mixdog 0.9.4 → 0.9.5

package/scripts/bench/round-r32-vs-codex-lead.json

@@ -0,0 +1,254 @@
+{
+  "round": "32",
+  "runner": "lead",
+  "opts": {
+    "provider": "openai-oauth",
+    "model": "gpt-5.5",
+    "effort": null,
+    "fast": false,
+    "workflow": null,
+    "env": {
+      "MIXDOG_DATA_DIR": "C:\\Users\\tempe\\.mixdog\\bench-data"
+    }
+  },
+  "tasks": 3,
+  "completed": 3,
+  "completion_rate": 100,
+  "sessions": [
+    "sess_24920_1783076179443_1783076179685_fbebe3e4d18c07e27c8a48fe8cd36028",
+    "sess_20680_1783076179459_1783076179699_20f64b7a9c29ae8b82bce96811957213",
+    "sess_3308_1783076179477_1783076179720_e7074a71c9936c9202b9c4095ddcaf89"
+  ],
+  "results": [
+    {
+      "id": "r4-debug-forwarding-gap",
+      "agent": "reviewer",
+      "ok": true,
+      "ms": 91356,
+      "sessionId": "sess_24920_1783076179443_1783076179685_fbebe3e4d18c07e27c8a48fe8cd36028",
+      "output": "sessionId: sess_24920_1783076179443_1783076179685_fbebe3e4d18c07e27c8a48fe8cd36028\n관련 경로를 기준으로 활성화부터 첫 포워딩까지의 흐름과 손실 지점을 확인하겠습니다.활성화 호출부와 inbound 바인딩 구간만 추가로 묶어서 확인하겠습니다.재영님, 조사 결과 bridge 활성화 직후의 지연은 일부 복구되지만, 기존 transcript에 이미 써진 출력은 실제 유실 가능성이 있습니다.\n\n- **최고 영향: 기존 transcript pre-bind 출력 유실** — `setContext()`가 기본/불일치 persisted 경로에서 cursor를 EOF로 둡니다; 이후 `forwardNewText()`가 과거 bytes를 보지 않습니다. `src/runtime/channels/lib/output-forwarder.mjs:117`, `src/runtime/channels/index.mjs:625`, `src/runtime/channels/index.mjs:1597` — backlog catch-up: **복구 안 됨**.\n\n- **binding 없음/잘못된 이전 binding** — 활성화는 `bridge-state.json`을 먼저 쓰고 ownership/binding을 뒤따라 수행합니다; 이 사이 현재 세션 transcript 발견이 실패하면 이전 path 또는 무binding 상태가 될 수 있습니다. `src/runtime/channels/lib/tool-dispatch.mjs:45`, `src/runtime/channels/index.mjs:577`, `src/runtime/channels/index.mjs:503` — backlog catch-up: **같은 persisted cursor일 때만 복구, 새 기존 파일은 유실**.\n\n- **not-yet-created transcript watch 실패** — `fs.watch()`는 없는 파일에서 실패하지만 pending re-arm poll이 파일 생성 후 watch 설치와 `forwardNewText()`를 호출합니다. `src/runtime/channels/index.mjs:340`, `src/runtime/channels/index.mjs:486`, `src/runtime/channels/lib/output-forwarder.mjs:647` — backlog catch-up: **60초 내 생성이면 복구, timeout 후에는 다음 rebind 전까지 지연/잠재 유실**.\n\n- **ownership/backend 연결 전 flush 무시** — `_isOwner()`가 false면 읽지도 않고 반환하므로 cursor는 보존됩니다; connect 후 `forwardNewText()` 또는 re-arm이 다시 실행됩니다. `src/runtime/channels/lib/output-forwarder.mjs:351`, `src/runtime/channels/index.mjs:758`, `src/runtime/channels/index.mjs:801` — backlog catch-up: **startup 지연은 복구, ownership을 잃으면 새 owner binding에 의존**.\n\n- **watch 설치 직전 write race** — `setContext()` stat 이후 `fs.watch()` 설치 사이 bytes는 즉시 `scheduleWatchFlush()`로 커버됩니다. `src/runtime/channels/lib/output-forwarder.mjs:665` — backlog catch-up: **복구됨**.\n\nVerdict: missing-file/ownership startup window는 대체로 지연 후 복구되지만, 이미 존재하는 새 transcript를 EOF baseline으로 bind하는 경로는 첫 assistant 출력을 조용히 유실할 수 있습니다.",
+      "card": {
+        "session": "sess_24920_1783076179443_1783076179685_fbebe3e4d18c07e27c8a48fe8cd36028",
+        "provider": "openai-oauth",
+        "model": "gpt-5.5",
+        "wall_ms": 80858,
+        "turns": 6,
+        "tool_calls": 7,
+        "tools_per_turn": 1.2,
+        "total_tool_ms": 1065,
+        "llm_stream_ms": 79802,
+        "cache_ratio": 0.6660624556860569,
+        "cached_tokens": 108032,
+        "prompt_tokens": 162195,
+        "uncached_tokens": 54163,
+        "output_tokens": 8103,
+        "cache_weighted_input_10": 64966.2,
+        "cache_weighted_input_25": 81171,
+        "prompt_growth": 38414,
+        "antipatterns": 1,
+        "issues": 10,
+        "issue_types": [
+          "cache",
+          "cache_break",
+          "broad_tool_result",
+          "service_tier",
+          "tool_churn_cluster"
+        ],
+        "read_fragmentation_paths": [],
+        "grep_sweep_count": 0
+      }
+    },
+    {
+      "id": "r4-impl-usage-summary",
+      "agent": "worker",
+      "ok": true,
+      "ms": 125526,
+      "sessionId": "sess_20680_1783076179459_1783076179699_20f64b7a9c29ae8b82bce96811957213",
+      "output": "sessionId: sess_20680_1783076179459_1783076179699_20f64b7a9c29ae8b82bce96811957213\n이미 일부 구현이 보이므로 현재 변경 상태를 확인한 뒤 필요한 보완만 하겠습니다.구현은 이미 작업 트리에 반영되어 있어 검증부터 진행하겠습니다.완료되었습니다. `usage:today` 스크립트와 `scripts/usage-today.mjs`가 추가되어 오늘 자정 이후 usage를 모델별/총합으로 집계합니다.\n\n- **Diff 요약**: `package.json`에 `usage:today` 추가, `scripts/usage-today.mjs` 신규 추가.\n- **검증**: `node --check scripts/usage-today.mjs` 통과, `npm run usage:today` 실행 완료.\n- **참고**: 기존 작업 트리에 unrelated 수정 파일 2개가 이미 있었습니다: `src/runtime/channels/lib/event-queue.mjs`, `src/runtime/channels/lib/output-forwarder.mjs`.",
+      "card": {
+        "session": "sess_20680_1783076179459_1783076179699_20f64b7a9c29ae8b82bce96811957213",
+        "provider": "openai-oauth",
+        "model": "gpt-5.5",
+        "wall_ms": 115026,
+        "turns": 14,
+        "tool_calls": 16,
+        "tools_per_turn": 1.1,
+        "total_tool_ms": 2818,
+        "llm_stream_ms": 112089,
+        "cache_ratio": 0.7775056333888508,
+        "cached_tokens": 238080,
+        "prompt_tokens": 306210,
+        "uncached_tokens": 68130,
+        "output_tokens": 7341,
+        "cache_weighted_input_10": 91938,
+        "cache_weighted_input_25": 127650,
+        "prompt_growth": 21818,
+        "antipatterns": 1,
+        "issues": 15,
+        "issue_types": [
+          "cache",
+          "cache_break",
+          "cache_miss",
+          "service_tier",
+          "missed_parallelism",
