@kodax-ai/kodax 0.7.39 → 0.7.41

package/CHANGELOG.md

@@ -4,6 +4,106 @@ All notable changes to this project will be documented in this file.
 
 > Full history for versions prior to v0.7.0: [CHANGELOG_ARCHIVE.md](docs/CHANGELOG_ARCHIVE.md)
 
+## [0.7.41] - 2026-05-19
+
+### Theme
+
+**KodaX Team Mode + AMA Reliability + Source-Tree Modularization + REPL Render & TTFB Perf** — The release lands a third-axis differentiator (multi-instance auto coordination) alongside three AMA-path reliability fixes (mid-turn inject, pending-children handoff gate, post-handoff label flip, terminal-verdict fallback), the Todo V2 per-task CRUD migration with extension hooks, and the largest source-tree refactor since v0.7.25 (`runner-driven.ts` 6406 → 1897 lines, -70.4%, byte-identical). FEATURE_125 KodaX Team Mode is the headline: zero-cognitive-load multi-session awareness (no `/team create`, no `team_id`) — each KodaX instance writes per-pid state to `<configHome>/instances/<pid>/`, every LLM round injects a sibling-snapshot block into all 5 managed roles' system prompts (Scout / Planner / Generator / Evaluator / Worker), and a runtime content-hash safety net catches the only genuine data-race surface (concurrent overwrite of a file another session already read). The LLM-First design contrasts with claude code Team Mode's mode-based 4-stage workflow and no-conflict-resolution semantics. FEATURE_167 closes a structural `signal:'COMPLETE'` false-positive on V2 Evaluator turns (3-layer probe-gated defense: B0 parser SKIP / B1 retry cap / B2 synthesized verdict accept). FEATURE_165 + 166 land together as a Worker→Evaluator handoff hardening pair — runtime gate blocks `emit_handoff` while child registry is non-empty (covers V1+V2 shared `handoffEmit` path), and the REPL surface flips role labels immediately on `agentSwitched` (was lagged a turn). FEATURE_170 migrates the todo subsystem from monolithic init/replace to per-task add/patch/remove with extension events + hooks + new `todo_create` tool; Layer 2 LLM-judge eval (Layer A + Layer B 3.2% disagreement → DATA VALID) clears gate (a)+(b) MET, (c) saturation-artifact noted. FEATURE_171 extracts 12 submodules from `runner-driven.ts` across R1–R4 — zero behavior change, 4 reviewer APPROVE rounds, 4314/4314 tests pass each commit, ADR-026 + HLD §3.5.1 documented.
+
+### Added
+
+- **FEATURE_125 — KodaX Team Mode (Multi-Instance Auto Coordination)**. 11 commits S1–S7 + W1–W4 (`acef3c5e` → `9225ad31` → S7 `e2916675` + `e6bc5d7b` audit) + release-prep wiring `0cfc8bc4`. KodaX 自创的多 session 自动协调机制：用户**零认知负担**（无 `/team create`、无 `team_id` 概念），KodaX 自动感知本机其他 KodaX session 状态，把状态注入 LLM system prompt 让 LLM 自决避让/协作/调度；runtime 仅在 race condition 物理边界（content hash mismatch）兜底，不强制 lock、不强制等待。这是与 claude code Team Mode（mode-based 4-stage workflow + 完全无 conflict resolution）的核心差异化。**5 layers**: (S1) per-instance state writer at `<configHome>/instances/<pid>/{state.json,meta.json,heartbeat}` with atomic writes + 1s heartbeat + register/refresh/shutdown lifecycle; (S2) sibling-instance discovery + stale detection + reap with `PersistedSessionState v1` version guard + per-instance failure isolation; (S3) pure system-prompt formatter for the `=== Other active KodaX sessions ===` block (LLM-First wording, truncation, no behavior dictation); (S4) `KodaXToolExecutionContext.contentHashCache?` sha256-based stale-write detection with `recordRead` / `checkStale` / `recordWrite` per-task lifetime; (S5) tool-time soft-warning formatter for exact-path overlap match (no blocking, just an informational banner). **Wiring** (W1–W4): Read tool records sha256 on every successful read up to 5 MB (size cap so huge files don't pay the hash cost); Edit / Write / MultiEdit pre-mutation `checkStale` block + post-mutation `recordWrite` + sibling-overlap warning banner via `ctx.siblingSnapshot`; REPL bootstrap helper `bootstrapTeamMode()` with process-level singleton + `/exit` + SIGTERM lifecycle hooks (mirror wiring landed for both `runInteractiveMode` legacy path AND `runInkInteractiveMode` Ink REPL path — the latter was the release-blocker fix in commit `0cfc8bc4`); runner-driven adapter does per-LLM-round sibling discovery, injects `teamModeSection` into all 5 managed roles' system prompts (Scout / Planner / Generator / Evaluator / Worker) via a mutable `siblingSnapshot` ref + `Object.defineProperty` getter so tool ctx always reads the freshest snapshot. **S7 Layer 2 panel + audit**: `tests/feature-125-team-mode-awareness.eval.ts` + `benchmark/datasets/feature-125-team-mode-awareness/cases.ts`; 5 aliases × 2 cases × 5 runs = 50 LLM calls; **SHIP** per pre-registered matrix after audit-corrected regex extension (`buildToolNamePatterns` expanded from 4 to 9 syntax variants to capture kimi `read:0>{...}` and zhipu `<tool_name>read</tool_name>` forms). Layer A + Layer B audit-corrected primary verdict: case 1 84% / case 2 60% overall (4/5 aliases ≥60% — kimi case 2 narrate-without-tool documented as `feedback_model_structural_floor_not_prompt_tunable`; not addressable via prompt iteration). Design doc: [docs/features/v0.7.41.md#feature_125-kodax-team-mode--multi-instance-auto-coordination](docs/features/v0.7.41.md#feature_125-kodax-team-mode--multi-instance-auto-coordination). Test guide: `docs/test-guides/FEATURE_125_v0.7.41_TEST_GUIDE.md`.
+- **FEATURE_165 — Worker `emit_handoff` pending-children gate**. Commit `0ebeb15f`. Runtime gate at `runner-driven.ts:2402` blocks `emit_handoff` when the child registry is non-empty (covers both V1 and V2's shared `handoffEmit` path). 9 unit tests + 1 integration test pin the gate semantics across both paths. **Prompt addition PARTIAL/dropped**: Layer 2 probe (250 calls × 5 aliases) showed negative-case D/E already 100% on the baseline (`Δ=0pp`), so the pre-registered SHIP condition (2) failed mathematically; the runtime gate is the production-load-bearing change. Probe also confirmed zhipu intent-vs-action floor reproduces in canned-history sessions (structural, not context-length-driven). Design doc: [docs/features/v0.7.41.md#feature_165--worker-emit_handoff-pending-children-gatev0741-hotfix](docs/features/v0.7.41.md#feature_165--worker-emit_handoff-pending-children-gatev0741-hotfix).
+- **FEATURE_166 — Post-handoff role label flip**. Commit `0ebeb15f`. New `onAgentSwitched` hook on agent-runtime + `ObserverBridge.agentSwitched(role)` on coding-side. Fixes the V2 Worker→Evaluator handoff label-lag (`[Worker]` would persist on the next Evaluator turn until the assistant produced output). Production session `20260515_185354` gave a directly reproducible verdict trace. 7 unit tests + 1 pre-existing test corrected. Same session also surfaced FEATURE_167 (Evaluator text-only termination leaves `recorder.verdict === undefined`, V2 runner-driven never wired the `parseManagedTaskVerdictDirectiveFromJson` fallback — landed in FEATURE_167 below). Design doc: [docs/features/v0.7.41.md#feature_166--post-handoff-role-label-flipshipped](docs/features/v0.7.41.md#feature_166--post-handoff-role-label-flipshipped).
+- **FEATURE_167 — Evaluator terminal-verdict fallback (B0 parser + B1 retry + B2 synthesized accept)**. Commit `d537c784` 2026-05-15. Three-layer probe-gated defense closes the structural `signal:'COMPLETE'` false-positive on V2 Evaluator turns where the model exits text-only without calling `emit_verdict` and `recorder.verdict` therefore stays `undefined`. **Layer B0** — parser SKIP path (regex+JSON parse on the assistant's terminal text looking for `{"signal":"COMPLETE","grade":...}` directives); **Layer B1** — retry gate with per-alias cap (default 2, zhipu cap 1 to avoid amplifying the intent-vs-action floor — see `project_zhipu_send_message_floor` memory); **Layer B2** — synthesized verdict accept (fabricates a `{signal:'NEEDS_REVISION', grade:'C', summary:'…inferred from terminal text…'}` envelope so the V2 task engine can complete instead of hanging on the missing verdict). Reviewer-suggested change "include `revise` in the gate" was rejected — correct invariant is `recorder.verdict` object identity comparison (NOT status comparison), otherwise a stale `revise` from a prior turn would falsely satisfy the gate. 29 tests (16 retry-config + 9 predicate + 4 integration); audit panel 0/75 disagreement → DATA VALID. Design doc: [docs/features/v0.7.41.md#feature_167--evaluator-terminal-verdict-兜底shipped](docs/features/v0.7.41.md#feature_167--evaluator-terminal-verdict-兜底shipped).
+- **FEATURE_170 — Todo V2 Migration (per-task CRUD + extension hooks + `todo_create` tool)**. C1–C6 across 8 commits (`e45ddaa8` → `20e02103`). Replaces v0.7.x's monolithic init/replace todo-store API with per-task `add` / `patch` / `remove` operations + monotonic counter + metadata + extension events (`todo:added` / `todo:patched` / `todo:removed` / `todo:before-complete`) + before-complete hook for downstream consumers. New `todo_create` tool added to the registry + role wiring + throttle reset. Worker / legacy / throttle prompts updated to teach the per-item API (C5) with activeForm parity fix (C5 follow-up). **Layer 2 LLM-judge eval (Layer A + Layer B)**: 250-call panel + Layer A 5-sub-agent self-judge + Layer B 3-judge majority (750 calls), Layer B 3.2% disagreement → DATA VALID; gate (a)+(b) MET; gate (c) FAIL as a **pre-registered SHIP gate saturation artifact** (C2 baseline 96% / C3 100% — mathematically unable to add +20pp from a near-saturated baseline). C1 +32pp / mmx 0→100% are direct prompt-cause evidence — SHIP, keep the prompt rewrite. Lessons captured in two new memory entries: `feedback_pre_registered_gate_saturation` (pilot for baseline ceiling before deferring on Δ ≥+N pp) and `feedback_simplifying_prompt_can_regress` (Prefer over X when Y comparative clauses are load-bearing). Design doc: [docs/features/v0.7.41.md#feature_170--todo-v2-migration-per-task-crud--extension-hookssshipped-2026-05-16](docs/features/v0.7.41.md#feature_170--todo-v2-migration-per-task-crud--extension-hookssshipped-2026-05-16).
+- **FEATURE_164 — Mid-turn user-input injection** (shipped as part of commit `0ebeb15f`, the FEATURE_164+165+166 triple). Closes the gap where a user prompt typed during an active LLM round was queued but only delivered as a synthetic `[user]` banner on the next idle-yield wake — semantically incorrect for the user's intent ("inject as if I'd typed it mid-turn"). Now the runner-driven adapter checks the `MessageQueue` snapshot before each LLM call and prepends any queued real-user messages as proper non-synthetic user-bubble messages within the same round.
+
+### Fixed
+
+- **FEATURE_125 W3 — Ink REPL Team Mode bootstrap wiring** (release-blocker fix). Commit `0cfc8bc4`. Discovered during v0.7.41 release prep audit: FEATURE_125 W3 (commit `1a073ecc`) wired `bootstrapTeamMode` into the legacy `runInteractiveMode` path but never into `runInkInteractiveMode`, so the Ink REPL (the default REPL path on all platforms since v0.7.25) ran with Team Mode dormant — `<configHome>/instances/<pid>/` was never created, no heartbeat thread started, sibling discovery returned empty, and the system-prompt `teamModeSection` was a no-op for every Ink-launched session. Mirror-wires the bootstrap + `process.on('exit')` + `process.on('SIGTERM')` + clean-exit cleanup into `runInkInteractiveMode` at the same insertion point (after `gitRoot` resolution, before render). 37 lines net.
+- **Issue 132 — h2-boundary `session.jsonl` ENOENT race**. Commit `bf3006fb`. Eager-read in `agent-task-runner` resolves the timing window where benchmark h2-boundary cases would call `tail -f session.jsonl` before the file existed on disk; pre-reads on task start instead of awaiting the first append.
+- **FEATURE_166 stale-test correction** (1 pre-existing test): `agent-runtime.test.ts` had been asserting the buggy label-lag behavior as-correct — corrected to pin the fixed semantics so future regressions surface immediately.
+- **FEATURE_171 build break + decl emit** (covered transitively by the R1–R4 chain test-pass discipline): every refactor commit ran `tsc -b tsconfig.build.json` + 4314 tests green; no stage shipped a partial transform.
+- **Bundle SDK `.d.ts` so consumer `tsc` resolves types** (commit `af623000`). Footgun caught at the SDK consumer surface: tarball shipped `dist/index.js` + subpath bundles but no matching `.d.ts`, so `import { runKodaX } from '@kodax-ai/kodax'` worked at runtime while consumer `tsc` reported missing types. Build pipeline now layers `tsc --emitDeclarationOnly` on top of the esbuild bundle so every published subpath ships real types.
+- **`KODAX_RENDER_TRACE` default path uses `os.tmpdir()` not `homedir()`** (commit `54a59caa`). Phase A.0 review follow-up — `homedir()` pollutes the user's home with per-pid trace files; `os.tmpdir()` is the conventional location for ephemeral diagnostic output and gets cleaned up by the OS.
+
+### Performance
+
+- **FEATURE_172 — REPL Render Path Optimization (Phase 1 + Phase A.0/A.1)**. Triggered by user SSH long-session (`kodax -c` with 200+ history items) reporting "every 2-3s a frame refresh" during streaming. Two-phase work, with a mid-feature scope correction.
+  - **Phase 1 (data layer)** — 5 commits `19c6aff3` → `26d47084`. Split `transcript-layout.ts` into pure static/dynamic helpers (`buildTranscriptStaticPortion` / `buildTranscriptDynamicPortion` / `composeTranscriptRenderModel`); split `promptMainScreenRenderModel` + `transcriptMainScreenRenderModel` `useMemo` into static + dynamic with a static-cache-key invariant (streaming-state changes no longer invalidate the static portion); added `React.memo` `areTranscriptRowPropsEqual` comparator on `TranscriptRowRenderer`. **Data-layer bench** (`baseline-26d47084.json`, 800 items): streaming-tick p95 94.18ms → 0.52ms (-99.4%).
+  - **Phase 1 scope correction (2026-05-19)** — Phase 1 ship review with 3 parallel Explore-agent traces + claudecode end-to-end pipeline comparison revealed the data-layer bench (`benchmark/perf/repl-render-perf.bench.ts`) only measured `buildTranscriptRenderModel` inner function (~3-5% of total per-frame cost). The real ~80% lives in `tui/substrate/ink/` rendering substrate: `renderNodeToOutput` full-tree recursion (~55%), `setCellAt` `cells.slice()` O(N²) (~12%), `Output.getGrid()` rebuild (~12%), `diffEach` full-screen walk (~10%), `markDirty` propagation gap (~5%). **Lesson** captured to feedback memory: bench must measure end-to-end wall-time, not isolated inner functions; static analysis of a hot loop can miss the actual cost center.
+  - **Phase A.0 — `KODAX_RENDER_TRACE` env-gated per-frame trace + end-to-end bench scaffold** (commits `5ca91970` + `54a59caa` + `dae85141` + `99e7f2af`). Env-gated trace writes one `frame=N renderTime=X bytes_per_frame=Y writes=Z` line per render to `<tmpdir>/kodax-render-trace-<pid>.log`; bench scaffold parametrizes viewport at the user's real SSH dimensions (148×43) and measures the full engine `onRender` pipeline with a mock stdout so `setCellAt` / `outputToScreen` / `diff` costs are real.
+  - **Phase A.1 — `ScreenBuilder` eliminates `setCellAt` O(N²) `cells.slice()`** (commit `25bf0f52`). New mutable builder pattern at `output-to-screen.ts:211`: original `setCellAt(screen, ...)` did `screen.cells.slice()` (full width×height ref copy) + `{...screen, cells}` per non-empty cell — on a 148×43 viewport with ~500 non-empty cells/frame that's ~3.18M element-copies + 500 fresh arrays + 500 fresh Screen objects per frame. `createScreenBuilder(width, height)` exposes O(1) `setCellAt` writes + one-shot `build()` that returns a frozen Screen; only the `outputToScreen` hot loop migrated, public `setCellAt(Screen, ...)` API preserved for tests + future immutable callers. **End-to-end bench delta** (148×43, `mainscreen-windowed-800` scenario): renderer p95 14.804ms → 3.095ms (-79%, 4.78× speedup). 193 substrate-ink tests + 7 new ScreenBuilder unit tests (byte-equal vs `setCellAt`, OOB rejection, post-build-write rejection, 10k-write soft budget) + last-write-wins test `1105a181` close the review loop. 1426/1427 full repl PASS.
