@ngockhoale/ukit 2.7.10 → 2.7.12

package/templates/docs/UKIT_INTERNALS.md

@@ -127,22 +127,6 @@ the root contract is the only carrier — keep both in sync with
   turn it into an end-user workflow.
 - Always preserve the CoDev priority: quality > safety > speed > token discipline.
 
-## Post-Edit Sidecar Review — detail
-
-- If routed state's `routeSummary.line` includes a `review=code-reviewer(diff)` segment, a
-  `local-build` or `shared-edit` task qualifies for a non-blocking second opinion — this
-  exists because the daily-flow executor model can miss edge cases.
-- Only launch it once write evidence AND verification evidence already exist for the task
-  (never before; never as a substitute for either).
-- Launch the `code-reviewer` agent (see the harness table under 3-Tier Model Routing) with
-  `REVIEW_TARGET_TYPE=diff`, in the background, on the `smart` tier per
-  `subagents.diffReviewModel`. Do not wait for it — continue and report the task as done
-  using the normal completion rules.
-- Its findings are advisory only: never re-open, block, or delay the already-reported
-  completion on their account. Surface them to the user as a follow-up note if/when they
-  arrive.
-- This is internal orchestration — end users never invoke it directly; `ukit install` plus
-  natural language remains the whole surface. No new commands.
 
