 `MIXDOG_DATA_DIR` 우선으로 `history/agent-trace.jsonl`을 읽습니다.\n\n- **동작**: 로컬 자정 이후 `usage_raw`/`usage` 행을 필터링해 모델별 calls, input/cached/output tokens, `costUsd`와 `GRAND TOTAL`을 출력하며 `--trace`, `--json`도 지원합니다.\n\n- **검증**: `node --check scripts/usage-today.mjs` 통과, `npm run usage:today` 실행 완료로 총 1369 calls / `$89.6468` 출력 확인했습니다.",
+      "card": {
+        "session": "sess_54744_1783058755944_1783058756177_365b216b85254be92613100c34872249",
+        "provider": "openai-oauth",
+        "model": "gpt-5.5",
+        "wall_ms": 98171,
+        "turns": 11,
+        "tool_calls": 12,
+        "tools_per_turn": 1.1,
+        "total_tool_ms": 2881,
+        "llm_stream_ms": 95194,
+        "cache_ratio": 0.8461555996261597,
+        "cached_tokens": 148480,
+        "prompt_tokens": 175476,
+        "uncached_tokens": 26996,
+        "output_tokens": 5479,
+        "cache_weighted_input_10": 41844,
+        "cache_weighted_input_25": 64116,
+        "prompt_growth": 17107,
+        "antipatterns": 1,
+        "issues": 15,
+        "issue_types": [
+          "cache",
+          "cache_break",
+          "cache_miss",
+          "service_tier",
+          "missed_parallelism",
+          "tool_churn_cluster"
+        ],
+        "read_fragmentation_paths": [],
+        "grep_sweep_count": 0
+      }
+    },
+    {
+      "id": "r4-explain-compact-engine#2",
+      "agent": "explore",
+      "ok": true,
+      "ms": 85624,
+      "sessionId": "sess_54388_1783058842867_1783058843114_396cad8719025805d423bfec1e2af48c",
+      "output": "sessionId: sess_54388_1783058842867_1783058843114_396cad8719025805d423bfec1e2af48c\n완료했습니다 — `engine.mjs` 중심 compaction 서브시스템 맵입니다.\n\n- **엔트리/호출자:** 실시간 루프는 `askSession → agentLoop`로 들어오고 pre-send 직전에 compaction을 검사한다 (`src/runtime/agent/orchestrator/session/manager.mjs:1379`, `src/runtime/agent/orchestrator/session/loop.mjs:324`).\n- **수동/clear 경로:** `compactSessionMessages`와 clear-before-compact가 `runSessionCompaction(... force:true)`를 호출한다 (`src/runtime/agent/orchestrator/session/manager.mjs:1849`, `src/runtime/agent/orchestrator/session/manager.mjs:1950`).\n- **발화 조건:** auto가 켜진 policy만 검사하고, 토큰 추정+reserve 압력이 trigger/boundary 이상이거나 reactive overflow면 실행한다 (`src/runtime/agent/orchestrator/session/loop.mjs:329`, `src/runtime/agent/orchestrator/session/loop.mjs:351`, `src/runtime/agent/orchestrator/session/loop/compact-policy.mjs:170`).\n- **예산 결정:** compact 목표는 boundary 안에서 target+reserve로 잡고, 실행 시 `compactTargetBudget` 또는 boundary를 쓴다 (`src/runtime/agent/orchestrator/session/loop.mjs:362`, `src/runtime/agent/orchestrator/session/loop/compact-policy.mjs:163`).\n- **타입 분기:** policy가 recall-fasttrack이면 memory recall 기반 compaction, 아니면 semantic compaction으로 간다 (`src/runtime/agent/orchestrator/session/loop.mjs:414`, `src/runtime/agent/orchestrator/session/loop.mjs:439`).\n- **공통 전처리:** engine은 tool pair 정리, 중복 tool result 정리, protected/system prefix 분리, 이전 summary 탐색, live 메시지 추출을 먼저 한다 (`src/runtime/agent/orchestrator/session/compact/engine.mjs:100`).\n- **semantic tail:** user turn이 있으면 최근 `tailTurns`를 `preserveRecentBudget` 안에서 보존하고, 최신 turn이 너무 크면 suffix split을 시도한다 (`src/runtime/agent/orchestrator/session/compact/engine.mjs:134`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:151`).