mixdog 0.9.4 → 0.9.6

package/scripts/bench/round-r25-serial.json

@@ -0,0 +1,464 @@
+{
+  "round": "25",
+  "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_53408_1783063642444_1783063642668_3cfee1069f85e2bb12e8d28e8818547c",
+    "sess_6356_1783063737425_1783063737661_1619e4bb11b96b6c6f8a2015039eca61",
+    "sess_54836_1783063801897_1783063802131_358c27f874c909a61e1055fa674cc778",
+    "sess_37028_1783063871406_1783063871636_34284b6e9fe43e25a279de7888dc1dc1",
+    "sess_35624_1783063944146_1783063944381_cbf25d084f0f52db125bb2a739ccdc6d",
+    "sess_47024_1783064101403_1783064101639_70166643f319fb0dc14080ac610d3d19"
+  ],
+  "results": [
+    {
+      "id": "r4-debug-forwarding-gap#1",
+      "agent": "reviewer",
+      "ok": true,
+      "ms": 94983,
+      "sessionId": "sess_53408_1783063642444_1783063642668_3cfee1069f85e2bb12e8d28e8818547c",
+      "output": "sessionId: sess_53408_1783063642444_1783063642668_3cfee1069f85e2bb12e8d28e8818547c\n조사 완료했습니다, 재영님. 활성화 후 첫 전달 경로는 `writeBridgeState(true)` → ownership claim/refresh → `bindPersistedTranscriptIfAny()` → `applyTranscriptBinding()` → `OutputForwarder.startWatch()` → `forwardNewText()` → queue drain/send 입니다.\n\n- **높음: 이미 생성된 transcript를 처음 바인딩하면 기존 출력이 EOF 처리됩니다.** `catchUpFromPersisted`가 같은 persisted path를 못 찾으면 `currentSize`에서 시작하므로, 바인딩 전에 기록된 assistant 출력은 backlog catch-up으로 복구되지 않습니다. `src/runtime/channels/lib/output-forwarder.mjs:117`, `src/runtime/channels/lib/output-forwarder.mjs:122`, `src/runtime/channels/lib/output-forwarder.mjs:126`\n\n- **중간: transcript가 아직 없으면 `fs.watch` 설치가 실패하지만 re-arm poll이 복구합니다.** missing file bind는 `startWatch()`가 catch/close되고, `schedulePendingTranscriptRearm()`이 생성 감지 후 `startWatch()`와 `forwardNewText()`를 호출해 backlog를 보냅니다. 단 60초 timeout 이후에는 다음 rebind 없이는 복구되지 않습니다. `src/runtime/channels/index.mjs:339`, `src/runtime/channels/index.mjs:376`, `src/runtime/channels/index.mjs:378`, `src/runtime/channels/index.mjs:384`\n\n- **중간: ownership/connect 전에는 forward가 fail-closed로 지연됩니다.** `_isOwner()`가 false면 `forwardNewText()`는 읽기 전 반환하므로 cursor는 안 밀리고, connect 후 `forwardNewText()` 또는 re-arm 재등록으로 transcript backlog는 복구됩니다. `src/runtime/channels/lib/output-forwarder.mjs:36`, `src/runtime/channels/lib/output-forwarder.mjs:351`, `src/runtime/channels/index.mjs:776`, `src/runtime/channels/index.mjs:803`\n\n- **낮음: watch 설치 직전/직후 기록은 지연만 됩니다.** `startWatch()`가 설치 직후 `scheduleWatchFlush()`를 호출하고, flush가 `forwardNewText()`로 읽어 보내므로 newline이 완성된 backlog는 복구됩니다. `src/runtime/channels/lib/output-forwarder.mjs:657`, `src/runtime/channels/lib/output-forwarder.mjs:665`, `src/runtime/channels/lib/output-forwarder.mjs:724`\n\n판정: 가장 큰 실제 손실 위험은 “처음 바인딩 시 이미 존재하는 transcript를 EOF로 잡는 경우”이며, missing-file/watch/ownership 창은 대부분 지연 후 catch-up됩니다.",
+      "card": {
+        "session": "sess_53408_1783063642444_1783063642668_3cfee1069f85e2bb12e8d28e8818547c",
+        "provider": "openai-oauth",
+        "model": "gpt-5.5",
+        "wall_ms": 84498,
+        "turns": 8,
+        "tool_calls": 9,
+        "tools_per_turn": 1.1,
+        "total_tool_ms": 780,
+        "llm_stream_ms": 83431,
+        "cache_ratio": 0.8006575913711181,
+        "cached_tokens": 199680,
+        "prompt_tokens": 249395,
+        "uncached_tokens": 49715,
+        "output_tokens": 7234,
+        "cache_weighted_input_10": 69683,
+        "cache_weighted_input_25": 99635,
+        "prompt_growth": 38233,
+        "antipatterns": 1,
+        "issues": 13,
+        "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#1",
+      "agent": "worker",
+      "ok": true,
+      "ms": 64470,
+      "sessionId": "sess_6356_1783063737425_1783063737661_1619e4bb11b96b6c6f8a2015039eca61",