 ## Selective Subagent Policy — detail
 
@@ -238,3 +222,81 @@ Phase-4 telemetry → advisory tuning loop (FR-206, TASK-232).
   - `learning/suggestions.json` — result of `computeTuningSuggestions(projectRoot)` (`src/learning/tuning.js`).
   - `cache/retriever-lanes.jsonl` — per-query lane hit/weight telemetry consumed by `collectLaneStats`.
 - **Advisory-only contract**: `suggestions` carry `{target, current, suggested, evidence}`; `applied` is always `[]`. `applyMode` is restricted to `manual|off` — **no writer ever mutates `codeIntel.retriever.weights` or `orchestration.escalation.debugLoopThreshold`**. Rules: lane weight ±0.1 step (clamped `[0.1, 2.0]`) when events ≥ 50 with a >0.5 dominant lane and a <0.02 starved lane; `debugLoopThreshold − 1` (min 1) when repeat-stall events ≥ 10. `ukit metrics` prints a `learning` section (pending count + targets; `n/a` when the artifact is absent).
+
+## Permission Modes + Unsafe Eval — detail
+
+C38 unified permission model. `orchestration.permissionMode` (optional-present; absent tolerated for pre-C35 configs) declares one of three reserved modes — only `unattended` is implemented end-to-end today; `interactive`/`safe-auto` are modelled in the compiler so doctor can report on them but are not rendered modes.
+
+- **`interactive`** — prompts are expected; mode fallback for unclassified entries is `prompt`. Doctor reports reachable prompts as INFO, never FAIL.
+- **`safe-auto`** — same `prompt` fallback, modelled as "auto-allow safe classes, prompt the rest"; reachable prompts remain INFO.
+- **`unattended`** — no human in the loop; mode fallback is `allow` for unclassified entries and any reachable `prompt`/`ask` on a managed surface is a doctor FAIL (`unattended-prompt`), because a prompt dead-ends the run.
+- **Unattended ≠ all-destructive-allowed.** The canonical deny set (`CANONICAL_DENY` in `src/core/permissionPolicy.js`) applies in every mode; `hard-deny` entries (root/home/parent/dot recursive deletes, raw-device writes, mkfs, fork-bomb) are denied on every host including report-only codex, and `protected-guard` hooks (release credentials, protected files, stale-spec guard) outrank any host-deny. A deny never becomes a question — the agent picks the next safer alternative that satisfies the intent and records the substitution.
+- **Host-internal waits are unchanged** — announce long waits, surface API/tool errors verbatim, name network/API as the suspect on stalls; the permission model does not alter the Stall & Wait Reporting contract.
+
+`orchestration.allowUnsafeEval` (default `false`, optional-present boolean, migration never enables): only an explicit `true` renders `tools.approval.eval: allow` in `.omp/config.yml` via the `omp.evalApproval` render variable; every other state renders `deny` — fail-closed, never `prompt`. Residual risk stays documented: eval can carry subprocess payloads (`os.remove`, `shutil.rmtree`) that text-level `bash.patterns` denies cannot see.
+
+Internals: `src/core/permissionPolicy.js` compiles the semantic policy (precedence hard-deny > protected-guard > host-deny > explicit-allow > mode-fallback; `UKIT_EPOLICY` on unknown mode). `src/core/permissionDoctor.js` backs `ukit doctor --permissions [--json]` — FAIL classes `unattended-prompt`, `deny-loss`, `unsafe-eval-mismatch`, `unrendered-token` exit 1; JSON shape `{ permissions: { declared, effective, hosts, checks, failures } }`. Template render, doctor, and the no-freeze matrix consume the same compiler so surfaces cannot drift.
+
+## Hook budget
+
+- **Cap: 8 chain steps per hook group** (current max is 6). Any new step must also
+  carry a `:N` suffix with N > 0 — the runner never falls back to an implicit budget
+  for a declared chain.
+- **Outer timeout = Σ(:N) + 10s minimum.** `settings.json` `timeout` is host-enforced
+  and cannot be derived at runtime, so the margin is enforced by
+  `tests/hooks/hookChainBudget.test.js` — forgetting to bump the outer fails loudly.
+  Every shipped group currently sits at exactly +10s.
+- **Why env knobs can't violate the invariant:** `UKIT_HOOK_*_MS` only affect fallback
+  budgets for args that lack `:N`. On all-declared chains the inner budget is Σ:N by
+  construction, so inner ≤ outer − 10s holds regardless of env (the omp bridge shares
+  the `hook-chain-budget.mjs` resolver).
+- **Convention: hot-path hooks ship as `.mjs` module steps, not new `.sh` processes.**
+  Each new `.sh` step costs a fresh spawn (~30–100ms on the hot path); `.mjs` steps run
+  in-proc in the runner with identical `:N` semantics. Keep the `.sh` thin wrapper
+  only when a direct/debug invocation path is needed.
+- **Module-step contract:** a `"<path>.mjs[:N]"` arg is imported in-proc and must
+  export `runHook(ctx) -> {code, stdout, stderr}`; `ctx` =
+  `{payload, payloadText, projectRoot, env, deadlineMs, signal}`. A throw reports
+  `error`; a `deadlineMs`/`signal` breach reports `timeout` + `killed`. Export
+  `hookFailClosed = true` to join the fail-closed registry (infra failure → exit 2,
+  else 1) — the name-based `FAIL_CLOSED_SCRIPTS` list covers
+  `sensitive-data-guard.mjs` + `block-dangerous.mjs`. `isDirectRun()` gates a CLI
+  mode so the same file stays directly invocable; `UKIT_HOOK_DEADLINE_MS` is
+  scrubbed from `process.env`/`ctx.env` around import. Shipped modules:
+  `sensitive-data-guard.mjs`, `block-dangerous.mjs`, `record-execution.mjs`
+  (`hookFailClosed = false`; uses `setLedgerMessageEmitter` on
+  `execution-ledger.mjs` for the `noteSweepFailure` systemMessage).
+- **Failure taxonomy + `--emit-verdict` mapping.** Every `results[]` entry carries a
+  `failureKind` from one closed set — `ok` (child exited 0), `exit-code` (child exited
+  non-zero on its own), `output-overflow` (capture exceeded the 2 MiB maxBuffer and the
+  tree was culled), `timeout` (the runner killed the child at its deadline), `signal`
+  (the child died from a signal the runner did not send), `error` (the child never
+  spawned), `budget-exhausted` (the script never ran — the chain total budget was spent
+  first). The `--emit-verdict` contract maps them onto one exit code: `ok`/`exit-code 0`
+  and any non-fail-closed failure → exit 0 (fail OPEN); a real `exit-code 2` block, or
+  ANY failure on a fail-closed gate (`output-overflow`, `timeout`, `signal`, `error`,
+  `budget-exhausted`) → exit 2 (fail CLOSED). An overflowed capture is dropped whole
+  before it can be replayed, so a truncated stream can never masquerade as a verdict.
+  `tests/hooks/hookChainFailureSurface.test.js` pins these surfaces.
+- **stderr is a diagnostic channel, not verdict payload.** Under `--emit-verdict` the
+  runner drops the stderr of every non-verdict script: stdout is the verdict and Claude
+  Code parses it as ONE JSON document, so interleaving arbitrary hook diagnostics there
+  would break the contract (and leak whatever a hook happened to log). This is a
+  deliberate decision, not an open gap — stderr is replayed only where it carries the
+  verdict itself: the decision owner's stderr, the last executed script's stderr (block,
+  fail-closed-infra, and pass paths), and the runner's own synthesized diagnostics (the
+  unrun-gate line, `... could not produce a verdict (<failureKind>)`, the empty-chain
+  line). To surface diagnostics from an earlier step, have the step that owns the
+  verdict include them.
+- The same test also guards the settings↔bridge mirror: `HOOK_EVENT_MAP` in
+  `templates/.omp/hooks/pre/ukit-bridge.js` must equal the settings chain args 1:1
+  per event/matcher.
+
+## Compressed contract detail (cycle 45 diet)
+
+Detail dropped from `templates/instructions/core.md` during the ~25% system-prompt diet; semantics unchanged.
+
+- **Prompt Caching one-liners** — CTX-01 deterministic segment bytes · CTX-02 keep roles and order · CTX-03 keep tool IDs and continuation state · CTX-04 no clock/random IDs in static blocks · CTX-05 compaction starts a new epoch · CTX-06 never change data to match a cache · CTX-07 no unconfirmed cache fields · CTX-08 tool-result reuse needs valid freshness · CTX-09 missing usage is unknown, not zero · CTX-10 never cut a required check to reduce calls. Full ruleset: `docs/PROMPT_CACHING.md`.
+- **Common skill triggers** — review/audit/diff/PR → `code-review` · bug/error/crash/triage → `debugging-toolkit` · test/spec/coverage → `testing-quality` · docs/README/changelog/`docs/` edits → `docs-quality` · open-ended/status/continue/queued task → `next-step` · wrap up/`docs/STATUS.md` → `update-status` · auth/security/token/permission/risky → `discover-security` · stale workspace/reinstall/cleanup → `repo-maintenance` (all under `.claude/skills/<name>/SKILL.md`).
+- **3-Tier typical tasks** — lite: reads, git queries, bash summaries, small doc edits · code: normal coding, shared edits, builds, debugging, impact mapping · smart: release review/audit, escalated deep reasoning after repeated failure.
+- **UNATTENDED-02 tightened** — `maxAttemptsPerFailure` 5 → 2, `maxRecoveryStrategies` 3 → 2 (retry the same edit, then fix the cause differently; never loop one strategy).