\n- **no-user semantic tail:** user turn이 없으면 tool result로 시작하지 않는 cut point를 최신→과거로 넓히며 budget에 맞는 tail을 고른다 (`src/runtime/agent/orchestrator/session/compact/engine.mjs:121`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:193`).\n- **recall tail:** recall-fasttrack은 live를 turn으로 나누고 최신 최대 2 user turns를 보존하되 cap 초과 시 오래된 turn을 버리고 마지막 turn을 truncate한다 (`src/runtime/agent/orchestrator/session/compact/engine.mjs:593`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:601`).\n- **tail budget:** 보존 recent budget은 explicit `preserveRecentTokens/keepTokens` 우선, 아니면 budget의 25%와 기본 keep token을 80% 상한 안에서 선택한다 (`src/runtime/agent/orchestrator/session/compact/budget.mjs:274`).\n- **redaction 위치:** fits-budget no-op은 원본 sanitized를 그대로 반환하지만, 실제 compaction은 window selection 전에 preserved system/tail까지 `redactToolCallSecretsInMessages`로 scrub한다 (`src/runtime/agent/orchestrator/session/compact/engine.mjs:286`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:306`).\n- **redaction 범위:** preserved tail redaction은 assistant `toolCalls` argument value만 shape-preserving으로 바꾸며, raw `authorization/password/token`류도 `[redacted]` 처리한다 (`src/runtime/agent/orchestrator/session/compact/text-utils.mjs:217`, `src/runtime/agent/orchestrator/session/compact/text-utils.mjs:323`).\n- **summary prompt:** semantic은 head+previous summary+preserved facts로 prompt를 만들고, tool call metadata는 prompt-side normalizer가 별도 redaction/truncation한다 (`src/runtime/agent/orchestrator/session/compact/summary.mjs:93`, `src/runtime/agent/orchestrator/session/compact/text-utils.mjs:121`).\n- **결과 조립:** semantic은 system + summary + redacted tail을 재정리하고 budget 초과 시 실패하며, recall도 system + recall summary + redacted tail을 같은 방식으로 검증한다 (`src/runtime/agent/orchestrator/session/compact/engine.mjs:392`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:695`).\n- **budget fallback:** mandatory system+tail이 목표 budget보다 크면 실패하지 않고 summary 최소 공간만큼 budget을 올린다 (`src/runtime/agent/orchestrator/session/compact/engine.mjs:328`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:661`).\n- **semantic 실패:** provider 없음, empty live, compactable prior 없음, prompt fit 실패, empty summary, summary fit 실패, final budget 초과가 hard error다 (`src/runtime/agent/orchestrator/session/compact/engine.mjs:279`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:315`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:335`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:372`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:385`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:395`).\n- **recall fallback:** manual/session runner에서 recall-fasttrack 실패 시 semantic이 가능하면 한 번 degraded fallback을 시도하고 성공한 경우만 history를 교체한다 (`src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:237`, `src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:245`).\n- **prune fallback:** semantic/non-recall 실패는 unanchored old tool output prune으로 한 번 salvage하고, recall-fasttrack 실패에는 이 prune을 적용하지 않는다 (`src/runtime/agent/orchestrator/session/loop.mjs:500`, `src/runtime/agent/orchestrator/session/loop.mjs:517`, `src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:322`).\n- **최종 실패:** 루프 pre-send 실패는 telemetry/event 기록 후 `AgentContextOverflowError`로 승격하고, runner 실패는 session compaction state에 `manual_failed/post_turn_failed`로 기록한다 (`src/runtime/agent/orchestrator/session/loop.mjs:546`, `src/runtime/agent/orchestrator/session/loop.mjs:618`, `src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:348`).",
+      "card": {