+      "output": "sessionId: sess_6356_1783063737425_1783063737661_1619e4bb11b96b6c6f8a2015039eca61\n요청 범위가 작고 실행 승인까지 포함된 것으로 보고 바로 반영하겠습니다.이미 관련 파일이 있어 요구사항과 맞는지 확인한 뒤 필요한 부분만 정리하겠습니다.구현 상태가 이미 작업 트리에 반영되어 있어 검증만 진행합니다.재영님, `usage:today` 기능이 작업 트리에 반영되어 있고 검증까지 완료했습니다.\n\n- **Diff 요약**: `package.json`에 `usage:today` npm script 추가, `scripts/usage-today.mjs` 신규 추가.\n- **동작**: `MIXDOG_DATA_DIR`를 존중해 `history/agent-trace.jsonl`을 읽고, 로컬 자정 이후 usage를 모델별/총합으로 집계합니다.\n- **검증**: `node --check scripts/usage-today.mjs` 통과, `npm run usage:today` 실행 성공.\n- **참고**: 범위 밖 기존 수정 `src/runtime/channels/lib/event-queue.mjs`, `output-forwarder.mjs`는 건드리지 않았습니다.",
+      "card": {
+        "session": "sess_6356_1783063737425_1783063737661_1619e4bb11b96b6c6f8a2015039eca61",
+        "provider": "openai-oauth",
+        "model": "gpt-5.5",
+        "wall_ms": 53970,
+        "turns": 7,
+        "tool_calls": 12,
+        "tools_per_turn": 1.7,
+        "total_tool_ms": 2323,
+        "llm_stream_ms": 51748,
+        "cache_ratio": 0.7647431722461031,
+        "cached_tokens": 88064,
+        "prompt_tokens": 115155,
+        "uncached_tokens": 27091,
+        "output_tokens": 4646,
+        "cache_weighted_input_10": 35897.4,
+        "cache_weighted_input_25": 49107,
+        "prompt_growth": 12328,
+        "antipatterns": 0,
+        "issues": 9,
+        "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": 69508,
+      "sessionId": "sess_54836_1783063801897_1783063802131_358c27f874c909a61e1055fa674cc778",
+      "output": "sessionId: sess_54836_1783063801897_1783063802131_358c27f874c909a61e1055fa674cc778\n완료했습니다. `src/runtime/agent/orchestrator/session/compact/engine.mjs` 기준 compaction 구조 요약입니다.\n\n- **Public entry**: `compact.mjs` re-export가 기존 import surface를 유지하며 `semanticCompactMessages`/`recallFastTrackCompactMessages`를 노출한다 (`src/runtime/agent/orchestrator/session/compact.mjs:2`, `src/runtime/agent/orchestrator/session/compact.mjs:52`).\n- **Main callers**: pre-send auto loop가 recall-fasttrack 또는 semantic을 호출한다 (`src/runtime/agent/orchestrator/session/loop.mjs:414`, `src/runtime/agent/orchestrator/session/loop.mjs:441`).\n- **Manual/session callers**: `runSessionCompaction`이 manual/clear/auto-clear 경로에서 type별 engine을 호출한다 (`src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:143`, `src/runtime/agent/orchestrator/session/manager.mjs:1849`, `src/runtime/agent/orchestrator/session/manager.mjs:1950`).\n- **Trigger policy**: auto policy는 boundary/context/autoLimit에서 boundary를 정하고 buffer를 뺀 `triggerTokens`를 만든다 (`src/runtime/agent/orchestrator/session/loop/compact-policy.mjs:94`, `src/runtime/agent/orchestrator/session/loop/compact-policy.mjs:117`).\n- **When it fires**: pre-send에서 `messageTokens + reserveTokens >= triggerTokens`이거나 reactive overflow retry면 compaction이 실행된다 (`src/runtime/agent/orchestrator/session/loop.mjs:351`, `src/runtime/agent/orchestrator/session/loop/compact-policy.mjs:170`).\n- **Type choice**: agent-owned sessions are forced semantic; otherwise env/config selects semantic vs recall-fasttrack (`src/runtime/agent/orchestrator/session/loop/compact-policy.mjs:39`, `src/runtime/agent/orchestrator/session/loop/compact-policy.mjs:45`).\n- **No-op path**: both semantic and recall-fasttrack return sanitized original messages when already within effective budget and not forced (`src/runtime/agent/orchestrator/session/compact/engine.mjs:283`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:627`).\n- **Pre-window cleanup**: engine first sanitizes tool pairs, dedups tool results, reconciles stubs, and splits protected prefix from live conversation (`src/runtime/agent/orchestrator/session/compact/engine.mjs:100`, `src/runtime/agent/orchestrator/session/compact/messages.mjs:56`).\n- **Redaction hook**: if compaction proceeds, tool-call argument secrets are redacted before window selection, so preserved system/tail are measured and emitted redacted (`src/runtime/agent/orchestrator/session/compact/engine.mjs:306`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:312`).