+          "tool_churn_cluster"
+        ],
+        "read_fragmentation_paths": [],
+        "grep_sweep_count": 0
+      }
+    },
+    {
+      "id": "r4-explain-compact-engine",
+      "agent": "explore",
+      "ok": true,
+      "ms": 71766,
+      "sessionId": "sess_3308_1783076179477_1783076179720_e7074a71c9936c9202b9c4095ddcaf89",
+      "output": "sessionId: sess_3308_1783076179477_1783076179720_e7074a71c9936c9202b9c4095ddcaf89\n재영님, compaction subsystem architecture brief입니다.\n\n- `compact.mjs` is the public facade; it re-exports `semanticCompactMessages` and `recallFastTrackCompactMessages` from `compact/engine.mjs`. `src/runtime/agent/orchestrator/session/compact.mjs:1`, `src/runtime/agent/orchestrator/session/compact.mjs:52`\n- Runtime pre-send auto path calls recall-fasttrack first when policy says so, otherwise semantic compaction. `src/runtime/agent/orchestrator/session/loop.mjs:329`, `src/runtime/agent/orchestrator/session/loop.mjs:414`, `src/runtime/agent/orchestrator/session/loop.mjs:439`\n- Manual/session paths enter via `runSessionCompaction`; clear-before-clear and explicit manual compact both call it. `src/runtime/agent/orchestrator/session/manager.mjs:1849`, `src/runtime/agent/orchestrator/session/manager.mjs:1950`\n- Auto fires when policy is enabled, a boundary exists, and pressure reaches trigger; reactive overflow forces it. `src/runtime/agent/orchestrator/session/loop/compact-policy.mjs:89`, `src/runtime/agent/orchestrator/session/loop/compact-policy.mjs:170`\n- Pressure is message tokens plus reserve; target budget is compact target plus reserve, capped by boundary. `src/runtime/agent/orchestrator/session/loop/compact-policy.mjs:151`, `src/runtime/agent/orchestrator/session/loop/compact-policy.mjs:163`\n- Manual compaction is forced; non-forced session compaction returns “below threshold” if pressure is under trigger. `src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:146`, `src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:176`\n- Semantic engine no-ops if sanitized transcript fits and not forced; otherwise it redacts, selects a window, prompts the provider, fits a summary, then emits system + summary + tail. `src/runtime/agent/orchestrator/session/compact/engine.mjs:286`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:312`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:314`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:367`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:384`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:392`\n- Window selection splits protected system context, previous summary, and live non-summary messages. `src/runtime/agent/orchestrator/session/compact/engine.mjs:100`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:102`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:110`\n- Semantic preserved tail keeps recent user turns within `preserveRecentBudget`; if newest turn alone is too large it keeps a fitting suffix. `src/runtime/agent/orchestrator/session/compact/engine.mjs:134`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:141`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:151`\n- If no user turns exist, semantic tail boundaries are any non-`tool` message, avoiding orphan tool results. `src/runtime/agent/orchestrator/session/compact/engine.mjs:121`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:193`\n- Recall-fasttrack preserves newest structured turns, default 2 users / 8k cap, dropping old turns then truncating if needed. `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`\n- Recall-fasttrack accepts no-user tails and computes head as everything before the preserved tail. `src/runtime/agent/orchestrator/session/compact/engine.mjs:613`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:617`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:618`\n- Redaction hooks after no-op checks and before window selection, so preserved system/tail are measured and emitted redacted. `src/runtime/agent/orchestrator/session/compact/engine.mjs:286`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:306`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:389`\n- Redaction is shape-preserving for verbatim tool-call arguments; prompt-side summaries also redact/truncate tool args. `src/runtime/agent/orchestrator/session/compact/text-utils.mjs:217`, `src/runtime/agent/orchestrator/session/compact/text-utils.mjs:261`, `src/runtime/agent/orchestrator/session/compact/summary.mjs:85`\n- Semantic failures in loop can fall back to unanchored tool-output prune unless recall-fasttrack already failed. `src/runtime/agent/orchestrator/session/loop.mjs:500`, `src/runtime/agent/orchestrator/session/loop.mjs:517`, `src/runtime/agent/orchestrator/session/loop.mjs:527`\n- Recall-fasttrack failure in session runner tries semantic degraded fallback; if still no compacted messages, failure metadata is recorded. `src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:231`, `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:348`\n- Hard failures in the loop emit compact failure telemetry and throw `agentContextOverflowError`. `src/runtime/agent/orchestrator/session/loop.mjs:552`, `src/runtime/agent/orchestrator/session/loop.mjs:597`, `src/runtime/agent/orchestrator/session/loop.mjs:618`",
+      "card": {
+        "session": "sess_3308_1783076179477_1783076179720_e7074a71c9936c9202b9c4095ddcaf89",
+        "provider": "openai-oauth",
+        "model": "gpt-5.5",