+  - **Phase A.2-E deferred pending user SSH trace measurement after A.1 ship.** ADR-028 documents the full claudecode port plan (Phase B nodeCache + markDirty / Phase C screen.damage bounding box / Phase D Output.charCache + StylePool / Phase E FRAME_INTERVAL + viewport culling). Layer 0 G1 (transcript render goldens, `925a4d77`) + G2 (perf bench + baseline, `4641ebb9`) + G4 (hit-test + selection 22 edge tests, `4fb590f3`) shipped as Phase 0 planning artifacts; ADR-027 + ADR-028 + `docs/test-guides/FEATURE_172_v0.7.41_REGRESSION_GUIDE.md` document the full pipeline.
+- **First-round TTFB compression — drop `refresh:true` tax + parallel pre-LLM + REPL-mount prewarm** (commit `e8b336ed`). Triggered by user observation: review-type prompts on a medium repo paid ~24s pre-LLM wall-time (after parallel/memoize work) before any LLM token streamed. Compressed via L1+L2 to ~10-15s (LLM-TTFB-bound). 5 stacked changes:
+  - **L1 — `middleware/repo-intelligence.ts` first-round NEVER forces `refresh:true`**. 4 sites of `refresh: isNewSession` → `refresh: false`. The 30s `PREMIUM_REFRESH_TIMEOUT_MS` budget was paid on every new session, but the daemon's own background polling keeps its on-disk state fresh; the 4s budget path returns daemon's already-cached state immediately. Single biggest savings (~10-15s).
+  - **L2 — REPL-mount prewarm** (new `prewarmRepoIntelligenceCaches` helper exported from `@kodax-ai/coding` + Ink-REPL `useEffect`). Fires `getRepoRoutingSignals` + `getRepoPreturnBundle` with refresh:false at REPL mount, fire-and-forget. Cache-coherent with L1 (both refresh:false) so user-path either coalesces onto in-flight prewarm Promise (~2s) or hits warmed P3+ cache (~0ms). Default-on; opt-out via `KODAX_PREWARM_REPO_INTELLIGENCE=0`.
+  - **P1.a — middleware parallel fan-out**. Two-phase `Promise.all`: Phase 1 races OSS overview (git+fs) with premium preturn (daemon); Phase 2 races module + impact direct-call fallbacks ONLY for slots not already filled by preturn. Behavioral pins preserved (preturn gating + `.catch(() => null)` error isolation + emit order: preturn → module → impact).
+  - **P1.b — run-substrate parallel**. `hydrateSession` (MCP state restore) and `getRepoRoutingSignals` collapsed to one wall-time slot via `Promise.all`; hydration error propagation unchanged, routing has independent `.catch(() => null)`.
+  - **P2 / P3 / P3+ — multi-tier cache stack**. P2 in-flight Promise sharing in `tryPremiumPreturn` (1.5s TTL, cacheKey DELIBERATELY includes `refresh` so explicit `refresh:true` callers — `/repointel warm`, eval harness — get their own daemon work). P3/P3+ session-scoped caches (60s TTL on routing signals + preturn bundle, cacheKey OMITS refresh so prewarm + first-round share one entry under the "data within 60s is fresh by definition" semantics). `normalizeCachePath` helper makes cacheKey robust to Windows drive-letter case + relative-vs-absolute caller variations + Promise rejection paths.
+  - **Default repo-intelligence mode preserved as `'auto'`**. Briefly experimented with flipping default to `'oss'` for users without repointel; cost analysis showed `'auto'` fallback path is ~10ms localhost TCP RST + 2s `PREMIUM_FAILURE_TTL_MS` cache → 0ms within TTL + ~5-10ms per >2s gap (negligible vs LLM TTFB). Auto-detection of installed repointel is the right default per README:182.
+- **Inline spinner-row stats tail — elapsed + tokens (claudecode parity)** (commit `58682cbf`). REPL spinner row gains an inline `Xs · Y tokens` running tail (matches claudecode's status indicator). Frontline of a sequence of claudecode-parity surface improvements; documented in ADR-027 Phase 0.
+
+### Internal / architecture
+
+- **FEATURE_171 — `runner-driven.ts` modular split**. R1 `2fef1c31` (4 leaf modules + `types.ts`) → R2 `f0be2d4e` (4 mid-coupling modules) → R3 `bfb2b818` (agent-chain + llm-adapter) → R4 `62dc1c58` (payload-builder + checkpoint-flow). 12 submodule extraction; **6406 → 1897 lines (-70.4%)**; **zero behavior change**; **4 reviewer APPROVE rounds**; **4314/4314 tests pass each commit**. ADR-026 + HLD.md §3.5.1 documented in R5 (`4d108af9`). The refactor preserves the closure pattern around `baseCtx` / `siblingSnapshot` / `contextTokenSnapshotRef` — what was a 6400-line monolith is now a stack of named factories each under 800 lines. Module map: `types.ts`, `agent-chain.ts`, `payload-builder.ts`, `checkpoint-flow.ts`, `llm-adapter.ts`, `compaction-bridge.ts`, `manager-input-builder.ts`, `result-projection.ts`, `tool-ctx-builder.ts`, `child-task-orchestration.ts`, `recorder-bridge.ts`, plus the residual `runner-driven.ts` entry. Side benefit: faster IDE hover-pop on the public surface; the public export shape is unchanged so all consumers are byte-equivalent.
+- **`bootstrapTeamMode` + `TeamModeHandle` exports added to `@kodax-ai/agent`** so the Ink REPL can import them without depending on legacy-CLI internals. The handle exposes `shutdown()` and is opaque otherwise (per the layer-independence guarantee — REPL has no business poking at the per-instance writer's internals).
+- **`KodaXToolExecutionContext.contentHashCache?`** field added with `recordRead` / `checkStale` / `recordWrite` API surface. Per-task lifetime (created at task start, destroyed at completion). Wired into Read / Edit / Write / MultiEdit tool implementations so the FEATURE_125 race-detection works without per-tool plumbing.
+- **`KodaXToolExecutionContext.siblingSnapshot?`** field added (as a mutable ref) with `Object.defineProperty` getter on the tool ctx so each tool invocation reads the freshest snapshot from the runner-driven adapter's per-round refresh. Avoids stale-snapshot reads when the LLM stream spans multiple seconds.
+- **`buildToolNamePatterns` extended from 4 to 9 syntax variants** in the benchmark harness regex tooling (`benchmark/datasets/feature-125-team-mode-awareness/cases.ts` + downstream). Captures kimi `read:0>{...}`, zhipu `<tool_name>read</tool_name>` and 3 other non-canonical syntaxes; lesson saved as `feedback_regex_audit_per_new_eval`.
+- **`JudgeContext.toolCalls?`** plumbed through `benchmark/harness/judges.ts` + both call sites in `benchmark/harness/harness.ts`. Optional `judge(output, context?)` arg lets binding-only providers (zhipu/glm51, mmx/m27, etc. — they emit `text=""` and put the tool call in the structured `tool_calls` field) be judged on what the harness actually captured, not on the empty raw text. Existing text-only judges ignore the arg and continue to work unchanged. Per `feedback_audit_must_see_binding` + `feedback_audit_binding_priority_in_prompt`: also requires the audit judge prompt to label the binding as "ABSOLUTE GROUND TRUTH" + a `CRITICAL RULE` system prompt section, or judges over-anchor on the empty raw text.
+- **2 prompt-eval datasets** added under `benchmark/datasets/`: `feature-125-team-mode-awareness/` (S7 Layer 2 panel: peer-active-file-acknowledge-read-first + peer-recently-modified-reread) and `tool-schema-slim/` (Layer 2 eval of v2_slim ~half + v3_aggressive ~quarter description variants for `ask_user_question` + `todo_create` — see "Tool schema slim eval" below).
+- **Tool schema slim eval (DEFER both v2 + v3)**. Commit `d68141ea`. Designed + ran the largest two-Scout-tool slim attempt: `ask_user_question` (2760 B / ~690 tok) + `todo_create` (2384 B / ~596 tok) — combined ~785–990 tokens potentially saved. 4-alias panel × 9 cases × 5 runs + panel-internal majority audit (initial 85–97% disagreement on AUQ_6 / 18–30% on TC_1 fixed by switching to v2 `CRITICAL RULE` prompt → 0% disagreement, data validated). Both variants **DEFER**: v2 gate (a) violations AUQ_1 zhipu −20pp + TC_1 zhipu/ds/kimi −20 to −40pp; v3 gate (a) violations AUQ_1 zhipu/ds −20 to −40pp + TC_1 zhipu/ds −40pp. Reason: `"For X use Y, NOT Z"` comparative clauses in schema descriptions are load-bearing disambiguation priors — slimming caused zhipu/ds to mis-classify simple cases. Pattern matches existing `feedback_simplifying_prompt_can_regress` + `feedback_model_structural_floor_not_prompt_tunable`. Future schema-slim work: don't touch "use X for ... NOT for ..." clauses; safe to slim version prefixes + return-value descriptions + "use sparingly" style instructions + property description secondary detail. Net cost ~$23 within ~$27 budget.
+
+### Test coverage delta
+
+- New: 17 (S1) + 20 (S2) + 20 (S3) + 15 (S4) + 15 (S5) + 7 (S6 integration) + 4 (W1) + 14 (W2) + 10 (W3) + 9 (W4) = 131 FEATURE_125 tests; 9 (FEATURE_165) + 7 (FEATURE_166) + 29 (FEATURE_167) + ~30 (FEATURE_170 C1–C6 follow-ups) = ~75 reliability tests; 50 (FEATURE_171 R4 tool wiring contract) + 0 net new for R1–R3 (all R-series ran the full pre-existing 4314 each round); ~95 FEATURE_172 Phase 1 (65 transcript-layout helpers + 8 golden snapshot + 22 hit-test/selection edge) + 17 React.memo comparator + 7 ScreenBuilder + 1 last-write-wins + 1 KODAX_RENDER_TRACE = ~120 FEATURE_172 tests; 4 cache-coalesce regression tests for the TTFB stack (P2 in-flight, P3 cross-call, P3+ multi-round, refresh:true-within-TTL).
+- Total green at HEAD: **5,081 tests pass + 23 todo + 1 skipped across 8 workspaces** (agent 477 / coding 2712 / llm 276 / mcp 28 / repl 1419 / repo-intel & skills 136 / repointel-protocol & session-lineage 18 / tracing 15). `tsc -p packages/coding/tsconfig.json --noEmit` + `tsc -p packages/repl/tsconfig.json --noEmit` both clean.
+
+## [0.7.40] - 2026-05-13
+
+### Theme
+
+**Envelope Spillover + Vision Bridge** — Two parallel-developed features close gaps in the child-agent communication path (FEATURE_121) and the REPL input path (FEATURE_134). FEATURE_121 routes child task summaries through the existing `tool-result-policy.ts` spillover system (50KB per-banner + 200KB envelope aggregate cap, mirror of claudecode's `MAX_TOOL_RESULTS_PER_MESSAGE_CHARS`), removing two prior hard truncations (`orchestration.ts:1033` 1600-char slice + `dispatch-child-tasks.ts:256` 200-char slice) that were silently losing 95%+ of child output bytes. A follow-up LLM blob summarizer fallback handles the residual `spillFailed + content > 100KB` data-loss edge that the main slice's inline fallback would otherwise resolve by blowing the Worker context window. FEATURE_134 adds 5 paste sources (bracketed paste / `@<path>` file refs / macOS Cmd+V auto-link / Windows Alt+V explicit / macOS-Linux Ctrl+V backup) on top of the existing `KodaXImageBlock` AI-layer vision serialization, completing the round-trip from screenshot in clipboard to multimodal `user` message at any of the 12 KodaX providers. A late P0 regression in REPL transcript rendering (transcript items invisible during agent execution under AMA mode on Windows ConPTY) was diagnosed as `useDeferredValue` starvation in Ink under Node.js (no DOM idle-scheduling bridge), fixed by removing the deferral indirection — the 200-item UUID-anchored cap from FEATURE_060 Tier 2 retains the perf protection that was the actual fix for SSH-resume O(N) blow-up.
+
+### Added
+
+- **FEATURE_121 — Envelope Spillover Gap-Fix + LLM Blob Summarizer Fallback**. Two-commit landing: main slice (`0a0f844e`) + follow-up LLM blob summarizer (`ba0c82f9` + review fixes `05259ab2`). Removes the `orchestration.ts:1033` 1600-char `truncateText` + `dispatch-child-tasks.ts:256` 200-char `slice` two-layer hard truncation that silently dropped 95%+ of child task output (a 25KB audit report reached the Worker as ~50 tokens). Routes every `<task-completed>` banner through `applyToolResultGuardrail('child_task_summary', ...)` — 50KB head + spill-to-file under `getAgentConfigPath('tool-results')/<id>.txt`, banner now carries a preview + spill path that the Worker reads via standard Read tool. `composeIdleYieldUserMessage` in `@kodax/agent/orchestration/idle-yield.ts` gains a 200_000-chars aggregate envelope cap (mirror of claudecode's `MAX_TOOL_RESULTS_PER_MESSAGE_CHARS=200_000`) — when N banners individually fit but together exceed envelope budget, the enforcer calls `applyToolResultGuardrail(..., { forceSpill: true })` to reclaim space. Capability sections include a `LARGE CHILD OUTPUT (FEATURE_121 v0.7.40)` block teaching the Worker the spillover-path Read pattern. 4-alias × 3-case × 5-runs Layer 2 eval clears PARTIAL SHIP (3/4 aliases ≥80% on each case; mmx/m27 weak floor 60-80% documented in test guide). **Follow-up — LLM Blob Summarizer (last-resort fallback)**: when `persistToolOutput` fails (ENOSPC / EACCES / EROFS / SELinux denial) AND raw content > `LARGE_CONTENT_THRESHOLD_BYTES` (100KB), `dispatch-child-tasks` now calls `ctx.summarizeBlob(content, {maxChars: 8000})` to compress to a ~2-8KB lossy summary preserving file paths / line numbers / error codes / identifiers / findings verbatim, banner-wrapped with `[SPILL FAILED — original ${size} compressed via LLM summarizer; raw content unavailable. Worker: treat this summary as LOSSY...]`. The summarizer callback is injected into `KodaXToolExecutionContext` via lazy-once memoization in `runner-driven.ts` bound to the Worker's own provider/model — layer-independent (`@kodax/agent` stays unaware of LLM client). If the summarizer itself fails (provider error / abort), falls back to inline full content with an emergency banner `[SPILL FAILED AND LLM SUMMARIZER FAILED — original ${size} inlined as last-resort emergency dump...]` so the Worker never silently receives oversized opaque content. Honors the FEATURE_121 contract: **silent data loss is the worst outcome; over-budget but observable is acceptable**. 4-alias × 2-case × 3-runs Layer 2 eval clears SHIP gate (4/4 aliases × 100% retention on audit_report + grep_findings cases). Test guide: `docs/test-guides/FEATURE_121_v0.7.40_TEST_GUIDE.md`. Design doc: `docs/features/v0.7.40.md#feature_121-envelope-spillover-gap-fix--child-task-summary-接入-tool-result-policy`.
+- **FEATURE_159 — MessageQueue as Single Source of Truth + Idle-Yield Mode-Split**. Main commit `948b8879` (Phase 3 mode-split synthetic + unified queued-followup predicate) + design `daa8e846` + follow-up `9d4c6ae4` (queue filter scope + verdict.summary echo) + test-isolation `29369a2a` (MessageQueue test isolation + compaction flake fix). User-reported during v0.7.40 RC: after Worker dispatched 3 child tasks and the main agent was still tool-calling, user typed a follow-up "你是派出了子Agent再做嘛？" — status bar showed `Queue 1` / `Queued follow-ups: 1`, but the prompt looked **swallowed** when the agent finished its investigation (not folded into the answer). Forensic showed two stacked failures: (1) `waitForWakeEvent` consumed messages via `MessageQueue.dequeue()` with no reverse notification to REPL, so the React `state.pendingInputs` retained stale entries and the `Queue N` indicator never cleared; (2) drained prompts were marked `_synthetic: true` and hidden from transcript — the user prompt appeared dropped even though it had landed in the model context. Fix flips the substrate: `MessageQueue` becomes the **single source of truth** with typed event subscribe + frozen snapshot read + `mode`/`id`/`predicate` filtering APIs; REPL reverses its sync direction (queue→React mirror via `subscribe`); `composeIdleYieldUserMessage` branches on `msg.mode` — real user prompts emit as **non-synthetic** user-bubble messages (visible in transcript), task-notifications stay synthetic (silent background framing). Net result: claudecode-parity messageQueueManager semantics + SDK-grade observable substrate, without importing claudecode's `commandLifecycle` module, `recordQueueOperation` file sink, or 3-tier priority (kept KodaX's binary user/background). Net −70 LoC, zero new modules. Test guide: `docs/test-guides/FEATURE_159_v0.7.40_TEST_GUIDE.md`. Design doc: `docs/features/v0.7.40.md#feature_159-messagequeue-as-single-source-of-truth--idle-yield-mode-split`.