\n- **Redaction behavior**: redaction only touches assistant `toolCalls` argument values, preserving shape/order/pairing; sensitive keys and raw `key:value` secrets become `[redacted]` (`src/runtime/agent/orchestrator/session/compact/text-utils.mjs:217`, `src/runtime/agent/orchestrator/session/compact/text-utils.mjs:323`).\n- **Semantic tail choice**: with user turns, keep newest `tailTurns` turns within `preserveRecentBudget`, splitting newest turn suffix if needed (`src/runtime/agent/orchestrator/session/compact/engine.mjs:134`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:151`).\n- **No-user semantic tail**: without user turns, choose newest valid non-tool cut that fits; never start tail on a `tool` message (`src/runtime/agent/orchestrator/session/compact/engine.mjs:121`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:193`).\n- **Semantic window**: `selectCompactionWindow` returns protected system, compactable head, preserved tail, previous summary, and preserved facts (`src/runtime/agent/orchestrator/session/compact/engine.mjs:235`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:258`).\n- **Semantic execution**: head + previous summary are fit into a compaction prompt, provider is called with compact send options, response schema is repaired, then summary replaces head (`src/runtime/agent/orchestrator/session/compact/engine.mjs:333`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:367`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:376`).\n- **Mandatory budget fallback**: if system+tail plus minimum summary room exceeds target, working budget is raised instead of refusing compaction (`src/runtime/agent/orchestrator/session/compact/engine.mjs:319`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:328`).\n- **Recall pipeline caller**: recall-fasttrack hydrates/dumps memory roots, drains raw pending cycle1 rows, caps recall injection to 5% context floor, then calls engine (`src/runtime/agent/orchestrator/session/loop/recall-fasttrack.mjs:148`, `src/runtime/agent/orchestrator/session/loop/recall-fasttrack.mjs:157`, `src/runtime/agent/orchestrator/session/loop/recall-fasttrack.mjs:161`).\n- **Recall tail choice**: preserve newest two user-anchored turns by default, drop oldest turns to fit cap, then truncate oldest surviving content if still oversized (`src/runtime/agent/orchestrator/session/compact/engine.mjs:467`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:601`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:609`).\n- **Recall no-user tail**: no-user recent assistant/tool structure is valid and preserved; `sanitizeToolPairs` repairs tool-call/tool-result structure after trimming (`src/runtime/agent/orchestrator/session/compact/engine.mjs:613`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:617`).\n- **Recall summary**: recall text plus prior summary are fit into one synthetic summary message, then emitted as system + summary + preserved tail (`src/runtime/agent/orchestrator/session/compact/engine.mjs:652`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:684`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:695`).\n- **Hard engine failures**: semantic throws on missing provider, empty live window, no compactable head without prior summary, prompt/summary fit failure, empty provider summary, or final over-budget result (`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:395`).\n- **Recall hard failures**: recall-fasttrack throws on no compactable prior context without recall/prior, empty recall when not allowed, summary fit failure, or final over-budget result (`src/runtime/agent/orchestrator/session/compact/engine.mjs:653`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:666`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:698`).