+        "session": "sess_54388_1783058842867_1783058843114_396cad8719025805d423bfec1e2af48c",
+        "provider": "openai-oauth",
+        "model": "gpt-5.5",
+        "wall_ms": 75113,
+        "turns": 8,
+        "tool_calls": 8,
+        "tools_per_turn": 1,
+        "total_tool_ms": 540,
+        "llm_stream_ms": 74441,
+        "cache_ratio": 0.5597193267910714,
+        "cached_tokens": 88064,
+        "prompt_tokens": 157336,
+        "uncached_tokens": 69272,
+        "output_tokens": 3912,
+        "cache_weighted_input_10": 78078.4,
+        "cache_weighted_input_25": 91288,
+        "prompt_growth": 44302,
+        "antipatterns": 1,
+        "issues": 15,
+        "issue_types": [
+          "cache",
+          "cache_break",
+          "cache_miss",
+          "broad_tool_result",
+          "service_tier",
+          "missed_parallelism",
+          "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": 128283,
+    "turns": 8.5,
+    "tool_calls": 9.8,
+    "tools_per_turn": 1.2,
+    "total_tool_ms": 1433.7,
+    "llm_stream_ms": 95811.5,
+    "cache_ratio": 0.7,
+    "cached_tokens": 131328,
+    "prompt_tokens": 196417.3,
+    "uncached_tokens": 65089.3,
+    "output_tokens": 6579,
+    "cache_weighted_input_10": 78222.1,
+    "cache_weighted_input_25": 97921.3,
+    "prompt_growth": 37241.3,
+    "antipatterns": 0.7,
+    "issues": 12.8,
+    "grep_sweep_count": 0
+  },
+  "cards": [
+    {
+      "session": "sess_20052_1783058612106_1783058612356_363946685eccb22a98de415cf55e2b6b",
+      "provider": "openai-oauth",
+      "model": "gpt-5.5",
+      "wall_ms": 133321,
+      "turns": 7,
+      "tool_calls": 9,
+      "tools_per_turn": 1.3,
+      "total_tool_ms": 776,
+      "llm_stream_ms": 131748,
+      "cache_ratio": 0.44157442640570693,
+      "cached_tokens": 101888,
+      "prompt_tokens": 230738,
+      "uncached_tokens": 128850,
+      "output_tokens": 9586,
+      "cache_weighted_input_10": 139038.8,
+      "cache_weighted_input_25": 154322,
+      "prompt_growth": 44194,
+      "antipatterns": 0,
+      "issues": 11,
+      "issue_types": [
+        "cache",
+        "cache_break",
+        "cache_miss",
+        "service_tier"
+      ],
+      "read_fragmentation_paths": [],
+      "grep_sweep_count": 0
+    },
+    {
+      "session": "sess_6636_1783058612119_1783058612368_878b7998c6a201833b0d90d7db428756",
+      "provider": "openai-oauth",
+      "model": "gpt-5.5",
+      "wall_ms": 253360,
+      "turns": 9,
+      "tool_calls": 11,
+      "tools_per_turn": 1.2,
+      "total_tool_ms": 2515,
+      "llm_stream_ms": 65504,
+      "cache_ratio": 0.6846480771224215,
+      "cached_tokens": 132096,
+      "prompt_tokens": 192940,
+      "uncached_tokens": 60844,
+      "output_tokens": 5124,
+      "cache_weighted_input_10": 74053.6,
+      "cache_weighted_input_25": 93868,
+      "prompt_growth": 27537,
+      "antipatterns": 1,
+      "issues": 14,
+      "issue_types": [
+        "cache",
+        "cache_break",
+        "cache_miss",
+        "service_tier",
+        "missed_parallelism",
+        "tool_churn_cluster"
+      ],
+      "read_fragmentation_paths": [],
+      "grep_sweep_count": 0
+    },
+    {
+      "session": "sess_32108_1783058612133_1783058612379_16813f439160d71cffa545f046f32444",
+      "provider": "openai-oauth",
+      "model": "gpt-5.5",
+      "wall_ms": 106845,
