mixdog 0.9.18 → 0.9.20

package/scripts/bench/round-r5-gpt-lead.json

@@ -1,259 +0,0 @@
-{
-  "round": "5",
-  "runner": "lead",
-  "opts": {
-    "provider": null,
-    "model": null,
-    "effort": null,
-    "fast": false,
-    "env": {
-      "MIXDOG_DATA_DIR": "C:\\Users\\tempe\\.mixdog\\bench-data"
-    }
-  },
-  "tasks": 3,
-  "completed": 3,
-  "completion_rate": 100,
-  "sessions": [
-    "sess_59600_1783037462018_1783037462266_c37bce6b74c830c56d6c7aff52efeca7",
-    "sess_54696_1783037462027_1783037462276_0dfadaccfebd2d42d0c14e5053f61d4a",
-    "sess_59440_1783037462045_1783037462291_6eb45d2479f73ed04b11f9f18b0b8109"
-  ],
-  "results": [
-    {
-      "id": "r4-debug-forwarding-gap",
-      "agent": "reviewer",
-      "ok": true,
-      "ms": 99020,
-      "sessionId": "sess_59600_1783037462018_1783037462266_c37bce6b74c830c56d6c7aff52efeca7",
-      "output": "sessionId: sess_59600_1783037462018_1783037462266_c37bce6b74c830c56d6c7aff52efeca7\nInvestigated: the bridge has recovery for the known pending-file/owner-start race, but there are still EOF-binding cases that can permanently skip early output.\n\n**Path traced**: `start()` writes `bridge-state.json` before runtime readiness, then ownership/startup binds a transcript and `OutputForwarder.startWatch()` triggers `forwardNewText()` → `backend.sendMessage()` (`src/runtime/channels/index.mjs:1734`, `src/runtime/channels/index.mjs:710`, `src/runtime/channels/index.mjs:317`, `src/runtime/channels/lib/output-forwarder.mjs:647`, `src/runtime/channels/lib/output-forwarder.mjs:351`).\n\n**Highest impact — existing transcript + no matching persisted cursor drops backlog**: `catchUpFromPersisted` falls back to `currentSize` when the persisted path does not match, so bytes already written before binding become the baseline and are never reread (`src/runtime/channels/lib/output-forwarder.mjs:117`); backlog catch-up does **not** recover this.\n\n**No channel binding also drops pre-bind existing backlog**: if no persisted/main channel exists, activation returns without binding (`src/runtime/channels/index.mjs:624`), and the first inbound bind uses default EOF cursoring (`src/runtime/channels/index.mjs:1597`, `src/runtime/channels/lib/output-forwarder.mjs:125`); backlog catch-up does **not** recover already-written output in an existing transcript.\n\n**Pending transcript file is mostly recovered**: binding a not-yet-created path makes `fs.watch` fail/close (`src/runtime/channels/lib/output-forwarder.mjs:656`), but `schedulePendingTranscriptRearm()` polls, installs the watcher, and calls `forwardNewText()` once the file appears (`src/runtime/channels/index.mjs:357`, `src/runtime/channels/index.mjs:376`); backlog catch-up **does** recover it if the path is known and appears within 60s, otherwise recovery is not automatic (`src/runtime/channels/index.mjs:384`).\n\n**Ownership/startup probe window delays, not drops**: pre-connect forwarding is owner-gated off because `_isOwner()` requires `bridgeRuntimeConnected && currentOwnerState().owned` (`src/runtime/channels/index.mjs:316`, `src/runtime/channels/lib/output-forwarder.mjs:351`), but startup explicitly forwards or re-arms after connect (`src/runtime/channels/index.mjs:801`); backlog catch-up **does** recover unless the earlier binding already set the cursor to EOF.\n\nVerdict: the pending-file bridge race is covered, but activation can still silently lose early assistant output when an existing transcript is bound without a matching persisted cursor.",
-      "card": {
-        "session": "sess_59600_1783037462018_1783037462266_c37bce6b74c830c56d6c7aff52efeca7",
-        "provider": "openai-oauth",
-        "model": "gpt-5.5",
-        "wall_ms": 88475,
-        "turns": 5,
-        "tool_calls": 5,
-        "tools_per_turn": 1,
-        "total_tool_ms": 428,
-        "llm_stream_ms": 87399,
-        "cache_ratio": 0.4883110541414881,
-        "cached_tokens": 64000,
-        "prompt_tokens": 131064,
-        "uncached_tokens": 67064,
-        "output_tokens": 9669,
-        "cache_weighted_input_10": 73464,
-        "cache_weighted_input_25": 83064,