package/templates/instructions/core.md

@@ -1,167 +1,145 @@
 ## Core Rule
 <!-- RULES: CORE-01 CORE-02 -->
-- Human-facing UKit workflow collapses to one remembered command: `ukit install`.
-- After install, default to natural-language work inside **Claude Code / Codex / omp**.
+- One command: `ukit install`; then work in natural language inside **Claude Code / Codex / omp**.
 - **Quality first, then speed, then token discipline.**
-- **Never stop after read-only steps** — implement/apply/fix requests continue to Edit/Write + verification in the same turn.
+- **Never stop after read-only steps** — implement/apply/fix continues to Edit/Write + verification same turn.
 
 ## Fast Classification
 <!-- RULE: CLS-01 -->
-- **Trivial** — typo, label, small rename, spacing, toggle flag, obvious config change. Act directly. No doc reads, planning, index, or agents.
+- **Trivial** — typo, label, rename, spacing, flag, obvious config. Act directly; no docs/index/agents.
 <!-- RULE: CLS-02 -->
-- **Simple** — 1-2 files, clear scope, existing pattern. Handle directly; pull only the smallest useful context via resolver or targeted read.
+- **Simple** — 1-2 files, clear scope, existing pattern. Act directly; smallest useful context only.
 <!-- RULE: CLS-03 -->
-- **Non-trivial / Risky** — auth, security, migration, uninstall, shared runtime, race/flaky, data-loss. Read deeper, verify harder; index-first loop → skill activation → targeted verification.
+- **Non-trivial / Risky** — auth, security, migration, uninstall, shared runtime, race/flaky, data-loss. Index-first → skill → targeted verification.
 