\n- **Loop fallback**: semantic auto failure can recover via `pruneToolOutputsUnanchored`; recall-fasttrack failures deliberately skip this prune fallback (`src/runtime/agent/orchestrator/session/loop.mjs:501`, `src/runtime/agent/orchestrator/session/loop.mjs:517`).\n- **Session fallback**: manual/session recall-fasttrack failure tries semantic once if enabled/provider exists; semantic/manual failure then may use unanchored prune (`src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:237`, `src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:327`).\n- **Failure persistence**: unrecovered failures record `lastStage`/errors and return unchanged messages; ask failure can persist sanitized compacted outgoing best-effort (`src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:348`, `src/runtime/agent/orchestrator/session/manager.mjs:1098`, `src/runtime/agent/orchestrator/session/manager.mjs:1134`).",
+      "card": {
+        "session": "sess_54836_1783063801897_1783063802131_358c27f874c909a61e1055fa674cc778",
+        "provider": "openai-oauth",
+        "model": "gpt-5.5",
+        "wall_ms": 59002,
+        "turns": 7,
+        "tool_calls": 7,
+        "tools_per_turn": 1,
+        "total_tool_ms": 931,
+        "llm_stream_ms": 57686,
+        "cache_ratio": 0.6654929577464789,
+        "cached_tokens": 96768,
+        "prompt_tokens": 145408,
+        "uncached_tokens": 48640,
+        "output_tokens": 4198,
+        "cache_weighted_input_10": 58316.8,
+        "cache_weighted_input_25": 72832,
+        "prompt_growth": 36676,
+        "antipatterns": 1,
+        "issues": 13,
+        "issue_types": [
+          "cache",
+          "cache_break",
+          "broad_tool_result",
+          "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": 72748,
+      "sessionId": "sess_37028_1783063871406_1783063871636_34284b6e9fe43e25a279de7888dc1dc1",
+      "output": "sessionId: sess_37028_1783063871406_1783063871636_34284b6e9fe43e25a279de7888dc1dc1\n재영님, 조사 완료했습니다. activation→첫 전송 경로에는 복구되는 지연 구간과 복구 안 되는 skip 구간이 섞여 있습니다.\n\n- **최대 영향: 기존 transcript를 EOF로 바인딩** — `setContext()`는 `catchUpFromPersisted`라도 persisted path가 다르면 현재 크기부터 읽어 이미 쓴 assistant 출력은 영구 skip됩니다; backlog catch-up **불가**. `src/runtime/channels/lib/output-forwarder.mjs:117`\n\n- **pre-connect ownership gate** — `bridge-state.json`은 먼저 쓰지만 `bridgeRuntimeConnected` 전에는 `_isOwner()`가 false라 watch flush/forward가 no-op입니다; post-connect `forwardNewText()` 또는 re-arm으로 backlog catch-up **가능**. `src/runtime/channels/index.mjs:1735`, `src/runtime/channels/index.mjs:776`, `src/runtime/channels/lib/output-forwarder.mjs:351`\n\n- **없는 transcript에 대한 fs.watch 실패** — `startWatch()`는 없는 파일에서 catch 후 닫히며, `schedulePendingTranscriptRearm()`이 250ms poll로 파일 생성 후 `forwardNewText()`를 호출합니다; 60초 안이면 backlog catch-up **가능**, timeout 뒤에는 다음 rebind 없이는 **불가**. `src/runtime/channels/index.mjs:357`, `src/runtime/channels/lib/output-forwarder.mjs:656`\n\n- **pre-connect re-arm 조기 종료** — re-arm tick이 `!bridgeRuntimeConnected`면 중단되지만 start path가 연결 후 다시 arm합니다; backlog catch-up **가능하지만 연결 완료까지 지연**됩니다. `src/runtime/channels/index.mjs:372`, `src/runtime/channels/index.mjs:803`\n\n- **첫 inbound의 stale binding 방어** — parent-chain transcript 발견이 늦으면 500ms steal poll 뒤 pending rebind/re-arm으로 전환합니다; 성공 시 backlog catch-up **가능**, poll 실패 후 기존 transcript를 EOF 바인딩하면 해당 window 출력은 **불가**입니다. `src/runtime/channels/index.mjs:1526`, `src/runtime/channels/index.mjs:1571`\n\nVerdict: 가장 위험한 실제 손실은 “현재 transcript를 기존 파일로 늦게 발견했지만 persisted path가 달라 EOF에서 시작하는” 경로이고, not-yet-created/watch/ownership window 자체는 대부분 catch-up으로 지연 복구됩니다.",
+      "card": {
+        "session": "sess_37028_1783063871406_1783063871636_34284b6e9fe43e25a279de7888dc1dc1",
+        "provider": "openai-oauth",
+        "model": "gpt-5.5",
+        "wall_ms": 62249,
+        "turns": 6,
+        "tool_calls": 7,
