mixdog 0.9.4 → 0.9.5

package/scripts/bench/round-r20-prewarm-clean.json

@@ -0,0 +1,475 @@
+{
+  "round": "20",
+  "runner": "lead",
+  "opts": {
+    "provider": "openai-oauth",
+    "model": "gpt-5.5",
+    "effort": null,
+    "fast": false,
+    "env": {
+      "MIXDOG_DATA_DIR": "C:\\Users\\tempe\\.mixdog\\bench-data",
+      "MIXDOG_OPENAI_OAUTH_WS_WARMUP": "1"
+    }
+  },
+  "tasks": 6,
+  "completed": 6,
+  "completion_rate": 100,
+  "sessions": [
+    "sess_24484_1783061209581_1783061209828_780131e53fd6bc2b4d00e76277ef0eac",
+    "sess_50164_1783061209590_1783061209828_64514b47078973e8ff313ba80fddc6d6",
+    "sess_6636_1783061209603_1783061209840_309bbb037fe8492a38533175c6dec7e7",
+    "sess_12456_1783061273178_1783061273417_b75f3cff29775e56257b108f96388d21",
+    "sess_34340_1783061299082_1783061299306_c1576ad2a7e9cafad2329f5c0bf5a041",
+    "sess_40856_1783061322058_1783061322304_6da5edceb48f481a206a016ccc63c5fc"
+  ],
+  "results": [
+    {
+      "id": "r4-debug-forwarding-gap#1",
+      "agent": "reviewer",
+      "ok": true,
+      "ms": 112485,
+      "sessionId": "sess_24484_1783061209581_1783061209828_780131e53fd6bc2b4d00e76277ef0eac",
+      "output": "sessionId: sess_24484_1783061209581_1783061209828_780131e53fd6bc2b4d00e76277ef0eac\n조사 결과, 활성화 직후 출력 유실 가능성은 “바인딩 실패/EOF 바인딩” 쪽에 있고, `fs.watch`/ownership 창은 대부분 지연 후 catch-up됩니다.\n\n- **경로:** `writeBridgeState()`가 `bridge-state.json`을 쓰고(`src/runtime/channels/index.mjs:218`), ownership 갱신이 `startOwnedRuntime()`로 이어져 바인딩/연결/후속 catch-up을 수행합니다(`index.mjs:881`, `index.mjs:710`, `index.mjs:801`).\n\n- **최대 영향 — 바인딩 후보 없음:** 초기 `rebindTranscriptContext()`가 30×150ms 안에 transcript/previous path를 못 찾으면 바인딩 없이 종료됩니다(`index.mjs:431`, `index.mjs:503`, `index.mjs:515`). 이후 이미 존재하는 transcript에 일반 바인딩되면 cursor가 EOF로 잡혀 그 사이 출력은 복구되지 않습니다(`src/runtime/channels/lib/output-forwarder.mjs:125`).\n\n- **기존 transcript + persisted cursor 불일치:** `catchUpFromPersisted`라도 같은 transcript가 아니면 현재 파일 크기부터 시작해 활성화~바인딩 사이에 이미 쓰인 출력은 건너뜁니다(`output-forwarder.mjs:117`, `output-forwarder.mjs:122`). backlog catch-up은 같은 persisted path이거나 `replayFromStart`인 경우에만 복구합니다.\n\n- **미생성 transcript `fs.watch` 실패:** missing file watch는 실패하지만(`output-forwarder.mjs:647`), pending re-arm이 생성 감지 후 watch와 `forwardNewText()`를 실행합니다(`index.mjs:357`, `index.mjs:376`, `index.mjs:486`). 60초 내 생성은 복구되고, timeout 후 추가 rebind가 없으면 지연/미복구입니다.\n\n- **ownership/pre-connect 창:** `_isOwner()`가 `bridgeRuntimeConnected && owned`라 pre-connect flush는 no-op입니다(`index.mjs:316`, `output-forwarder.mjs:351`). cursor를 전진시키지 않고 post-connect catch-up/re-arm이 다시 실행되어 보통 복구됩니다(`index.mjs:801`).\n\n- **watch 설치 race:** `setContext()` stat 이후 watch 설치 전 쓰인 bytes는 `startWatch()`가 즉시 flush를 예약해 복구합니다(`output-forwarder.mjs:665`, `output-forwarder.mjs:703`).\n\n판정: 가장 실제적인 유실 위험은 “초기 discovery miss 후 EOF 바인딩”이며, pending-file watch와 ownership 창은 제한적으로 catch-up됩니다.",
+      "card": {
+        "session": "sess_24484_1783061209581_1783061209828_780131e53fd6bc2b4d00e76277ef0eac",
+        "provider": "openai-oauth",
+        "model": "gpt-5.5",
+        "wall_ms": 101982,
+        "turns": 11,