-        "prompt_growth": 32405,
-        "antipatterns": 0,
-        "issues": 4,
-        "issue_types": [
-          "cache",
-          "cache_break",
-          "broad_tool_result",
-          "service_tier"
-        ],
-        "read_fragmentation_paths": [],
-        "grep_sweep_count": 0
-      }
-    },
-    {
-      "id": "r4-impl-usage-summary",
-      "agent": "worker",
-      "ok": true,
-      "ms": 125876,
-      "sessionId": "sess_54696_1783037462027_1783037462276_0dfadaccfebd2d42d0c14e5053f61d4a",
-      "output": "sessionId: sess_54696_1783037462027_1783037462276_0dfadaccfebd2d42d0c14e5053f61d4a\n구현 위치와 기존 스크립트 스타일을 확인한 뒤 바로 반영하겠습니다.재영님, 구현 완료했고 `usage:today` 실행까지 확인했습니다.\n\n- **바뀐 점**: `package.json`에 `usage:today`가 추가됐고, 새 `scripts/usage-today.mjs`가 `MIXDOG_DATA_DIR` 또는 기본 `~/.mixdog/data/history/agent-trace.jsonl`을 읽습니다.\n\n- **동작**: 로컬 자정 이후 `usage`/`usage_raw` 행을 모델별로 집계해 `calls`, `input`, `cached`, `output`, `costUsd`와 `GRAND TOTAL`을 출력합니다.\n\n- **검증**: `node --check scripts/usage-today.mjs` 통과, `npm run usage:today` 실행 성공. 현재 실행 결과 총합은 `858 calls`, `60.3242 costUsd`였습니다.\n\n- **diff 요약**: `package.json` 1줄 추가, `scripts/usage-today.mjs` 신규 파일. 기존 작업물로 보이는 `src/runtime/channels/lib/*` 수정 2건은 건드리지 않았습니다.",
-      "card": {
-        "session": "sess_54696_1783037462027_1783037462276_0dfadaccfebd2d42d0c14e5053f61d4a",
-        "provider": "openai-oauth",
-        "model": "gpt-5.5",
-        "wall_ms": 115327,
-        "turns": 13,
-        "tool_calls": 13,
-        "tools_per_turn": 1,
-        "total_tool_ms": 2688,
-        "llm_stream_ms": 112157,
-        "cache_ratio": 0.7477417701432699,
-        "cached_tokens": 212992,
-        "prompt_tokens": 284847,
-        "uncached_tokens": 71855,
-        "output_tokens": 8740,
-        "cache_weighted_input_10": 93154.2,
-        "cache_weighted_input_25": 125103,
-        "prompt_growth": 27149,
-        "antipatterns": 1,
-        "issues": 7,
-        "issue_types": [
-          "readonly_stall",
-          "cache",
-          "cache_break",
-          "tool_error",
-          "service_tier",
-          "missed_parallelism",
-          "tool_churn_cluster"
-        ],
-        "read_fragmentation_paths": [],
-        "grep_sweep_count": 0
-      }
-    },
-    {
-      "id": "r4-explain-compact-engine",
-      "agent": "explore",
-      "ok": true,
-      "ms": 98391,
-      "sessionId": "sess_59440_1783037462045_1783037462291_6eb45d2479f73ed04b11f9f18b0b8109",
-      "output": "ession/loop/compact-policy.mjs:89`, `src/runtime/agent/orchestrator/session/loop/compact-policy.mjs:100`\n\n- **Threshold math:** boundary comes from `compactBoundaryTokens/contextWindow/autoCompactTokenLimit`, trigger is explicit sub-boundary limit or boundary minus buffer, and reserve includes request tools plus configured reserve. `src/runtime/agent/orchestrator/session/loop/compact-policy.mjs:94`, `src/runtime/agent/orchestrator/session/loop/compact-policy.mjs:117`, `src/runtime/agent/orchestrator/session/loop/compact-policy.mjs:118`\n\n- **Fire condition:** compaction fires when forced reactive overflow is pending, token estimate is unknown, or transcript pressure plus reserve reaches trigger. `src/runtime/agent/orchestrator/session/loop/compact-policy.mjs:170`\n\n- **No-fire path:** if below threshold, telemetry records `pre_send_check` without mutating messages. `src/runtime/agent/orchestrator/session/loop.mjs:365`, `src/runtime/agent/orchestrator/session/loop.mjs:366`\n\n- **Reactive path:** provider context overflow marks a one-shot reactive retry and loops back through the normal pre-send compact path. `src/runtime/agent/orchestrator/session/loop.mjs:982`, `src/runtime/agent/orchestrator/session/loop.mjs:995`, `src/runtime/agent/orchestrator/session/loop.mjs:1010`\n\n- **Mode selection:** agent-owned sessions are forced semantic; otherwise type comes from env/config and resolves to semantic or recall-fasttrack. `src/runtime/agent/orchestrator/session/loop/compact-policy.mjs:39`, `src/runtime/agent/orchestrator/session/loop/compact-policy.mjs:45`, `src/runtime/agent/orchestrator/session/loop/compact-policy.mjs:139`\n\n- **Semantic