+        "tools_per_turn": 1.2,
+        "total_tool_ms": 686,
+        "llm_stream_ms": 61369,
+        "cache_ratio": 0.704984040222811,
+        "cached_tokens": 101376,
+        "prompt_tokens": 143799,
+        "uncached_tokens": 42423,
+        "output_tokens": 5270,
+        "cache_weighted_input_10": 52560.6,
+        "cache_weighted_input_25": 67767,
+        "prompt_growth": 29079,
+        "antipatterns": 1,
+        "issues": 9,
+        "issue_types": [
+          "cache",
+          "cache_break",
+          "service_tier",
+          "tool_churn_cluster"
+        ],
+        "read_fragmentation_paths": [],
+        "grep_sweep_count": 0
+      }
+    },
+    {
+      "id": "r4-impl-usage-summary#2",
+      "agent": "worker",
+      "ok": true,
+      "ms": 157204,
+      "sessionId": "sess_35624_1783063944146_1783063944381_cbf25d084f0f52db125bb2a739ccdc6d",
+      "output": "sessionId: sess_35624_1783063944146_1783063944381_cbf25d084f0f52db125bb2a739ccdc6d\n요청 범위가 명확해 바로 구현하고 검증까지 진행하겠습니다.구현은 반영했고, 문법 검사와 실제 실행으로 확인하겠습니다.검증은 통과했고, 변경 요약을 위해 diff를 확인하겠습니다.재영님, 완료했습니다.\n\n- **Diff 요약**: `package.json`에 `usage:today` 추가, `scripts/usage-today.mjs` 신규 작성.\n- **동작**: `MIXDOG_DATA_DIR`를 반영해 `history/agent-trace.jsonl`을 읽고, 오늘 0시 이후 `usage` 행을 모델별/총합으로 출력합니다.\n- **검증**: `node --check scripts/usage-today.mjs` 통과, `npm run usage:today` 실행 완료.\n- **참고**: 기존 변경 파일 `src/runtime/channels/lib/*`는 건드리지 않았습니다.",
+      "card": {
+        "session": "sess_35624_1783063944146_1783063944381_cbf25d084f0f52db125bb2a739ccdc6d",
+        "provider": "openai-oauth",
+        "model": "gpt-5.5",
+        "wall_ms": 146718,
+        "turns": 22,
+        "tool_calls": 25,
+        "tools_per_turn": 1.1,
+        "total_tool_ms": 4794,
+        "llm_stream_ms": 141892,
+        "cache_ratio": 0.919211593309697,
+        "cached_tokens": 548864,
+        "prompt_tokens": 597103,
+        "uncached_tokens": 48239,
+        "output_tokens": 13661,
+        "cache_weighted_input_10": 103125.4,
+        "cache_weighted_input_25": 185455,
+        "prompt_growth": 30611,
+        "antipatterns": 5,
+        "issues": 18,
+        "issue_types": [
+          "cache",
+          "cache_break",
+          "readonly_stall",
+          "tool_error",
+          "service_tier",
+          "duplicate_tool",
+          "missed_parallelism",
+          "tool_churn_cluster"
+        ],
+        "read_fragmentation_paths": [],
+        "grep_sweep_count": 0
+      }
+    },
+    {
+      "id": "r4-explain-compact-engine#2",
+      "agent": "explore",
+      "ok": true,
+      "ms": 90079,
+      "sessionId": "sess_47024_1783064101403_1783064101639_70166643f319fb0dc14080ac610d3d19",
+      "output": "sessionId: sess_47024_1783064101403_1783064101639_70166643f319fb0dc14080ac610d3d19\nArchitecture brief completed.\n\n- **Entry points**: `semanticCompactMessages` and `recallFastTrackCompactMessages` are exported from `engine.mjs`, then re-exported by `compact.mjs` as the public surface. `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`\n\n- **Runtime callers**: the loop calls recall-fasttrack via `runRecallFastTrackCompact` or semantic directly; the manager calls recall-fasttrack, falls back to semantic, or runs semantic directly. `src/runtime/agent/orchestrator/session/loop.mjs:414`, `src/runtime/agent/orchestrator/session/loop.mjs:441`, `src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:210`, `src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:245`, `src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:288`\n\n- **Fire decision**: auto policy resolves boundary, trigger, buffer, reserve, type, tail and keep budgets; compaction fires when reactive retry is pending, token estimate is unavailable, or transcript+reserve pressure reaches