+        "tool_calls": 10,
+        "tools_per_turn": 0.9,
+        "total_tool_ms": 989,
+        "llm_stream_ms": 100233,
+        "cache_ratio": 0.793172745612974,
+        "cached_tokens": 203264,
+        "prompt_tokens": 256267,
+        "uncached_tokens": 53003,
+        "output_tokens": 8121,
+        "cache_weighted_input_10": 73329.4,
+        "cache_weighted_input_25": 103819,
+        "prompt_growth": 35245,
+        "antipatterns": 1,
+        "issues": 16,
+        "issue_types": [
+          "cache",
+          "cache_break",
+          "cache_miss",
+          "readonly_stall",
+          "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": 89491,
+      "sessionId": "sess_50164_1783061209590_1783061209828_64514b47078973e8ff313ba80fddc6d6",
+      "output": "sessionId: sess_50164_1783061209590_1783061209828_64514b47078973e8ff313ba80fddc6d6\n요청 범위가 명확해서 바로 구현하고 검증하겠습니다.기능 파일과 스크립트 엔트리는 이미 작업 트리에 반영되어 있어, 관련 diff만 확인하겠습니다.검증은 통과했고, 최종 상태만 확인합니다.재영님, 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, 새 스크립트는 `MIXDOG_DATA_DIR`/기본 trace 경로를 읽고 local midnight 이후 `usage` 행을 집계합니다.\n- **검증**: `node --check scripts/usage-today.mjs` 통과, `npm run usage:today` 실행 완료.\n- **실행 결과**: GRAND TOTAL 기준 `1629` calls, input `60.99M`, cached `56.76M`, output `1.83M`, costUsd `101.8728`.\n- **참고**: 요청 범위 밖 기존 수정 파일 `src/runtime/channels/lib/*` 2개는 건드리지 않았습니다.",
+      "card": {
+        "session": "sess_50164_1783061209590_1783061209828_64514b47078973e8ff313ba80fddc6d6",
+        "provider": "openai-oauth",
+        "model": "gpt-5.5",
+        "wall_ms": 79004,
+        "turns": 11,
+        "tool_calls": 13,
+        "tools_per_turn": 1.2,
+        "total_tool_ms": 3127,
+        "llm_stream_ms": 75555,
+        "cache_ratio": 0.8285965759064194,
+        "cached_tokens": 145920,
+        "prompt_tokens": 176105,
+        "uncached_tokens": 30185,
+        "output_tokens": 6181,
+        "cache_weighted_input_10": 44777,
+        "cache_weighted_input_25": 66665,
+        "prompt_growth": 16826,
+        "antipatterns": 1,
+        "issues": 14,
+        "issue_types": [
+          "cache",
+          "cache_break",
+          "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": 63576,
+      "sessionId": "sess_6636_1783061209603_1783061209840_309bbb037fe8492a38533175c6dec7e7",
+      "output": "sessionId: sess_6636_1783061209603_1783061209840_309bbb037fe8492a38533175c6dec7e7\nCompaction architecture brief:\n\n- Entry points: `semanticCompactMessages` and `recallFastTrackCompactMessages`; loop pre-send calls recall at `loop.mjs:416` or semantic at `loop.mjs:441`, while manual/clear paths call `runSessionCompaction` at `manager.mjs:1849` and `manager.mjs:1950`.\n\n- Fire decision: auto policy resolves boundary/trigger/reserve/type at `loop/compact-policy.mjs:89`; pre-send compacts when reactive retry is pending or pressure reaches trigger at `loop.mjs:351` and `loop/compact-policy.mjs:170`.\n\n- Manual decision: `runSessionCompaction` forces manual mode, skips small/unforced sessions, requires a boundary, and returns “below threshold” when pressure is under trigger at `manager/compaction-runner.mjs:143`.