+        "wall_ms": 61265,
+        "turns": 8,
+        "tool_calls": 11,
+        "tools_per_turn": 1.4,
+        "total_tool_ms": 1377,
+        "llm_stream_ms": 59937,
+        "cache_ratio": 0.6204795716575684,
+        "cached_tokens": 112640,
+        "prompt_tokens": 181537,
+        "uncached_tokens": 68897,
+        "output_tokens": 4296,
+        "cache_weighted_input_10": 80161,
+        "cache_weighted_input_25": 97057,
+        "prompt_growth": 33315,
+        "antipatterns": 0,
+        "issues": 11,
+        "issue_types": [
+          "cache",
+          "cache_break",
+          "cache_miss",
+          "service_tier"
+        ],
+        "read_fragmentation_paths": [],
+        "grep_sweep_count": 0
+      }
+    }
+  ],
+  "task_complete": true,
+  "task_errors": [],
+  "score_complete": true,
+  "score_errors": [],
+  "group": {
+    "n": 3,
+    "wall_ms": 85716.3,
+    "turns": 9.3,
+    "tool_calls": 11.3,
+    "tools_per_turn": 1.2,
+    "total_tool_ms": 1753.3,
+    "llm_stream_ms": 83942.7,
+    "cache_ratio": 0.7,
+    "cached_tokens": 152917.3,
+    "prompt_tokens": 216647.3,
+    "uncached_tokens": 63730,
+    "output_tokens": 6580,
+    "cache_weighted_input_10": 79021.7,
+    "cache_weighted_input_25": 101959.3,
+    "prompt_growth": 31182.3,
+    "antipatterns": 0.7,
+    "issues": 12,
+    "grep_sweep_count": 0
+  },
+  "cards": [
+    {
+      "session": "sess_24920_1783076179443_1783076179685_fbebe3e4d18c07e27c8a48fe8cd36028",
+      "provider": "openai-oauth",
+      "model": "gpt-5.5",
+      "wall_ms": 80858,
+      "turns": 6,
+      "tool_calls": 7,
+      "tools_per_turn": 1.2,
+      "total_tool_ms": 1065,
+      "llm_stream_ms": 79802,
+      "cache_ratio": 0.6660624556860569,
+      "cached_tokens": 108032,
+      "prompt_tokens": 162195,
+      "uncached_tokens": 54163,
+      "output_tokens": 8103,
+      "cache_weighted_input_10": 64966.2,
+      "cache_weighted_input_25": 81171,
+      "prompt_growth": 38414,
+      "antipatterns": 1,
+      "issues": 10,
+      "issue_types": [
+        "cache",
+        "cache_break",
+        "broad_tool_result",
+        "service_tier",
+        "tool_churn_cluster"
+      ],
+      "read_fragmentation_paths": [],
+      "grep_sweep_count": 0
+    },
+    {
+      "session": "sess_20680_1783076179459_1783076179699_20f64b7a9c29ae8b82bce96811957213",
+      "provider": "openai-oauth",
+      "model": "gpt-5.5",
+      "wall_ms": 115026,
+      "turns": 14,
+      "tool_calls": 16,
+      "tools_per_turn": 1.1,
+      "total_tool_ms": 2818,
+      "llm_stream_ms": 112089,
+      "cache_ratio": 0.7775056333888508,
+      "cached_tokens": 238080,
+      "prompt_tokens": 306210,
+      "uncached_tokens": 68130,
+      "output_tokens": 7341,
+      "cache_weighted_input_10": 91938,
+      "cache_weighted_input_25": 127650,
+      "prompt_growth": 21818,
+      "antipatterns": 1,
+      "issues": 15,
+      "issue_types": [
+        "cache",
+        "cache_break",
+        "cache_miss",
+        "service_tier",
+        "missed_parallelism",
+        "tool_churn_cluster"
+      ],
+      "read_fragmentation_paths": [],
+      "grep_sweep_count": 0
+    },
+    {
+      "session": "sess_3308_1783076179477_1783076179720_e7074a71c9936c9202b9c4095ddcaf89",
+      "provider": "openai-oauth",
+      "model": "gpt-5.5",
+      "wall_ms": 61265,
+      "turns": 8,
+      "tool_calls": 11,
+      "tools_per_turn": 1.4,
+      "total_tool_ms": 1377,
+      "llm_stream_ms": 59937,
+      "cache_ratio": 0.6204795716575684,
+      "cached_tokens": 112640,
+      "prompt_tokens": 181537,
+      "uncached_tokens": 68897,
+      "output_tokens": 4296,
+      "cache_weighted_input_10": 80161,
+      "cache_weighted_input_25": 97057,
+      "prompt_growth": 33315,
+      "antipatterns": 0,
+      "issues": 11,
+      "issue_types": [
+        "cache",
+        "cache_break",
+        "cache_miss",
+        "service_tier"
+      ],
+      "read_fragmentation_paths": [],
+      "grep_sweep_count": 0
+    }
+  ]
+}

package/scripts/bench/round-r33-vs-codex-codex.json

@@ -0,0 +1,115 @@
+{
+  "round": "33",
+  "runner": "codex",
+  "opts": {
+    "provider": "openai-oauth",
+    "model": "gpt-5.5",
+    "effort": null,
+    "fast": false,
+    "workflow": null,
+    "env": {}
+  },
+  "tasks": 3,
+  "completed": 3,
+  "completion_rate": 100,
+  "sessions": [
+    "019f27a1-7dc4-7eb0-8195-f06a7c0022a3",
+    "019f27a1-7c61-7e70-bea3-392ebbe5e725",
+    "019f27a1-7df6-7690-aa6d-6734860ac723"
+  ],
+  "results": [
+    {
+      "id": "r4-debug-forwarding-gap",
+      "agent": "reviewer",
+      "ok": true,
+      "ms": 240320,
+      "sessionId": "019f27a1-7dc4-7eb0-8195-f06a7c0022a3",
+      "card": {
+        "session": "019f27a1-7dc4-7eb0-8195-f06a7c0022a3",
+        "wall_ms": 240320,
+        "uncached_in": 114654,
+        "cached_in": 981120,
+        "out_tokens": 14675,
+        "cache_ratio": 0.895,
+        "completed": 1
+      },
