mixdog 0.9.3 → 0.9.5

package/scripts/bench/round-r21-delta-retry.json

@@ -0,0 +1,473 @@
+{
+  "round": "21",
+  "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_OAI_WS_DELTA": "1"
+    }
+  },
+  "tasks": 6,
+  "completed": 6,
+  "completion_rate": 100,
+  "sessions": [
+    "sess_53040_1783061772419_1783061772667_52530e4a94c00ff1a9af3420ba33999d",
+    "sess_44948_1783061772432_1783061772678_1dc2eccd0ea785768ca34e3a5ae803d0",
+    "sess_39556_1783061772446_1783061772692_aa271e435633f4354e8a22e39cf2c817",
+    "sess_44472_1783061844580_1783061844803_7cea5dcd79cef6e363bcecd4e5901e9b",
+    "sess_29908_1783061880433_1783061880659_3a0ccb5163abe4322c4be7c7cabec26e",
+    "sess_59932_1783061888249_1783061888489_ce57d1706ce48640a8a1d9b85bc7b315"
+  ],
+  "results": [
+    {
+      "id": "r4-debug-forwarding-gap#1",
+      "agent": "reviewer",
+      "ok": true,
+      "ms": 115828,
+      "sessionId": "sess_53040_1783061772419_1783061772667_52530e4a94c00ff1a9af3420ba33999d",
+      "output": "sessionId: sess_53040_1783061772419_1783061772667_52530e4a94c00ff1a9af3420ba33999d\nI’ll trace the activation-to-forwarding path directly in the channel runtime files and verify the recovery behavior against the implemented backlog logic.조사 완료: 활성화 직후 출력은 일부 경로에서 지연 후 회수되지만, “이미 존재하는 새 transcript를 처음 EOF로 바인딩”하거나 “잘못된 transcript를 500ms 안에 못 훔친” 경우는 첫 출력이 조용히 누락될 수 있습니다.\n\n- **최대 영향 — 첫 바인딩이 EOF로 시작**: `setContext()`는 기본/비매칭 `catchUpFromPersisted`에서 현재 파일 크기를 커서로 잡습니다(`src/runtime/channels/lib/output-forwarder.mjs:111`–`127`). 활성화 후 바인딩 전에 이미 쓰인 출력은 `replayFromStart`나 같은 transcript의 persisted cursor가 없으면 회수되지 않습니다.\n\n- **잘못된/stale transcript 바인딩**: 첫 inbound에서 self transcript 발견이 늦으면 500ms만 재탐색합니다(`src/runtime/channels/index.mjs:1526`–`1531`, `1571`–`1596`). 이후 실제 transcript가 이미 존재한 상태로 훔치면 일반 `applyTranscriptBinding()`이 EOF로 잡아 이전 출력은 회수되지 않을 수 있습니다(`src/runtime/channels/index.mjs:1597`–`1602`).\n\n- **소유권/백엔드 연결 전 watch flush**: owner probe가 false면 `forwardNewText()`가 읽지 않고 종료합니다(`src/runtime/channels/index.mjs:316`, `src/runtime/channels/lib/output-forwarder.mjs:351`–`354`). 커서는 안 움직여 손실은 아니며, 연결 후 `forwardNewText()`로 회수됩니다(`src/runtime/channels/index.mjs:801`–`812`).\n\n- **아직 없는 transcript 파일**: `fs.watch` 설치 실패는 re-arm poll이 파일 생성 후 watch 설치와 즉시 catch-up을 수행합니다(`src/runtime/channels/index.mjs:339`–`346`, `376`–`381`). 단 60초 제한 이후 생성되면 다음 rebind 전까지 지연/미회수입니다(`src/runtime/channels/index.mjs:348`–`385`).\n\n- **stat→watch 설치 사이 쓰기**: watch 설치 직후 flush를 예약해 그 사이의 backlog를 읽습니다(`src/runtime/channels/lib/output-forwarder.mjs:647`–`667`, `703`–`731`). 이 구간은 회수됩니다.\n\n판정: bridge 활성화 창의 지연은 대부분 backlog catch-up으로 복구되지만, EOF 초기 바인딩/늦은 self-transcript 탈환 경로에는 실제 silent drop 가능성이 남아 있습니다.",
+      "card": {
+        "session": "sess_53040_1783061772419_1783061772667_52530e4a94c00ff1a9af3420ba33999d",
+        "provider": "openai-oauth",
+        "model": "gpt-5.5",
+        "wall_ms": 105316,
+        "turns": 7,
+        "tool_calls": 8,
+        "tools_per_turn": 1.1,
+        "total_tool_ms": 657,
+        "llm_stream_ms": 104332,
+        "cache_ratio": 0.5201565655306525,
+        "cached_tokens": 83456,
+        "prompt_tokens": 160444,
+        "uncached_tokens": 76988,
+        "output_tokens": 10039,
+        "cache_weighted_input_10": 85333.6,
+        "cache_weighted_input_25": 97852,
+        "prompt_growth": 32398,
+        "antipatterns": 0,
+        "issues": 4,
+        "issue_types": [
+          "cache_break",
+          "cache_miss",
+          "cache",
+          "service_tier"
+        ],
+        "read_fragmentation_paths": [],
+        "grep_sweep_count": 0
+      }
+    },
+    {