\n\n- Tail selection: semantic splits protected prefix/live/prior summary at `engine.mjs:100`, keeps recent user turns if any at `engine.mjs:240`, otherwise uses non-tool cut points to preserve assistant/tool-only tails at `engine.mjs:193`.\n\n- Recall tail: recall-fasttrack keeps the newest structured turns, drops oldest turns to fit cap, truncates only if needed, and permits no-user tails at `engine.mjs:593`.\n\n- Redaction: if the transcript already fits, no-op returns sanitized messages before redaction at `engine.mjs:286`; actual compaction redacts tool-call secret values before window selection/output at `engine.mjs:306` and `engine.mjs:644`, implemented shape-preserving in `text-utils.mjs:319`.\n\n- Failure/fallback: semantic throws on missing provider, empty live/no prior head, prompt/summary fit failures, empty provider summary, or final overflow at `engine.mjs:278`, `engine.mjs:315`, `engine.mjs:335`, `engine.mjs:371`, `engine.mjs:384`, and `engine.mjs:395`.\n\n- Degradation: loop fallback prunes old tool outputs only for non-recall failures at `loop.mjs:500`; manager recall-fasttrack can fall back to semantic at `compaction-runner.mjs:237`, then non-recall failures can use unanchored prune at `compaction-runner.mjs:322`, otherwise failure is recorded at `compaction-runner.mjs:348`.",
+      "card": {
+        "session": "sess_6636_1783061209603_1783061209840_309bbb037fe8492a38533175c6dec7e7",
+        "provider": "openai-oauth",
+        "model": "gpt-5.5",
+        "wall_ms": 53078,
+        "turns": 7,
+        "tool_calls": 7,
+        "tools_per_turn": 1,
+        "total_tool_ms": 549,
+        "llm_stream_ms": 51861,
+        "cache_ratio": 0.47149308599591866,
+        "cached_tokens": 64000,
+        "prompt_tokens": 135739,
+        "uncached_tokens": 71739,
+        "output_tokens": 2800,
+        "cache_weighted_input_10": 78139,
+        "cache_weighted_input_25": 87739,
+        "prompt_growth": 41335,
+        "antipatterns": 1,
+        "issues": 14,
+        "issue_types": [
+          "cache",
+          "cache_break",
+          "cache_miss",
+          "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": 115365,
+      "sessionId": "sess_12456_1783061273178_1783061273417_b75f3cff29775e56257b108f96388d21",
+      "output": "sessionId: sess_12456_1783061273178_1783061273417_b75f3cff29775e56257b108f96388d21\n조사 결과, 활성화 직후 출력은 일부 경로에서 지연 복구되지만, “기존 파일을 EOF로 바인딩”하거나 “잘못된 transcript에 붙는” 경우는 영구 유실될 수 있습니다.\n\n**경로 요약**\n- `activate_channel_bridge`/`start()`가 `channelBridgeActive=true`와 `bridge-state.json`을 먼저 기록합니다 (`src/runtime/channels/lib/tool-dispatch.mjs:46`, `src/runtime/channels/index.mjs:1735`).\n- 이후 ownership claim → `refreshBridgeOwnership()` → `startOwnedRuntime()` → `bindPersistedTranscriptIfAny()` → `rebindTranscriptContext()` → `applyTranscriptBinding()` → `OutputForwarder.startWatch()` → `forwardNewText()` 순서로 첫 전달이 열립니다 (`src/runtime/channels/index.mjs:710`, `src/runtime/channels/index.mjs:577`, `src/runtime/channels/index.mjs:394`, `src/runtime/channels/index.mjs:317`, `src/runtime/channels/lib/output-forwarder.mjs:647`).