+      "output": "조사 결과, 활성화부터 첫 forward까지의 경로는 이렇게 이어집니다: `bridge-state.json` 기록 → ownership claim/refresh → backend connect → persisted/current transcript bind → `fs.watch` 또는 explicit `forwardNewText()` → transcript parse → backend send. 핵심 앵커는 [index.mjs:1735](C:/Project/mixdog-bench/src/runtime/channels/index.mjs:1735), [index.mjs:1744](C:/Project/mixdog-bench/src/runtime/channels/index.mjs:1744), [index.mjs:759](C:/Project/mixdog-bench/src/runtime/channels/index.mjs:759), [index.mjs:625](C:/Project/mixdog-bench/src/runtime/channels/index.mjs:625), [output-forwarder.mjs:351](C:/Project/mixdog-bench/src/runtime/channels/lib/output-forwarder.mjs:351)입니다.\n\n**Findings**\n\n1. **High: 새 transcript가 이미 존재하지만 persisted path와 다르면 EOF로 바인딩되어 기존 출력이 조용히 스킵됩니다.**  \n   `bindPersistedTranscriptIfAny()`는 `catchUpFromPersisted: true`만 넘깁니다 [index.mjs:625](C:/Project/mixdog-bench/src/runtime/channels/index.mjs:625). 기존 파일 발견 branch는 그대로 `applyTranscriptBinding()`을 호출합니다 [index.mjs:435](C:/Project/mixdog-bench/src/runtime/channels/index.mjs:435), [index.mjs:441](C:/Project/mixdog-bench/src/runtime/channels/index.mjs:441). 그런데 `setContext()`는 persisted transcript가 같은 파일일 때만 persisted offset을 쓰고, 아니면 `currentSize`로 cursor를 둡니다 [output-forwarder.mjs:117](C:/Project/mixdog-bench/src/runtime/channels/lib/output-forwarder.mjs:117), [output-forwarder.mjs:122](C:/Project/mixdog-bench/src/runtime/channels/lib/output-forwarder.mjs:122), [output-forwarder.mjs:125](C:/Project/mixdog-bench/src/runtime/channels/lib/output-forwarder.mjs:125).  \n   Catch-up: **No**, 단 persisted path가 같은 경우만 recovered.\n\n2. **High: activation rebind 때 session record/transcript를 못 찾으면 binding 자체가 없고, 이후 일반 inbound bind가 과거 출력을 EOF 처리할 수 있습니다.**  \n   rebind는 30회 poll 후 pending path도 previous path도 없으면 빈 문자열로 종료합니다 [index.mjs:431](C:/Project/mixdog-bench/src/runtime/channels/index.mjs:431), [index.mjs:515](C:/Project/mixdog-bench/src/runtime/channels/index.mjs:515). post-connect도 `forwarder.transcriptPath`가 없으면 실질적으로 읽을 파일이 없습니다 [index.mjs:801](C:/Project/mixdog-bench/src/runtime/channels/index.mjs:801), [output-forwarder.mjs:150](C:/Project/mixdog-bench/src/runtime/channels/lib/output-forwarder.mjs:150). 나중에 inbound가 이미 존재하는 transcript를 잡으면 `applyTranscriptBinding()`은 catch-up 옵션 없이 호출되고 [index.mjs:1597](C:/Project/mixdog-bench/src/runtime/channels/index.mjs:1597), `setContext()` 기본값은 EOF입니다 [output-forwarder.mjs:125](C:/Project/mixdog-bench/src/runtime/channels/lib/output-forwarder.mjs:125).  \n   Catch-up: **No automatic catch-up**.\n\n3. **Medium: transcript path는 알지만 파일이 아직 없으면 `fs.watch`는 실패하고 re-arm poll로 지연 복구됩니다.**  \n   `startWatch()`는 없는 파일에서 catch 후 watcher를 닫습니다 [output-forwarder.mjs:656](C:/Project/mixdog-bench/src/runtime/channels/lib/output-forwarder.mjs:656), [output-forwarder.mjs:667](C:/Project/mixdog-bench/src/runtime/channels/lib/output-forwarder.mjs:667). pending branch는 `replayFromStart` 또는 persisted cursor를 설정한 뒤 [index.mjs:485](C:/Project/mixdog-bench/src/runtime/channels/index.mjs:485), [index.mjs:486](C:/Project/mixdog-bench/src/runtime/channels/index.mjs:486), 250ms poll을 60초까지 돌려 파일 생성 시 `startWatch()`와 `forwardNewText()`를 호출합니다 [index.mjs:348](C:/Project/mixdog-bench/src/runtime/channels/index.mjs:348), [index.mjs:349](C:/Project/mixdog-bench/src/runtime/channels/index.mjs:349), [index.mjs:376](C:/Project/mixdog-bench/src/runtime/channels/index.mjs:376).  \n   Catch-up: **Yes within 60s**, 그 이후는 자동 복구 없음.\n\n4. **Medium: bridge-state는 backend/owner 준비 전에 active가 되므로 그 사이 forward 시도는 fail-closed로 지연됩니다.**  \n   state write가 먼저 일어나고 [index.mjs:1735](C:/Project/mixdog-bench/src/runtime/channels/index.mjs:1735), [index.mjs:1736](C:/Project/mixdog-bench/src/runtime/channels/index.mjs:1736), binding task는 `backend.connect()` 전에 시작됩니다 [index.mjs:759](C:/Project/mixdog-bench/src/runtime/channels/index.mjs:759), [index.mjs:769](C:/Project/mixdog-bench/src/runtime/channels/index.mjs:769). 이때 owner probe는 `bridgeRuntimeConnected && currentOwnerState().owned`라 false입니다 [index.mjs:316](C:/Project/mixdog-bench/src/runtime/channels/index.mjs:316), 그래서 `forwardNewText()`/`forwardFinalText()`는 읽지 않고 반환합니다 [output-forwarder.mjs:351](C:/Project/mixdog-bench/src/runtime/channels/lib/output-forwarder.mjs:351), [output-forwarder.mjs:491](C:/Project/mixdog-bench/src/runtime/channels/lib/output-forwarder.mjs:491). connect 후에는 existing file이면 `forwardNewText()`, pending file이면 re-arm을 다시 겁니다 [index.mjs:803](C:/Project/mixdog-bench/src/runtime/channels/index.mjs:803), [index.mjs:807](C:/Project/mixdog-bench/src/runtime/channels/index.mjs:807), [index.mjs:809](C:/Project/mixdog-bench/src/runtime/channels/index.mjs:809).  \n   Catch-up: **Yes if a valid binding exists**, 아니면 finding 1/2로 떨어집니다.\n\n5. **Low/intentional: activation 중 ownership을 잃으면 이 프로세스의 output은 포워딩되지 않습니다.**  \n   CAS가 실패하면 owned runtime start가 중단됩니다 [index.mjs:731](C:/Project/mixdog-bench/src/runtime/channels/index.mjs:731), [index.mjs:736](C:/Project/mixdog-bench/src/runtime/channels/index.mjs:736). 다른 active owner가 보이면 조용히 standby로 갑니다 [index.mjs:916](C:/Project/mixdog-bench/src/runtime/channels/index.mjs:916).  \n   Catch-up: **No for the losing session**, last-wins 설계상 의도된 동작입니다.\n\n6. **Low: 기존 파일에서 stat 이후 watch 설치 전 쓰인 bytes는 100ms flush로 복구됩니다.**  \n   `startWatch()`는 watcher 설치 직후 `scheduleWatchFlush()`를 호출합니다 [output-forwarder.mjs:665](C:/Project/mixdog-bench/src/runtime/channels/lib/output-forwarder.mjs:665), [output-forwarder.mjs:666](C:/Project/mixdog-bench/src/runtime/channels/lib/output-forwarder.mjs:666). flush는 stat 후 `forwardNewText()`를 실행합니다 [output-forwarder.mjs:703](C:/Project/mixdog-bench/src/runtime/channels/lib/output-forwarder.mjs:703), [output-forwarder.mjs:724](C:/Project/mixdog-bench/src/runtime/channels/lib/output-forwarder.mjs:724).  \n   Catch-up: **Yes**, 짧은 지연만 있습니다.\n\n판정: bridge activation window는 pending-file 케이스는 대체로 복구하지만, “새 transcript가 이미 존재하거나 최초 discovery가 빈손인 경우”에는 backlog가 EOF로 조용히 유실될 수 있습니다."