+      "id": "r4-impl-usage-summary#1",
+      "agent": "worker",
+      "ok": true,
+      "ms": 72147,
+      "sessionId": "sess_44948_1783061772432_1783061772678_1dc2eccd0ea785768ca34e3a5ae803d0",
+      "output": "sessionId: sess_44948_1783061772432_1783061772678_1dc2eccd0ea785768ca34e3a5ae803d0\n요청 범위가 작고 명확해서 바로 구현한 뒤 검증까지 진행하겠습니다.완료했습니다, 재영님.\n\n- **추가**: `usage:today` npm script와 `scripts/usage-today.mjs`가 구성됐습니다.\n- **동작**: `MIXDOG_DATA_DIR`를 우선 사용하고, 오늘 0시 이후 `usage_raw/usage`를 모델별·총합으로 집계합니다.\n- **검증**: `node --check scripts/usage-today.mjs` 통과, `npm run usage:today` 실행 완료.\n- **참고**: 워킹트리에 기존 수정 `src/runtime/channels/lib/*` 2개가 있어 건드리지 않았습니다.",
+      "card": {
+        "session": "sess_44948_1783061772432_1783061772678_1dc2eccd0ea785768ca34e3a5ae803d0",
+        "provider": "openai-oauth",
+        "model": "gpt-5.5",
+        "wall_ms": 61635,
+        "turns": 12,
+        "tool_calls": 12,
+        "tools_per_turn": 1,
+        "total_tool_ms": 2827,
+        "llm_stream_ms": 58371,
+        "cache_ratio": 0.5649887302779865,
+        "cached_tokens": 96256,
+        "prompt_tokens": 170368,
+        "uncached_tokens": 74112,
+        "output_tokens": 3491,
+        "cache_weighted_input_10": 83737.6,
+        "cache_weighted_input_25": 98176,
+        "prompt_growth": 12579,
+        "antipatterns": 1,
+        "issues": 9,
+        "issue_types": [
+          "cache_break",
+          "cache_miss",
+          "cache",
+          "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": 107990,
+      "sessionId": "sess_39556_1783061772446_1783061772692_aa271e435633f4354e8a22e39cf2c817",
+      "output": "sessionId: sess_39556_1783061772446_1783061772692_aa271e435633f4354e8a22e39cf2c817\n재영님, compaction subsystem 맵입니다.\n\n- **Entry points**: engine exports `semanticCompactMessages` and `recallFastTrackCompactMessages`; facade `compact.mjs` re-exports them as the stable public API. `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- **Main callers**: `agentLoop` runs auto pre-send compaction; manager uses `runSessionCompaction` for manual compact and auto-clear pre-compact. `src/runtime/agent/orchestrator/session/loop.mjs:324-330`, `src/runtime/agent/orchestrator/session/manager.mjs:1841-1850`, `src/runtime/agent/orchestrator/session/manager.mjs:1943-1957`\n\n- **Policy/trigger**: auto is disabled by `cfg.auto === false`; boundary comes from explicit boundary/context/autoLimit, trigger is explicit sub-boundary limit or `boundary - buffer`. `src/runtime/agent/orchestrator/session/loop/compact-policy.mjs:89-117`\n\n- **Fire decision**: compaction fires when `forceReactive`, token estimate is unavailable, or `messageTokens + reserve >= trigger`; otherwise telemetry records `pre_send_check`. `src/runtime/agent/orchestrator/session/loop/compact-policy.mjs:170-174`, `src/runtime/agent/orchestrator/session/loop.mjs:351-371`\n\n- **Type routing**: agent-owned sessions force semantic; otherwise config/env selects semantic vs recall-fasttrack, and loop dispatches recall first or semantic. `src/runtime/agent/orchestrator/session/loop/compact-policy.mjs:39-51`, `src/runtime/agent/orchestrator/session/loop.mjs:414-465`\n\n- **Semantic flow**: sanitize → no-op if already within budget unless forced → redact → select window → prompt summary provider → insert summary between protected prefix and preserved tail. `src/runtime/agent/orchestrator/session/compact/engine.mjs:284-314`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:367-392`\n\n- **Window split**: protected system/reminder prefix is separated, prior summary is detected, live excludes summary messages, and head/tail are cut from