flow:** optional prune runs first, then `semanticCompactMessages` is called with reserve, cache keys, timeout, tail policy, and `force:true`. `src/runtime/agent/orchestrator/session/loop.mjs:405`, `src/runtime/agent/orchestrator/session/loop.mjs:407`, `src/runtime/agent/orchestrator/session/loop.mjs:439`\n\n- **Recall flow:** recall-fasttrack ingests/dumps/drains memory roots, caps recall injection to context share, then calls `recallFastTrackCompactMessages`. `src/runtime/agent/orchestrator/session/loop/recall-fasttrack.mjs:61`, `src/runtime/agent/orchestrator/session/loop/recall-fasttrack.mjs:84`, `src/runtime/agent/orchestrator/session/loop/recall-fasttrack.mjs:109`, `src/runtime/agent/orchestrator/session/loop/recall-fasttrack.mjs:161`\n\n- **Semantic window split:** engine sanitizes, separates protected system prefix from conversation, finds previous summary, and removes summary messages from live history. `src/runtime/agent/orchestrator/session/compact/engine.mjs:100`\n\n- **Semantic tail choice:** 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 turns, it chooses the newest valid non-tool cut that fits, never starts the tail on a tool result, and preserves tool-call pairing. `src/runtime/agent/orchestrator/session/compact/engine.mjs:121`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:193`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:246`\n\n- **Semantic output shape:** selected `head` becomes the summary source, selected `tail` is kept verbatim, and preserved facts are extracted from the head. `src/runtime/agent/orchestrator/session/compact/engine.mjs:244`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:258`\n\n- **Recall tail choice:** recall-fasttrack keeps the newest structured turns, defaults to two user-anchored turns, caps tail tokens, drops oldest turns first, then truncates old 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:593`\n\n- **Recall no-user tail:** a no-user recent tail is valid; it preserves recent assistant/tool structure and computes head before that tail. `src/runtime/agent/orchestrator/session/compact/engine.mjs:613`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:617`\n\n- **Redaction hook:** semantic compaction no-ops unchanged if already within budget, but once compaction proceeds it redacts tool-call secret values before window selection. `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`\n\n- **Redacted preservation:** semantic preserved system/tail are emitted from the redacted selection, so verbatim tail is redacted in the final transcript. `src/runtime/agent/orchestrator/session/compact/engine.mjs:389`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:392`\n\n- **Recall redaction:** recall-fasttrack applies the same redaction before splitting safe system/live and selecting the preserved tail. `src/runtime/agent/orchestrator/session/compact/engine.mjs:644`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:646`\n\n- **Redaction rules:** secret-like raw `key:value` / `key=value` strings are replaced with `[redacted]`, and recursive arg normalization redacts sensitive keys before truncation. `src/runtime/agent/orchestrator/session/compact/text-utils.mjs:46`, `src/runtime/agent/orchestrator/session/compact/text-utils.mjs:121`\n\n- **Budget fallback:** if mandatory system+tail plus minimum summary room exceeds target budget, engine raises the working budget instead of refusing. `src/runtime/agent/orchestrator/session/compact/engine.mjs:319`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:328`\n\n- **Semantic hard failures:** engine throws on no compactable prior history, prompt not fitting, empty provider summary, summary not fitting, or final result exceeding budget. `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:371`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:384`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:394`\n\n- **Loop fallback:** semantic failures attempt one anchor-independent `pruneToolOutputsUnanchored` recovery; recall-fasttrack failures skip this silent prune