trigger. `src/runtime/agent/orchestrator/session/loop/compact-policy.mjs:89`, `src/runtime/agent/orchestrator/session/loop/compact-policy.mjs:117`, `src/runtime/agent/orchestrator/session/loop/compact-policy.mjs:123`, `src/runtime/agent/orchestrator/session/loop/compact-policy.mjs:170`\n\n- **Loop path**: pre-send estimates message tokens, computes `shouldCompact`, selects target budget, optionally prunes tool outputs, then runs recall-fasttrack if configured, else semantic. `src/runtime/agent/orchestrator/session/loop.mjs:351`, `src/runtime/agent/orchestrator/session/loop.mjs:353`, `src/runtime/agent/orchestrator/session/loop.mjs:362`, `src/runtime/agent/orchestrator/session/loop.mjs:407`, `src/runtime/agent/orchestrator/session/loop.mjs:414`, `src/runtime/agent/orchestrator/session/loop.mjs:439`\n\n- **Manual/manager path**: manual is forced, auto skips below threshold, missing boundary blocks forced compaction, and budget is computed before dispatch. `src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:143`, `src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:150`, `src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:176`, `src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:196`\n\n- **Semantic windowing**: engine sanitizes/dedups, separates protected prefix from live conversation, extracts previous summary, then chooses head/tail before prompting. `src/runtime/agent/orchestrator/session/compact/engine.mjs:100`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:102`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:103`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:235`\n\n- **Semantic tail choice**: with user turns, it keeps up to configured recent turns within `preserveRecentBudget`; without user turns, it cuts only at non-tool boundaries so tool results do not start the preserved tail. `src/runtime/agent/orchestrator/session/compact/engine.mjs:134`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:141`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:193`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:240`\n\n- **Recall tail choice**: recall-fasttrack keeps the newest structured turns, drops oldest kept turns until under cap, truncates the surviving turn if needed, and allows no-user assistant/tool tails. `src/runtime/agent/orchestrator/session/compact/engine.mjs:593`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:601`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:602`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:610`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:613`\n\n- **Redaction hook**: no-op compaction returns sanitized messages unchanged, but real semantic and recall compaction redact tool-call secret argument values before window selection, so preserved system/tail are measured and emitted redacted. `src/runtime/agent/orchestrator/session/compact/engine.mjs:286`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:306`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:312`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:644`\n\n- **Redaction scope**: redaction only touches assistant `toolCalls` argument values, preserving role/content/ids/order/pairing and leaving non-tool-bearing messages as the same reference. `src/runtime/agent/orchestrator/session/compact/text-utils.mjs:291`, `src/runtime/agent/orchestrator/session/compact/text-utils.mjs:294`, `src/runtime/agent/orchestrator/session/compact/text-utils.mjs:319`\n\n- **Semantic failure guards**: it throws for missing provider, empty live window, no compactable history without prior summary, prompt fit failure, empty provider summary, summary fit failure, or final budget overflow. `src/runtime/agent/orchestrator/session/compact/engine.mjs:279`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:254`, `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\n- **Recall failure guards**: it throws when there