live. `src/runtime/agent/orchestrator/session/compact/engine.mjs:100-111`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:235-267`\n\n- **Semantic tail choice**: with user turns, keep up to `tailTurns` recent turns within `preserveRecentBudget`, splitting newest suffix if needed; without user turns, cut at newest valid non-tool boundary that fits. `src/runtime/agent/orchestrator/session/compact/engine.mjs:134-185`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:193-232`, `src/runtime/agent/orchestrator/session/compact/budget.mjs:274-288`\n\n- **Recall tail choice**: recall-fasttrack preserves newest structured turns, drops oldest kept turns until cap, then truncates the oldest surviving turn/message while allowing no-user assistant/tool tails. `src/runtime/agent/orchestrator/session/compact/engine.mjs:593-619`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:525-575`\n\n- **Redaction hook**: no-op returns the sanitized transcript unchanged; actual compaction redacts tool-call argument values before window selection, so preserved prefix/tail are emitted redacted. `src/runtime/agent/orchestrator/session/compact/engine.mjs:286-312`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:389-392`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:644-651`\n\n- **Redaction behavior**: sensitive keys become `[redacted]`, raw secret pairs inside strings are scrubbed, and tool-call structure/order/pairing is preserved. `src/runtime/agent/orchestrator/session/compact/text-utils.mjs:219-238`, `src/runtime/agent/orchestrator/session/compact/text-utils.mjs:261-326`\n\n- **Budget fallback**: if mandatory prefix+tail plus summary room exceeds target, working budget is raised instead of refusing. `src/runtime/agent/orchestrator/session/compact/engine.mjs:319-331`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:658-664`\n\n- **Prompt/summary fallback**: semantic prompt fitting shrinks/drops prior summary/facts and can reduce older head messages; malformed provider summaries are schema-repaired before insertion. `src/runtime/agent/orchestrator/session/compact/summary.mjs:199-281`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:371-384`\n\n- **Runtime fallback**: semantic failures try anchor-independent `pruneToolOutputsUnanchored`; recall-fasttrack failures do not silently prune in loop, while manager can try semantic fallback after recall failure. `src/runtime/agent/orchestrator/session/loop.mjs:500-537`, `src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:231-278`\n\n- **Hard failure paths**: unresolved compact failure records `overflow_failed` and throws `agentContextOverflowError`; provider send overflow gets one reactive compact retry, then surfaces overflow. `src/runtime/agent/orchestrator/session/loop.mjs:552-626`, `src/runtime/agent/orchestrator/session/loop.mjs:982-1026`\n\n- **Persistence recovery**: stale `lastStage='compacting'` is normalized to `interrupted` on fresh ask/resume so UI does not stay stuck. `src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:28-43`, `src/runtime/agent/orchestrator/session/manager.mjs:1233-1237`, `src/runtime/agent/orchestrator/session/manager.mjs:1774-1781`",
+      "card": {
+        "session": "sess_39556_1783061772446_1783061772692_aa271e435633f4354e8a22e39cf2c817",
+        "provider": "openai-oauth",
+        "model": "gpt-5.5",
+        "wall_ms": 97475,
+        "turns": 13,
+        "tool_calls": 18,
+        "tools_per_turn": 1.4,
+        "total_tool_ms": 838,
+        "llm_stream_ms": 96234,
+        "cache_ratio": 0.8082769625109276,
+        "cached_tokens": 380928,
+        "prompt_tokens": 471284,
+        "uncached_tokens": 90356,
+        "output_tokens": 7636,
