mixdog 0.9.4 → 0.9.5

package/scripts/bench/round-r14-betafeatures.json

@@ -0,0 +1,466 @@
+{
+  "round": "14",
+  "runner": "lead",
+  "opts": {
+    "provider": "openai-oauth",
+    "model": "gpt-5.5",
+    "effort": null,
+    "fast": false,
+    "env": {
+      "MIXDOG_DATA_DIR": "C:\\Users\\tempe\\.mixdog\\bench-data"
+    }
+  },
+  "tasks": 6,
+  "completed": 6,
+  "completion_rate": 100,
+  "sessions": [
+    "sess_45552_1783058338359_1783058338607_a015a1cdab90e4a2d1f531080d1c286b",
+    "sess_40340_1783058338374_1783058338625_09d691094df16330e45e10c7a25fca4b",
+    "sess_41232_1783058338395_1783058338643_9c39c68d28de50be18441725abb6f037",
+    "sess_25396_1783058424363_1783058424619_5963443d3a6c9851961f027db3819c54",
+    "sess_22136_1783058427761_1783058427997_8ea3b1019cb8865372862f9df6344d43",
+    "sess_59476_1783058430062_1783058430294_c496f72ecc8cb5c98b93240b9e9ded45"
+  ],
+  "results": [
+    {
+      "id": "r4-debug-forwarding-gap#1",
+      "agent": "reviewer",
+      "ok": true,
+      "ms": 89407,
+      "sessionId": "sess_45552_1783058338359_1783058338607_a015a1cdab90e4a2d1f531080d1c286b",
+      "output": "sessionId: sess_45552_1783058338359_1783058338607_a015a1cdab90e4a2d1f531080d1c286b\n관련 경로를 직접 추적해 첫 전달까지의 손실/지연 구간만 확인하겠습니다.Found real activation windows: two can silently lose first output; the rest delay and are usually recovered.\n\n**Activation path**: `start()` sets `channelBridgeActive` and writes `bridge-state.json` before ownership/backend/binding are ready (`src/runtime/channels/index.mjs:1735-1744`), then `startOwnedRuntime()` binds/restores transcript around backend connect (`src/runtime/channels/index.mjs:758-813`), `applyTranscriptBinding()` installs context/watch (`src/runtime/channels/index.mjs:317-337`), and `OutputForwarder` forwards via `fs.watch → forwardNewText → deliverQueueItem` (`src/runtime/channels/lib/output-forwarder.mjs:646-731`, `351-410`, `283-340`).\n\n**Highest impact — wrong/stale transcript binding on first inbound**: if an old binding exists and self-session discovery is not active within the 500ms steal poll, inbound is delivered while the forwarder still tails the old transcript (`src/runtime/channels/index.mjs:1506-1531`, `1571-1617`, `1623-1628`). Backlog catch-up does **not reliably recover**: a later steal to an already-existing correct transcript calls `setContext()` without replay/catch-up, so it seeks to EOF and skips the missed first reply (`src/runtime/channels/lib/output-forwarder.mjs:102-129`).\n\n**High impact — existing transcript with no matching persisted cursor**: activation restore uses `catchUpFromPersisted`, but if the discovered transcript already exists and persisted state does not match it, `setContext()` starts at `currentSize`, dropping all output already written before binding (`src/runtime/channels/index.mjs:625-629`, `441-449`; `src/runtime/channels/lib/output-forwarder.mjs:117-127`). Backlog catch-up: **no**, because the cursor is initialized at EOF.\n\n**Medium impact — not-yet-created transcript**: `fs.watch()` cannot attach, `startWatch()` catches/closes, and re-arm polls until the file appears, then installs watch and calls `forwardNewText()` (`src/runtime/channels/lib/output-forwarder.mjs:646-670`; `src/runtime/channels/index.mjs:339-393`, `469-499`). Backlog catch-up: **yes within the re-arm window** and post-connect compensation (`src/runtime/channels/index.mjs:801-812`); **no automatic recovery after the 60s timeout** unless a later rebind happens.\n\n**Lower impact — ownership/backend gate after bridge-state**: before `bridgeRuntimeConnected && currentOwnerState().owned`, `forwardNewText()` returns without reading or advancing the cursor (`src/runtime/channels/index.mjs:316`, `776-781`; `src/runtime/channels/lib/output-forwarder.mjs:351-354`). Backlog catch-up: **yes**, because unread bytes remain and post-connect `forwardNewText()` runs, unless a prior EOF-seeking bind already skipped them.\n\n**Verdict**: the bridge has bounded-delay recovery for missing-file and pre-connect windows, but first-output loss is still possible when binding chooses an existing wrong/unpersisted transcript and initializes the cursor at EOF.",