+- **FEATURE_134 — Image / Screenshot Paste Input (REPL Vision Bridge)**. Commit `2e9674bb`. Adds the REPL paste entry point on top of KodaX's existing `KodaXImageBlock` AI-layer vision serialization (already implemented in `packages/ai/src/providers/anthropic.ts:770` + `openai.ts:904` + `image-serialization.ts`). 5 paste sources per claudecode `usePasteHandler.ts` parity: (1) bracketed paste base infra (DEC 2004 `ESC[200~/201~` wrapping — note FEATURE_134's redundant `enableBracketedPasteMode()` pre-render write was removed in `ca009b3a` since `KeypressContext.tsx:134` already owns DEC 2004 lifecycle via Ink's managed stdout); (2) file path paste — `extractImagePaths` splits on `/` or `[A-Za-z]:\\` + extension; (3) macOS Cmd+V auto-link via `osascript NSPasteboard` read; (4) Windows Alt+V explicit keybind (Ctrl+V is system-paste-reserved on Windows, same as claudecode); (5) macOS/Linux Ctrl+V backup via `wl-paste` / `xclip` / `osascript`. Building blocks in `packages/repl/src/paste/`: `bracketed-paste-mode.ts` (DEC 2004 lifecycle), `image-normalize.ts` (jimp decode + clamp 2000px + PNG→JPEG quality ladder 80/60/40 to fit 3.75MB), `clipboard-image.ts` (cross-platform reader, never throws), `persist-image.ts` (writes to `$TMPDIR/kodax-paste/` + returns path-based `KodaXImageBlock`), `paste-handler.ts` (5-source orchestrator). Library choice: jimp (~10MB pure JS) over sharp (~30MB native binary) — jimp install never fails on cross-platform CI; claudecode picked sharp for broader use cases but KodaX only paste-uses, so jimp is the simpler fit. Integration: paste event → `@<resolved-path>` text-token translation via existing `common/input-artifacts.ts:preparePromptInputArtifacts` pipeline which already converts `@<path>` refs to `KodaXInputArtifact[]` on submit — no changes needed downstream. Test guide: `docs/test-guides/FEATURE_134_v0.7.40_TEST_GUIDE.md`. Design doc: `docs/features/v0.7.40.md#feature_134-image--screenshot-paste-input--repl-vision-bridge`.
+
+### Fixed
+
+- **REPL transcript rendering starvation — `useDeferredValue` removed from `displayHistory` chain**. Commit `ca009b3a`. User-reported P0 (AMA mode, Windows ConPTY): pressing Enter on a query showed nothing in the transcript for the entire agent run — header banner + TodoListSurface + spinner + status bar rendered normally, but assistant thinking blocks / tool calls / tool results / assistant text remained invisible until task completion forced a re-render. Root cause: `useDeferredValue(displayHistory)` (introduced in v0.7.30 FEATURE_060 Tier 2 as a polish on top of the 200-item cap that was the real perf fix) marks transcript rebuild as low-priority React work. Under React DOM, this work flushes during browser idle time via the scheduler's idle-callback bridge. Ink uses react-reconciler without that bridge — under Node.js, high-priority work (spinner ticks @30fps, streaming setState bursts, tool-state updates) perpetually pre-empts the deferred work, so the low-priority track never flushes until a "big" state change like `setIsLoading(false)` forces a sync re-render. v0.7.40 surfaced the latent starvation: FEATURE_121's envelope spillover replaced `slice(0, 200)` child task truncation with up-to-50KB head + spill content, raising per-item `buildPromptSurfaceItems` cost from ~0.1ms (200 chars) to ~10-50ms (50KB) — single-item cost crossed the React scheduler's low-priority deferral threshold, starving the deferred update perpetually under high-frequency setState bursts. Fix: replace `useDeferredValue(displayHistory)` with direct passthrough. The 200-item UUID-anchored cap (the real perf protection from FEATURE_060 Tier 2) and the transcript-mode 30-message cap remain — together with React's built-in `useMemo` memoization (which prevents spinner ticks from triggering `buildPromptSurfaceItems` when history is unchanged), per-render cost stays bounded at O(min(N, 200)). Trade-off: long-session `kodax -c` resume on Windows-SSH may add ~10-40ms one-time first-paint cost — well below human perception threshold (~100ms). Length-thresholded fallback pattern documented inline (`displayHistory.length > 100 ? lazyDeferred : displayHistory`) for future repro.
+- **FEATURE_121 v0.7.40 follow-up build break + memoize + emergency banner**. Commit `05259ab2`. Fixes three review-uncovered issues from `ba0c82f9`: (1) **CRITICAL build break** — `blob-summarizer.ts` imported `KodaXProvider` (not exported from `@kodax-ai/llm`); replaced with `KodaXBaseProvider` everywhere. `tsc --noEmit` missed this; `tsc -b` with declaration emission caught it during `npm run build:packages`. (2) **HIGH memoize** — `runner-driven.ts` `summarizeBlob` was rebuilding the factory closure (including `resolveProvider`) on every call; changed to lazy-once memoize (cached on first invocation, reused for the rest of the Worker run). (3) **MEDIUM emergency banner** — when LLM summarizer ITSELF fails AND we fall back to inline 100KB+ content, the Worker received the raw blob with no banner; now prepends `[SPILL FAILED AND LLM SUMMARIZER FAILED — original ${size} inlined as last-resort emergency dump]` so the Worker sees a clear signal to expect possible downstream truncation and re-run upstream with narrower scope.
+- **`fix(build,v0.7.39): make 'npm run build' produce shippable dist/`** (commit `b77fa0a3`). Footgun caught during clean-room publish dry-run: `npm run build` previously ran `npm run build:packages && tsc` — the trailing bare `tsc` read root `tsconfig.json` (`outDir: ./dist`, `declaration: true`) and overwrote the esbuild bundle with unbundled `tsc` output containing bare `import '@kodax-ai/coding'` specifiers, which would have shipped a broken tarball — consumers running `npm install @kodax-ai/kodax` would hit `ERR_MODULE_NOT_FOUND`. Fix: `"build": "npm run build:packages && npm run build:bundle && tsc --emitDeclarationOnly"`. The `--emitDeclarationOnly` flag is the critical guard — tsc skips `.js` entirely and only writes `.d.ts` on top of the esbuild bundle. Now any caller (developer, CI, `release.mjs`) can run `npm run build` at any time and get a complete, shippable dist/. Side benefit: SDK consumers get real TypeScript types for all 5 subpath entries (was untyped before).
+- **`fix(release,v0.7.40): bake KODAX_VERSION into bundle via esbuild --define`** (commit `b70048b7`). The previous v0.7.39 release bundled with `process.env.KODAX_VERSION` left unresolved, so the runtime version was `undefined` in the published `dist/kodax_cli.js` — visible to users as `KodaX undefined` in the banner. Fix: `scripts/build-bundle.mjs` reads root `package.json` at build time and injects via esbuild `--define`, so the bundled CLI carries the correct version literal. `release.mjs` flow unchanged: bumps `package.json` first, then `npm run build` (which reads the bumped version).
+- **FEATURE_134 follow-up: Alt+V duplicate-image-file accumulation in temp dir**. User-reported regression during v0.7.40 RC validation: pressing Alt+V on the same screenshot created many identical-content files with different UUID names under `$TMPDIR/kodax-paste/`. Two root causes: (1) `prompt-input-controller.ts:triggerExplicitImagePaste` had no single-flight guard — OS-level key autorepeat (Windows ConPTY in particular) fires multiple Alt+V keypresses on a brief hold, each spawning a concurrent `readClipboardImage` + `persistImageAsBlock` pair; (2) `persist-image.ts` used `randomUUID()` filenames so even identical buffer content wrote N distinct files. Fix in [packages/repl/src/paste/persist-image.ts](packages/repl/src/paste/persist-image.ts): filename derived from `sha256(buffer).slice(0,16)` — identical content reuses one path, `writeFile` is idempotent on rewrite. Fix in [packages/repl/src/ui/utils/prompt-input-controller.ts](packages/repl/src/ui/utils/prompt-input-controller.ts): `explicitImagePasteInflightRef` boolean ref guard drops re-entrant invocations until the in-flight clipboard read settles. New tests: `persist-image.test.ts` adds "reuses the same path for identical content" + "produces distinct paths for different content"; `prompt-input-controller.test.ts` adds "Alt+V autorepeat fires the clipboard read only once (single-flight guard)". Together the two fixes cap temp-dir growth per unique screenshot at one file per session.
+- **FEATURE_134 follow-up: Gemini-CLI vision via `@<path>` token injection** (commit `71d45783`). The ACP base class (`packages/ai/src/providers/acp-base.ts`) was silently dropping image blocks at the prompt flatten step (`.filter(b => b.type === 'text')`) before forwarding to the CLI bridge. New extension point `serializeImageBlockToPromptToken(block)` on `KodaXAcpProvider` defaults to `null` (preserves silent-drop for Codex-CLI which has no `codex exec --json` image surface), and `KodaXGeminiCliProvider` overrides to return `@<absolutePath>` — Gemini CLI 2.x's file-include syntax inlines any readable file (including images) into the model context. The CLI-bridge capability profile gets a new sibling `IMAGE_INPUT_CLI_BRIDGE_PROVIDER_CAPABILITY_PROFILE` so the policy gate sees the right metadata. 2 new tests in `cli-bridge-providers.test.ts` pin the exact wire format for both Gemini (`@<path>`) and Codex (null).
+- **FEATURE_134 follow-up: paste tmp dir age-based GC at REPL bootstrap** (commit `0eb6cbb4`). Per-session content-hash dedup (commit `12589a46`) prevented Alt+V autorepeat from stacking duplicates within one session, but cross-session accumulation was still unbounded — files only cleared on OS tmpdir reboot. New `prunePasteTmpDir(now?)` exported from `packages/repl/src/paste/persist-image.ts` deletes `paste-*` files older than `PASTE_TMP_TTL_MS` (24h). Non-paste files in the same dir survive (e.g. user-dropped `notes.txt`); per-file errors swallowed so concurrent KodaX instance races don't break REPL startup. Wired in `InkREPL.tsx` bootstrap as a fire-and-forget dynamic import — never blocks first paint. Active-session paste files (always within TTL) are preserved across overnight idle. 4 new tests cover dir-not-exist / age-based deletion / non-paste preservation / TTL boundary.
+- **FEATURE_134 follow-up: custom provider vision opt-in documented** (commit `db36dc69`). README.md + README_CN.md gain a complete `customProviders` example showing the `capabilityProfile.multimodalSupport: 'image-input'` nested shape. Built-in 12 providers ship with the flag already enabled; only custom providers needed opt-in instructions because the field was implicitly available via `KodaXCustomProviderConfig.capabilityProfile` but undocumented.
+- **AMA in-turn compaction parity gap — three-phase lifecycle restored (microcompact + snapshot-aware trigger + graceful fallback)**. User-reported during v0.7.40 RC validation: status bar climbed from 124k → 138k → 150k across three AMA rounds with zhipu-coding/glm-5.1, no compaction / microcompaction / graceful prune ever firing despite crossing the 60% × 200k = 120k threshold. Latent since FEATURE_114 v0.7.36 (4→2 role consolidation) but masked by FEATURE_076's over-aggressive round-exit reshape collapsing context to ~1k after every round; once the v0.7.40 reshape fix above let context grow naturally, the gap surfaced. Three structural deltas vs SA path's `runCompactionLifecycle` ([compaction-orchestration.ts:358](packages/coding/src/agent-runtime/middleware/compaction-orchestration.ts#L358)): (1) **trigger metric mismatch** — the AMA `compactionHook` ([_internal/managed-task/compaction.ts](packages/coding/src/task-engine/_internal/managed-task/compaction.ts)) compared `estimateTokens(transcript)` to the threshold, but transcript-only estimate excludes the system prompt + tools schema (Worker role prompt + AGENTS.md + REPO INTELLIGENCE TOOLS teaching + repo-intel capsule + 12 tool definitions ≈ 20–35k tokens after FEATURE_114 worker consolidation + FEATURE_161 prompt growth). 200k-window 60% trigger needs API total > 120k, but the hook saw ~95–115k transcript and never fired — the status bar's 138k was the real total (system + tools + transcript) sourced from `streamResult.usage.totalTokens`. (2) **no microcompact phase** — SA path runs `microcompact(messages, DEFAULT_MICROCOMPACTION_CONFIG)` every turn at zero LLM cost ([run-substrate.ts:603](packages/coding/src/agent-runtime/run-substrate.ts#L603)) to prune old tool_results / image blocks past `maxAge=20` turns; AMA hook bypassed it entirely. (3) **no graceful degradation fallback** — SA path's third phase `applyGracefulDegradationGate` ([compaction-orchestration.ts:250](packages/coding/src/agent-runtime/middleware/compaction-orchestration.ts#L250)) deterministically prunes tool_results when LLM compact threw / returned `compacted: false` / left context still above `triggerTokens × pruningGapRatio`; AMA hook bailed silently on LLM failure, letting context grow unbounded. Fix in [packages/coding/src/task-engine/_internal/managed-task/compaction.ts](packages/coding/src/task-engine/_internal/managed-task/compaction.ts): hook now mirrors SA path's three-phase lifecycle — (Phase 1) `microcompact` every call (free, prunes old tool blocks); (Phase 2) `intelligentCompact` LLM summary with snapshot-aware trigger check `resolveContextTokenCount(transcript, snapshot)` instead of raw `estimateTokens` — `snapshot.currentTokens` carries the LAST LLM call's API-reported `usage.totalTokens` (system + tools + transcript) so the threshold check uses the same metric the status bar displays; the snapshot is refreshed by the LLM adapter ([runner-driven.ts](packages/coding/src/task-engine/runner-driven.ts) `buildRunnerLlmAdapter`) after every stream completion via the new `contextTokenSnapshotRef` shared between adapter (writer) and hook (reader); (Phase 3) `gracefulCompactDegradation` deterministic prune fallback when LLM compact failed / partial / circuit-breaker-tripped — the gate's "still over" check also uses snapshot-aware accounting for symmetry with Phase 2. Snapshot rebases to `createEstimatedContextTokenSnapshot(compacted)` after any compaction (LLM or graceful) so subsequent delta corrections start from the compacted baseline rather than the stale pre-compaction API total. New test file [packages/coding/src/task-engine/_internal/managed-task/compaction.test.ts](packages/coding/src/task-engine/_internal/managed-task/compaction.test.ts) pins all three phases with 9 tests covering: microcompact-only return path (below trigger); microcompact identity (no diff returns undefined); snapshot-aware trigger fires when transcript-only estimate is sub-threshold but API-total crosses; no-snapshot cold-start falls back to transcript estimate (unchanged baseline); graceful fallback on LLM throw / `compacted: false` / partial success above gap-ratio; snapshot rebase after successful LLM compaction; circuit-breaker semantics (3 strikes skip LLM but graceful still fires). Side fix in [packages/coding/vitest.config.ts](packages/coding/vitest.config.ts) + [packages/repl/vitest.config.ts](packages/repl/vitest.config.ts): `@kodax-ai/llm` alias updated from stale `packages/ai/` to `packages/llm/` (rename leftover from v0.7.40 directory rename above). **Behaviour delta vs pre-fix**: AMA mode compaction now triggers at the same effective API-total threshold as SA mode; graceful fallback prevents zhipu/kimi/mimo provider-side LLM summary failures from leading to monotonic context growth; microcompact's per-turn pruning recovers ~5–15% of context per long-running Worker turn even without crossing the LLM trigger threshold. Full coding suite green: **2446 tests pass** (239 test files, 23 todo; +9 new compaction hook tests).