+        "cache_weighted_input_10": 128448.8,
+        "cache_weighted_input_25": 185588,
+        "prompt_growth": 77840,
+        "antipatterns": 1,
+        "issues": 8,
+        "issue_types": [
+          "cache_break",
+          "readonly_stall",
+          "cache",
+          "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": 88433,
+      "sessionId": "sess_44472_1783061844580_1783061844803_7cea5dcd79cef6e363bcecd4e5901e9b",
+      "output": "sessionId: sess_44472_1783061844580_1783061844803_7cea5dcd79cef6e363bcecd4e5901e9b\nTraced the activation-to-forward path; the main true-loss cases are binding/cursor selection, while owner/backend/watch races are mostly delayed and caught up.\n\n- **Highest impact — no usable binding:** if no `channelId` is restored or auto-bound, no context is installed, and a later normal bind starts at EOF; backlog catch-up does **not** recover it. `src/runtime/channels/index.mjs:610-629`, `src/runtime/channels/lib/output-forwarder.mjs:93-129`\n\n- **Existing transcript before first bind:** `catchUpFromPersisted` only replays from a persisted cursor for the same transcript; a fresh existing file is cursored to current size, so earlier assistant bytes are skipped. Recovery: **only if same persisted path exists**. `src/runtime/channels/lib/output-forwarder.mjs:117-126`\n\n- **Bridge active before owner/backend ready:** `bridge-state.json` is written before `refreshBridgeOwnership`; `_isOwner()` blocks forwarding but does not advance the cursor, then post-connect `forwardNewText()` catches up. Recovery: **yes, delayed**. `src/runtime/channels/index.mjs:1735-1747`, `src/runtime/channels/index.mjs:801-812`, `src/runtime/channels/lib/output-forwarder.mjs:351-358`\n\n- **Transcript file not yet created:** initial `fs.watch()` fails, but pending re-arm polls and forwards backlog once the file appears. Recovery: **yes within 60s; no after timeout unless another rebind occurs**. `src/runtime/channels/index.mjs:339-389`, `src/runtime/channels/lib/output-forwarder.mjs:647-670`\n\n- **Pre-connect re-arm exits on non-owner:** a re-arm tick can stop while `bridgeRuntimeConnected` is still false, but startup re-arms or forwards after connect. Recovery: **yes**. `src/runtime/channels/index.mjs:369-379`, `src/runtime/channels/index.mjs:803-812`\n\n- **Watch/debounce race after binding:** bytes written between `setContext()` stat and watch install are covered by scheduled flush; flush debounce adds latency. Recovery: **yes, delayed about 100ms+send time**. `src/runtime/channels/lib/output-forwarder.mjs:665-666`, `src/runtime/channels/lib/output-forwarder.mjs:703-731`\n\nVerdict: the activation window is mostly catch-up safe, except when the first bind has no destination or binds an already-existing unpersisted transcript at EOF.",
+      "card": {
+        "session": "sess_44472_1783061844580_1783061844803_7cea5dcd79cef6e363bcecd4e5901e9b",
+        "provider": "openai-oauth",
+        "model": "gpt-5.5",
+        "wall_ms": 77953,
+        "turns": 7,
+        "tool_calls": 8,
+        "tools_per_turn": 1.1,
+        "total_tool_ms": 658,
+        "llm_stream_ms": 76661,
+        "cache_ratio": 0.336732434720323,
+        "cached_tokens": 51712,
+        "prompt_tokens": 153570,
+        "uncached_tokens": 101858,
+        "output_tokens": 8026,
+        "cache_weighted_input_10": 107029.2,
+        "cache_weighted_input_25": 114786,
+        "prompt_growth": 35224,
+        "antipatterns": 1,
+        "issues": 8,
+        "issue_types": [
+          "cache_break",
+          "cache_miss",
+          "cache",
+          "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": 68113,