+      "card": {
+        "session": "sess_45552_1783058338359_1783058338607_a015a1cdab90e4a2d1f531080d1c286b",
+        "provider": "openai-oauth",
+        "model": "gpt-5.5",
+        "wall_ms": 78893,
+        "turns": 4,
+        "tool_calls": 3,
+        "tools_per_turn": 0.8,
+        "total_tool_ms": 20,
+        "llm_stream_ms": 78392,
+        "cache_ratio": 0.11257956255063072,
+        "cached_tokens": 9728,
+        "prompt_tokens": 86410,
+        "uncached_tokens": 76682,
+        "output_tokens": 7512,
+        "cache_weighted_input_10": 77654.8,
+        "cache_weighted_input_25": 79114,
+        "prompt_growth": 34226,
+        "antipatterns": 1,
+        "issues": 6,
+        "issue_types": [
+          "cache",
+          "cache_miss",
+          "broad_tool_result",
+          "service_tier",
+          "tool_churn_cluster"
+        ],
+        "read_fragmentation_paths": [],
+        "grep_sweep_count": 0
+      }
+    },
+    {
+      "id": "r4-impl-usage-summary#1",
+      "agent": "worker",
+      "ok": true,
+      "ms": 85967,
+      "sessionId": "sess_40340_1783058338374_1783058338625_09d691094df16330e45e10c7a25fca4b",
+      "output": "sessionId: sess_40340_1783058338374_1783058338625_09d691094df16330e45e10c7a25fca4b\n요청 범위가 명확해서 바로 구현하고 검증까지 진행하겠습니다.구현 파일과 npm script가 준비된 상태라 문법 검사와 실제 실행을 확인하겠습니다.완료됐습니다. `usage:today` 스크립트와 `scripts/usage-today.mjs`가 추가되어 오늘 로컬 자정 이후 usage를 모델별로 집계합니다.\n\n- **Diff 요약**: `package.json`에 `usage:today` 추가, `scripts/usage-today.mjs` 신규 추가.\n- **동작**: `MIXDOG_DATA_DIR` 우선, 기본값은 `~/.mixdog/data/history/agent-trace.jsonl`; 모델별 calls/input/cached/output/costUsd와 `GRAND TOTAL` 출력.\n- **검증**: `node --check scripts/usage-today.mjs` 통과, `npm run usage:today` 실행 완료.\n- **참고**: 작업 전부터 `src/runtime/channels/lib/event-queue.mjs`, `src/runtime/channels/lib/output-forwarder.mjs`에 별도 수정 상태가 있었습니다.",
+      "card": {
+        "session": "sess_40340_1783058338374_1783058338625_09d691094df16330e45e10c7a25fca4b",
+        "provider": "openai-oauth",
+        "model": "gpt-5.5",
+        "wall_ms": 75469,
+        "turns": 10,
+        "tool_calls": 13,
+        "tools_per_turn": 1.3,
+        "total_tool_ms": 2431,
+        "llm_stream_ms": 70209,
+        "cache_ratio": 0.6937749383029269,
+        "cached_tokens": 120320,
+        "prompt_tokens": 173428,
+        "uncached_tokens": 53108,
+        "output_tokens": 5186,
+        "cache_weighted_input_10": 65140,
+        "cache_weighted_input_25": 83188,
+        "prompt_growth": 21498,
+        "antipatterns": 1,
+        "issues": 5,
+        "issue_types": [
+          "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": 91674,
+      "sessionId": "sess_41232_1783058338395_1783058338643_9c39c68d28de50be18441725abb6f037",
+      "output": "sessionId: sess_41232_1783058338395_1783058338643_9c39c68d28de50be18441725abb6f037\nMapped; concise architecture brief follows.\n\n- **Public surface:** `src/runtime/agent/orchestrator/session/compact.mjs:52-55` re-exports `semanticCompactMessages` and `recallFastTrackCompactMessages` from `compact/engine.mjs`.\n\n- **Callers:** normal turns enter through `askSession → agentLoop` at `src/runtime/agent/orchestrator/session/manager.mjs:1379-1380`; manual compact and clear paths call `runSessionCompaction` at `src/runtime/agent/orchestrator/session/manager.mjs:1849` and `src/runtime/agent/orchestrator/session/manager.mjs:1950-1957`.\n\n- **Policy gate:** loop compaction is enabled by `resolveWorkerCompactPolicy`, which derives `auto`, boundary, trigger, reserve, type, and tail settings from session/env config at `src/runtime/agent/orchestrator/session/loop/compact-policy.mjs:89-148`.