+      "turns": 8,
+      "tool_calls": 10,
+      "tools_per_turn": 1.3,
+      "total_tool_ms": 1070,
+      "llm_stream_ms": 105884,
+      "cache_ratio": 0.7214462101076363,
+      "cached_tokens": 153088,
+      "prompt_tokens": 212196,
+      "uncached_tokens": 59108,
+      "output_tokens": 4515,
+      "cache_weighted_input_10": 74416.8,
+      "cache_weighted_input_25": 97380,
+      "prompt_growth": 52013,
+      "antipatterns": 0,
+      "issues": 10,
+      "issue_types": [
+        "cache",
+        "cache_break",
+        "service_tier"
+      ],
+      "read_fragmentation_paths": [],
+      "grep_sweep_count": 0
+    },
+    {
+      "session": "sess_60200_1783058729488_1783058729715_0addf90688066bfe101c1e912ae350eb",
+      "provider": "openai-oauth",
+      "model": "gpt-5.5",
+      "wall_ms": 102888,
+      "turns": 8,
+      "tool_calls": 9,
+      "tools_per_turn": 1.1,
+      "total_tool_ms": 820,
+      "llm_stream_ms": 102098,
+      "cache_ratio": 0.7833074378747296,
+      "cached_tokens": 164352,
+      "prompt_tokens": 209818,
+      "uncached_tokens": 45466,
+      "output_tokens": 10858,
+      "cache_weighted_input_10": 61901.2,
+      "cache_weighted_input_25": 86554,
+      "prompt_growth": 38295,
+      "antipatterns": 1,
+      "issues": 12,
+      "issue_types": [
+        "cache",
+        "cache_break",
+        "service_tier",
+        "missed_parallelism",
+        "tool_churn_cluster"
+      ],
+      "read_fragmentation_paths": [],
+      "grep_sweep_count": 0
+    },
+    {
+      "session": "sess_54744_1783058755944_1783058756177_365b216b85254be92613100c34872249",
+      "provider": "openai-oauth",
+      "model": "gpt-5.5",
+      "wall_ms": 98171,
+      "turns": 11,
+      "tool_calls": 12,
+      "tools_per_turn": 1.1,
+      "total_tool_ms": 2881,
+      "llm_stream_ms": 95194,
+      "cache_ratio": 0.8461555996261597,
+      "cached_tokens": 148480,
+      "prompt_tokens": 175476,
+      "uncached_tokens": 26996,
+      "output_tokens": 5479,
+      "cache_weighted_input_10": 41844,
+      "cache_weighted_input_25": 64116,
+      "prompt_growth": 17107,
+      "antipatterns": 1,
+      "issues": 15,
+      "issue_types": [
+        "cache",
+        "cache_break",
+        "cache_miss",
+        "service_tier",
+        "missed_parallelism",
+        "tool_churn_cluster"
+      ],
+      "read_fragmentation_paths": [],
+      "grep_sweep_count": 0
+    },
+    {
+      "session": "sess_54388_1783058842867_1783058843114_396cad8719025805d423bfec1e2af48c",
+      "provider": "openai-oauth",
+      "model": "gpt-5.5",
+      "wall_ms": 75113,
+      "turns": 8,
+      "tool_calls": 8,
+      "tools_per_turn": 1,
+      "total_tool_ms": 540,
+      "llm_stream_ms": 74441,
+      "cache_ratio": 0.5597193267910714,
+      "cached_tokens": 88064,
+      "prompt_tokens": 157336,
+      "uncached_tokens": 69272,
+      "output_tokens": 3912,
+      "cache_weighted_input_10": 78078.4,
+      "cache_weighted_input_25": 91288,
+      "prompt_growth": 44302,
+      "antipatterns": 1,
+      "issues": 15,
+      "issue_types": [
+        "cache",
+        "cache_break",
+        "cache_miss",
+        "broad_tool_result",
+        "service_tier",
+        "missed_parallelism",
+        "tool_churn_cluster"
+      ],
+      "read_fragmentation_paths": [],
+      "grep_sweep_count": 0
+    }
+  ]
+}