+      "sessionId": "sess_29908_1783061880433_1783061880659_3a0ccb5163abe4322c4be7c7cabec26e",
+      "output": "sessionId: sess_29908_1783061880433_1783061880659_3a0ccb5163abe4322c4be7c7cabec26e\n구현 기준을 확인한 뒤 바로 반영하고 검증하겠습니다.이미 일부 구현이 있어 실제 기록 포맷과 맞는지 확인한 뒤 필요한 보정만 하겠습니다.기능 파일은 이미 작업 트리에 준비되어 있어 문법 검사와 실제 실행으로 검증하겠습니다.완료했습니다, 재영님. `usage:today` 스크립트와 `scripts/usage-today.mjs`가 준비됐고 검증도 통과했습니다.\n\n- **Diff 요약**: `package.json`에 `usage:today` 추가, 새 스크립트가 오늘 0시 이후 usage를 모델별/총합으로 집계합니다.\n- **검증**: `node --check scripts/usage-today.mjs` 통과, `node scripts/usage-today.mjs` 실행 완료.\n- **참고**: 기존 수정 파일 `src/runtime/channels/...` 2개는 건드리지 않았습니다.",
+      "card": {
+        "session": "sess_29908_1783061880433_1783061880659_3a0ccb5163abe4322c4be7c7cabec26e",
+        "provider": "openai-oauth",
+        "model": "gpt-5.5",
+        "wall_ms": 57635,
+        "turns": 11,
+        "tool_calls": 10,
+        "tools_per_turn": 0.9,
+        "total_tool_ms": 1834,
+        "llm_stream_ms": 55184,
+        "cache_ratio": 0.7788073227427078,
+        "cached_tokens": 145408,
+        "prompt_tokens": 186706,
+        "uncached_tokens": 41298,
+        "output_tokens": 3729,
+        "cache_weighted_input_10": 55838.8,
+        "cache_weighted_input_25": 77650,
+        "prompt_growth": 18346,
+        "antipatterns": 1,
+        "issues": 6,
+        "issue_types": [
+          "cache_break",
+          "cache_miss",
+          "cache",
+          "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": 72005,
+      "sessionId": "sess_59932_1783061888249_1783061888489_ce57d1706ce48640a8a1d9b85bc7b315",
+      "output": "sessionId: sess_59932_1783061888249_1783061888489_ce57d1706ce48640a8a1d9b85bc7b315\n compaction 구조를 `engine.mjs` 기준으로 호출자·분기·실패 경로까지 추적해 간결히 정리하겠습니다.완료했습니다. `engine.mjs` 기준 compaction 아키텍처 요약입니다.\n\n- **Entry points**: `semanticCompactMessages`와 `recallFastTrackCompactMessages`가 public engine API이고 facade가 그대로 re-export한다. `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- **Callers**: live loop는 recall-fasttrack 또는 semantic 중 하나를 호출하고, manager runner는 manual/auto session compaction에서 같은 두 경로를 호출한다. `src/runtime/agent/orchestrator/session/loop.mjs:414`, `src/runtime/agent/orchestrator/session/loop.mjs:439`, `src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:207`, `src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:280`\n\n- **Fire decision**: worker policy computes boundary/trigger/buffer/reserve/type; auto fires when reactive, estimate is null, or `messages+reserve >= 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:170`\n\n- **Loop trigger path**: loop estimates transcript tokens, calls `shouldCompactForSession`, then targets `compactTargetBudget` before running compaction. `src/runtime/agent/orchestrator/session/loop.mjs:351`, `src/runtime/agent/orchestrator/session/loop.mjs:353`, `src/runtime/agent/orchestrator/session/loop.mjs:362`\n\n- **Manager trigger path**: manual sets `force`, auto can return `below threshold`, and missing boundary only throws for forced compaction. `src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:143`, `src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:153`, `src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:176`\n\n- **Common preprocessing**: engine sanitizes tool pairs, dedups tool bodies, separates protected prefix, finds previous summary, and treats non-summary messages as live history. `src/runtime/agent/orchestrator/session/compact/engine.mjs:100`\n\n- **Semantic no-op**: if sanitized input already fits and `force` is not true, semantic returns unchanged without preserved-tail