is no compactable prior history and no recall/prior/empty allowance, when recall text is empty, when the recall summary cannot fit, or when final output exceeds budget. `src/runtime/agent/orchestrator/session/compact/engine.mjs:653`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:666`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:691`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:698`\n\n- **Fallback paths**: manager recall-fasttrack failure can degrade to semantic; non-recall failures in loop/manager can recover with unanchored tool-output prune; unrecovered loop failure records telemetry and throws context overflow. `src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:237`, `src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:245`, `src/runtime/agent/orchestrator/session/loop.mjs:501`, `src/runtime/agent/orchestrator/session/loop.mjs:527`, `src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:322`, `src/runtime/agent/orchestrator/session/loop.mjs:552`, `src/runtime/agent/orchestrator/session/loop.mjs:618`",
+      "card": {
+        "session": "sess_47024_1783064101403_1783064101639_70166643f319fb0dc14080ac610d3d19",
+        "provider": "openai-oauth",
+        "model": "gpt-5.5",
+        "wall_ms": 79532,
+        "turns": 7,
+        "tool_calls": 8,
+        "tools_per_turn": 1.1,
+        "total_tool_ms": 532,
+        "llm_stream_ms": 78523,
+        "cache_ratio": 0.685222081345972,
+        "cached_tokens": 136192,
+        "prompt_tokens": 198756,
+        "uncached_tokens": 62564,
+        "output_tokens": 4446,
+        "cache_weighted_input_10": 76183.2,
+        "cache_weighted_input_25": 96612,
+        "prompt_growth": 48415,
+        "antipatterns": 1,
+        "issues": 11,
+        "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": 80994.8,
+    "turns": 9.5,
+    "tool_calls": 11.3,
+    "tools_per_turn": 1.2,
+    "total_tool_ms": 1674.3,
+    "llm_stream_ms": 79108.2,
+    "cache_ratio": 0.8,
+    "cached_tokens": 195157.3,
+    "prompt_tokens": 241602.7,
+    "uncached_tokens": 46445.3,
+    "output_tokens": 6575.8,
+    "cache_weighted_input_10": 65961.1,
+    "cache_weighted_input_25": 95234.7,
+    "prompt_growth": 32557,
+    "antipatterns": 1.5,
+    "issues": 12.2,
+    "grep_sweep_count": 0
+  },
+  "cards": [
+    {
+      "session": "sess_53408_1783063642444_1783063642668_3cfee1069f85e2bb12e8d28e8818547c",
+      "provider": "openai-oauth",
+      "model": "gpt-5.5",
+      "wall_ms": 84498,
+      "turns": 8,
+      "tool_calls": 9,
+      "tools_per_turn": 1.1,
+      "total_tool_ms": 780,
+      "llm_stream_ms": 83431,
+      "cache_ratio": 0.8006575913711181,
+      "cached_tokens": 199680,
+      "prompt_tokens": 249395,
+      "uncached_tokens": 49715,
+      "output_tokens": 7234,
+      "cache_weighted_input_10": 69683,
+      "cache_weighted_input_25": 99635,
+      "prompt_growth": 38233,
+      "antipatterns": 1,
+      "issues": 13,
+      "issue_types": [
+        "cache",
+        "cache_break",
+        "broad_tool_result",
+        "service_tier",
+        "missed_parallelism",
+        "tool_churn_cluster"
+      ],
+      "read_fragmentation_paths": [],
+      "grep_sweep_count": 0
+    },
+    {
+      "session": "sess_6356_1783063737425_1783063737661_1619e4bb11b96b6c6f8a2015039eca61",
+      "provider": "openai-oauth",
+      "model": "gpt-5.5",
+      "wall_ms": 53970,
+      "turns": 7,
+      "tool_calls": 12,
+      "tools_per_turn": 1.7,
+      "total_tool_ms": 2323,
+      "llm_stream_ms": 51748,
+      "cache_ratio": 0.7647431722461031,
+      "cached_tokens": 88064,
+      "prompt_tokens": 115155,
+      "uncached_tokens": 27091,
+      "output_tokens": 4646,
+      "cache_weighted_input_10": 35897.4,
+      "cache_weighted_input_25": 49107,
+      "prompt_growth": 12328,