\n\n**영구 유실: 기존 transcript를 persisted cursor 없이 EOF로 바인딩**\n- `setContext()`는 `catchUpFromPersisted`라도 같은 persisted transcript가 아니면 `currentSize`로 커서를 둡니다 (`src/runtime/channels/lib/output-forwarder.mjs:117`). 활성화와 바인딩 사이에 이미 쓰인 assistant 출력은 이후 `forwardNewText()`가 읽지 않습니다.\n- **backlog catch-up:** 복구 안 됨. 같은 persisted path의 `lastFileSize`가 있거나 `replayFromStart`일 때만 복구됩니다.\n\n**영구 유실 가능: 활성화 직후 self transcript 발견 전 stale/wrong transcript에 바인딩**\n- 첫 inbound에서 기존 binding을 재사용하거나 500ms steal poll이 실패하면 stale path를 `exists:true`로 취급해 바인딩할 수 있습니다 (`src/runtime/channels/index.mjs:1506`, `src/runtime/channels/index.mjs:1571`, `src/runtime/channels/index.mjs:1597`).\n- **backlog catch-up:** 실제 self transcript가 나중에 “이미 존재” 상태로 재바인딩되면 EOF 바인딩이라 복구 안 됨. “pending/not-yet-created”로 잡힌 경우만 `replayFromStart`/rearm으로 복구됩니다 (`src/runtime/channels/index.mjs:486`).\n\n**지연 복구: transcript 파일이 아직 없어서 `fs.watch` 설치 실패**\n- `startWatch()`는 없는 파일에서 `watch()` 실패 후 watcher를 닫습니다 (`src/runtime/channels/lib/output-forwarder.mjs:657`, `src/runtime/channels/lib/output-forwarder.mjs:667`). 이를 `schedulePendingTranscriptRearm()`가 250ms 폴링으로 다시 설치하고 즉시 `forwardNewText()`를 호출합니다 (`src/runtime/channels/index.mjs:357`, `src/runtime/channels/index.mjs:376`).\n- **backlog catch-up:** 60초 안에 파일이 생기면 복구됨. 60초 timeout 이후 생성되면 복구 안 됨 (`src/runtime/channels/index.mjs:384`).\n\n**지연 복구: ownership/backend 연결 전 forwarder가 fail-closed**\n- owner probe는 `bridgeRuntimeConnected && currentOwnerState().owned`가 아니면 false이고, `forwardNewText()`는 읽기 전에 반환합니다 (`src/runtime/channels/index.mjs:316`, `src/runtime/channels/lib/output-forwarder.mjs:36`, `src/runtime/channels/lib/output-forwarder.mjs:351`).\n- **backlog catch-up:** 커서를 전진시키지 않으므로 보통 복구됨. 연결 후 `post-connect forwardNewText()` 또는 rearm이 다시 읽습니다 (`src/runtime/channels/index.mjs:776`, `src/runtime/channels/index.mjs:803`).\n\n**지연 복구: `setContext()` stat과 `fs.watch` 설치 사이의 write**\n- `startWatch()`가 watch 설치 직후 `scheduleWatchFlush()`를 호출해 그 사이 쓰인 bytes를 읽도록 되어 있습니다 (`src/runtime/channels/lib/output-forwarder.mjs:665`, `src/runtime/channels/lib/output-forwarder.mjs:703`).\n- **backlog catch-up:** 복구됨. 다만 이 flush가 owner false로 빠지면 post-connect forward가 맡습니다.\n\n평결: 활성화 창의 missing-file/ownership 지연은 대부분 catch-up되지만, 기존 transcript를 EOF로 잡는 바인딩과 wrong-transcript 바인딩은 첫 assistant 출력 영구 유실 위험이 남아 있습니다.",
+      "card": {
+        "session": "sess_12456_1783061273178_1783061273417_b75f3cff29775e56257b108f96388d21",
+        "provider": "openai-oauth",
+        "model": "gpt-5.5",
+        "wall_ms": 104870,
+        "turns": 10,
+        "tool_calls": 12,
+        "tools_per_turn": 1.2,
+        "total_tool_ms": 866,
+        "llm_stream_ms": 103586,
+        "cache_ratio": 0.8181367561503744,
+        "cached_tokens": 250880,
+        "prompt_tokens": 306648,
+        "uncached_tokens": 55768,
+        "output_tokens": 10163,
+        "cache_weighted_input_10": 80856,
+        "cache_weighted_input_25": 118488,
+        "prompt_growth": 44684,
+        "antipatterns": 1,
+        "issues": 15,
+        "issue_types": [
+          "cache",