redaction. `src/runtime/agent/orchestrator/session/compact/engine.mjs:284`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:289`\n\n- **Semantic tail**: with user turns, it keeps up to `tailTurns` recent turns within `preserveRecentBudget`, splitting the newest turn suffix if needed. `src/runtime/agent/orchestrator/session/compact/engine.mjs:134`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:151`\n\n- **No-user semantic tail**: without user anchors, valid cut points exclude `tool` messages so assistant/tool pairs stay together, then newest-fitting suffix is chosen. `src/runtime/agent/orchestrator/session/compact/engine.mjs:121`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:193`\n\n- **Semantic window**: `selectCompactionWindow` chooses user-turn or cut-point path, splits `head/tail`, repairs tail pairs, and only empty live history is fatal. `src/runtime/agent/orchestrator/session/compact/engine.mjs:235`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:240`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:250`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:254`\n\n- **Semantic execution**: mandatory cost is `system+tail`; budget is raised to fit mandatory plus summary room; prompt is fit and sent to provider. `src/runtime/agent/orchestrator/session/compact/engine.mjs:319`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:328`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:333`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:367`\n\n- **Semantic output**: provider text is schema-repaired, summary is fit into remaining room, then result is `system + summary + tail` with final budget check. `src/runtime/agent/orchestrator/session/compact/engine.mjs:371`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:376`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:384`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:392`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:395`\n\n- **Recall tail**: recall-fasttrack preserves newest structured turns, default max two user turns and 8k cap, dropping oldest turns then truncating oldest surviving content. `src/runtime/agent/orchestrator/session/compact/engine.mjs:449`, `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:610`\n\n- **Recall no-user handling**: no-user recent tail is valid, and orphan tool results are repaired by extending backward or rebuilding from newest non-tool. `src/runtime/agent/orchestrator/session/compact/engine.mjs:525`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:558`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:613`\n\n- **Recall execution**: recall path redacts, splits live context, selects preserved tail, requires compactable prior/recall unless explicitly allowed, caps recall injection room, then emits `system + recall summary + tail`. `src/runtime/agent/orchestrator/session/compact/engine.mjs:644`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:646`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:651`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:653`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:681`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:695`\n\n- **Redaction hook**: semantic and recall both call `redactToolCallSecretsInMessages` after fit/no-op check but before window/tail 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`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:389`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:644`\n\n- **Engine failures**: semantic throws for missing provider, empty live, no compactable prior, prompt not fitting, empty provider summary, summary not fitting, 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- **Fallbacks**: loop semantic failure can recover via unanchored prune, but recall-fasttrack failures skip that prune and escalate; manager recall-fasttrack failure first tries semantic fallback, then semantic/prune failure records `*_failed`. `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:237`, `src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:245`, `src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:327`, `src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:348`",