+    },
+    {
+      "id": "r4-impl-usage-summary",
+      "agent": "worker",
+      "ok": true,
+      "ms": 121182,
+      "sessionId": "019f27a1-7c61-7e70-bea3-392ebbe5e725",
+      "card": {
+        "session": "019f27a1-7c61-7e70-bea3-392ebbe5e725",
+        "wall_ms": 121182,
+        "uncached_in": 51502,
+        "cached_in": 392320,
+        "out_tokens": 6537,
+        "cache_ratio": 0.884,
+        "completed": 1
+      },
+      "output": "현재 워크스페이스가 읽기 전용이라 최종 수정은 적용하지 못했습니다. `apply_patch`가 “writing is blocked by read-only sandbox”로 거절됐습니다.\n\n확인한 현재 diff 요약:\n\n- `package.json`: `usage:today` npm script 1줄 추가됨\n- `scripts/usage-today.mjs`: 신규 파일, 202 lines\n- 별도 기존 변경: `src/runtime/channels/lib/event-queue.mjs`, `src/runtime/channels/lib/output-forwarder.mjs`\n\n검증:\n\n- `node --check scripts/usage-today.mjs`: 통과\n- `npm run usage:today`: 실행 성공, 표 출력됨  \n  단, npm.ps1 접근 경고가 한 줄 있었지만 종료 코드는 0이었습니다.\n\n중요한 점: 현재 `usage-today.mjs`는 `kind:\"usage\"`와 `kind:\"usage_raw\"`를 둘 다 집계해서 비용/토큰이 중복될 수 있습니다. 적용해야 할 최소 수정은 아래 1줄입니다.\n\n```diff\n-  if (kind === 'usage' || kind === 'usage_raw') return true;\n+  if (kind === 'usage') return true;\n```\n\n쓰기 권한이 열리면 이 한 줄만 반영한 뒤 다시 `node --check`와 `npm run usage:today`를 돌리면 됩니다."
+    },
+    {
+      "id": "r4-explain-compact-engine",
+      "agent": "explore",
+      "ok": true,
+      "ms": 139884,
+      "sessionId": "019f27a1-7df6-7690-aa6d-6734860ac723",
+      "card": {
+        "session": "019f27a1-7df6-7690-aa6d-6734860ac723",
+        "wall_ms": 139884,
+        "uncached_in": 97389,
+        "cached_in": 1032704,
+        "out_tokens": 7491,
+        "cache_ratio": 0.914,
+        "completed": 1
+      },
+      "output": "**Entry Points**\n- `session/compact.mjs:52-55` re-exports `semanticCompactMessages` and `recallFastTrackCompactMessages` from `compact/engine.mjs`; the engine exports semantic at `compact/engine.mjs:278` and recall-fasttrack at `compact/engine.mjs:445-446`.\n- The live agent loop imports semantic compaction and recall runner at `session/loop.mjs:10-16` and `session/loop.mjs:32`, then calls recall at `session/loop.mjs:414-428` or semantic at `session/loop.mjs:439-465`.\n- Manual/auto-clear compaction goes through `runSessionCompaction`, imported by `manager.mjs:32-35`, invoked before clear at `manager.mjs:1841-1850`, and exposed as `compactSessionMessages` at `manager.mjs:1943-1957`.\n\n**When It Fires**\n- The loop drains queued steering before deciding, so compaction sees the exact next provider payload: `session/loop.mjs:324-329`.\n- Policy resolution requires auto enabled, derives boundary/trigger/buffer/reserve/type, and marks semantic vs recall-fasttrack: `loop/compact-policy.mjs:89-148`.\n- Pressure is `messageTokensEst + reserveTokens`; compaction fires on forced reactive, unknown estimate, or pressure >= trigger: `loop/compact-policy.mjs:150-174`.\n- If below threshold, the loop only records telemetry: `session/loop.mjs:365-371`; if above, it computes target budget and enters compacting stage: `session/loop.mjs:362-397`.\n- Manager-side compaction skips closed/tiny/non-forced sessions, skips auto below threshold, and otherwise computes target budget: `manager/compaction-runner.mjs:143-198`.\n\n**Preserved Tail**\n- Semantic compaction first sanitizes/dedups, splits protected prefix, finds prior summary, and excludes summary messages from live history: `compact/engine.mjs:100-111`.\n- Summary messages are synthetic `role:'user'` anchors, not real turns: `compact/messages.mjs:19-28`; protected system/reminder prefix is split out at `compact/messages.mjs:56-81`.\n- Semantic tail selection uses user-turn slicing when users exist, otherwise cut-points that never start on `tool`: `compact/engine.mjs:121-128`, `compact/engine.mjs:235-242`.\n- User-turn mode keeps newest turns within recent budget, splitting the newest turn suffix if needed: `compact/engine.mjs:134-185`.\n- No-user mode walks valid cuts newest-to-oldest, tolerates oversized/minimal tails, and avoids empty-head failure cases: `compact/engine.mjs:193-233`.\n- Final semantic window is `head = live[..tailStart]`, repaired `tail = live[tailStart..]`, with preserved facts extracted from head: `compact/engine.mjs:244-258`.\n- Recall-fasttrack keeps newest turns under cap, drops oldest kept turns first, or fits one turn; no-user tails are valid: `compact/engine.mjs:593-619`.\n\n**Redaction**\n- If semantic input already fits and is not forced, it returns sanitized messages unchanged, with no preserved-tail redaction: `compact/engine.mjs:286-312`.\n- When semantic proceeds, tool-call secrets are redacted before window selection, so preserved system/tail are measured and emitted redacted: `compact/engine.mjs:306-314`, `compact/engine.mjs:389-392`.\n- Recall-fasttrack has the same no-op-then-redact-before-split pattern: `compact/engine.mjs:627-646`.\n- Redaction is shape-preserving for verbatim tool-call args and maps only tool-bearing messages: `compact/text-utils.mjs:261-266`, `compact/text-utils.mjs:319-325`.\n\n**Failures / Fallbacks**\n- Semantic throws for missing provider, empty compactable head, unfit prompt, empty summary, unfit summary, or final budget overflow: `compact/engine.mjs:278-280`, `compact/engine.mjs:315-317`, `compact/engine.mjs:335-337`, `compact/engine.mjs:371-397`.\n- Mandatory preserved context can raise the working budget instead of refusing: `compact/engine.mjs:319-331`; recall-fasttrack mirrors this at `compact/engine.mjs:658-664`.\n- Loop failures try one non-LLM unanchored prune only for non-recall failures: `session/loop.mjs:500-537`; recall failures are explicitly excluded from that fallback: `session/loop.mjs:511-517`.\n- If fallback prune fails, loop records telemetry/trace/event and throws `AgentContextOverflowError`: `session/loop.mjs:546-626`.\n- Provider context overflow triggers one reactive compact retry via the normal pre-send path; after retry it surfaces overflow: `session/loop.mjs:982-1026`.\n- Manager runner degrades recall-fasttrack failure to semantic if possible, then tries non-recall prune, then records failure or replaces `session.messages` and clears provider state: `manager/compaction-runner.mjs:237-278`, `manager/compaction-runner.mjs:322-347`, `manager/compaction-runner.mjs:348-411`."