+          "cache_break",
+          "readonly_stall",
+          "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": 68832,
+      "sessionId": "sess_34340_1783061299082_1783061299306_c1576ad2a7e9cafad2329f5c0bf5a041",
+      "output": "sessionId: sess_34340_1783061299082_1783061299306_c1576ad2a7e9cafad2329f5c0bf5a041\n작업 범위가 작고 실행 요청까지 포함되어 있어 바로 반영하고 검증하겠습니다.구현 파일과 npm script가 이미 작업트리에 있어 검증과 diff 확인을 진행합니다.완료했습니다. `usage:today` 스크립트와 `scripts/usage-today.mjs`가 작업트리에 반영되어 있습니다.\n\n- **Diff 요약**: `package.json`에 `\"usage:today\": \"node scripts/usage-today.mjs\"` 추가, 새 스크립트가 오늘 자정 이후 usage를 모델별/총합으로 집계합니다.\n- **검증**: `node --check scripts/usage-today.mjs` 통과, `npm run usage:today` 실행 성공.\n- **참고**: 기존 작업트리에 `src/runtime/channels/lib/event-queue.mjs`, `src/runtime/channels/lib/output-forwarder.mjs` 수정도 남아 있으며 이번 변경 범위와 별개입니다.",
+      "card": {
+        "session": "sess_34340_1783061299082_1783061299306_c1576ad2a7e9cafad2329f5c0bf5a041",
+        "provider": "openai-oauth",
+        "model": "gpt-5.5",
+        "wall_ms": 58353,
+        "turns": 9,
+        "tool_calls": 14,
+        "tools_per_turn": 1.6,
+        "total_tool_ms": 2016,
+        "llm_stream_ms": 55759,
+        "cache_ratio": 0.6363588621444202,
+        "cached_tokens": 72704,
+        "prompt_tokens": 114250,
+        "uncached_tokens": 41546,
+        "output_tokens": 3718,
+        "cache_weighted_input_10": 48816.4,
+        "cache_weighted_input_25": 59722,
+        "prompt_growth": 11717,
+        "antipatterns": 0,
+        "issues": 12,
+        "issue_types": [
+          "cache",
+          "cache_break",
+          "cache_miss",
+          "service_tier"
+        ],
+        "read_fragmentation_paths": [],
+        "grep_sweep_count": 0
+      }
+    },
+    {
+      "id": "r4-explain-compact-engine#2",
+      "agent": "explore",
+      "ok": true,
+      "ms": 98300,
+      "sessionId": "sess_40856_1783061322058_1783061322304_6da5edceb48f481a206a016ccc63c5fc",
+      "output": "sessionId: sess_40856_1783061322058_1783061322304_6da5edceb48f481a206a016ccc63c5fc\n- **Entry points**: `semanticCompactMessages` and `recallFastTrackCompactMessages` live in `engine.mjs` and are re-exported by `compact.mjs` for loop/manager callers (`engine.mjs:278`, `engine.mjs:445`, `compact.mjs:52-55`).\n\n- **Callers**: auto pre-send compaction runs inside `agentLoop`; manual/auto-clear uses `runSessionCompaction`; recall-fasttrack has a loop pipeline wrapper before calling the engine (`loop.mjs:329-365`, `manager.mjs:1849-1957`, `loop/recall-fasttrack.mjs:161-172`).\n\n- **Fire decision**: policy resolves `auto`, boundary, trigger, reserve, type, tail budget; trigger is explicit sub-boundary limit or `boundary - buffer`, with default buffer from shared context utils (`compact-policy.mjs:89-148`, `context-utils.mjs:221-227`).\n\n- **Auto threshold**: compaction fires when forced reactive, token estimate is unavailable, or `messageTokens + reserve >= trigger`; loop drains queued steering before checking (`compact-policy.mjs:170-175`, `loop.mjs:324-354`).