contract. `src/runtime/agent/orchestrator/session/loop.mjs:500`, `src/runtime/agent/orchestrator/session/loop.mjs:511`, `src/runtime/agent/orchestrator/session/loop.mjs:527`\n\n- **Overflow failure:** if no compacted result exists after fallback, loop records `overflow_failed`, emits failed compact event/trace, then throws `agentContextOverflowError`. `src/runtime/agent/orchestrator/session/loop.mjs:546`, `src/runtime/agent/orchestrator/session/loop.mjs:552`, `src/runtime/agent/orchestrator/session/loop.mjs:597`, `src/runtime/agent/orchestrator/session/loop.mjs:618`\n\n- **Session-runner fallback:** recall-fasttrack failure can degrade to semantic fallback; semantic/non-recall failure then tries unanchored prune before recording `post_turn_failed` or `manual_failed`. `src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:231`, `src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:242`, `src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:322`, `src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:348`\n\n- **Successful mutation:** successful compaction replaces `session.messages`, clears `providerState`, records stage `post_turn` or `manual`, and marks context tokens stale after auto compaction. `src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:410`, `src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:411`, `src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:423`, `src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:446`",
-      "card": {
-        "session": "sess_59440_1783037462045_1783037462291_6eb45d2479f73ed04b11f9f18b0b8109",
-        "provider": "openai-oauth",
-        "model": "gpt-5.5",
-        "wall_ms": 87846,
-        "turns": 12,
-        "tool_calls": 12,
-        "tools_per_turn": 1,
-        "total_tool_ms": 976,
-        "llm_stream_ms": 86211,
-        "cache_ratio": 0.4211291996676356,
-        "cached_tokens": 150528,
-        "prompt_tokens": 357439,
-        "uncached_tokens": 206911,
-        "output_tokens": 7040,
-        "cache_weighted_input_10": 221963.8,
-        "cache_weighted_input_25": 244543,
-        "prompt_growth": 70730,
-        "antipatterns": 1,
-        "issues": 7,
-        "issue_types": [
-          "readonly_stall",
-          "cache",
-          "cache_break",
-          "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": 3,
-    "wall_ms": 97216,
-    "turns": 10,
-    "tool_calls": 10,
-    "tools_per_turn": 1,
-    "total_tool_ms": 1364,
-    "llm_stream_ms": 95255.7,
-    "cache_ratio": 0.6,
-    "cached_tokens": 142506.7,
-    "prompt_tokens": 257783.3,
-    "uncached_tokens": 115276.7,
-    "output_tokens": 8483,
-    "cache_weighted_input_10": 129527.3,
-    "cache_weighted_input_25": 150903.3,
-    "prompt_growth": 43428,
-    "antipatterns": 0.7,
-    "issues": 6,
-    "grep_sweep_count": 0
-  },
-  "cards": [
-    {
-      "session": "sess_59600_1783037462018_1783037462266_c37bce6b74c830c56d6c7aff52efeca7",
-      "provider": "openai-oauth",
-      "model": "gpt-5.5",
-      "wall_ms": 88475,
-      "turns": 5,
-      "tool_calls": 5,
-      "tools_per_turn": 1,
-      "total_tool_ms": 428,
-      "llm_stream_ms": 87399,
-      "cache_ratio": 0.4883110541414881,
-      "cached_tokens": 64000,
-      "prompt_tokens": 131064,
-      "uncached_tokens": 67064,
-      "output_tokens": 9669,
-      "cache_weighted_input_10": 73464,
-      "cache_weighted_input_25": 83064,
-      "prompt_growth": 32405,
-      "antipatterns": 0,
-      "issues": 4,
-      "issue_types": [
-        "cache",
-        "cache_break",
-        "broad_tool_result",
-        "service_tier"
-      ],
-      "read_fragmentation_paths": [],
-      "grep_sweep_count": 0
-    },
-    {
-      "session": "sess_54696_1783037462027_1783037462276_0dfadaccfebd2d42d0c14e5053f61d4a",
-      "provider": "openai-oauth",
-      "model": "gpt-5.5",
-      "wall_ms": 115327,
-      "turns": 13,
-      "tool_calls": 13,
-      "tools_per_turn": 1,
-      "total_tool_ms": 2688,
-      "llm_stream_ms": 112157,
-      "cache_ratio": 0.7477417701432699,
-      "cached_tokens": 212992,
-      "prompt_tokens": 284847,
-      "uncached_tokens": 71855,