+      "antipatterns": 0,
+      "issues": 9,
+      "issue_types": [
+        "cache",
+        "cache_break",
+        "service_tier"
+      ],
+      "read_fragmentation_paths": [],
+      "grep_sweep_count": 0
+    },
+    {
+      "session": "sess_54836_1783063801897_1783063802131_358c27f874c909a61e1055fa674cc778",
+      "provider": "openai-oauth",
+      "model": "gpt-5.5",
+      "wall_ms": 59002,
+      "turns": 7,
+      "tool_calls": 7,
+      "tools_per_turn": 1,
+      "total_tool_ms": 931,
+      "llm_stream_ms": 57686,
+      "cache_ratio": 0.6654929577464789,
+      "cached_tokens": 96768,
+      "prompt_tokens": 145408,
+      "uncached_tokens": 48640,
+      "output_tokens": 4198,
+      "cache_weighted_input_10": 58316.8,
+      "cache_weighted_input_25": 72832,
+      "prompt_growth": 36676,
+      "antipatterns": 1,
+      "issues": 13,
+      "issue_types": [
+        "cache",
+        "cache_break",
+        "broad_tool_result",
+        "service_tier",
+        "missed_parallelism",
+        "tool_churn_cluster"
+      ],
+      "read_fragmentation_paths": [],
+      "grep_sweep_count": 0
+    },
+    {
+      "session": "sess_37028_1783063871406_1783063871636_34284b6e9fe43e25a279de7888dc1dc1",
+      "provider": "openai-oauth",
+      "model": "gpt-5.5",
+      "wall_ms": 62249,
+      "turns": 6,
+      "tool_calls": 7,
+      "tools_per_turn": 1.2,
+      "total_tool_ms": 686,
+      "llm_stream_ms": 61369,
+      "cache_ratio": 0.704984040222811,
+      "cached_tokens": 101376,
+      "prompt_tokens": 143799,
+      "uncached_tokens": 42423,
+      "output_tokens": 5270,
+      "cache_weighted_input_10": 52560.6,
+      "cache_weighted_input_25": 67767,
+      "prompt_growth": 29079,
+      "antipatterns": 1,
+      "issues": 9,
+      "issue_types": [
+        "cache",
+        "cache_break",
+        "service_tier",
+        "tool_churn_cluster"
+      ],
+      "read_fragmentation_paths": [],
+      "grep_sweep_count": 0
+    },
+    {
+      "session": "sess_35624_1783063944146_1783063944381_cbf25d084f0f52db125bb2a739ccdc6d",
+      "provider": "openai-oauth",
+      "model": "gpt-5.5",
+      "wall_ms": 146718,
+      "turns": 22,
+      "tool_calls": 25,
+      "tools_per_turn": 1.1,
+      "total_tool_ms": 4794,
+      "llm_stream_ms": 141892,
+      "cache_ratio": 0.919211593309697,
+      "cached_tokens": 548864,
+      "prompt_tokens": 597103,
+      "uncached_tokens": 48239,
+      "output_tokens": 13661,
+      "cache_weighted_input_10": 103125.4,
+      "cache_weighted_input_25": 185455,
+      "prompt_growth": 30611,
+      "antipatterns": 5,
+      "issues": 18,
+      "issue_types": [
+        "cache",
+        "cache_break",
+        "readonly_stall",
+        "tool_error",
+        "service_tier",
+        "duplicate_tool",
+        "missed_parallelism",
+        "tool_churn_cluster"
+      ],
+      "read_fragmentation_paths": [],
+      "grep_sweep_count": 0
+    },
+    {
+      "session": "sess_47024_1783064101403_1783064101639_70166643f319fb0dc14080ac610d3d19",
+      "provider": "openai-oauth",
+      "model": "gpt-5.5",
+      "wall_ms": 79532,
+      "turns": 7,
+      "tool_calls": 8,
+      "tools_per_turn": 1.1,
+      "total_tool_ms": 532,
+      "llm_stream_ms": 78523,
+      "cache_ratio": 0.685222081345972,
+      "cached_tokens": 136192,
+      "prompt_tokens": 198756,
+      "uncached_tokens": 62564,
+      "output_tokens": 4446,
+      "cache_weighted_input_10": 76183.2,
+      "cache_weighted_input_25": 96612,
+      "prompt_growth": 48415,
+      "antipatterns": 1,
+      "issues": 11,
+      "issue_types": [
+        "cache",
+        "cache_break",
+        "broad_tool_result",
+        "service_tier",
+        "tool_churn_cluster"
+      ],
+      "read_fragmentation_paths": [],
+      "grep_sweep_count": 0
+    }
+  ]
+}