\n\n- **Fire decision:** before provider send, the loop drains queued steering, estimates transcript tokens, then calls `shouldCompactForSession`; compaction fires on reactive retry, null estimate, or pressure ≥ trigger at `src/runtime/agent/orchestrator/session/loop.mjs:324-364` and `src/runtime/agent/orchestrator/session/loop/compact-policy.mjs:170-174`.\n\n- **Target budget:** the compact target is bounded by policy boundary and adds reserve back to the effective target at `src/runtime/agent/orchestrator/session/loop/compact-policy.mjs:163-168`; loop uses that target when `shouldCompact` is true at `src/runtime/agent/orchestrator/session/loop.mjs:362-364`.\n\n- **Path choice:** loop optionally prunes tool outputs first, then runs recall-fasttrack when configured, semantic otherwise, and rejects unavailable types at `src/runtime/agent/orchestrator/session/loop.mjs:405-498`.\n\n- **Session/manual gate:** `runSessionCompaction` skips closed sessions, disabled auto, too-few messages, missing boundary, or below-threshold pressure unless forced/manual at `src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:143-198`.\n\n- **Semantic window split:** `splitLiveCompactionContext` repairs tool pairs, separates protected prefix, finds prior summary, and removes summary messages from live history at `src/runtime/agent/orchestrator/session/compact/engine.mjs:100-111`.\n\n- **Protected prefix:** system messages plus leading `<system-reminder>` user messages and their `.` acknowledgements are preserved outside the compactable conversation at `src/runtime/agent/orchestrator/session/compact/messages.mjs:56-80`.\n\n- **Semantic tail budget:** `selectCompactionWindow` computes recent-budget via `preserveRecentBudget`, then chooses user-turn or cut-point selection at `src/runtime/agent/orchestrator/session/compact/engine.mjs:235-250` and `src/runtime/agent/orchestrator/session/compact/budget.mjs:274-287`.\n\n- **User-tail policy:** with user turns, semantic compaction keeps up to `tailTurns` newest turns within budget and can split the newest oversized turn suffix at `src/runtime/agent/orchestrator/session/compact/engine.mjs:134-185`.\n\n- **No-user policy:** without user turns, semantic compaction cuts only at non-`tool` messages, grows the newest fitting suffix, and keeps a head to summarize when possible at `src/runtime/agent/orchestrator/session/compact/engine.mjs:121-127` and `src/runtime/agent/orchestrator/session/compact/engine.mjs:193-232`.\n\n- **Recall tail policy:** recall-fasttrack keeps the newest structured turns, drops oldest kept turns until under cap, truncates the newest turn if needed, and allows no-user tails at `src/runtime/agent/orchestrator/session/compact/engine.mjs:593-619`.\n\n- **Recall pipeline:** recall-fasttrack ingests session memory, dumps roots/raw rows, optionally drains cycle1, combines recall text, caps recall injection to context share, then calls `recallFastTrackCompactMessages` at `src/runtime/agent/orchestrator/session/loop/recall-fasttrack.mjs:61-172`.\n\n- **Redaction hook:** semantic redacts after base sanitation and before window selection, so preserved system/tail are measured and emitted redacted at `src/runtime/agent/orchestrator/session/compact/engine.mjs:306-314` and `src/runtime/agent/orchestrator/session/compact/engine.mjs:389-392`.\n\n- **Recall redaction:** recall-fasttrack applies the same pre-window redaction before splitting protected context and choosing the preserved tail at `src/runtime/agent/orchestrator/session/compact/engine.mjs:622-651`.\n\n- **Redaction scope:** `redactToolCallSecretsInMessages` only changes assistant `toolCalls` argument values, preserving role/content/tool IDs/order/pairing at `src/runtime/agent/orchestrator/session/compact/text-utils.mjs:291-325`.\n\n- **Prompt-side scrub:** compaction prompt tool metadata and preserved facts also redact raw secret strings/tool-call args before summarization at `src/runtime/agent/orchestrator/session/compact/text-utils.mjs:151-154`, `src/runtime/agent/orchestrator/session/compact/summary.mjs:85-90`, and `src/runtime/agent/orchestrator/session/compact/budget.mjs:151-162`.