+      "card": {
+        "session": "sess_59932_1783061888249_1783061888489_ce57d1706ce48640a8a1d9b85bc7b315",
+        "provider": "openai-oauth",
+        "model": "gpt-5.5",
+        "wall_ms": 61515,
+        "turns": 8,
+        "tool_calls": 9,
+        "tools_per_turn": 1.1,
+        "total_tool_ms": 318,
+        "llm_stream_ms": 60614,
+        "cache_ratio": 0.5796200062741247,
+        "cached_tokens": 84992,
+        "prompt_tokens": 146634,
+        "uncached_tokens": 61642,
+        "output_tokens": 4622,
+        "cache_weighted_input_10": 70141.2,
+        "cache_weighted_input_25": 82890,
+        "prompt_growth": 24384,
+        "antipatterns": 1,
+        "issues": 6,
+        "issue_types": [
+          "cache_break",
+          "cache_miss",
+          "cache",
+          "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": 76921.5,
+    "turns": 9.7,
+    "tool_calls": 10.8,
+    "tools_per_turn": 1.1,
+    "total_tool_ms": 1188.7,
+    "llm_stream_ms": 75232.7,
+    "cache_ratio": 0.6,
+    "cached_tokens": 140458.7,
+    "prompt_tokens": 214834.3,
+    "uncached_tokens": 74375.7,
+    "output_tokens": 6257.2,
+    "cache_weighted_input_10": 88421.5,
+    "cache_weighted_input_25": 109490.3,
+    "prompt_growth": 33461.8,
+    "antipatterns": 0.8,
+    "issues": 6.8,
+    "grep_sweep_count": 0
+  },
+  "cards": [
+    {
+      "session": "sess_53040_1783061772419_1783061772667_52530e4a94c00ff1a9af3420ba33999d",
+      "provider": "openai-oauth",
+      "model": "gpt-5.5",
+      "wall_ms": 105316,
+      "turns": 7,
+      "tool_calls": 8,
+      "tools_per_turn": 1.1,
+      "total_tool_ms": 657,
+      "llm_stream_ms": 104332,
+      "cache_ratio": 0.5201565655306525,
+      "cached_tokens": 83456,
+      "prompt_tokens": 160444,
+      "uncached_tokens": 76988,
+      "output_tokens": 10039,
+      "cache_weighted_input_10": 85333.6,
+      "cache_weighted_input_25": 97852,
+      "prompt_growth": 32398,
+      "antipatterns": 0,
+      "issues": 4,
+      "issue_types": [
+        "cache_break",
+        "cache_miss",
+        "cache",
+        "service_tier"
+      ],
+      "read_fragmentation_paths": [],
+      "grep_sweep_count": 0
+    },
+    {
+      "session": "sess_44948_1783061772432_1783061772678_1dc2eccd0ea785768ca34e3a5ae803d0",
+      "provider": "openai-oauth",
+      "model": "gpt-5.5",
+      "wall_ms": 61635,
+      "turns": 12,
+      "tool_calls": 12,
+      "tools_per_turn": 1,
+      "total_tool_ms": 2827,
+      "llm_stream_ms": 58371,
+      "cache_ratio": 0.5649887302779865,
+      "cached_tokens": 96256,
+      "prompt_tokens": 170368,
+      "uncached_tokens": 74112,
+      "output_tokens": 3491,
+      "cache_weighted_input_10": 83737.6,
+      "cache_weighted_input_25": 98176,
+      "prompt_growth": 12579,
+      "antipatterns": 1,
+      "issues": 9,
+      "issue_types": [
+        "cache_break",
+        "cache_miss",
+        "cache",
+        "service_tier",
+        "missed_parallelism",
+        "tool_churn_cluster"
+      ],
+      "read_fragmentation_paths": [],
+      "grep_sweep_count": 0
+    },
+    {