+    }
+  ],
+  "task_complete": true,
+  "task_errors": [],
+  "score_complete": true,
+  "score_errors": [],
+  "group": {
+    "n": 3,
+    "wall_ms": 167128.7,
+    "uncached_in": 87848.3,
+    "cached_in": 802048,
+    "out_tokens": 9567.7,
+    "cache_ratio": 0.9,
+    "completed": 1
+  },
+  "cards": [
+    {
+      "session": "019f27a1-7dc4-7eb0-8195-f06a7c0022a3",
+      "wall_ms": 240320,
+      "uncached_in": 114654,
+      "cached_in": 981120,
+      "out_tokens": 14675,
+      "cache_ratio": 0.895,
+      "completed": 1
+    },
+    {
+      "session": "019f27a1-7c61-7e70-bea3-392ebbe5e725",
+      "wall_ms": 121182,
+      "uncached_in": 51502,
+      "cached_in": 392320,
+      "out_tokens": 6537,
+      "cache_ratio": 0.884,
+      "completed": 1
+    },
+    {
+      "session": "019f27a1-7df6-7690-aa6d-6734860ac723",
+      "wall_ms": 139884,
+      "uncached_in": 97389,
+      "cached_in": 1032704,
+      "out_tokens": 7491,
+      "cache_ratio": 0.914,
+      "completed": 1
+    }
+  ]
+}

package/scripts/bench/round-r34-orchestrated.json

@@ -0,0 +1,120 @@
+{
+  "round": "34",
+  "runner": "lead",
+  "opts": {
+    "provider": "openai-oauth",
+    "model": "gpt-5.5",
+    "effort": null,
+    "fast": false,
+    "workflow": "default",
+    "env": {
+      "MIXDOG_DATA_DIR": "C:\\Users\\tempe\\.mixdog\\bench-data"
+    }
+  },
+  "tasks": 1,
+  "completed": 1,
+  "completion_rate": 100,
+  "sessions": [
+    "sess_23752_1783076753474_1783076753723_263c02f99dbb98fd9e99bc46c41bb5b8"
+  ],
+  "results": [
+    {
+      "id": "r5-orchestrated-audit",
+      "agent": "lead",
+      "ok": true,
+      "ms": 320527,
+      "sessionId": "sess_23752_1783076753474_1783076753723_263c02f99dbb98fd9e99bc46c41bb5b8",
+      "output": "sessionId: sess_23752_1783076753474_1783076753723_263c02f99dbb98fd9e99bc46c41bb5b8\nScope B 교차검토는 완료됐고, 주요 결론은 유지됩니다.\n\n- **확정:** handshake 5회, midstream 4/2회 예산, retry마다 fresh acquire, HTTP 5xx 누적 재시도 가능성이 확인됐습니다.\n- **추가 발견:** `openai-oauth-ws.mjs:507-508` 주석이 현재 코드의 4/2회 midstream 예산과 맞지 않는 stale comment입니다.\n- **대기 중:** Scope A reviewer 완료 후 두 범위를 합쳐 최종 보고서를 내겠습니다.분석 완료 — 두 scope와 리뷰 결과를 합친 최종 보고서입니다. 파일 수정은 하지 않았습니다.\n\n- **Classifier bucket inventory:** `classifyError()`는 `auth|permanent|transient|unknown`이고, WS midstream은 `first_byte_timeout`, `ws_send_failed`, `http_${status}`, `agent_stall`, `stream_stalled`, `ws_1006`, `ws_1011`, `ws_1012`, `ws_4000`, `ws_1000`, `response_failed_disconnected`, `response_failed_network`, `response_failed_auth_expired`, `null`을 반환합니다 (`retry-classifier.mjs:100`, `:293-338`).\n\n- **Reachability:** 대부분은 local producer가 있습니다. 예: `ws_send_failed`는 `_sendFrame` 실패 태그(`openai-ws-pool.mjs:106-134`), first-byte timeout은 `err.firstByteTimeout`(`openai-ws-stream.mjs:459-485`), response.failed payload는 `midState.responseFailedPayload`(`openai-ws-stream.mjs:953-963`)를 통해 소비됩니다.\n\n- **No local producer:** `response_failed_auth_expired`는 `response.failed` payload 문자열 `\"auth context expired\"`에만 의존하며(`retry-classifier.mjs:332-338`), repo 내 해당 문자열 producer는 classifier 자신뿐입니다. 즉 로컬 코드 경로는 만들 수 없고 upstream 서버 문자열에만 의존합니다.\n\n- **Stale docs/comments:** `_classifyMidstreamError` 주석은 실제 bucket 중 `http_${status}`, `ws_send_failed`, `response_failed_auth_expired`를 누락합니다(`openai-oauth-ws.mjs:124-139`). `ws_4000` 설명도 first-byte path가 실제로는 `first_byte_timeout`에서 먼저 잡히므로 오해 소지가 있습니다(`retry-classifier.mjs:293-329`, `openai-ws-stream.mjs:467-480`).\n\n- **Handshake loop:** `_acquireWithRetry`는 최대 5회 시도(`stall-policy.mjs:200-201`, `openai-oauth-ws.mjs:104`, `:389`)하고, `500ms * 2^(attempt-1)`를 5000ms cap으로 jitter 처리합니다(`openai-oauth-ws.mjs:175-178`, `:427-459`). 401/403/404/429는 retry하지 않고, 5xx와 transient socket 계열만 classifier로 재시도됩니다(`retry-classifier.mjs:477-489`).\n\n- **Midstream loop:** `sendViaWebSocket`는 `attemptIndex <= MAX_MIDSTREAM_RETRIES`로 최대 5번 stream attempt를 돌며(`openai-oauth-ws.mjs:514`, `:557`), 각 retry는 fresh acquire를 다시 수행합니다(`:564-572`). `ws_1006/ws_1011`은 4회 retry, 그 외 bucket은 2회 retry이고 backoff는 `[250,1000,2000,4000]` jitter입니다(`retry-classifier.mjs:245-268`, `openai-oauth-ws.mjs:164-172`, `:843-860`).\n\n- **Double-retry risk:** HTTP 5xx는 handshake에서 `http_${status}`로 재시도되고(`retry-classifier.mjs:480-482`), midstream에서도 같은 5xx 상태가 retry bucket이 됩니다(`:306-310`). 따라서 persistent/flapping 5xx는 midstream 재시도마다 fresh acquire 5회 예산을 다시 얻어 `sendViaWebSocket` 내부에서 최대 25 acquire 시도까지 누적될 수 있습니다(`openai-oauth-ws.mjs:557`, `:564-572`).\n\n- **Caller layer:** WS 실패 뒤 401/403은 live text가 없을 때 auth refresh 후 WS 1회 재실행이고(`openai-oauth.mjs:951-976`), fallback 대상이면 HTTP/SSE로 1회 전환합니다(`openai-oauth.mjs:989-995`, `retry-classifier.mjs:398-413`). 401/403/429는 handshake/midstream 내부 retry와 겹치지 않습니다(`retry-classifier.mjs:477-479`, `:307`).",
+      "card": {
+        "session": "sess_23752_1783076753474_1783076753723_263c02f99dbb98fd9e99bc46c41bb5b8",
+        "provider": "openai-oauth",
+        "model": "gpt-5.5",
+        "wall_ms": 309924,
+        "turns": 56,
+        "tool_calls": 65,
+        "tools_per_turn": 1.2,
+        "total_tool_ms": 5015,
+        "llm_stream_ms": 522741,
+        "cache_ratio": 0.8561893645226001,