+- **FEATURE_076 follow-up: round-boundary reshape now preserves tool_use / tool_result chains across rounds (cache + re-read regression fix)**. User-reported during v0.7.40 release validation: status bar dropped from `121k/200k` to `1.1k/200k` after an AMA round, and follow-up rounds were re-reading files the prior worker had already read. Root cause: FEATURE_076 (v0.7.25)'s `reshapeToUserConversation` replaced `result.messages` wholesale with a synthetic `[user, assistant]` dialog at every round exit, discarding the entire `tool_use` / `tool_result` chain. The replacement was the correct fix for v0.7.25's actual problem (cross-round role-prompt pollution — Evaluator role prompts leaking into the next round's worker context as user messages) but over-corrected by stripping structurally useful content, with two concrete production costs: (1) **cross-round file re-reads** — round 2 worker had no visible `tool_result` for files round 1 already read, so common follow-ups like "now modify that file you reviewed" forced re-reads; (2) **provider prompt-cache miss on the dialog prefix** — round 1's first LLM call prefix was `[system, user, assistant(tool_use), user(tool_result), …]`, round 2's is `[system, user, assistant(final), user_2, …]`; the prefixes diverge immediately after the first user message, so the dialog portion gets zero cache reuse across rounds. Fix in [packages/coding/src/task-engine/_internal/round-boundary.ts](packages/coding/src/task-engine/_internal/round-boundary.ts): new `preserveTranscriptForRoundExit` helper runs a 4-step pipeline — (Step 1) strip the leading stale role-prompt system message (Runner.run leaves the last-active agent's role prompt at `transcript[0]`; round 2's entry agent injects its own at position 0, so keeping the previous one would create two conflicting `system` instructions back-to-back); CompactionSummary system messages are preserved via the now-exported `COMPACTION_SUMMARY_PREFIX` discriminator from `@kodax-ai/session-lineage`. (Step 2) Apply `normalizeLoadedSessionMessages` to strip V1-legacy role-prompt-wrapped trailing `{user, assistant}` pairs (`"You are the Evaluator role..."` phrased as a user message); no-op for V2 AMA where role prompts are system-message-shaped. (Step 3) Ensure the round's user prompt is observable — V2 sessions retain it via `runnerInput`, V1 paths may have lost it when normalisation stripped the wrapper. (Step 4) Ensure the transcript ends with a plain-text assistant carrying the sanitised final answer; **replaces** (not appends) when the last message is an assistant with array content (typically `emit_verdict` / `emit_handoff` tool_use blocks — KodaX protocol machinery whose user-facing payload is captured in `result.lastText`). The replace avoids two consecutive `role: 'assistant'` messages which Anthropic's API rejects on the next request. Net behaviour: round 2 worker sees what round 1 read/edited, prompt-cache prefix stays continuous, status bar reflects the actual context size (e.g. ~50k instead of 1.1k after a heavy worker round) — a side-effect users will perceive as the bar "no longer collapsing after each turn". Updated [packages/coding/src/task-engine/_internal/round-boundary.test.ts](packages/coding/src/task-engine/_internal/round-boundary.test.ts) with 7 new cases covering: V2 worker-shape preservation, terminal `emit_verdict` tool_use replacement (no consecutive-assistant violation), `CompactionSummary` system message preservation, dedup of already-correct trailing assistant, empty `result.messages`, compaction-system-only transcript, terminal `thinking`-only assistant. `COMPACTION_SUMMARY_PREFIX` exported from [packages/session-lineage/src/compaction/compaction.ts](packages/session-lineage/src/compaction/compaction.ts) so the producer-consumer pair no longer maintains the literal in two places. Test guide: regression case lives in [docs/test-guides/FEATURE_134_v0.7.40_TEST_GUIDE.md](docs/test-guides/FEATURE_134_v0.7.40_TEST_GUIDE.md) once that doc's "round-boundary cache + re-read" section is appended. **Behaviour delta vs FEATURE_076 baseline**: the v0.7.25 cross-round-coherence guarantee is preserved (stale role prompts still get stripped); only the over-corrective full-transcript replacement is reverted. Full test suite green at HEAD: **5745 tests pass** (was 5742; +3 new edge-case tests).
+- **FEATURE_168 — AMA agent tool wiring (exclude-based, registry as source of truth)**. Commit `56330d1c` + doc `e902b194`. Root cause investigation surfaced from production trace 2026-05-15 (zhipu/glm51 Worker blocked at `emit_handoff` pending-children gate, told by the gate's own error message to call `task_stop`, model honestly responded that `task_stop` "is not registered as a callable tool" — and the model was correct). `runner-driven.ts::buildRunnerAgentChain` had used **include-mode hand-written `agent.tools` arrays** for all 5 AMA roles (Scout / Planner / Generator / Evaluator / Worker) since v0.7.26, while the SA path defaults to `listToolDefinitions()` minus `excludeTools`. The two paths drifted across three features: **FEATURE_120 v0.7.39** registered `send_message` + `task_stop`, taught them in the Worker prompt, gated `emit_handoff` with a "call `task_stop` first" error message — but never wired the `RunnableTool` instances into any AMA agent's tools array. **FEATURE_161 v0.7.40 prompt teaching** taught the Worker 8 repo-intel pull tools (`module_context` / `symbol_context` / `process_context` / `impact_estimate` + 4 shallow ones), `worker-role-prompt.ts:127` comment claimed "the 8 pull tools get stripped from the LLM-visible tool list (see agent-runtime/tool-resolution.ts)", but AMA path doesn't go through `tool-resolution.ts` — it reads `agent.tools` directly, so only 4 of the 8 ever landed in the schema. **Web tools / ask_user_question / worktree_create / worktree_remove / insert_after_anchor / undo**: registered + `permission` system gave them entries, but no AMA agent could see them. Total **17 registered tools silently dropped** from production AMA; no test layer caught it because no test asserted "agent.tools actually contains a schema entry with this name". Fix: switch AMA path to exclude-based wiring matching SA semantics. New `buildAgentToolsFromRegistry(role, ctx, budget, events, overrides)` helper enumerates `listToolDefinitions()` and applies `AMA_BASELINE_EXCLUDE ∪ <ROLE>_EXTRA_EXCLUDE`. Role-specific wraps (mutation-guarded bash/write/edit/multi_edit for Generator/Worker, read-only bash for Evaluator, `dispatch_child_task` per-role drain wrappers, FEATURE_097 throttle-aware `todo_update`) flow through the `overrides` map. Evaluator security boundary made **architectural, not prompt-dependent**: write/edit/multi_edit/insert_after_anchor/undo/dispatch_child_task/send_message/task_stop/worktree_create/worktree_remove/exit_plan_mode/todo_update/ask_user_question all hard-excluded from `EVALUATOR_EXTRA_EXCLUDE` — a prompt-jailbroken or tool-confused Evaluator is physically unable to mutate, dispatch, or change plan state. Planner kept as read-only inspection role (no bash, no mutation, no dispatch, no user interaction). Scout/Worker carry the full execution surface. Generator unchanged behaviorally; new tools (web/repo-intel/coordinator) added to schema. New contract test [packages/coding/src/task-engine/runner-driven-tool-wiring.test.ts](packages/coding/src/task-engine/runner-driven-tool-wiring.test.ts) pins each role's full tool-name set against `getAmaRoleExpectedToolNames(role)` derivation plus spot-checks for FEATURE_120 / FEATURE_161 coverage / Evaluator boundary / Planner boundary / no-orphan invariant — 50 assertions total. Any future EXCLUDE-set change or registry addition surfaces as a concrete test failure, not a silent production schema gap. **Test impact**: 2507/2507 coding-package tests pass (240 files, +50 new contract assertions). No regression in `runner-driven.test.ts` (130 tests including FEATURE_165 race-regression), `child-executor.test.ts` (30), `task-stop` / `send-message` handler tests (28), `worker-role-prompt.test.ts`. Design doc: [docs/features/v0.7.40.md#feature_168--ama-agent-tool-wiringexclude-based-registry-as-source-of-truth](docs/features/v0.7.40.md#feature_168--ama-agent-tool-wiringexclude-based-registry-as-source-of-truth).
+- **FEATURE_169 — Pull-Tool Prompt Adoption Hardening** (commit `519af4b9`). Production trace after FEATURE_161 wiring + FEATURE_168 schema fix surfaced 3 residual adoption gaps the wiring alone did not close: (1) **Worker hand-feeding bash in `dispatch_child_task.objective`** — 18% of production dispatch (3/17 in 2026-05-15 audit) embedded literal `git diff v0.7.39..HEAD`-style command directives, overriding the child's prompt-side tool teaching; 0/17 objectives recommended a pull-tool family. (2) **Child agent prompt stayed read+grep-first** — `CHILD_AGENT_SYSTEM_PROMPT` had taught "3-8 parallel tool calls (glob + grep + key file reads)" since v0.7.18 with no mention of pull-tools; children defaulted to grep/read in review and exploration tasks. (3) **Worker self-review still picked `bash git diff` first** — F7 taught tool existence but not "for review tasks, use which one". Three localized prompt strengthenings (F0a / F0b / F1v2 / F3) ship; F2 (3-tier order injection) **rejected** post-eval as zero-value churn. F0a teaches Worker to keep dispatch objective as data ("scope: v0.7.39..HEAD") not command ("使用 `git diff v0.7.39..HEAD`"); F0b adds an explicit pull-tool-family recommendation to the dispatch objective teaching; F1v2 adds child-agent reverse-steering toward pull-tools when the task is review/exploration-shaped; F3 reframes the Worker change-review surface from `bash git diff` first to `module_context` / `symbol_context` first. No wiring change, no new tool, no new permission. Prompt eval (`tests/feature-169-pull-tool-adoption.eval.ts`) clears SHIP gate (4/4 aliases, F0a/F0b/F1v2/F3 each ≥80% on 3 cases). Design doc: `docs/features/v0.7.40.md#feature_169--pull-tool-prompt-adoption-hardening-worker-dispatch-objective--child-reverse-steering--change-review-reframe`.
+- **FEATURE_134 follow-up: vision capability flag widened to 9 additional providers**. User empirically validated 2026-05-13 that `kimi-code` accepts and processes image input despite its v0.7.40 RC snapshot flag claiming `multimodalSupport: 'none'`. Root cause: AMA path (`runner-driven.ts` `provider.stream` direct call) bypasses the SA-path `applyProviderPolicyGate` (`run-substrate.ts:660`) where multimodal block enforcement lives, so the latent flag mismatch was never observable in production. The pre-v0.7.40 RC flag was over-conservative: every Anthropic-compat clone inherits the image-block forwarding serializer at `anthropic.ts:770`, and every OpenAI-compat clone inherits the `image_url` forwarding serializer at `openai.ts:904`. Flag widened in `packages/ai/src/providers/registry.ts` from `NATIVE_PROVIDER_CAPABILITY_PROFILE` to `IMAGE_INPUT_NATIVE_PROVIDER_CAPABILITY_PROFILE` for: **Anthropic-compat clones** — `kimi-code`, `zhipu-coding`, `mimo-coding`, `ark-coding`, `minimax-coding`; **OpenAI-compat clones** — `deepseek`, `kimi`, `qwen`, `zhipu`. Plus a separate follow-up commit (`71d45783`) wires **Gemini-CLI** vision via the new `serializeImageBlockToPromptToken` extension point on `KodaXAcpProvider` (Gemini CLI 2.x `@<path>` file-include syntax), bringing the final post-release total to **12 vision-capable providers (was 2)** — only `codex-cli` remains text-only because `codex exec --json --full-auto` has no image-input surface today. Production semantics: the flag controls only KodaX's SA-path artificial block; the actual model-level vision contract remains the upstream provider's responsibility — if a specific model alias is text-only, users now see the real API error from the provider instead of a KodaX-side `[Provider Policy] multimodal requests are unsupported` rejection. `capability-profile.test.ts` gains an explicit pin test asserting all 12 vision-capable providers report `multimodalSupport === 'image-input'` and `codex-cli` reports `'none'` so future regressions are caught.
+
+### Internal / architecture
+
+- **Directory rename: `packages/ai/` → `packages/llm/` for package-name parity**. Closes the directory-vs-npm-name discrepancy left behind by FEATURE_147 v0.7.37, which renamed the npm package from `@kodax/ai` to `@kodax-ai/llm` but left the on-disk directory at `packages/ai/`. New contributors had to mentally translate "ai is actually llm"; prompt examples and tool-description hints saying "Audit packages/ai" pointed to a path that disagreed with the package name. Scope: `git mv packages/ai packages/llm` + 8 tsconfig path/reference updates (root + agent + coding + mcp + session-lineage) + 2 real relative imports in `tests/feature-116-active-cache-control.eval.ts` + ~30 prompt/comment/eval-fixture string updates (`worker-role-prompt.ts` / `role-prompt.ts` / `tools/registry.ts` / `tools/todo-list.test.ts` / `paste/persist-image.ts` / 5 `tests/*.eval.ts` files / 5 `benchmark/datasets/*` fixture files including `h2-plan-execute-boundary/cases.ts` path-prefix `mustNotTouchFiles` assertions). Frozen historical artifacts NOT touched: `benchmark/results/**` snapshots, `.agent/**` runner cache, `.repointel/**` index cache, `docs/CHANGELOG_ARCHIVE.md` + `docs/features/v0.7.0-39.md` historical design records, and `.claude/settings.local.json` user-private allowlist. Active operational docs (CHANGELOG, README, README_CN, DD.md, KNOWN_ISSUES.md, root CLAUDE.md, docs/CLAUDE.md) all updated. No behavior change; git rename detection preserves blame across the move. Pure naming-consistency refactor.
+- **`@kodax/coding` blob summarizer module** (`packages/coding/src/tools/blob-summarizer.ts`, 187 lines). `createBlobSummarizer({provider, model, timeoutMs?})` returns a `SummarizeBlob` callback. Combines caller's `abortSignal` with a 30-second timeout via a fresh `AbortController` + listener fan-out. Throws `BlobSummarizerError` on empty input / empty output / provider error. System prompt + user prompt builder exported as constants (`SUMMARIZER_SYSTEM_PROMPT`, `buildSummarizerUserMessage`) so the Layer 2 eval can pin the EXACT production prompt text. 8 deterministic-shell unit tests cover the contract surface (timeout / abort / empty rejection / error wrapping / threshold constants).
+- **`KodaXToolExecutionContext.summarizeBlob?` field** (`packages/coding/src/types.ts`). Optional callback shape `(content: string, options?: {readonly maxChars?: number; readonly abortSignal?: AbortSignal}) => Promise<string>`. Wired in `runner-driven.ts` baseCtx with lazy-once memoization bound to the Worker's own provider/model. The dispatch tool calls it via `ctx.summarizeBlob` without owning provider construction — preserves `@kodax/agent` layer independence (agent stays unaware of LLM client).
+- **`GuardedToolResult.spillFailed?` flag** (`packages/coding/src/tools/tool-result-policy.ts`). Set when `persistToolOutput` throws and content was returned inline as the data-loss-guard fallback. Callers that need an LLM-summary follow-up (`dispatch-child-tasks` for `child_task_summary` >100KB) branch on this flag. The `console.warn` is intentionally NOT gated on `KODAX_DEBUG_TOOL_GUARDRAILS` — disk failure is a severe operational event an operator must see immediately.
+- **`applyChildSummaryGuardrailWithSummarizer` helper in `dispatch-child-tasks.ts`**. Unifies the 4 dispatch call sites (async success / async crash / sync success / sync failure) through a single helper that chains: `applyToolResultGuardrail` → if `spillFailed` + content > 100KB + `ctx.summarizeBlob` exists, attempt LLM summarize → on success, banner-wrap with LOSSY marker → on summarizer failure, console.warn + fall back to inline full content with emergency banner. Symmetric error handling across all 4 sites; banner strings consistent.
+- **3 new tracker entries marked Released**: FEATURE_121 (Envelope Spillover Gap-Fix), FEATURE_134 (Image / Screenshot Paste Input). Both originally `Planned` → now `v0.7.40 Released`. FEATURE_LIST.md `Current released version` bumped from v0.7.39 to v0.7.40.
+- **Layer 2 eval datasets** at `benchmark/datasets/feature-121-envelope-spillover/` (3 cases: `preview_sufficient` / `detail_required` / `inline_no_spillover`) and `benchmark/datasets/feature-121-blob-summarizer/` (2 cases: `audit_report` / `grep_findings`). Eval drivers at `tests/feature-121-envelope-spillover.eval.ts` and `tests/feature-121-blob-summarizer.eval.ts` (gated on `KODAX_EVAL_F121_SUMMARIZER=1`). Raw outputs preserved per `EVAL_GUIDELINES.md` §"Raw output preservation" under `os.tmpdir()/kodax-eval-dumps/feature-121-*/`.
+
+### Test coverage delta
+
+- New: 8 blob-summarizer unit tests + 25 blob-summarizer dataset shape tests + 27 envelope-spillover dataset shape tests + 9 tool-output-gc tests + 4 worker-role-prompt new tests (LARGE CHILD OUTPUT block + spill-path hint). Plus FEATURE_134's 52 new tests (45 foundation + 7 integration in `prompt-input-controller`).
+- Total green at HEAD: **5745 tests pass** (507 test files, 1 skipped, 23 todo).
+
 ## [0.7.39] - 2026-05-12
 