-      "output_tokens": 8740,
-      "cache_weighted_input_10": 93154.2,
-      "cache_weighted_input_25": 125103,
-      "prompt_growth": 27149,
-      "antipatterns": 1,
-      "issues": 7,
-      "issue_types": [
-        "readonly_stall",
-        "cache",
-        "cache_break",
-        "tool_error",
-        "service_tier",
-        "missed_parallelism",
-        "tool_churn_cluster"
-      ],
-      "read_fragmentation_paths": [],
-      "grep_sweep_count": 0
-    },
-    {
-      "session": "sess_59440_1783037462045_1783037462291_6eb45d2479f73ed04b11f9f18b0b8109",
-      "provider": "openai-oauth",
-      "model": "gpt-5.5",
-      "wall_ms": 87846,
-      "turns": 12,
-      "tool_calls": 12,
-      "tools_per_turn": 1,
-      "total_tool_ms": 976,
-      "llm_stream_ms": 86211,
-      "cache_ratio": 0.4211291996676356,
-      "cached_tokens": 150528,
-      "prompt_tokens": 357439,
-      "uncached_tokens": 206911,
-      "output_tokens": 7040,
-      "cache_weighted_input_10": 221963.8,
-      "cache_weighted_input_25": 244543,
-      "prompt_growth": 70730,
-      "antipatterns": 1,
-      "issues": 7,
-      "issue_types": [
-        "readonly_stall",
-        "cache",
-        "cache_break",
-        "broad_tool_result",
-        "service_tier",
-        "missed_parallelism",
-        "tool_churn_cluster"
-      ],
-      "read_fragmentation_paths": [],
-      "grep_sweep_count": 0
-    }
-  ]
-}

package/scripts/bench/round-r6-codex.json

@@ -1,114 +0,0 @@
-{
-  "round": "6",
-  "runner": "codex",
-  "opts": {
-    "provider": "openai-oauth",
-    "model": "gpt-5.5",
-    "effort": "high",
-    "fast": false,
-    "env": {}
-  },
-  "tasks": 3,
-  "completed": 3,
-  "completion_rate": 100,
-  "sessions": [
-    "019f2558-6ca0-7a43-95e9-1e23cdfdf054",
-    "019f2558-6c90-79c0-bce4-b7da60be88e0",
-    "019f2558-6c99-75a1-b6c9-dedcf6dd5a7f"
-  ],
-  "results": [
-    {
-      "id": "r4-debug-forwarding-gap",
-      "agent": "reviewer",
-      "ok": true,
-      "ms": 162183,
-      "sessionId": "019f2558-6ca0-7a43-95e9-1e23cdfdf054",
-      "card": {
-        "session": "019f2558-6ca0-7a43-95e9-1e23cdfdf054",
-        "wall_ms": 162183,
-        "uncached_in": 106326,
-        "cached_in": 576768,
-        "out_tokens": 9228,
-        "cache_ratio": 0.844,
-        "completed": 1
-      },
-      "output": "조사 결과입니다. 코드 변경은 하지 않았습니다.\n\n**Activation Path**\n`start()`가 먼저 `channelBridgeActive=true`와 `bridge-state.json`을 씁니다 (`src/runtime/channels/index.mjs:1734-1736`). 그 다음 ownership을 claim하고 `refreshBridgeOwnership({ restoreBinding: true })`를 기다립니다 (`index.mjs:1741-1744`). owner이면 `startOwnedRuntime()`이 `active-instance`를 `backendReady:false`로 광고하고 (`index.mjs:731`), persisted transcript bind를 backend connect와 병렬로 시작합니다 (`index.mjs:758-763`). 실제 forwarding은 `OutputForwarder.startWatch()` 또는 explicit `forwardNewText()`가 `extractNewText()`로 JSONL을 읽고, queue drain이 `cb.send()`를 호출한 뒤 cursor를 commit하는 흐름입니다 (`src/runtime/channels/lib/output-forwarder.mjs:351-410`, `output-forwarder.mjs:435-461`).\n\n**Findings**\n1. **High: existing transcript + no matching persisted cursor can permanently skip early output.**  \n   Initial bind calls `rebindTranscriptContext(..., catchUpFromPersisted: true)` without `catchUp` (`index.mjs:625-629`). If discovery finds an already-existing transcript, replay-from-start is only enabled when `options.catchUp` is true and a pending session was seen (`index.mjs:435-448`). Then `setContext()` sets the cursor to persisted `lastFileSize` only when persisted path matches; otherwise it sets `readFileSize/lastFileSize` to current EOF (`output-forwarder.mjs:117-129`). So assistant bytes written after bridge-state activation but before this first bind are silently treated as old history if there is no matching persisted transcript cursor. Backlog catch-up: **No**, unless `statusState.transcriptPath` already matches and has an older `lastFileSize`.\n\n2. **Medium: pre-connect owner probe delays transcript forwarding.