+        "cached_tokens": 1459683,
+        "prompt_tokens": 1704860,
+        "uncached_tokens": 51582,
+        "output_tokens": 24458,
+        "cache_weighted_input_10": 197550.30000000002,
+        "cache_weighted_input_25": 416502.75,
+        "prompt_growth": 45489,
+        "antipatterns": 3,
+        "issues": 23,
+        "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": 1,
+    "wall_ms": 309924,
+    "turns": 56,
+    "tool_calls": 65,
+    "tools_per_turn": 1.2,
+    "total_tool_ms": 5015,
+    "llm_stream_ms": 522741,
+    "cache_ratio": 0.9,
+    "cached_tokens": 1459683,
+    "prompt_tokens": 1704860,
+    "uncached_tokens": 51582,
+    "output_tokens": 24458,
+    "cache_weighted_input_10": 197550.3,
+    "cache_weighted_input_25": 416502.8,
+    "prompt_growth": 45489,
+    "antipatterns": 3,
+    "issues": 23,
+    "grep_sweep_count": 0
+  },
+  "cards": [
+    {
+      "session": "sess_23752_1783076753474_1783076753723_263c02f99dbb98fd9e99bc46c41bb5b8",
+      "provider": "openai-oauth",
+      "model": "gpt-5.5",
+      "wall_ms": 309924,
+      "turns": 56,
+      "tool_calls": 65,
+      "tools_per_turn": 1.2,
+      "total_tool_ms": 5015,
+      "llm_stream_ms": 522741,
+      "cache_ratio": 0.8561893645226001,
+      "cached_tokens": 1459683,
+      "prompt_tokens": 1704860,
+      "uncached_tokens": 51582,
+      "output_tokens": 24458,
+      "cache_weighted_input_10": 197550.30000000002,
+      "cache_weighted_input_25": 416502.75,
+      "prompt_growth": 45489,
+      "antipatterns": 3,
+      "issues": 23,
+      "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-r35-orchestrated-codex.json

@@ -0,0 +1,61 @@
+{
+  "round": "35",
+  "runner": "codex",
+  "opts": {
+    "provider": "openai-oauth",
+    "model": "gpt-5.5",
+    "effort": null,
+    "fast": false,
+    "workflow": null,
+    "env": {}
+  },
+  "tasks": 1,
+  "completed": 1,
+  "completion_rate": 100,
+  "sessions": [
+    "019f27b1-8fef-71e0-ba3d-9c104a0eb324"
+  ],
+  "results": [
+    {
+      "id": "r5-orchestrated-audit",
+      "agent": "lead",
+      "ok": true,
+      "ms": 327028,
+      "sessionId": "019f27b1-8fef-71e0-ba3d-9c104a0eb324",
+      "card": {
+        "session": "019f27b1-8fef-71e0-ba3d-9c104a0eb324",
+        "wall_ms": 327028,
+        "uncached_in": 74996,
+        "cached_in": 997888,
+        "out_tokens": 14950,
+        "cache_ratio": 0.93,
+        "completed": 1
+      },
+      "output": "분석 전용으로 진행했고 파일은 수정하지 않았습니다. A/B를 별도 워커에 병렬 위임한 뒤, 제가 원문 라인과 대조해 병합했습니다.\n\n**Scope A**\n`openai-oauth-ws.mjs`가 직접 소비하는 shared classifier는 `classifyHandshakeError`와 `classifyMidstreamError`뿐입니다. import는 [openai-oauth-ws.mjs](C:/Project/mixdog/src/runtime/agent/orchestrator/providers/openai-oauth-ws.mjs:33), wrapper는 [openai-oauth-ws.mjs](C:/Project/mixdog/src/runtime/agent/orchestrator/providers/openai-oauth-ws.mjs:112)와 [openai-oauth-ws.mjs](C:/Project/mixdog/src/runtime/agent/orchestrator/providers/openai-oauth-ws.mjs:157)에 있습니다.\n\n`classifyError()` bucket:\n`auth`, `permanent`, `transient`, `unknown`입니다. 각각 401/403, unsafe/429/permanent 4xx, truncation/5xx/socket errno, fallback unknown으로 생산됩니다: [retry-classifier.mjs](C:/Project/mixdog/src/runtime/agent/orchestrator/providers/retry-classifier.mjs:100), [retry-classifier.mjs](C:/Project/mixdog/src/runtime/agent/orchestrator/providers/retry-classifier.mjs:105), [retry-classifier.mjs](C:/Project/mixdog/src/runtime/agent/orchestrator/providers/retry-classifier.mjs:113), [retry-classifier.mjs](C:/Project/mixdog/src/runtime/agent/orchestrator/providers/retry-classifier.mjs:117). 단, `openai-oauth-ws.mjs`는 이 classifier를 직접 호출하지 않습니다.\n\nHandshake bucket:\n`timeout`, `reset`, `dns`, `refused`, `network`, `acquire_timeout`, numeric `http_<status>` 예: `http_503`, 그리고 `null`입니다. 생산 근거는 [retry-classifier.mjs](C:/Project/mixdog/src/runtime/agent/orchestrator/providers/retry-classifier.mjs:477)부터 [retry-classifier.mjs](C:/Project/mixdog/src/runtime/agent/orchestrator/providers/retry-classifier.mjs:494)입니다. 소비는 `_acquireWithRetry()`의 [openai-oauth-ws.mjs](C:/Project/mixdog/src/runtime/agent/orchestrator/providers/openai-oauth-ws.mjs:403)에서 하며, `null`은 즉시 throw, non-null은 retry/tag 처리됩니다.\n\nWS midstream bucket:\n`first_byte_timeout`, `ws_send_failed`, numeric `http_<status>`, `agent_stall`, `stream_stalled`, `ws_1006`, `ws_1011`, `ws_1012`, `ws_4000`, `ws_1000`, `response_failed_disconnected`, `response_failed_network`, `response_failed_auth_expired`, `null`입니다. 분류기는 [retry-classifier.mjs](C:/Project/mixdog/src/runtime/agent/orchestrator/providers/retry-classifier.mjs:286)부터 [retry-classifier.mjs](C:/Project/mixdog/src/runtime/agent/orchestrator/providers/retry-classifier.mjs:342)이고, 소비는 [openai-oauth-ws.mjs](C:/Project/mixdog/src/runtime/agent/orchestrator/providers/openai-oauth-ws.mjs:843)부터 [openai-oauth-ws.mjs](C:/Project/mixdog/src/runtime/agent/orchestrator/providers/openai-oauth-ws.mjs:860)입니다. 