 ### Theme

package/README.md

@@ -81,6 +81,37 @@ If you need a custom base URL or an OpenAI/Anthropic-compatible endpoint, define
 
 `userAgentMode` defaults to `"compat"`, which sends `KodaX` instead of the official SDK User-Agent. Switch it to `"sdk"` only when your gateway expects the upstream SDK header.
 
+#### Opting a custom provider into image / vision input (FEATURE_134 v0.7.40)
+
+If your custom provider's underlying model supports image input (vision), add a `capabilityProfile.multimodalSupport: "image-input"` block so KodaX does not artificially block multimodal requests at the SA-path policy gate. The 12 built-in vision-capable providers (Anthropic, OpenAI, the 9 Anthropic-/OpenAI-compat clones — DeepSeek, Kimi, Kimi-code, Qwen, Zhipu, Zhipu-coding, MiniMax-coding, MiMo-coding, Ark-coding — plus Gemini-CLI via the CLI's `@<path>` file-include syntax) already ship with this flag enabled by default; only Codex-CLI and custom providers need to opt in.
+
+```json
+{
+  "customProviders": [
+    {
+      "name": "my-vision-provider",
+      "protocol": "openai",
+      "baseUrl": "https://example.com/v1",
+      "apiKeyEnv": "MY_LLM_API_KEY",
+      "model": "my-vision-model",
+      "capabilityProfile": {
+        "transport": "native-api",
+        "conversationSemantics": "full-history",
+        "mcpSupport": "none",
+        "contextFidelity": "full",
+        "toolCallingFidelity": "full",
+        "sessionSupport": "full",
+        "longRunningSupport": "full",
+        "multimodalSupport": "image-input",
+        "evidenceSupport": "full"
+      }
+    }
+  ]
+}
+```
+
+The serializer layer (`packages/llm/src/providers/anthropic.ts:770` for Anthropic-compat, `openai.ts:904` for OpenAI-compat) forwards image blocks automatically through base-class inheritance. The flag only gates whether KodaX's policy layer pre-rejects multimodal requests — the model-level vision contract remains your upstream provider's responsibility. If the model is actually text-only, you'll see the real upstream API error instead of a KodaX-side rejection.
+
 ### 3. Start in REPL or run a one-shot task
 
 ```bash
@@ -674,6 +705,33 @@ kodax -h team          # Multi-agent parallel execution
 kodax -h print         # Print configuration
 ```
 
+### Environment Variables
+
+KodaX recognizes a number of environment variables for tuning runtime behavior. The most commonly used ones are listed below; for the full list, search the repo for `process.env.KODAX_`.
+
+#### `KODAX_MAX_OUTPUT_TOKENS`
+
+Overrides the per-turn `max_tokens` value sent to **every** provider (Anthropic, OpenAI, Zhipu, Kimi, MiniMax, Qwen, DeepSeek, MiMo, Gemini, Codex, …). Set to a positive integer; unset or non-numeric values are ignored. This is an **explicit user intent**: when set, it wins over the provider's model descriptor cap, over the provider config default, and over the global `KODAX_MAX_TOKENS` fallback. The runtime's automatic safety caps (e.g. the v0.7.28 P2b RST-prone write-turn cap that limits write/edit turns to 8K tokens on Zhipu/Kimi/MiniMax) are **bypassed** when this variable is set, so the user override is also a way to opt out of those caps.
+
+```bash
+# Allow up to 48K output tokens per turn (use a higher cap when generating long files)
+export KODAX_MAX_OUTPUT_TOKENS=48000
+kodax "generate the full implementation"
+
+# Unset to restore default behavior
+unset KODAX_MAX_OUTPUT_TOKENS
+```
+
+Precedence used by every provider's `getEffectiveMaxOutputTokens()` (see `packages/llm/src/providers/base.ts`):
+
+1. One-shot per-request override (agent-loop escalation / context-overflow recovery — internal)
+2. **`KODAX_MAX_OUTPUT_TOKENS`** (this variable, explicit user intent)
+3. Active model descriptor's `maxOutputTokens` (FEATURE_098 per-model cap)
+4. Provider config default
+5. Global `KODAX_MAX_TOKENS` fallback
+
+Related variables: `KODAX_MAX_TOKENS` (global fallback when no provider/model cap applies), `KODAX_RST_PRONE_PROVIDERS` and `KODAX_WRITE_TURN_MAX_TOKENS` (v0.7.28 P2b write-turn safety cap configuration), `KODAX_ESCALATED_MAX_OUTPUT_TOKENS` (escalation budget used by the agent loop when a turn returns `stop_reason: max_tokens`).
+
 ## Advanced Library Usage
 
 #### Simple Mode (runKodaX)

package/README_CN.md

@@ -131,6 +131,37 @@ import { loadConfig } from '@kodax-ai/kodax/repl';     // REPL 配置 / session
 
 `userAgentMode` 默认 `"compat"`（发送 `KodaX` 而非上游 SDK 的 User-Agent）；如果你的网关要求原生 SDK header，再切到 `"sdk"`。
 