\n\n- **Reactive path**: provider context-overflow sets `reactiveOverflowRetryPending`, loops back, and forces the next pre-send compact once; second failure surfaces overflow (`loop.mjs:981-1026`).\n\n- **Semantic flow**: sanitize/dedup first, no-op if already under budget and not forced, otherwise redact, select window, summarize head, then emit `system + summary + tail` (`engine.mjs:284-314`, `engine.mjs:367-397`).\n\n- **Tail selection**: protected system/reminder prefix is separated, old summaries are excluded from live history, and previous summary is captured for merge (`engine.mjs:100-111`, `messages.mjs:56-82`).\n\n- **User-tail path**: with real user turns, keep up to `tailTurns` recent turns within `preserveRecentBudget`, splitting the newest turn suffix if needed (`engine.mjs:134-185`, `engine.mjs:235-245`).\n\n- **No-user tail path**: without user turns, tail boundaries are any non-tool message so assistant/tool pairs stay valid; it grows newest-to-oldest until budget overflow (`engine.mjs:121-128`, `engine.mjs:193-232`).\n\n- **Recall tail**: recall-fasttrack preserves newest structured turns, drops whole oldest turns over cap, then truncates oldest surviving content; no-user recent structure is valid (`engine.mjs:593-619`, `engine.mjs:601-610`).\n\n- **Redaction hook**: compaction redacts sensitive tool-call args before window selection, so preserved system/tail are measured and emitted redacted; no-op fast path intentionally skips redaction (`engine.mjs:286-312`, `engine.mjs:389-392`, `engine.mjs:644-651`).\n\n- **Redaction mechanics**: sensitive keys/raw `key:value` secrets become `[redacted]`; verbatim redaction preserves object shape/order and only changes secret values (`text-utils.mjs:12-17`, `text-utils.mjs:46-80`, `text-utils.mjs:217-230`).\n\n- **Semantic failures**: provider missing, empty live/no compactable head, prompt fit failure, empty provider summary, summary fit failure, or final over-budget all throw (`engine.mjs:279-280`, `engine.mjs:254-256`, `engine.mjs:315-337`, `engine.mjs:371-397`).\n\n- **Budget fallback**: if mandatory `system + tail` plus summary room exceeds target, engine raises the working budget instead of refusing (`engine.mjs:319-331`, `constants.mjs:33-40`).\n\n- **Runtime fallback**: recall-fasttrack failure may fall back to semantic in manager; semantic/non-recall failure may fall back to non-LLM prune; unrecovered failures record `*_failed` telemetry (`compaction-runner.mjs:231-279`, `compaction-runner.mjs:327-374`, `loop.mjs:500-627`).",
+      "card": {
+        "session": "sess_40856_1783061322058_1783061322304_6da5edceb48f481a206a016ccc63c5fc",
+        "provider": "openai-oauth",
+        "model": "gpt-5.5",
+        "wall_ms": 87800,
+        "turns": 10,
+        "tool_calls": 13,
+        "tools_per_turn": 1.3,
+        "total_tool_ms": 816,
+        "llm_stream_ms": 86336,
+        "cache_ratio": 0.6981522354599766,
+        "cached_tokens": 181248,
+        "prompt_tokens": 259611,
+        "uncached_tokens": 78363,
+        "output_tokens": 4217,
+        "cache_weighted_input_10": 96487.8,
+        "cache_weighted_input_25": 123675,
+        "prompt_growth": 60577,