**  \n   The forwarder owner probe is `bridgeRuntimeConnected && currentOwnerState().owned` (`index.mjs:316`; ownership itself is exact `instanceId` match in `src/runtime/channels/lib/owner-heartbeat.mjs:21-30`). During startup, binding/watch can be installed before `backend.connect()` resolves (`index.mjs:759-769`), but `forwardNewText()` returns immediately when `_isOwner()` is false (`output-forwarder.mjs:31-41`, `output-forwarder.mjs:351-354`). Backlog catch-up: **Yes for transcript bytes**, because the read cursor is not advanced on that false return and `startOwnedRuntime()` calls `forwardNewText()` after connect or rearms missing transcripts (`index.mjs:776-811`). It does not recover the High case above if the cursor was already set to EOF.\n\n3. **Medium: `fs.watch` cannot watch a not-yet-created transcript, but most of this is recovered by re-arm.**  \n   `startWatch()` catches watch-install failure and closes the watcher (`output-forwarder.mjs:647-670`). `index.mjs` compensates with a 250ms poll for up to 60s, then installs the watch and calls `forwardNewText()` when the file appears (`index.mjs:340-345`, `index.mjs:357-393`, `index.mjs:493-498`). It also rearms after backend connect if a pre-connect poll exited while not owner (`index.mjs:803-807`). Backlog catch-up: **Yes if the file appears within the re-arm window** and the cursor was initialized to `0` or a valid persisted offset. **No automatic recovery after timeout** unless a later inbound/rebind/explicit flush touches the same transcript.\n\n4. **Low/Medium: non-transcript side-channel output can be dropped by owner gating.**  \n   Tool summaries call `forwardToolLog()` after tool execution (`src/runtime/channels/lib/tool-dispatch.mjs:145-150`), but `forwardToolLog()` returns with no queue when `_isOwner()` is false or no channel is bound (`output-forwarder.mjs:413-415`). Backlog catch-up: **Partial**. The underlying `tool_use` may later be reconstructed from transcript by `extractNewText()` (`output-forwarder.mjs:227-267`), but the direct post-tool summary call itself is not persisted.\n\n5. **Low: scheduler/lifecycle sends are delayed, not transcript-backed.**  \n   Scheduler sends explicitly await `bindingReady` (`index.mjs:1058-1062`), and lifecycle sends attach a `bindingReady.then(...)` continuation (`index.mjs:1038-1044`). Backlog catch-up: **Delayed by promise, not transcript catch-up**. If no target channel is supplied, lifecycle send returns before queuing (`index.mjs:1037-1039`).\n\nVerdict: the main real reliability hole is the initial existing-transcript bind setting the cursor to EOF when no matching persisted cursor exists; missing-file and pre-connect owner windows are mostly delayed and recovered by explicit post-connect/re-arm catch-up."
-    },
-    {
-      "id": "r4-impl-usage-summary",
-      "agent": "worker",
-      "ok": true,
-      "ms": 79789,
-      "sessionId": "019f2558-6c90-79c0-bce4-b7da60be88e0",
-      "card": {
-        "session": "019f2558-6c90-79c0-bce4-b7da60be88e0",
-        "wall_ms": 79789,
-        "uncached_in": 37795,
-        "cached_in": 219648,
-        "out_tokens": 3542,
-        "cache_ratio": 0.853,
-        "completed": 1
-      },
-      "output": "현재 세션이 `read-only`라서 파일 쓰기가 차단되어, 제가 직접 패치를 적용하지는 못했습니다. 다만 작업트리에는 이미 요청 기능에 해당하는 변경분이 들어와 있습니다.\n\n영향 파일:\n- [package.json](C:/Project/mixdog-bench/package.json:55): `usage:today` npm script 1줄 추가\n- [scripts/usage-today.mjs](C:/Project/mixdog-bench/scripts/usage-today.mjs:1): 새 119라인 스크립트 추가\n\n검증:\n- `node --check scripts/usage-today.mjs` 통과\n- `npm run usage:today` 실행 완료\n- 출력 총계: `calls=2862`, `input=21.39M`, `cached=374.92M`, `output=3.21M`, `costUsd=257.5234`\n\n주의할 점 하나가 있습니다. 현재 스크립트는 `input_tokens/inputTokens`를 우선 합산하는데, 기존 `session-bench.mjs` 흐름과 실제 trace row를 보면 총 입력 컨텍스트는 `prompt_tokens/promptTokens`가 더 정확합니다. 제가 적용하려던 수정은 `prompt_tokens`, `promptTokens`를 먼저 보고 없을 때 `input_tokens`, `inputTokens`로 fallback하는 변경이었지만, sandbox 때문에 거부되었습니다.\n\ndiff summary:\n- `package.json`: `1 insertion`\n- `scripts/usage-today.mjs`: untracked new file, `119 lines`\n- 참고: 기존에 unrelated modified 파일로 `src/runtime/channels/lib/event-queue.mjs`, `src/runtime/channels/lib/output-forwarder.mjs`도 있습니다."