+#### 给自定义 provider 开图片 / vision 输入（FEATURE_134 v0.7.40）
+
+如果你的自定义 provider 后面的模型支持 vision，加 `capabilityProfile.multimodalSupport: "image-input"` 显式开启，KodaX 的 SA-path policy gate 就不会人为拦截多模态请求。内置的 12 个 vision-capable provider（Anthropic、OpenAI、9 个 Anthropic-/OpenAI-compat clone：DeepSeek / Kimi / Kimi-code / Qwen / Zhipu / Zhipu-coding / MiniMax-coding / MiMo-coding / Ark-coding，加 Gemini-CLI 通过 CLI 的 `@<path>` file-include 语法）已经默认开了这个 flag。只有 Codex-CLI 和自定义 provider 需要手动 opt-in。
+
+```json
+{
+  "customProviders": [
+    {
+      "name": "my-vision-provider",
+      "protocol": "openai",
+      "baseUrl": "https://example.com/v1",
+      "apiKeyEnv": "MY_LLM_API_KEY",
+      "model": "my-vision-model",
+      "capabilityProfile": {
+        "transport": "native-api",
+        "conversationSemantics": "full-history",
+        "mcpSupport": "none",
+        "contextFidelity": "full",
+        "toolCallingFidelity": "full",
+        "sessionSupport": "full",
+        "longRunningSupport": "full",
+        "multimodalSupport": "image-input",
+        "evidenceSupport": "full"
+      }
+    }
+  ]
+}
+```
+
+序列化层（Anthropic-compat 走 `packages/llm/src/providers/anthropic.ts:770`，OpenAI-compat 走 `openai.ts:904`）通过基类继承自动转发 image block。这个 flag 只控制 KodaX 自身是否预先拒绝多模态请求 —— 上游模型到底支不支持 vision 由 provider 自己决定。如果模型实际是 text-only，你会看到真实的上游 API 错误，而不是 KodaX 一侧的 `[Provider Policy] multimodal requests are unsupported` 预拦截。
+
 库模式下用 `registerCustomProviders()` 显式注册：
 
 ```typescript

package/dist/chunks/{chunk-SF7WD7E5.js → chunk-5TFLMGER.js}

@@ -1,2 +1,2 @@
 // @kodax-ai/kodax — bundled distribution. See docs/ADR.md ADR-022 + ADR-024.
-import{ic as A,kc as M}from"./chunk-7LQ2NCHF.js";import{ia as O}from"./chunk-N2VZ2MJF.js";import{$ as m,S as w,T as S,U as x,V as I,W as L,X as h,_ as k,ca as b,ja as R}from"./chunk-XI75LZIO.js";import{a as i}from"./chunk-4E76FLZ3.js";import a from"fs";import c from"path";import{exec as F,spawnSync as G}from"child_process";import{promisify as V}from"util";import{fileURLToPath as q}from"url";var E="en";function U(){return(process.env.LC_ALL||process.env.LC_MESSAGES||process.env.LANG||"").toLowerCase().startsWith("zh")?"zh":"en"}i(U,"detectSystemLocale");function N(e){if(!e||e==="auto"){E=U();return}E=e.toLowerCase().replace(/[-_].*/,"")==="zh"?"zh":"en"}i(N,"setLocale");var T={en:{"dialog.confirm":"[Confirm]","dialog.select":"[Select]","dialog.input":"[Input]","confirm.instruction.basic":"Press (y) yes, (n) no","confirm.instruction.always":"Press (y) yes, (a) always yes for this tool, (n) no","confirm.instruction.protected":"Press (y) to confirm, (n) to cancel (protected path)","confirm.result.approved":"Approved","confirm.result.approved_always":"Approved (always)","confirm.result.denied":"Denied","tool.bash.title":"Execute bash command?","tool.shell.title":"Execute shell command?","tool.write.title":"Write to file?","tool.edit.title":"Edit file?","tool.generic.title":"Execute {tool}?","field.reason":"Reason","field.intent":"Intent","field.target":"Target","field.scope":"Scope","field.risk":"Risk","field.summary":"Summary","intent.read":"Read project files","intent.delete":"Delete files","intent.deps":"Modify dependencies or environment","intent.modify":"Modify files","intent.execute":"Execute command","intent.write_file":"Write file","intent.edit_file":"Edit file","intent.use_tool":"Use {tool}","risk.destructive":"Destructive change","risk.deps":"May change dependencies or local tools","risk.modify":"May modify files","risk.unknown":"Command effects depend on its arguments","risk.network":"May access network","scope.outside":"Outside project","scope.protected":"Protected path","waiting.confirm":"Waiting: approval required","waiting.select":"Waiting: choose an option","waiting.input":"Waiting: answer the prompt","placeholder.confirm":"Respond to the approval prompt above...","placeholder.select":"Choose an option above...","placeholder.input":"Answer the prompt above...","placeholder.busy":"Agent is busy...","placeholder.queue":"Queue a follow-up for the next round...","placeholder.idle":"Type a message...","select.choice":"Choice:","select.type_number":"(type a number)","select.more":"{count} more choices...","select.more_above":"\u2191 {count} more above","select.more_below":"\u2193 {count} more below","select.confirm_hint":"Press Enter to confirm, Esc to cancel","select.navigate_hint":"Use \u2191\u2193 to navigate, Enter to confirm, Esc to cancel","select.multiselect_hint":"Use \u2191\u2193 to navigate, Space to toggle, Enter to confirm, Esc to cancel","select.multiselect_empty":"Select at least one option with Space before confirming.","select.back_prev":"\u2190 Back to previous question","input.default":"Default:","input.value":"Value:","input.type_response":"(type your response)","managed.completed":"Task completed","managed.completed.blocked":"Task blocked","managed.completed.continuation":"Task needs continuation",cancelled:"[Cancelled] Operation cancelled by user"},zh:{"dialog.confirm":"[\u786E\u8BA4]","dialog.select":"[\u9009\u62E9]","dialog.input":"[\u8F93\u5165]","confirm.instruction.basic":"\u6309 (y) \u786E\u8BA4, (n) \u62D2\u7EDD","confirm.instruction.always":"\u6309 (y) \u786E\u8BA4, (a) \u59CB\u7EC8\u5141\u8BB8\u6B64\u5DE5\u5177, (n) \u62D2\u7EDD","confirm.instruction.protected":"\u6309 (y) \u786E\u8BA4, (n) \u53D6\u6D88 (\u53D7\u4FDD\u62A4\u8DEF\u5F84)","confirm.result.approved":"\u5DF2\u6279\u51C6","confirm.result.approved_always":"\u5DF2\u6279\u51C6 (\u59CB\u7EC8\u5141\u8BB8)","confirm.result.denied":"\u5DF2\u62D2\u7EDD","tool.bash.title":"\u6267\u884C bash \u547D\u4EE4\uFF1F","tool.shell.title":"\u6267\u884C shell \u547D\u4EE4\uFF1F","tool.write.title":"\u5199\u5165\u6587\u4EF6\uFF1F","tool.edit.title":"\u7F16\u8F91\u6587\u4EF6\uFF1F","tool.generic.title":"\u6267\u884C {tool}\uFF1F","field.reason":"\u539F\u56E0","field.intent":"\u610F\u56FE","field.target":"\u76EE\u6807","field.scope":"\u8303\u56F4","field.risk":"\u98CE\u9669","field.summary":"\u6458\u8981","intent.read":"\u8BFB\u53D6\u9879\u76EE\u6587\u4EF6","intent.delete":"\u5220\u9664\u6587\u4EF6","intent.deps":"\u4FEE\u6539\u4F9D\u8D56\u6216\u73AF\u5883","intent.modify":"\u4FEE\u6539\u6587\u4EF6","intent.execute":"\u6267\u884C\u547D\u4EE4","intent.write_file":"\u5199\u5165\u6587\u4EF6","intent.edit_file":"\u7F16\u8F91\u6587\u4EF6","intent.use_tool":"\u4F7F\u7528 {tool}","risk.destructive":"\u7834\u574F\u6027\u53D8\u66F4","risk.deps":"\u53EF\u80FD\u4FEE\u6539\u4F9D\u8D56\u6216\u672C\u5730\u5DE5\u5177","risk.modify":"\u53EF\u80FD\u4FEE\u6539\u6587\u4EF6","risk.unknown":"\u547D\u4EE4\u6548\u679C\u53D6\u51B3\u4E8E\u53C2\u6570","risk.network":"\u53EF\u80FD\u8BBF\u95EE\u7F51\u7EDC","scope.outside":"\u9879\u76EE\u5916\u90E8","scope.protected":"\u53D7\u4FDD\u62A4\u8DEF\u5F84","waiting.confirm":"\u7B49\u5F85\u4E2D\uFF1A\u9700\u8981\u5BA1\u6279","waiting.select":"\u7B49\u5F85\u4E2D\uFF1A\u8BF7\u9009\u62E9","waiting.input":"\u7B49\u5F85\u4E2D\uFF1A\u8BF7\u56DE\u7B54","placeholder.confirm":"\u8BF7\u56DE\u5E94\u4E0A\u65B9\u7684\u5BA1\u6279\u63D0\u793A...","placeholder.select":"\u8BF7\u5728\u4E0A\u65B9\u9009\u62E9\u4E00\u4E2A\u9009\u9879...","placeholder.input":"\u8BF7\u56DE\u7B54\u4E0A\u65B9\u7684\u63D0\u793A...","placeholder.busy":"\u4EE3\u7406\u6B63\u5728\u5DE5\u4F5C\u4E2D...","placeholder.queue":"\u6392\u961F\u7B49\u5F85\u4E0B\u4E00\u8F6E\u8DDF\u8FDB...","placeholder.idle":"\u8F93\u5165\u6D88\u606F...","select.choice":"\u9009\u9879\uFF1A","select.type_number":"(\u8F93\u5165\u7F16\u53F7)","select.more":"\u8FD8\u6709 {count} \u4E2A\u9009\u9879...","select.more_above":"\u2191 \u4E0A\u65B9\u8FD8\u6709 {count} \u4E2A","select.more_below":"\u2193 \u4E0B\u65B9\u8FD8\u6709 {count} \u4E2A","select.confirm_hint":"\u6309 Enter \u786E\u8BA4\uFF0CEsc \u53D6\u6D88","select.navigate_hint":"\u4F7F\u7528 \u2191\u2193 \u5BFC\u822A\uFF0CEnter \u786E\u8BA4\uFF0CEsc \u53D6\u6D88","select.multiselect_hint":"\u4F7F\u7528 \u2191\u2193 \u5BFC\u822A\uFF0C\u7A7A\u683C \u5207\u6362\u9009\u4E2D\uFF0CEnter \u786E\u8BA4\uFF0CEsc \u53D6\u6D88","select.multiselect_empty":"\u8BF7\u5148\u4F7F\u7528\u7A7A\u683C\u9009\u62E9\u81F3\u5C11\u4E00\u4E2A\u9009\u9879\u3002","select.back_prev":"\u2190 \u8FD4\u56DE\u4E0A\u4E00\u9898","input.default":"\u9ED8\u8BA4\u503C\uFF1A","input.value":"\u503C\uFF1A","input.type_response":"(\u8F93\u5165\u4F60\u7684\u56DE\u7B54)","managed.completed":"\u4EFB\u52A1\u5B8C\u6210","managed.completed.blocked":"\u4EFB\u52A1\u53D7\u963B","managed.completed.continuation":"\u4EFB\u52A1\u9700\u8981\u7EE7\u7EED",cancelled:"[\u5DF2\u53D6\u6D88] \u64CD\u4F5C\u5DF2\u88AB\u7528\u6237\u53D6\u6D88"}};function pe(e,t){let o=T[E][e]??T.en[e]??e;if(t)for(let[r,l]of Object.entries(t))o=o.replace(`{${r}}`,String(l));return o}i(pe,"t");var W=V(F),X=O(),Ce=c.join(X,"sessions"),d=c.join(X,"config.json"),we=60,s=null,_=!1,J="dumb";function Q(e){let t=c.basename(e).toLowerCase(),n="__KODAX_SHELL_ENV_START__",o=`printf '%s\\0' '${n}'; env -0`;return t==="fish"?{args:["-i","-c",o],sentinel:n}:{args:t==="bash"||t==="zsh"?["-ic",o]:["-lc",o],sentinel:n}}i(Q,"buildShellEnvCommand");function Y(e,t){let n=`${t}\0`,o=e.lastIndexOf(n);if(o===-1)return{};let r=e.slice(o+n.length),l={};for(let u of r.split("\0")){if(!u)continue;let p=u.indexOf("=");p<=0||(l[u.slice(0,p)]=u.slice(p+1))}return l}i(Y,"parseNullDelimitedShellEnv");function Z(e={}){let t=e.env??process.env;if((e.platform??process.platform)==="win32"||t.KODAX_DISABLE_SHELL_ENV_HYDRATION==="1")return!1;let o=e.shell??t.SHELL;if(!o||!c.isAbsolute(o))return!1;let{args:r,sentinel:l}=Q(o),u=e.run??G,p={...t,TERM:J},f=u(o,r,{encoding:"utf8",env:p,maxBuffer:1024*1024,timeout:5e3,windowsHide:!0,detached:!0,stdio:["ignore","pipe","pipe"]});if(f.status!==0||!f.stdout)return!1;let H=typeof f.stdout=="string"?f.stdout:f.stdout.toString("utf8"),j=Y(H,l),C=!1;for(let[v,z]of Object.entries(j))v!=="TERM"&&t[v]===void 0&&(t[v]=z,C=!0);return C}i(Z,"hydrateProcessEnvFromShell");function ee(){if(!_){_=!0;try{Z()}catch{}}}i(ee,"ensureShellEnvironmentHydrated");function Se(){_=!1}i(Se,"resetShellEnvironmentHydrationForTesting");function te(e){k(e.customProviders??[])}i(te,"registerConfiguredCustomProviders");function $(e){if(!Array.isArray(e))return;let t=e.filter(n=>typeof n=="string").map(n=>n.trim()).filter(n=>n.length>0);return t.length>0?t:[]}i($,"normalizeConfiguredExtensions");function ne(e){if(e.permissionMode!=="default")return e;let t={...e,permissionMode:"accept-edits"};try{a.mkdirSync(c.dirname(d),{recursive:!0}),a.writeFileSync(d,JSON.stringify(t,null,2))}catch{}return t}i(ne,"migrateLegacyPermissionModeInConfig");function oe(){if(s)return s;let e=process.env.KODAX_VERSION;if(e)return s=e,s;let t=c.join(c.dirname(q(import.meta.url)),"../../package.json");if(a.existsSync(t))try{return s=JSON.parse(a.readFileSync(t,"utf-8")).version??"0.0.0",s??"0.0.0"}catch{}return s="0.0.0",s}i(oe,"getVersion");var xe=oe();function Ie(e){return S(e)}i(Ie,"getProviderModel");function P(e,t){let n=new Set(e.map(r=>r.toLowerCase())),o=[...e];for(let r of t)n.has(r.toLowerCase())||o.push(r);return o}i(P,"mergeModels");function ie(e,t){t||(t=y().providerModels);let n=t?.[e];if(n&&n.length>0){try{let o=h(e);if(o.length>0)return P(n,o)}catch{}try{let o=m(e);if(o)return P(n,o.getAvailableModels())}catch{}return n}try{let o=h(e);if(o.length>0)return o}catch{}try{let o=m(e);if(o)return o.getAvailableModels()}catch{}return[]}i(ie,"getProviderAvailableModels");function re(e,t){let n=x(e,t);if(n!