실제 생산 근거는 first-byte timeout [openai-ws-stream.mjs](C:/Project/mixdog/src/runtime/agent/orchestrator/providers/openai-ws-stream.mjs:466), send failure [openai-ws-pool.mjs](C:/Project/mixdog/src/runtime/agent/orchestrator/providers/openai-ws-pool.mjs:106), response.failed [openai-ws-stream.mjs](C:/Project/mixdog/src/runtime/agent/orchestrator/providers/openai-ws-stream.mjs:953), close code [openai-ws-stream.mjs](C:/Project/mixdog/src/runtime/agent/orchestrator/providers/openai-ws-stream.mjs:1001)입니다.\n\nOpenAI WS에서 도달 불가:\nSSE path bucket인 `sse_idle_timeout`, `stream_terminated`, `fetch_failed`는 `policy.mode === 'sse'`에서만 나옵니다: [retry-classifier.mjs](C:/Project/mixdog/src/runtime/agent/orchestrator/providers/retry-classifier.mjs:348). OpenAI WS는 항상 `mode: 'ws'`를 전달합니다: [openai-oauth-ws.mjs](C:/Project/mixdog/src/runtime/agent/orchestrator/providers/openai-oauth-ws.mjs:87).\n\nNo-producer bucket:\nliteral `http_5xx`는 주석/allow-list에는 있지만 실제 producer가 없습니다. allow-list는 [retry-classifier.mjs](C:/Project/mixdog/src/runtime/agent/orchestrator/providers/retry-classifier.mjs:386), 주석은 [retry-classifier.mjs](C:/Project/mixdog/src/runtime/agent/orchestrator/providers/retry-classifier.mjs:469), 실제 producer는 numeric `http_${status}`입니다: [retry-classifier.mjs](C:/Project/mixdog/src/runtime/agent/orchestrator/providers/retry-classifier.mjs:309), [retry-classifier.mjs](C:/Project/mixdog/src/runtime/agent/orchestrator/providers/retry-classifier.mjs:481). Regex fallback이 numeric form은 받습니다: [retry-classifier.mjs](C:/Project/mixdog/src/runtime/agent/orchestrator/providers/retry-classifier.mjs:410).\n\n**Scope B**\nHandshake retry budget은 5 attempts입니다. `HANDSHAKE_MAX_ATTEMPTS = PROVIDER_RETRY_MAX_ATTEMPTS`이고, 값은 backoff array length 5입니다: [openai-oauth-ws.mjs](C:/Project/mixdog/src/runtime/agent/orchestrator/providers/openai-oauth-ws.mjs:104), [stall-policy.mjs](C:/Project/mixdog/src/runtime/agent/orchestrator/stall-policy.mjs:200). Loop는 `attempt = 1..5`: [openai-oauth-ws.mjs](C:/Project/mixdog/src/runtime/agent/orchestrator/providers/openai-oauth-ws.mjs:389). Backoff raw는 500/1000/2000/4000ms, cap 5000ms, symmetric jitter 기본 0.2입니다: [openai-oauth-ws.mjs](C:/Project/mixdog/src/runtime/agent/orchestrator/providers/openai-oauth-ws.mjs:175), [retry-classifier.mjs](C:/Project/mixdog/src/runtime/agent/orchestrator/providers/retry-classifier.mjs:222), [stall-policy.mjs](C:/Project/mixdog/src/runtime/agent/orchestrator/stall-policy.mjs:202).\n\nMidstream retry budget은 `ws_1006`/`ws_1011`만 4 retries, 나머지 retryable WS bucket은 2 retries입니다: [retry-classifier.mjs](C:/Project/mixdog/src/runtime/agent/orchestrator/providers/retry-classifier.mjs:245), [retry-classifier.mjs](C:/Project/mixdog/src/runtime/agent/orchestrator/providers/retry-classifier.mjs:251), [openai-oauth-ws.mjs](C:/Project/mixdog/src/runtime/agent/orchestrator/providers/openai-oauth-ws.mjs:164). Loop 자체는 max 4 retries까지 돌 수 있지만 per-classifier gate가 적용됩니다: [openai-oauth-ws.mjs](C:/Project/mixdog/src/runtime/agent/orchestrator/providers/openai-oauth-ws.mjs:557), [retry-classifier.mjs](C:/Project/mixdog/src/runtime/agent/orchestrator/providers/retry-classifier.mjs:266). Backoff raw는 250/1000/2000/4000ms + jitter입니다: [openai-oauth-ws.mjs](C:/Project/mixdog/src/runtime/agent/orchestrator/providers/openai-oauth-ws.mjs:170), [openai-oauth-ws.mjs](C:/Project/mixdog/src/runtime/agent/orchestrator/providers/openai-oauth-ws.mjs:852).\n\nDouble-retry 결론:\n같은 error object가 handshake와 midstream 양쪽에서 동시에 retry되지는 않습니다. acquire 실패는 별도 catch에서 처리되고 midstream classifier로 가지 않습니다: [openai-oauth-ws.mjs](C:/Project/mixdog/src/runtime/agent/orchestrator/providers/openai-oauth-ws.mjs:563). midstream classification은 acquire 성공 이후 send/stream catch에서만 실행됩니다: [openai-oauth-ws.mjs](C:/Project/mixdog/src/runtime/agent/orchestrator/providers/openai-oauth-ws.mjs:668), [openai-oauth-ws.mjs](C:/Project/mixdog/src/runtime/agent/orchestrator/providers/openai-oauth-ws.mjs:843). 다만 midstream retry가 발생하면 다음 attempt가 fresh acquire를 하므로 handshake budget이 순차적으로 다시 붙을 수 있습니다: [openai-oauth-ws.mjs](C:/Project/mixdog/src/runtime/agent/orchestrator/providers/openai-oauth-ws.mjs:571).\n\n추가 개선 포인트:\n`openai-oauth-ws.mjs` 주석은 아직 `ws_1006/ws_1011 get two retries`와 `up to 2`라고 되어 있지만 실제 정책은 4 retries입니다: [openai-oauth-ws.mjs](C:/Project/mixdog/src/runtime/agent/orchestrator/providers/openai-oauth-ws.mjs:507), [openai-oauth-ws.mjs](C:/Project/mixdog/src/runtime/agent/orchestrator/providers/openai-oauth-ws.mjs:1113). 또 local inter-chunk timeout은 4000 close를 호출하지만 error에 `wsCloseCode`를 붙이지 않고 바로 `finish()`하므로 `ws_4000`이 아니라 `null`로 끝날 가능성이 큽니다: [openai-ws-stream.mjs](C:/Project/mixdog/src/runtime/agent/orchestrator/providers/openai-ws-stream.mjs:526), [openai-ws-stream.mjs](C:/Project/mixdog/src/runtime/agent/orchestrator/providers/openai-ws-stream.mjs:530).\n\n변경 파일: 없음. 테스트 실행: 없음, 정적 분석과 라인 근거 교차검증만 수행했습니다."
+    }
+  ],
+  "task_complete": true,
+  "task_errors": [],
+  "score_complete": true,
+  "score_errors": [],
+  "group": {
+    "n": 1,
+    "wall_ms": 327028,
+    "uncached_in": 74996,
+    "cached_in": 997888,
+    "out_tokens": 14950,
+    "cache_ratio": 0.9,
+    "completed": 1
+  },
+  "cards": [
+    {
+      "session": "019f27b1-8fef-71e0-ba3d-9c104a0eb324",
+      "wall_ms": 327028,
+      "uncached_in": 74996,
+      "cached_in": 997888,
+      "out_tokens": 14950,
+      "cache_ratio": 0.93,
+      "completed": 1
+    }
+  ]
+}