\n\n- **Semantic hard failures:** missing provider, empty compactable head without prior summary, unfit prompt/summary, empty provider summary, or final budget overflow throw from `src/runtime/agent/orchestrator/session/compact/engine.mjs:278-281`, `src/runtime/agent/orchestrator/session/compact/engine.mjs:315-337`, and `src/runtime/agent/orchestrator/session/compact/engine.mjs:371-397`.\n\n- **Semantic soft repair:** malformed non-empty semantic summaries are schema-repaired before insertion, not injected blindly, at `src/runtime/agent/orchestrator/session/compact/summary.mjs:302-312`.\n\n- **Mandatory-cost fallback:** both semantic and recall raise the effective budget when preserved system+tail plus minimum summary room cannot fit, instead of refusing immediately at `src/runtime/agent/orchestrator/session/compact/engine.mjs:319-331` and `src/runtime/agent/orchestrator/session/compact/engine.mjs:658-664`.\n\n- **Recall hard failures:** recall-fasttrack throws on no compactable prior+empty recall, empty recall text when not allowed, unfit summary, or final budget overflow at `src/runtime/agent/orchestrator/session/compact/engine.mjs:653-699`.\n\n- **Loop fallback:** semantic failures can be salvaged by anchor-independent tool-output pruning if the pruned transcript fits effective budget; recall failures skip this silent prune and escalate at `src/runtime/agent/orchestrator/session/loop.mjs:500-537`.\n\n- **Loop escalation:** unrecovered compact failure records telemetry/events and throws `agentContextOverflowError` at `src/runtime/agent/orchestrator/session/loop.mjs:546-626`.\n\n- **Session fallback:** manual/clear recall-fasttrack failure can degrade to semantic compaction, then non-recall failures can use unanchored prune before recording failure at `src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:237-347`.\n\n- **State mutation:** on success, loop replaces `messages`, clears `providerState`, resets iteration/usage epoch, and emits compact telemetry at `src/runtime/agent/orchestrator/session/loop.mjs:630-716`; session/manual paths replace `session.messages`, clear `providerState`, and persist compaction metadata at `src/runtime/agent/orchestrator/session/manager/compaction-runner.mjs:399-470`.",
+      "card": {
+        "session": "sess_41232_1783058338395_1783058338643_9c39c68d28de50be18441725abb6f037",
+        "provider": "openai-oauth",
+        "model": "gpt-5.5",
+        "wall_ms": 81153,
+        "turns": 7,
+        "tool_calls": 9,
+        "tools_per_turn": 1.3,
+        "total_tool_ms": 697,
+        "llm_stream_ms": 80277,
+        "cache_ratio": 0.6459696986035477,
+        "cached_tokens": 119808,
+        "prompt_tokens": 185470,
+        "uncached_tokens": 65662,
+        "output_tokens": 6640,
+        "cache_weighted_input_10": 77642.8,
+        "cache_weighted_input_25": 95614,
+        "prompt_growth": 52898,
+        "antipatterns": 1,
+        "issues": 4,
+        "issue_types": [
+          "cache",
+          "broad_tool_result",
+          "service_tier",
+          "tool_churn_cluster"
+        ],
+        "read_fragmentation_paths": [],
+        "grep_sweep_count": 0
+      }
+    },
+    {
+      "id": "r4-debug-forwarding-gap#2",
+      "agent": "reviewer",
+      "ok": true,
+      "ms": 129661,
+      "sessionId": "sess_25396_1783058424363_1783058424619_5963443d3a6c9851961f027db3819c54",