=="unknown")return n;try{let o=m(e);if(o)return o.getReasoningCapability(t)}catch{}return"unknown"}i(re,"getProviderReasoningCapability");function se(e){let t=I(e);if(t)return t;try{return b().find(o=>o.name===e)?.capabilityProfile??null}catch{return null}}i(se,"getProviderCapabilityProfile");function B(e,t){let n=se(e),o=re(e,t);if(n)return{capabilityProfile:n,reasoningCapability:o};try{let r=R(e);return{capabilityProfile:r.getCapabilityProfile(),reasoningCapability:r.getReasoningCapability(t)}}catch{return null}}i(B,"getProviderCapabilityMetadata");function Le(e,t){let n=B(e,t);return n?A({providerName:e,model:t,capabilityProfile:n.capabilityProfile,reasoningCapability:n.reasoningCapability==="unknown"?void 0:n.reasoningCapability}):null}i(Le,"getProviderCapabilitySnapshot");function ae(e,t,n,o){let r=B(e,t);return r?M({providerName:e,model:t,capabilityProfile:r.capabilityProfile,reasoningCapability:r.reasoningCapability==="unknown"?void 0:r.reasoningCapability,reasoningMode:n,hints:o}):null}i(ae,"getProviderPolicyDecision");function ke(e){let t=e.transport==="cli-bridge"?"CLI bridge":"Native API",n=e.conversationSemantics==="last-user-message"?"forwards only the latest user message":"preserves full conversation history",o=e.mcpSupport==="native"?"MCP available":"MCP unavailable";return`${t}; ${n}; ${o}`}i(ke,"describeProviderCapabilitySummary");function Re(e){switch(e){case"native-budget":return"B";case"native-effort":return"E";case"native-toggle":return"T";default:return"-"}}i(Re,"formatReasoningCapabilityShort");function ce(e){switch(e){case"native-budget":return"budget";case"native-effort":return"effort";case"native-toggle":return"toggle";default:return"none"}}i(ce,"describeReasoningCapabilityControl");function Oe(e,t){if(e==="off")return"Reasoning disabled";switch(t){case"native-budget":return"Uses native thinking budget control";case"native-effort":return"Uses native reasoning effort control";case"native-toggle":return"Uses provider-native thinking toggle only";case"none":return"Runs without native reasoning parameters";case"prompt-only":return"Uses prompt overlays only; no native reasoning parameter";default:return"Runs without native reasoning parameters"}}i(Oe,"describeReasoningExecution");function Ae(e){let t=[];e||(e=y().providerModels);for(let n of L())t.push({name:n.name,model:n.model,models:ie(n.name,e),configured:n.capabilityProfile.transport==="cli-bridge"?!0:n.configured,reasoningCapability:n.reasoningCapability,capabilityProfile:n.capabilityProfile});try{let n=b().map(o=>({...o,models:(()=>{let r=e?.[o.name];return r&&r.length>0?P(r,o.models):o.models})()}));t.push(...n)}catch{}return t}i(Ae,"getProviderList");function Me(e){if(w(e))return!0;try{return m(e)?.isConfigured()??!1}catch{return!1}}i(Me,"isProviderConfigured");function y(){try{if(a.existsSync(d)){let e=JSON.parse(a.readFileSync(d,"utf-8")),t=e.reasoningCeiling??e.reasoningMode;return ne({...e,reasoningMode:t,extensions:$(e.extensions)})}}catch{}return{}}i(y,"loadConfig");function le(e){e.streamIdleTimeoutMs&&!process.env.KODAX_STREAM_IDLE_TIMEOUT_MS&&(process.env.KODAX_STREAM_IDLE_TIMEOUT_MS=String(e.streamIdleTimeoutMs))}i(le,"applyResilienceRuntimeEnv");function ue(e){e.repoIntelligenceMode&&!process.env.KODAX_REPO_INTELLIGENCE_MODE&&(process.env.KODAX_REPO_INTELLIGENCE_MODE=e.repoIntelligenceMode),e.repointelEndpoint&&!process.env.KODAX_REPOINTEL_ENDPOINT&&(process.env.KODAX_REPOINTEL_ENDPOINT=e.repointelEndpoint),e.repointelBin&&!process.env.KODAX_REPOINTEL_BIN&&(process.env.KODAX_REPOINTEL_BIN=e.repointelBin),e.repoIntelligenceTrace===!0&&!process.env.KODAX_REPO_INTELLIGENCE_TRACE&&(process.env.KODAX_REPO_INTELLIGENCE_TRACE="1")}i(ue,"applyRepoIntelligenceRuntimeEnv");function Te(){ee();let e=y();return le(e),ue(e),te(e),N(e.locale),e}i(Te,"prepareRuntimeConfig");function Ne(e){let n={...y(),...e},o=$(n.extensions);o!==void 0&&(n.extensions=o);for(let r of Object.keys(e))e[r]===void 0&&delete n[r];a.mkdirSync(c.dirname(d),{recursive:!0}),a.writeFileSync(d,JSON.stringify(n,null,2))}i(Ne,"saveConfig");async function De(){try{let{stdout:e}=await W("git rev-parse --show-toplevel");return e.trim()}catch{return null}}i(De,"getGitRoot");function de(e){switch(e){case"builtin":return"Built-in";case"runtime":return"Runtime extension";case"custom":return"Custom config";default:return"Unknown"}}i(de,"formatProviderSourceKind");function Ke(e){let t=e.transport==="cli-bridge"?"CLI bridge":"Native API",n=e.conversationSemantics==="last-user-message"?"latest-user-message only":"full conversation history";return[`Source: ${de(e.sourceKind)}`,`Transport: ${t}`,`Conversation semantics: ${n}`,`Context fidelity: ${e.contextFidelity}`,`Tool calling: ${e.toolCallingFidelity}`,`Session behavior: ${e.sessionSupport}`,`Long-running support: ${e.longRunningSupport}`,`Evidence-heavy flows: ${e.evidenceSupport}`,`Multimodal support: ${e.multimodalSupport}`,`MCP support: ${e.mcpSupport}`,`Reasoning control: ${ce(e.reasoningCapability)}`]}i(Ke,"formatProviderCapabilityDetailLines");function Xe(e,t,n){return[{label:"General coding",hints:{}},{label:"Evidence-heavy review",hints:{evidenceHeavy:!0}},{label:"Long-running task",hints:{longRunning:!0}}].map(r=>({label:r.label,decision:ae(e,t,n,r.hints)})).filter(r=>r.decision!==null)}i(Xe,"getProviderCommonPolicyScenarios");var D=.5,K=0,g={locked:!1,queue:[]};async function $e(e){for(;g.locked;)await new Promise(t=>g.queue.push(t));g.locked=!0;try{let t=(Date.now()-K)/1e3;t<D&&await new Promise(o=>setTimeout(o,(D-t)*1e3));let n=await e();return K=Date.now(),n}finally{g.locked=!1;let t=g.queue.shift();t&&t()}}i($e,"rateLimitedCall");export{pe as a,X as b,Ce as c,d,we as e,Z as f,Se as g,te as h,oe as i,xe as j,Ie as k,ie as l,re as m,se as n,Le as o,ae as p,ke as q,Re as r,ce as s,Oe as t,Ae as u,Me as v,y as w,Te as x,Ne as y,De as z,de as A,Ke as B,Xe as C,$e as D};
+import{pc as A,rc as M}from"./chunk-HYWVRTFA.js";import{ia as O}from"./chunk-SX2IS5JP.js";import{$ as m,S as w,T as S,U as x,V as I,W as L,X as h,_ as k,ca as b,ja as R}from"./chunk-6QO6HWGU.js";import{a as i}from"./chunk-V4WSBIXB.js";import a from"fs";import c from"path";import{exec as F,spawnSync as G}from"child_process";import{promisify as V}from"util";import{fileURLToPath as q}from"url";var E="en";function U(){return(process.env.LC_ALL||process.env.LC_MESSAGES||process.env.LANG||"").toLowerCase().startsWith("zh")?"zh":"en"}i(U,"detectSystemLocale");function N(e){if(!e||e==="auto"){E=U();return}E=e.toLowerCase().replace(/[-_].*/,"")==="zh"?"zh":"en"}i(N,"setLocale");var T={en:{"dialog.confirm":"[Confirm]","dialog.select":"[Select]","dialog.input":"[Input]","confirm.instruction.basic":"Press (y) yes, (n) no","confirm.instruction.always":"Press (y) yes, (a) always yes for this tool, (n) no","confirm.instruction.protected":"Press (y) to confirm, (n) to cancel (protected path)","confirm.result.approved":"Approved","confirm.result.approved_always":"Approved (always)","confirm.result.denied":"Denied","tool.bash.title":"Execute bash command?","tool.shell.title":"Execute shell command?","tool.write.title":"Write to file?","tool.edit.title":"Edit file?","tool.generic.title":"Execute {tool}?","field.reason":"Reason","field.intent":"Intent","field.target":"Target","field.scope":"Scope","field.risk":"Risk","field.summary":"Summary","intent.read":"Read project files","intent.delete":"Delete files","intent.deps":"Modify dependencies or environment","intent.modify":"Modify files","intent.execute":"Execute command","intent.write_file":"Write file","intent.edit_file":"Edit file","intent.use_tool":"Use {tool}","risk.destructive":"Destructive change","risk.deps":"May change dependencies or local tools","risk.modify":"May modify files","risk.unknown":"Command effects depend on its arguments","risk.network":"May access network","scope.outside":"Outside project","scope.protected":"Protected path","waiting.confirm":"Waiting: approval required","waiting.select":"Waiting: choose an option","waiting.input":"Waiting: answer the prompt","placeholder.confirm":"Respond to the approval prompt above...","placeholder.select":"Choose an option above...","placeholder.input":"Answer the prompt above...","placeholder.busy":"Agent is busy...","placeholder.queue":"Queue a follow-up for the next round...","placeholder.idle":"Type a message...","select.choice":"Choice:","select.type_number":"(type a number)","select.more":"{count} more choices...","select.more_above":"\u2191 {count} more above","select.more_below":"\u2193 {count} more below","select.confirm_hint":"Press Enter to confirm, Esc to cancel","select.navigate_hint":"Use \u2191\u2193 to navigate, Enter to confirm, Esc to cancel","select.multiselect_hint":"Use \u2191\u2193 to navigate, Space to toggle, Enter to confirm, Esc to cancel","select.multiselect_empty":"Select at least one option with Space before confirming.","select.back_prev":"\u2190 Back to previous question","input.default":"Default:","input.value":"Value:","input.type_response":"(type your response)","managed.completed":"Task completed","managed.completed.blocked":"Task blocked","managed.completed.continuation":"Task needs continuation",cancelled:"[Cancelled] Operation cancelled by user"},zh:{"dialog.confirm":"[\u786E\u8BA4]","dialog.select":"[\u9009\u62E9]","dialog.input":"[\u8F93\u5165]","confirm.instruction.basic":"\u6309 (y) \u786E\u8BA4, (n) \u62D2\u7EDD","confirm.instruction.always":"\u6309 (y) \u786E\u8BA4, (a) \u59CB\u7EC8\u5141\u8BB8\u6B64\u5DE5\u5177, (n) \u62D2\u7EDD","confirm.instruction.protected":"\u6309 (y) \u786E\u8BA4, (n) \u53D6\u6D88 (\u53D7\u4FDD\u62A4\u8DEF\u5F84)","confirm.result.approved":"\u5DF2\u6279\u51C6","confirm.result.approved_always":"\u5DF2\u6279\u51C6 (\u59CB\u7EC8\u5141\u8BB8)","confirm.result.denied":"\u5DF2\u62D2\u7EDD","tool.bash.title":"\u6267\u884C bash \u547D\u4EE4\uFF1F","tool.shell.title":"\u6267\u884C shell \u547D\u4EE4\uFF1F","tool.write.title":"\u5199\u5165\u6587\u4EF6\uFF1F","tool.edit.title":"\u7F16\u8F91\u6587\u4EF6\uFF1F","tool.generic.title":"\u6267\u884C {tool}\uFF1F","field.reason":"\u539F\u56E0","field.intent":"\u610F\u56FE","field.target":"\u76EE\u6807","field.scope":"\u8303\u56F4","field.risk":"\u98CE\u9669","field.summary":"\u6458\u8981","intent.read":"\u8BFB\u53D6\u9879\u76EE\u6587\u4EF6","intent.delete":"\u5220\u9664\u6587\u4EF6","intent.deps":"\u4FEE\u6539\u4F9D\u8D56\u6216\u73AF\u5883","intent.modify":"\u4FEE\u6539\u6587\u4EF6","intent.execute":"\u6267\u884C\u547D\u4EE4","intent.write_file":"\u5199\u5165\u6587\u4EF6","intent.edit_file":"\u7F16\u8F91\u6587\u4EF6","intent.use_tool":"\u4F7F\u7528 {tool}","risk.destructive":"\u7834\u574F\u6027\u53D8\u66F4","risk.deps":"\u53EF\u80FD\u4FEE\u6539\u4F9D\u8D56\u6216\u672C\u5730\u5DE5\u5177","risk.modify":"\u53EF\u80FD\u4FEE\u6539\u6587\u4EF6","risk.unknown":"\u547D\u4EE4\u6548\u679C\u53D6\u51B3\u4E8E\u53C2\u6570","risk.network":"\u53EF\u80FD\u8BBF\u95EE\u7F51\u7EDC","scope.outside":"\u9879\u76EE\u5916\u90E8","scope.protected":"\u53D7\u4FDD\u62A4\u8DEF\u5F84","waiting.confirm":"\u7B49\u5F85\u4E2D\uFF1A\u9700\u8981\u5BA1\u6279","waiting.select":"\u7B49\u5F85\u4E2D\uFF1A\u8BF7\u9009\u62E9","waiting.input":"\u7B49\u5F85\u4E2D\uFF1A\u8BF7\u56DE\u7B54","placeholder.confirm":"\u8BF7\u56DE\u5E94\u4E0A\u65B9\u7684\u5BA1\u6279\u63D0\u793A...","placeholder.select":"\u8BF7\u5728\u4E0A\u65B9\u9009\u62E9\u4E00\u4E2A\u9009\u9879...","placeholder.input":"\u8BF7\u56DE\u7B54\u4E0A\u65B9\u7684\u63D0\u793A...","placeholder.busy":"\u4EE3\u7406\u6B63\u5728\u5DE5\u4F5C\u4E2D...","placeholder.queue":"\u6392\u961F\u7B49\u5F85\u4E0B\u4E00\u8F6E\u8DDF\u8FDB...","placeholder.idle":"\u8F93\u5165\u6D88\u606F...","select.choice":"\u9009\u9879\uFF1A","select.type_number":"(\u8F93\u5165\u7F16\u53F7)","select.more":"\u8FD8\u6709 {count} \u4E2A\u9009\u9879...","select.more_above":"\u2191 \u4E0A\u65B9\u8FD8\u6709 {count} \u4E2A","select.more_below":"\u2193 \u4E0B\u65B9\u8FD8\u6709 {count} \u4E2A","select.confirm_hint":"\u6309 Enter \u786E\u8BA4\uFF0CEsc \u53D6\u6D88","select.navigate_hint":"\u4F7F\u7528 \u2191\u2193 \u5BFC\u822A\uFF0CEnter \u786E\u8BA4\uFF0CEsc \u53D6\u6D88","select.multiselect_hint":"\u4F7F\u7528 \u2191\u2193 \u5BFC\u822A\uFF0C\u7A7A\u683C \u5207\u6362\u9009\u4E2D\uFF0CEnter \u786E\u8BA4\uFF0CEsc \u53D6\u6D88","select.multiselect_empty":"\u8BF7\u5148\u4F7F\u7528\u7A7A\u683C\u9009\u62E9\u81F3\u5C11\u4E00\u4E2A\u9009\u9879\u3002","select.back_prev":"\u2190 \u8FD4\u56DE\u4E0A\u4E00\u9898","input.default":"\u9ED8\u8BA4\u503C\uFF1A","input.value":"\u503C\uFF1A","input.type_response":"(\u8F93\u5165\u4F60\u7684\u56DE\u7B54)","managed.completed":"\u4EFB\u52A1\u5B8C\u6210","managed.completed.blocked":"\u4EFB\u52A1\u53D7\u963B","managed.completed.continuation":"\u4EFB\u52A1\u9700\u8981\u7EE7\u7EED",cancelled:"[\u5DF2\u53D6\u6D88] \u64CD\u4F5C\u5DF2\u88AB\u7528\u6237\u53D6\u6D88"}};function pe(e,t){let o=T[E][e]??T.en[e]??e;if(t)for(let[r,l]of Object.entries(t))o=o.replace(`{${r}}`,String(l));return o}i(pe,"t");var W=V(F),X=O(),Ce=c.join(X,"sessions"),d=c.join(X,"config.json"),we=60,s=null,_=!1,J="dumb";function Q(e){let t=c.basename(e).toLowerCase(),n="__KODAX_SHELL_ENV_START__",o=`printf '%s\\0' '${n}'; env -0`;return t==="fish"?{args:["-i","-c",o],sentinel:n}:{args:t==="bash"||t==="zsh"?["-ic",o]:["-lc",o],sentinel:n}}i(Q,"buildShellEnvCommand");function Y(e,t){let n=`${t}\0`,o=e.lastIndexOf(n);if(o===-1)return{};let r=e.slice(o+n.length),l={};for(let u of r.split("\0")){if(!u)continue;let p=u.indexOf("=");p<=0||(l[u.slice(0,p)]=u.slice(p+1))}return l}i(Y,"parseNullDelimitedShellEnv");function Z(e={}){let t=e.env??process.env;if((e.platform??process.platform)==="win32"||t.KODAX_DISABLE_SHELL_ENV_HYDRATION==="1")return!1;let o=e.shell??t.SHELL;if(!o||!c.isAbsolute(o))return!1;let{args:r,sentinel:l}=Q(o),u=e.run??G,p={...t,TERM:J},f=u(o,r,{encoding:"utf8",env:p,maxBuffer:1024*1024,timeout:5e3,windowsHide:!0,detached:!0,stdio:["ignore","pipe","pipe"]});if(f.status!==0||!f.stdout)return!1;let H=typeof f.stdout=="string"?f.stdout:f.stdout.toString("utf8"),j=Y(H,l),C=!1;for(let[v,z]of Object.entries(j))v!=="TERM"&&t[v]===void 0&&(t[v]=z,C=!0);return C}i(Z,"hydrateProcessEnvFromShell");function ee(){if(!_){_=!0;try{Z()}catch{}}}i(ee,"ensureShellEnvironmentHydrated");function Se(){_=!1}i(Se,"resetShellEnvironmentHydrationForTesting");function te(e){k(e.customProviders??[])}i(te,"registerConfiguredCustomProviders");function $(e){if(!Array.isArray(e))return;let t=e.filter(n=>typeof n=="string").map(n=>n.trim()).filter(n=>n.length>0);return t.length>0?t:[]}i($,"normalizeConfiguredExtensions");function ne(e){if(e.permissionMode!=="default")return e;let t={...e,permissionMode:"accept-edits"};try{a.mkdirSync(c.dirname(d),{recursive:!0}),a.writeFileSync(d,JSON.stringify(t,null,2))}catch{}return t}i(ne,"migrateLegacyPermissionModeInConfig");function oe(){if(s)return s;let e="0.7.41";if(e)return s=e,s;let t=c.join(c.dirname(q(import.meta.url)),"../../package.json");if(a.existsSync(t))try{return s=JSON.parse(a.readFileSync(t,"utf-8")).version??"0.0.0",s??"0.0.0"}catch{}return s="0.0.0",s}i(oe,"getVersion");var xe=oe();function Ie(e){return S(e)}i(Ie,"getProviderModel");function P(e,t){let n=new Set(e.map(r=>r.toLowerCase())),o=[...e];for(let r of t)n.has(r.toLowerCase())||o.push(r);return o}i(P,"mergeModels");function ie(e,t){t||(t=y().providerModels);let n=t?.[e];if(n&&n.length>0){try{let o=h(e);if(o.length>0)return P(n,o)}catch{}try{let o=m(e);if(o)return P(n,o.getAvailableModels())}catch{}return n}try{let o=h(e);if(o.length>0)return o}catch{}try{let o=m(e);if(o)return o.getAvailableModels()}catch{}return[]}i(ie,"getProviderAvailableModels");function re(e,t){let n=x(e,t);if(n!=="unknown")return n;try{let o=m(e);if(o)return o.getReasoningCapability(t)}catch{}return"unknown"}i(re,"getProviderReasoningCapability");function se(e){let t=I(e);if(t)return t;try{return b().find(o=>o.name===e)?.capabilityProfile??null}catch{return null}}i(se,"getProviderCapabilityProfile");function B(e,t){let n=se(e),o=re(e,t);if(n)return{capabilityProfile:n,reasoningCapability:o};try{let r=R(e);return{capabilityProfile:r.getCapabilityProfile(),reasoningCapability:r.getReasoningCapability(t)}}catch{return null}}i(B,"getProviderCapabilityMetadata");function Le(e,t){let n=B(e,t);return n?A({providerName:e,model:t,capabilityProfile:n.capabilityProfile,reasoningCapability:n.reasoningCapability==="unknown"?void 0:n.reasoningCapability}):null}i(Le,"getProviderCapabilitySnapshot");function ae(e,t,n,o){let r=B(e,t);return r?M({providerName:e,model:t,capabilityProfile:r.capabilityProfile,reasoningCapability:r.reasoningCapability==="unknown"?void 0:r.reasoningCapability,reasoningMode:n,hints:o}):null}i(ae,"getProviderPolicyDecision");function ke(e){let t=e.transport==="cli-bridge"?"CLI bridge":"Native API",n=e.conversationSemantics==="last-user-message"?"forwards only the latest user message":"preserves full conversation history",o=e.mcpSupport==="native"?"MCP available":"MCP unavailable";return`${t}; ${n}; ${o}`}i(ke,"describeProviderCapabilitySummary");function Re(e){switch(e){case"native-budget":return"B";case"native-effort":return"E";case"native-toggle":return"T";default:return"-"}}i(Re,"formatReasoningCapabilityShort");function ce(e){switch(e){case"native-budget":return"budget";case"native-effort":return"effort";case"native-toggle":return"toggle";default:return"none"}}i(ce,"describeReasoningCapabilityControl");function Oe(e,t){if(e==="off")return"Reasoning disabled";switch(t){case"native-budget":return"Uses native thinking budget control";case"native-effort":return"Uses native reasoning effort control";case"native-toggle":return"Uses provider-native thinking toggle only";case"none":return"Runs without native reasoning parameters";case"prompt-only":return"Uses prompt overlays only; no native reasoning parameter";default:return"Runs without native reasoning parameters"}}i(Oe,"describeReasoningExecution");function Ae(e){let t=[];e||(e=y().providerModels);for(let n of L())t.push({name:n.name,model:n.model,models:ie(n.name,e),configured:n.capabilityProfile.transport==="cli-bridge"?!0:n.configured,reasoningCapability:n.reasoningCapability,capabilityProfile:n.capabilityProfile});try{let n=b().map(o=>({...o,models:(()=>{let r=e?.[o.name];return r&&r.length>0?P(r,o.models):o.models})()}));t.push(...n)}catch{}return t}i(Ae,"getProviderList");function Me(e){if(w(e))return!0;try{return m(e)?.isConfigured()??!1}catch{return!1}}i(Me,"isProviderConfigured");function y(){try{if(a.existsSync(d)){let e=JSON.parse(a.readFileSync(d,"utf-8")),t=e.reasoningCeiling??e.reasoningMode;return ne({...e,reasoningMode:t,extensions:$(e.extensions)})}}catch{}return{}}i(y,"loadConfig");function le(e){e.streamIdleTimeoutMs&&!process.env.KODAX_STREAM_IDLE_TIMEOUT_MS&&(process.env.KODAX_STREAM_IDLE_TIMEOUT_MS=String(e.streamIdleTimeoutMs))}i(le,"applyResilienceRuntimeEnv");function ue(e){e.repoIntelligenceMode&&!process.env.KODAX_REPO_INTELLIGENCE_MODE&&(process.env.KODAX_REPO_INTELLIGENCE_MODE=e.repoIntelligenceMode),e.repointelEndpoint&&!process.env.KODAX_REPOINTEL_ENDPOINT&&(process.env.KODAX_REPOINTEL_ENDPOINT=e.repointelEndpoint),e.repointelBin&&!process.env.KODAX_REPOINTEL_BIN&&(process.env.KODAX_REPOINTEL_BIN=e.repointelBin),e.repoIntelligenceTrace===!0&&!process.env.KODAX_REPO_INTELLIGENCE_TRACE&&(process.env.KODAX_REPO_INTELLIGENCE_TRACE="1")}i(ue,"applyRepoIntelligenceRuntimeEnv");function Te(){ee();let e=y();return le(e),ue(e),te(e),N(e.locale),e}i(Te,"prepareRuntimeConfig");function Ne(e){let n={...y(),...e},o=$(n.extensions);o!==void 0&&(n.extensions=o);for(let r of Object.keys(e))e[r]===void 0&&delete n[r];a.mkdirSync(c.dirname(d),{recursive:!0}),a.writeFileSync(d,JSON.stringify(n,null,2))}i(Ne,"saveConfig");async function De(){try{let{stdout:e}=await W("git rev-parse --show-toplevel");return e.trim()}catch{return null}}i(De,"getGitRoot");function de(e){switch(e){case"builtin":return"Built-in";case"runtime":return"Runtime extension";case"custom":return"Custom config";default:return"Unknown"}}i(de,"formatProviderSourceKind");function Ke(e){let t=e.transport==="cli-bridge"?"CLI bridge":"Native API",n=e.conversationSemantics==="last-user-message"?"latest-user-message only":"full conversation history";return[`Source: ${de(e.sourceKind)}`,`Transport: ${t}`,`Conversation semantics: ${n}`,`Context fidelity: ${e.contextFidelity}`,`Tool calling: ${e.toolCallingFidelity}`,`Session behavior: ${e.sessionSupport}`,`Long-running support: ${e.longRunningSupport}`,`Evidence-heavy flows: ${e.evidenceSupport}`,`Multimodal support: ${e.multimodalSupport}`,`MCP support: ${e.mcpSupport}`,`Reasoning control: ${ce(e.reasoningCapability)}`]}i(Ke,"formatProviderCapabilityDetailLines");function Xe(e,t,n){return[{label:"General coding",hints:{}},{label:"Evidence-heavy review",hints:{evidenceHeavy:!0}},{label:"Long-running task",hints:{longRunning:!0}}].map(r=>({label:r.label,decision:ae(e,t,n,r.hints)})).filter(r=>r.decision!==null)}i(Xe,"getProviderCommonPolicyScenarios");var D=.5,K=0,g={locked:!1,queue:[]};async function $e(e){for(;g.locked;)await new Promise(t=>g.queue.push(t));g.locked=!0;try{let t=(Date.now()-K)/1e3;t<D&&await new Promise(o=>setTimeout(o,(D-t)*1e3));let n=await e();return K=Date.now(),n}finally{g.locked=!1;let t=g.queue.shift();t&&t()}}i($e,"rateLimitedCall");export{pe as a,X as b,Ce as c,d,we as e,Z as f,Se as g,te as h,oe as i,xe as j,Ie as k,ie as l,re as m,se as n,Le as o,ae as p,ke as q,Re as r,ce as s,Oe as t,Ae as u,Me as v,y as w,Te as x,Ne as y,De as z,de as A,Ke as B,Xe as C,$e as D};

package/dist/chunks/{chunk-HUAU4KB3.js → chunk-6OB4AJOM.js}

@@ -1,2 +1,2 @@
 // @kodax-ai/kodax — bundled distribution. See docs/ADR.md ADR-022 + ADR-024.
-import{ja as e}from"./chunk-N2VZ2MJF.js";import{a as n}from"./chunk-4E76FLZ3.js";import{readFile as i}from"fs/promises";var c={enabled:!0,triggerPercent:75};async function g(o){let t=e("config.json");try{let r=await a(t);if(r?.compaction)return{...c,...r.compaction}}catch{}return c}n(g,"loadCompactionConfig");async function a(o){try{let t=await i(o,"utf-8");return JSON.parse(t)}catch{return null}}n(a,"readConfigFile");export{g as a};
+import{ja as e}from"./chunk-SX2IS5JP.js";import{a as n}from"./chunk-V4WSBIXB.js";import{readFile as i}from"fs/promises";var c={enabled:!0,triggerPercent:75};async function g(o){let t=e("config.json");try{let r=await a(t);if(r?.compaction)return{...c,...r.compaction}}catch{}return c}n(g,"loadCompactionConfig");async function a(o){try{let t=await i(o,"utf-8");return JSON.parse(t)}catch{return null}}n(a,"readConfigFile");export{g as a};