+      "session": "sess_39556_1783061772446_1783061772692_aa271e435633f4354e8a22e39cf2c817",
+      "provider": "openai-oauth",
+      "model": "gpt-5.5",
+      "wall_ms": 97475,
+      "turns": 13,
+      "tool_calls": 18,
+      "tools_per_turn": 1.4,
+      "total_tool_ms": 838,
+      "llm_stream_ms": 96234,
+      "cache_ratio": 0.8082769625109276,
+      "cached_tokens": 380928,
+      "prompt_tokens": 471284,
+      "uncached_tokens": 90356,
+      "output_tokens": 7636,
+      "cache_weighted_input_10": 128448.8,
+      "cache_weighted_input_25": 185588,
+      "prompt_growth": 77840,
+      "antipatterns": 1,
+      "issues": 8,
+      "issue_types": [
+        "cache_break",
+        "readonly_stall",
+        "cache",
+        "broad_tool_result",
+        "service_tier",
+        "missed_parallelism",
+        "tool_churn_cluster"
+      ],
+      "read_fragmentation_paths": [],
+      "grep_sweep_count": 0
+    },
+    {
+      "session": "sess_44472_1783061844580_1783061844803_7cea5dcd79cef6e363bcecd4e5901e9b",
+      "provider": "openai-oauth",
+      "model": "gpt-5.5",
+      "wall_ms": 77953,
+      "turns": 7,
+      "tool_calls": 8,
+      "tools_per_turn": 1.1,
+      "total_tool_ms": 658,
+      "llm_stream_ms": 76661,
+      "cache_ratio": 0.336732434720323,
+      "cached_tokens": 51712,
+      "prompt_tokens": 153570,
+      "uncached_tokens": 101858,
+      "output_tokens": 8026,
+      "cache_weighted_input_10": 107029.2,
+      "cache_weighted_input_25": 114786,
+      "prompt_growth": 35224,
+      "antipatterns": 1,
+      "issues": 8,
+      "issue_types": [
+        "cache_break",
+        "cache_miss",
+        "cache",
+        "broad_tool_result",
+        "service_tier",
+        "missed_parallelism",
+        "tool_churn_cluster"
+      ],
+      "read_fragmentation_paths": [],
+      "grep_sweep_count": 0
+    },
+    {
+      "session": "sess_29908_1783061880433_1783061880659_3a0ccb5163abe4322c4be7c7cabec26e",
+      "provider": "openai-oauth",
+      "model": "gpt-5.5",
+      "wall_ms": 57635,
+      "turns": 11,
+      "tool_calls": 10,
+      "tools_per_turn": 0.9,
+      "total_tool_ms": 1834,
+      "llm_stream_ms": 55184,
+      "cache_ratio": 0.7788073227427078,
+      "cached_tokens": 145408,
+      "prompt_tokens": 186706,
+      "uncached_tokens": 41298,
+      "output_tokens": 3729,
+      "cache_weighted_input_10": 55838.8,
+      "cache_weighted_input_25": 77650,
+      "prompt_growth": 18346,
+      "antipatterns": 1,
+      "issues": 6,
+      "issue_types": [
+        "cache_break",
+        "cache_miss",
+        "cache",
+        "service_tier",
+        "missed_parallelism",
+        "tool_churn_cluster"
+      ],
+      "read_fragmentation_paths": [],
+      "grep_sweep_count": 0
+    },
+    {
+      "session": "sess_59932_1783061888249_1783061888489_ce57d1706ce48640a8a1d9b85bc7b315",
+      "provider": "openai-oauth",
+      "model": "gpt-5.5",
+      "wall_ms": 61515,
+      "turns": 8,
+      "tool_calls": 9,
+      "tools_per_turn": 1.1,
+      "total_tool_ms": 318,
+      "llm_stream_ms": 60614,
+      "cache_ratio": 0.5796200062741247,
+      "cached_tokens": 84992,
+      "prompt_tokens": 146634,
+      "uncached_tokens": 61642,
+      "output_tokens": 4622,
+      "cache_weighted_input_10": 70141.2,
+      "cache_weighted_input_25": 82890,
+      "prompt_growth": 24384,
+      "antipatterns": 1,
+      "issues": 6,
+      "issue_types": [
+        "cache_break",
+        "cache_miss",
+        "cache",
+        "service_tier",
+        "missed_parallelism",
+        "tool_churn_cluster"
+      ],
+      "read_fragmentation_paths": [],
+      "grep_sweep_count": 0
+    }
+  ]
+}