+      "output": "sessionId: sess_25396_1783058424363_1783058424619_5963443d3a6c9851961f027db3819c54\n관련 경로를 직접 추적해 누락/지연 지점을 확인하겠습니다.I found several real activation-window gaps; most are delayed-and-caught-up, but two can still silently skip output.\n\n**Highest impact — EOF bind can drop early output:** `bindPersistedTranscriptIfAny()` calls rebind with `catchUpFromPersisted` only, and `setContext()` starts at EOF when the persisted path does not match, so assistant bytes already written before binding are skipped; backlog catch-up does **not** recover this. Anchors: `src/runtime/channels/index.mjs:625-629`, `src/runtime/channels/index.mjs:441-448`, `src/runtime/channels/lib/output-forwarder.mjs:117-126`.\n\n**High impact — no transcript path found:** if discovery never yields a pending/existing path and no previous path is usable, activation ends with no binding/watch, so output is delayed indefinitely until a later rebind trigger; backlog catch-up is **not automatic**. Anchors: `src/runtime/channels/index.mjs:431-460`, `src/runtime/channels/index.mjs:515-516`.\n\n**Medium impact — missing file watch:** binding a known-but-not-created transcript makes `fs.watch()` fail, but the re-arm poll installs the watch and calls `forwardNewText()` when the file appears; backlog catch-up **recovers within 60s**, otherwise it waits for a later trigger. Anchors: `src/runtime/channels/index.mjs:340-349`, `src/runtime/channels/index.mjs:469-498`, `src/runtime/channels/lib/output-forwarder.mjs:647-670`.\n\n**Medium impact — ownership/backend gate:** `bridge-state.json` is written before `backend.connect()` and before `bridgeRuntimeConnected`, so watcher flushes return before reading; backlog catch-up **recovers after connect** via post-connect `forwardNewText()`/re-arm. Anchors: `src/runtime/channels/index.mjs:1734-1744`, `src/runtime/channels/index.mjs:759-812`, `src/runtime/channels/lib/output-forwarder.mjs:36-42`, `src/runtime/channels/lib/output-forwarder.mjs:351-354`.\n\n**Low impact — stat-to-watch and partial-line windows:** bytes written between `setContext()` stat and watch install are covered by an immediate scheduled flush, and partial JSON lines are re-read after newline; backlog catch-up **recovers** both. Anchors: `src/runtime/channels/lib/output-forwarder.mjs:111-129`, `src/runtime/channels/lib/output-forwarder.mjs:665-666`, `src/runtime/channels/lib/output-forwarder.mjs:172-179`.\n\nVerdict: the bridge is mostly catch-up safe after ownership/connect, but first-output loss is still possible when activation binds an already-existing transcript without a matching persisted cursor.",
+      "card": {
+        "session": "sess_25396_1783058424363_1783058424619_5963443d3a6c9851961f027db3819c54",
+        "provider": "openai-oauth",
+        "model": "gpt-5.5",
+        "wall_ms": 119136,
+        "turns": 10,
+        "tool_calls": 9,
+        "tools_per_turn": 0.9,
+        "total_tool_ms": 508,
+        "llm_stream_ms": 118332,
+        "cache_ratio": 0.6560918574934662,
+        "cached_tokens": 181248,
+        "prompt_tokens": 276254,
+        "uncached_tokens": 95006,
+        "output_tokens": 10229,
+        "cache_weighted_input_10": 113130.8,
+        "cache_weighted_input_25": 140318,
+        "prompt_growth": 42817,
+        "antipatterns": 1,
+        "issues": 9,
+        "issue_types": [
+          "cache_miss",
+          "readonly_stall",
+          "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": 90912,
+      "sessionId": "sess_22136_1783058427761_1783058427997_8ea3b1019cb8865372862f9df6344d43",
+      "output": "sessionId: sess_22136_1783058427761_1783058427997_8ea3b1019cb8865372862f9df6344d43\n요청 범위가 명확해서 바로 구현하고 검증까지 진행하겠습니다.구현 파일이 이미 작업 트리에 준비되어 있어 검증을 진행합니다.완료했습니다, 재영님. `usage:today` 스크립트와 `scripts/usage-today.mjs`가 추가되었습니다.\n\n- **바뀐 점** `package.json`에 `usage:today`를 추가했고, 새 스크립트가 `MIXDOG_DATA_DIR` 또는 기본 `~/.mixdog/data/history/agent-trace.jsonl`을 읽어 오늘 사용량을 모델별/총합으로 출력합니다.\n\n- **집계 내용** 모델별 `calls`, `input`, `cached`, `output`, `costUsd`와 `GRAND TOTAL` 라인을 출력하며 `--trace`, `--json`도 지원합니다.\n\n- **검증** `node --check scripts/usage-today.mjs` 통과, `npm run usage:today` 실행 완료: `GRAND TOTAL 1318 calls, 48.23M input, 47.96M cached, 1.57M output, 85.4306 costUsd`.\n\n- **참고** 작업 대상 외 기존 수정 파일 `src/runtime/channels/lib/event-queue.mjs`, `src/runtime/channels/lib/output-forwarder.mjs`는 건드리지 않았습니다.",