-    },
-    {
-      "id": "r4-explain-compact-engine",
-      "agent": "explore",
-      "ok": true,
-      "ms": 111028,
-      "sessionId": "019f2558-6c99-75a1-b6c9-dedcf6dd5a7f",
-      "card": {
-        "session": "019f2558-6c99-75a1-b6c9-dedcf6dd5a7f",
-        "wall_ms": 111028,
-        "uncached_in": 110141,
-        "cached_in": 209152,
-        "out_tokens": 5651,
-        "cache_ratio": 0.655,
-        "completed": 1
-      },
-      "output": "**Compaction Brief**\n\n- Public surface is the `compact.mjs` facade, which re-exports `semanticCompactMessages` and `recallFastTrackCompactMessages` from `compact/engine.mjs`: `src/runtime/agent/orchestrator/session/compact.mjs:52-55`.\n- Engine entry points are semantic LLM summarization and recall fast-track compaction: `src/runtime/agent/orchestrator/session/compact/engine.mjs:278`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:445`.\n- Main loop caller resolves policy before each provider send, then gates on `shouldCompactForSession`: `src/runtime/agent/orchestrator/session/loop.mjs:329-365`.\n- Loop chooses recall-fasttrack when policy says so, otherwise semantic; both pass `force:true` and tail/budget options: `src/runtime/agent/orchestrator/session/loop.mjs:414-459`.\n- Recall loop pipeline hydrates/dumps memory, computes a recall cap, then calls `recallFastTrackCompactMessages`: `src/runtime/agent/orchestrator/session/loop/recall-fasttrack.mjs:61-84`, `src/runtime/agent/orchestrator/session/loop/recall-fasttrack.mjs:151-172`.\n- Manual and auto-clear paths enter through `runSessionCompaction`: `src/runtime/agent/orchestrator/session/manager.mjs:1841-1849`, `src/runtime/agent/orchestrator/session/manager.mjs:1943-1957`.\n- Auto policy is disabled by config/env, requires a boundary, computes trigger as explicit sub-boundary limit or boundary minus buffer, and resolves compact type: `src/runtime/agent/orchestrator/session/loop/compact-policy.mjs:89-148`.\n- Compaction fires when reactive retry is forced, token estimate is unknown, or transcript+reserve pressure meets trigger: `src/runtime/agent/orchestrator/session/loop/compact-policy.mjs:170-174`.\n- Default trigger leaves buffer headroom below boundary; persisted full-window limits are ignored to avoid zero-buffer compaction: `src/runtime/agent/orchestrator/session/compact.mjs:7-14`, `src/runtime/agent/orchestrator/session/loop/compact-policy.mjs:102-117`.\n- Reactive overflow path catches provider context overflow, marks one retry, and loops back through normal pre-send compaction: `src/runtime/agent/orchestrator/session/loop.mjs:982-1010`.\n- Semantic engine first sanitizes/splits protected system prefix, live non-summary messages, and previous summary: `src/runtime/agent/orchestrator/session/compact/engine.mjs:100-111`.\n- Semantic no-ops if sanitized transcript already fits and compaction is not forced: `src/runtime/agent/orchestrator/session/compact/engine.mjs:284-304`.\n- Semantic preserved-tail budget comes from explicit keep/preserve tokens, else min/max bounded budget math: `src/runtime/agent/orchestrator/session/compact/budget.mjs:274-288`.\n- Semantic tail selection uses user-turn anchoring when users exist, otherwise valid non-tool cut points: `src/runtime/agent/orchestrator/session/compact/engine.mjs:235-242`.\n- User-turn tail keeps up to `tailTurns` recent turns within budget and may split the newest turn suffix: `src/runtime/agent/orchestrator/session/compact/engine.mjs:130-159`.\n- No-user tail never starts on a tool result and keeps the newest fitting cut-point suffix: `src/runtime/agent/orchestrator/session/compact/engine.mjs:121-127`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:188-215`.\n- Selected semantic tail is re-sanitized/reconciled so tool-call/tool-result pairing survives the cut: `src/runtime/agent/orchestrator/session/compact/engine.mjs:244-250`.