+        "antipatterns": 1,
+        "issues": 18,
+        "issue_types": [
+          "cache",
+          "cache_break",
+          "cache_miss",
+          "readonly_stall",
+          "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": 80847.8,
+    "turns": 9.7,
+    "tool_calls": 11.5,
+    "tools_per_turn": 1.2,
+    "total_tool_ms": 1393.8,
+    "llm_stream_ms": 78888.3,
+    "cache_ratio": 0.7,
+    "cached_tokens": 153002.7,
+    "prompt_tokens": 208103.3,
+    "uncached_tokens": 55100.7,
+    "output_tokens": 5866.7,
+    "cache_weighted_input_10": 70400.9,
+    "cache_weighted_input_25": 93351.3,
+    "prompt_growth": 35064,
+    "antipatterns": 0.8,
+    "issues": 14.8,
+    "grep_sweep_count": 0
+  },
+  "cards": [
+    {
+      "session": "sess_24484_1783061209581_1783061209828_780131e53fd6bc2b4d00e76277ef0eac",
+      "provider": "openai-oauth",
+      "model": "gpt-5.5",
+      "wall_ms": 101982,
+      "turns": 11,
+      "tool_calls": 10,
+      "tools_per_turn": 0.9,
+      "total_tool_ms": 989,
+      "llm_stream_ms": 100233,
+      "cache_ratio": 0.793172745612974,
+      "cached_tokens": 203264,
+      "prompt_tokens": 256267,
+      "uncached_tokens": 53003,
+      "output_tokens": 8121,
+      "cache_weighted_input_10": 73329.4,
+      "cache_weighted_input_25": 103819,
+      "prompt_growth": 35245,
+      "antipatterns": 1,
+      "issues": 16,
+      "issue_types": [
+        "cache",
+        "cache_break",
+        "cache_miss",
+        "readonly_stall",
+        "service_tier",
+        "missed_parallelism",
+        "tool_churn_cluster"
+      ],
+      "read_fragmentation_paths": [],
+      "grep_sweep_count": 0
+    },
+    {
+      "session": "sess_50164_1783061209590_1783061209828_64514b47078973e8ff313ba80fddc6d6",
+      "provider": "openai-oauth",
+      "model": "gpt-5.5",
+      "wall_ms": 79004,
+      "turns": 11,
+      "tool_calls": 13,
+      "tools_per_turn": 1.2,
+      "total_tool_ms": 3127,
+      "llm_stream_ms": 75555,
+      "cache_ratio": 0.8285965759064194,
+      "cached_tokens": 145920,
+      "prompt_tokens": 176105,
+      "uncached_tokens": 30185,
+      "output_tokens": 6181,
+      "cache_weighted_input_10": 44777,
+      "cache_weighted_input_25": 66665,
+      "prompt_growth": 16826,
+      "antipatterns": 1,
+      "issues": 14,
+      "issue_types": [
+        "cache",
+        "cache_break",
+        "service_tier",
+        "missed_parallelism",
+        "tool_churn_cluster"
+      ],
+      "read_fragmentation_paths": [],
+      "grep_sweep_count": 0
+    },
+    {
+      "session": "sess_6636_1783061209603_1783061209840_309bbb037fe8492a38533175c6dec7e7",
+      "provider": "openai-oauth",
+      "model": "gpt-5.5",
+      "wall_ms": 53078,
+      "turns": 7,
+      "tool_calls": 7,
+      "tools_per_turn": 1,
+      "total_tool_ms": 549,
+      "llm_stream_ms": 51861,
+      "cache_ratio": 0.47149308599591866,
+      "cached_tokens": 64000,
+      "prompt_tokens": 135739,
+      "uncached_tokens": 71739,
+      "output_tokens": 2800,
+      "cache_weighted_input_10": 78139,
+      "cache_weighted_input_25": 87739,