+      "card": {
+        "session": "sess_22136_1783058427761_1783058427997_8ea3b1019cb8865372862f9df6344d43",
+        "provider": "openai-oauth",
+        "model": "gpt-5.5",
+        "wall_ms": 80417,
+        "turns": 12,
+        "tool_calls": 13,
+        "tools_per_turn": 1.1,
+        "total_tool_ms": 3259,
+        "llm_stream_ms": 77037,
+        "cache_ratio": 0.6162420476577276,
+        "cached_tokens": 103936,
+        "prompt_tokens": 168661,
+        "uncached_tokens": 64725,
+        "output_tokens": 3898,
+        "cache_weighted_input_10": 75118.6,
+        "cache_weighted_input_25": 90709,
+        "prompt_growth": 13620,
+        "antipatterns": 1,
+        "issues": 7,
+        "issue_types": [
+          "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": 78464,
+      "sessionId": "sess_59476_1783058430062_1783058430294_c496f72ecc8cb5c98b93240b9e9ded45",
+      "output": "sessionId: sess_59476_1783058430062_1783058430294_c496f72ecc8cb5c98b93240b9e9ded45\n- **Entry points/callers:** `compact.mjs` re-exports `semanticCompactMessages` and `recallFastTrackCompactMessages`; live loop calls semantic at pre-send and recall via `runRecallFastTrackCompact`, while manager/manual paths call both through `runSessionCompaction` (`compact.mjs:52-55`, `engine.mjs:278`, `engine.mjs:445`, `loop.mjs:414-459`, `compaction-runner.mjs:207-318`).\n\n- **Fire decision:** live compaction runs before provider send after queued steering is drained, only when policy auto is enabled and `shouldCompactForSession` says pressure meets trigger, estimate is unknown, or reactive overflow forced it (`loop.mjs:324-354`, `compact-policy.mjs:89-93`, `compact-policy.mjs:170-174`).\n\n- **Budget/trigger:** pressure is transcript tokens plus reserve; trigger is explicit sub-boundary auto limit or `boundary - buffer`, with default buffer ratio `0.1` capped by `0.25`; target budget is compact target plus reserve capped at boundary (`compact-policy.mjs:150-168`, `compact-policy.mjs:101-117`, `context-utils.mjs:94-96`, `context-utils.mjs:206-228`).\n\n- **Type routing:** policy chooses `semantic` by default, forces agent-owned sessions to semantic, and routes recall-fasttrack vs semantic accordingly (`constants.mjs:42-44`, `compact-policy.mjs:39-51`, `compact-policy.mjs:123-140`, `loop.mjs:414-459`).\n\n- **Semantic window:** engine first sanitizes, separates protected system/reminder prefix from conversation, finds prior compact summary, excludes summaries from live history, then splits into head/tail (`engine.mjs:100-112`, `messages.mjs:56-81`).\n\n- **Semantic tail:** with user turns, it keeps up to `tailTurns` recent turns within `preserveRecentBudget`, splitting the newest turn suffix if needed; without user turns, it cuts only at non-tool messages so assistant/tool pairs stay together (`engine.mjs:130-185`, `engine.mjs:188-233`, `engine.mjs:235-250`).\n\n- **Recall tail:** recall-fasttrack keeps newest user-anchored turns, drops oldest kept turns until under cap, then truncates the oldest surviving message if still oversized; no-user tails keep newest fitting messages and repair tool pairs (`engine.mjs:449-467`, `engine.mjs:525-575`, `engine.mjs:593-619`).\n\n- **Redaction hook:** no-op fit returns sanitized transcript unchanged; once compaction proceeds, `redactToolCallSecretsInMessages` runs before window selection so preserved system/tail are measured and emitted redacted, while prompt-side tool summaries also redact/truncate args (`engine.mjs:284-312`, `engine.mjs:389-393`, `text-utils.mjs:219-326`, `summary.mjs:85-90`).\n\n- **Summary construction:** semantic sends a compact prompt to provider, repairs malformed summary schema, fits summary into remaining budget, then emits `[system, summary, tail]`; recall-fasttrack injects memory recall/prior summary into a fitted summary without an LLM call (`engine.mjs:333-386`, `summary.mjs:302-386`, `engine.mjs:666-695`, `summary.mjs:412-450`).