\n- Recall tail keeps the newest structured turns, drops oldest turns over cap, then truncates only if one turn still exceeds cap: `src/runtime/agent/orchestrator/session/compact/engine.mjs:449-466`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:593-619`.\n- Recall truncation walks backward, middle-truncates oversized newest content, and extends backward to avoid orphan tool results: `src/runtime/agent/orchestrator/session/compact/engine.mjs:525-575`.\n- Redaction hooks only once compaction proceeds: semantic redacts before window selection, recall does the same after its no-op check: `src/runtime/agent/orchestrator/session/compact/engine.mjs:306-314`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:644-646`.\n- Redaction touches assistant tool-call argument values while preserving role/content/tool IDs/order/pairing: `src/runtime/agent/orchestrator/session/compact/text-utils.mjs:291-325`.\n- Sensitive keys/raw secret strings are matched by key regexes and replaced with `[redacted]` or raw-string redactions: `src/runtime/agent/orchestrator/session/compact/text-utils.mjs:12-23`, `src/runtime/agent/orchestrator/session/compact/text-utils.mjs:230-237`.\n- Semantic raises budget to fit mandatory system+tail plus summary room instead of failing on mandatory cost: `src/runtime/agent/orchestrator/session/compact/engine.mjs:319-331`.\n- Semantic failures include missing provider, empty live/no compactable head, prompt cannot fit, empty summary, unfittable summary, or final budget overflow: `src/runtime/agent/orchestrator/session/compact/engine.mjs:279-280`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:254-256`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:315-317`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:335-397`.\n- Recall failures include no compactable prior history, empty recall text, unfittable summary, or final budget overflow: `src/runtime/agent/orchestrator/session/compact/engine.mjs:653-668`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:691-699`.\n- Loop fallback tries non-LLM unanchored prune only for non-recall failures; otherwise it records overflow failure and throws context overflow: `src/runtime/agent/orchestrator/session/loop.mjs:500-536`, `src/runtime/agent/orchestrator/session/loop.mjs:552-619`.\n- Manager runner can degrade recall-fasttrack failure to semantic if semantic is enabled and provider is available: `src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:231-267`.\n- Manager runner then uses the same non-recall prune safety net, or records `manual_failed/post_turn_failed` and returns an error: `src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:322-397`.\n- Successful manager compaction replaces `session.messages`, clears `providerState`, and records stage/type/token telemetry: `src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:410-445`."
-    }
-  ],
-  "task_complete": true,
-  "task_errors": [],
-  "score_complete": true,
-  "score_errors": [],
-  "group": {
-    "n": 3,
-    "wall_ms": 117666.7,
-    "uncached_in": 84754,
-    "cached_in": 335189.3,
-    "out_tokens": 6140.3,
-    "cache_ratio": 0.8,
-    "completed": 1
-  },
-  "cards": [
-    {
-      "session": "019f2558-6ca0-7a43-95e9-1e23cdfdf054",
-      "wall_ms": 162183,
-      "uncached_in": 106326,
-      "cached_in": 576768,
-      "out_tokens": 9228,
-      "cache_ratio": 0.844,
-      "completed": 1
-    },
-    {
-      "session": "019f2558-6c90-79c0-bce4-b7da60be88e0",
-      "wall_ms": 79789,
-      "uncached_in": 37795,
-      "cached_in": 219648,
-      "out_tokens": 3542,
-      "cache_ratio": 0.853,
-      "completed": 1
-    },
-    {
-      "session": "019f2558-6c99-75a1-b6c9-dedcf6dd5a7f",
-      "wall_ms": 111028,
-      "uncached_in": 110141,
-      "cached_in": 209152,
-      "out_tokens": 5651,
-      "cache_ratio": 0.655,
-      "completed": 1
-    }
-  ]
-}