+      "prompt_growth": 41335,
+      "antipatterns": 1,
+      "issues": 14,
+      "issue_types": [
+        "cache",
+        "cache_break",
+        "cache_miss",
+        "broad_tool_result",
+        "service_tier",
+        "missed_parallelism",
+        "tool_churn_cluster"
+      ],
+      "read_fragmentation_paths": [],
+      "grep_sweep_count": 0
+    },
+    {
+      "session": "sess_12456_1783061273178_1783061273417_b75f3cff29775e56257b108f96388d21",
+      "provider": "openai-oauth",
+      "model": "gpt-5.5",
+      "wall_ms": 104870,
+      "turns": 10,
+      "tool_calls": 12,
+      "tools_per_turn": 1.2,
+      "total_tool_ms": 866,
+      "llm_stream_ms": 103586,
+      "cache_ratio": 0.8181367561503744,
+      "cached_tokens": 250880,
+      "prompt_tokens": 306648,
+      "uncached_tokens": 55768,
+      "output_tokens": 10163,
+      "cache_weighted_input_10": 80856,
+      "cache_weighted_input_25": 118488,
+      "prompt_growth": 44684,
+      "antipatterns": 1,
+      "issues": 15,
+      "issue_types": [
+        "cache",
+        "cache_break",
+        "readonly_stall",
+        "service_tier",
+        "missed_parallelism",
+        "tool_churn_cluster"
+      ],
+      "read_fragmentation_paths": [],
+      "grep_sweep_count": 0
+    },
+    {
+      "session": "sess_34340_1783061299082_1783061299306_c1576ad2a7e9cafad2329f5c0bf5a041",
+      "provider": "openai-oauth",
+      "model": "gpt-5.5",
+      "wall_ms": 58353,
+      "turns": 9,
+      "tool_calls": 14,
+      "tools_per_turn": 1.6,
+      "total_tool_ms": 2016,
+      "llm_stream_ms": 55759,
+      "cache_ratio": 0.6363588621444202,
+      "cached_tokens": 72704,
+      "prompt_tokens": 114250,
+      "uncached_tokens": 41546,
+      "output_tokens": 3718,
+      "cache_weighted_input_10": 48816.4,
+      "cache_weighted_input_25": 59722,
+      "prompt_growth": 11717,
+      "antipatterns": 0,
+      "issues": 12,
+      "issue_types": [
+        "cache",
+        "cache_break",
+        "cache_miss",
+        "service_tier"
+      ],
+      "read_fragmentation_paths": [],
+      "grep_sweep_count": 0
+    },
+    {
+      "session": "sess_40856_1783061322058_1783061322304_6da5edceb48f481a206a016ccc63c5fc",
+      "provider": "openai-oauth",
+      "model": "gpt-5.5",
+      "wall_ms": 87800,
+      "turns": 10,
+      "tool_calls": 13,
+      "tools_per_turn": 1.3,
+      "total_tool_ms": 816,
+      "llm_stream_ms": 86336,
+      "cache_ratio": 0.6981522354599766,
+      "cached_tokens": 181248,
+      "prompt_tokens": 259611,
+      "uncached_tokens": 78363,
+      "output_tokens": 4217,
+      "cache_weighted_input_10": 96487.8,
+      "cache_weighted_input_25": 123675,
+      "prompt_growth": 60577,
+      "antipatterns": 1,
+      "issues": 18,
+      "issue_types": [
+        "cache",
+        "cache_break",
+        "cache_miss",
+        "readonly_stall",
+        "broad_tool_result",
+        "service_tier",
+        "missed_parallelism",
+        "tool_churn_cluster"
+      ],
+      "read_fragmentation_paths": [],
+      "grep_sweep_count": 0
+    }
+  ]
+}