\n\n- **Failure/fallback:** semantic throws on missing provider, empty live/window, no compactable head, prompt/summary fit failures, empty provider output, or final budget overflow; recall throws on no compactable history, empty recall, summary fit failure, or overflow (`engine.mjs:278-280`, `engine.mjs:252-256`, `engine.mjs:315-317`, `engine.mjs:333-337`, `engine.mjs:371-397`, `engine.mjs:653-699`).\n\n- **Recovery paths:** loop semantic failures can fall back to anchor-independent tool-output prune, but recall-fasttrack failures do not silently prune; manager recall-fasttrack failure first tries semantic fallback, then non-recall failures may prune, otherwise records failed stage/error (`loop.mjs:500-618`, `compaction-runner.mjs:237-347`, `compaction-runner.mjs:348-397`).",
+      "card": {
+        "session": "sess_59476_1783058430062_1783058430294_c496f72ecc8cb5c98b93240b9e9ded45",
+        "provider": "openai-oauth",
+        "model": "gpt-5.5",
+        "wall_ms": 67975,
+        "turns": 13,
+        "tool_calls": 16,
+        "tools_per_turn": 1.2,
+        "total_tool_ms": 920,
+        "llm_stream_ms": 66801,
+        "cache_ratio": 0.563987246794594,
+        "cached_tokens": 247296,
+        "prompt_tokens": 438478,
+        "uncached_tokens": 191182,
+        "output_tokens": 4369,
+        "cache_weighted_input_10": 215911.6,
+        "cache_weighted_input_25": 253006,
+        "prompt_growth": 75759,
+        "antipatterns": 1,
+        "issues": 11,
+        "issue_types": [
+          "cache_miss",
+          "readonly_stall",
+          "cache",
+          "broad_tool_result",
+          "service_tier",
+          "missed_parallelism",
+          "tool_churn_cluster"
+        ],
+        "read_fragmentation_paths": [],
+        "grep_sweep_count": 0
+      }
+    }
+  ],
+  "task_complete": true,
+  "task_errors": [],
+  "score_complete": true,
+  "score_errors": [],
+  "group": {
+    "n": 6,
+    "wall_ms": 83840.5,
+    "turns": 9.3,
+    "tool_calls": 10.5,
+    "tools_per_turn": 1.1,
+    "total_tool_ms": 1305.8,
+    "llm_stream_ms": 81841.3,
+    "cache_ratio": 0.5,
+    "cached_tokens": 130389.3,
+    "prompt_tokens": 221450.2,
+    "uncached_tokens": 91060.8,
+    "output_tokens": 6305.7,
+    "cache_weighted_input_10": 104099.8,
+    "cache_weighted_input_25": 123658.2,
+    "prompt_growth": 40136.3,
+    "antipatterns": 1,
+    "issues": 7,
+    "grep_sweep_count": 0
+  },
+  "cards": [
+    {
+      "session": "sess_45552_1783058338359_1783058338607_a015a1cdab90e4a2d1f531080d1c286b",
+      "provider": "openai-oauth",
+      "model": "gpt-5.5",
+      "wall_ms": 78893,
+      "turns": 4,
+      "tool_calls": 3,
+      "tools_per_turn": 0.8,
+      "total_tool_ms": 20,
+      "llm_stream_ms": 78392,
+      "cache_ratio": 0.11257956255063072,
+      "cached_tokens": 9728,
+      "prompt_tokens": 86410,
+      "uncached_tokens": 76682,
+      "output_tokens": 7512,
+      "cache_weighted_input_10": 77654.8,
+      "cache_weighted_input_25": 79114,
+      "prompt_growth": 34226,
+      "antipatterns": 1,
+      "issues": 6,
+      "issue_types": [
+        "cache",
+        "cache_miss",
+        "broad_tool_result",
+        "service_tier",
+        "tool_churn_cluster"
+      ],
+      "read_fragmentation_paths": [],
+      "grep_sweep_count": 0
+    },
+    {
+      "session": "sess_40340_1783058338374_1783058338625_09d691094df16330e45e10c7a25fca4b",
+      "provider": "openai-oauth",
+      "model": "gpt-5.5",
+      "wall_ms": 75469,
+      "turns": 10,
+      "tool_calls": 13,
+      "tools_per_turn": 1.3,
+      "total_tool_ms": 2431,
+      "llm_stream_ms": 70209,
+      "cache_ratio": 0.6937749383029269,
+      "cached_tokens": 120320,
+      "prompt_tokens": 173428,
+      "uncached_tokens": 53108,
+      "output_tokens": 5186,
+      "cache_weighted_input_10": 65140,
+      "cache_weighted_input_25": 83188,
+      "prompt_growth": 21498,
+      "antipatterns": 1,
+      "issues": 5,
+      "issue_types": [
+        "cache_miss",
+        "cache",
+        "service_tier",
+        "missed_parallelism",
+        "tool_churn_cluster"
+      ],
+      "read_fragmentation_paths": [],
+      "grep_sweep_count": 0
+    },
+    {
+      "session": "sess_41232_1783058338395_1783058338643_9c39c68d28de50be18441725abb6f037",
+      "provider": "openai-oauth",
+      "model": "gpt-5.5",
+      "wall_ms": 81153,
+      "turns": 7,
+      "tool_calls": 9,
+      "tools_per_turn": 1.3,
+      "total_tool_ms": 697,
+      "llm_stream_ms": 80277,
+      "cache_ratio": 0.6459696986035477,
+      "cached_tokens": 119808,
+      "prompt_tokens": 185470,
+      "uncached_tokens": 65662,
+      "output_tokens": 6640,
+      "cache_weighted_input_10": 77642.8,
+      "cache_weighted_input_25": 95614,
+      "prompt_growth": 52898,
+      "antipatterns": 1,
+      "issues": 4,
+      "issue_types": [
+        "cache",
+        "broad_tool_result",
+        "service_tier",
+        "tool_churn_cluster"
+      ],
+      "read_fragmentation_paths": [],
+      "grep_sweep_count": 0
+    },
+    {
+      "session": "sess_25396_1783058424363_1783058424619_5963443d3a6c9851961f027db3819c54",
+      "provider": "openai-oauth",
+      "model": "gpt-5.5",
+      "wall_ms": 119136,
+      "turns": 10,
+      "tool_calls": 9,
+      "tools_per_turn": 0.9,
+      "total_tool_ms": 508,
+      "llm_stream_ms": 118332,
+      "cache_ratio": 0.6560918574934662,
+      "cached_tokens": 181248,
+      "prompt_tokens": 276254,
+      "uncached_tokens": 95006,
+      "output_tokens": 10229,
+      "cache_weighted_input_10": 113130.8,
+      "cache_weighted_input_25": 140318,
+      "prompt_growth": 42817,
+      "antipatterns": 1,
+      "issues": 9,
+      "issue_types": [
+        "cache_miss",
+        "readonly_stall",
+        "cache",
+        "broad_tool_result",
+        "service_tier",
+        "missed_parallelism",
+        "tool_churn_cluster"
+      ],
+      "read_fragmentation_paths": [],
+      "grep_sweep_count": 0
+    },
+    {
+      "session": "sess_22136_1783058427761_1783058427997_8ea3b1019cb8865372862f9df6344d43",
+      "provider": "openai-oauth",
+      "model": "gpt-5.5",
+      "wall_ms": 80417,
+      "turns": 12,
+      "tool_calls": 13,
+      "tools_per_turn": 1.1,
+      "total_tool_ms": 3259,
+      "llm_stream_ms": 77037,
+      "cache_ratio": 0.6162420476577276,
+      "cached_tokens": 103936,
+      "prompt_tokens": 168661,
+      "uncached_tokens": 64725,
+      "output_tokens": 3898,
+      "cache_weighted_input_10": 75118.6,
+      "cache_weighted_input_25": 90709,
+      "prompt_growth": 13620,
+      "antipatterns": 1,
+      "issues": 7,
+      "issue_types": [
+        "cache_miss",
+        "cache",
+        "service_tier",
+        "missed_parallelism",
+        "tool_churn_cluster"
+      ],
+      "read_fragmentation_paths": [],
+      "grep_sweep_count": 0
+    },
+    {
+      "session": "sess_59476_1783058430062_1783058430294_c496f72ecc8cb5c98b93240b9e9ded45",
+      "provider": "openai-oauth",
+      "model": "gpt-5.5",
+      "wall_ms": 67975,
+      "turns": 13,
+      "tool_calls": 16,
+      "tools_per_turn": 1.2,
+      "total_tool_ms": 920,
+      "llm_stream_ms": 66801,
+      "cache_ratio": 0.563987246794594,
+      "cached_tokens": 247296,
+      "prompt_tokens": 438478,
+      "uncached_tokens": 191182,
+      "output_tokens": 4369,
+      "cache_weighted_input_10": 215911.6,
+      "cache_weighted_input_25": 253006,
+      "prompt_growth": 75759,
+      "antipatterns": 1,
+      "issues": 11,
+      "issue_types": [
+        "cache_miss",
+        "readonly_stall",
+        "cache",
+        "broad_tool_result",
+        "service_tier",
+        "missed_parallelism",
+        "tool_churn_cluster"
+      ],
+      "read_fragmentation_paths": [],
+      "grep_sweep_count": 0
+    }
+  ]
+}