 ## Execution Contract (mandatory)
 <!-- RULE: EXEC-01 -->
-- For explicit implement/apply/fix requests, **continue until the actual edit is made** or a real blocker is found — never stop after a read-only inspection step.
+- Implement/apply/fix: continue until the actual edit is made or a real blocker — never stop after read-only inspection.
 <!-- RULE: EXEC-03 -->
-- Routed states like `pull-indexed-context` or `continuation required` — treat it as an internal continuation step, not a stopping point; finish the named milestone before widening reads.
+- Routed states (`pull-indexed-context`, `continuation required`): treat it as an internal continuation step, not a stopping point; finish the named milestone first.
 <!-- RULE: EXEC-02 -->
-- **Do NOT say "done"/"applied"/"fixed" after Read/Grep/analysis alone** — completion wording requires concrete Edit/Write evidence this turn, plus verification when scope is risky.
+- **No "done"/"applied"/"fixed" after Read/Grep/analysis alone** — needs concrete Edit/Write evidence this turn, plus verification when risky.
 <!-- RULE: EXEC-04 -->
-- **Every stop says why — no silent idle.** Turns ending on a user-only action open with `WAITING ON YOU: <command/action>` plus a one-shot wakeup (~20-30 min) when available — an ended turn cannot observe external changes, so without it idle looks identical to a stall. Report any error verbatim the same turn.
+- **Every stop says why — no silent idle.** User-only-action stops open with `WAITING ON YOU: <command/action>` + one-shot wakeup (~20-30 min) when available; report errors verbatim.
 
 ## Stall & Wait Reporting
 <!-- RULE: STALL-01 -->
-- Announce long waits as they happen: `Waiting for API response / tool result — will keep retrying; check your network/API provider if this persists`. Surface any API/tool error **verbatim** in the same turn, and name network/API as the suspect on long stalls — **never blame UKit**. This guidance makes the model's side of a wait visible; it cannot detect a stall inside the host's own request loop.
+- Announce long waits: `Waiting for API response / tool result — will keep retrying; check your network/API provider if this persists`. Surface errors verbatim; suspect network/API — **never blame UKit**.
 <!-- RULE: STALL-02 -->
-- Every stop still says why (Execution Contract EXEC-04) — the stall rules above extend it, they do not replace it.
+- Stall rules extend EXEC-04 (every stop says why); they do not replace it.
 <!-- RULE: STALL-03 -->
-- **Conditional workaround only:** fall back to one tool call per assistant message ONLY when malformed/concatenated tool-input errors are observed or the harness version is known-affected — otherwise keep using legitimate parallel calls; serializing healthy hosts contradicts batching guidance.
+- **Conditional workaround only:** one tool call per assistant message only when malformed/concatenated tool-input errors are observed or the harness is known-affected — healthy hosts keep parallel calls.
 
 ## Long-Run Continuity
 <!-- RULES: LONG-01 LONG-02 -->
-- Near token-cap: **LAND one thing** end-to-end (edit + verify, ≤3 tool calls), **DEFER** the rest into `docs/STATUS.md` or bounded `docs/AI_HANDOFF/` tasks, **DELEGATE** broad work to subagents. Only then compact.
-- After any compact or handoff: continue from persisted disk state — never reread pre-compact context; delegate broad work, keep replies short.
-- A run ends only on completion evidence, a genuine blocker, or a user-only action — every stop names its reason. Detail: `docs/UKIT_INTERNALS.md`.
+- Near token-cap: **LAND one thing** end-to-end (edit + verify, ≤3 tool calls), **DEFER** the rest to `docs/STATUS.md` or bounded `docs/AI_HANDOFF/` tasks, **DELEGATE** broad work to subagents. Then compact.
+- After compact/handoff: continue from persisted disk state — never reread pre-compact context; delegate broadly, keep replies short.
+- A run ends only on completion evidence, a genuine blocker, or a user-only action — every stop names its reason.
 
 ## Index-First Loop
-For any task needing code context:
-
 <!-- RULE: IDX-01 -->
-1. Check the index is fresh (`.cache/index/`); if stale/missing, refresh via `node .claude/ukit/index/refresh-index.mjs`.
-2. Query files: `node .claude/ukit/index/query-index.mjs "<error|symbol|path>"`; bug signatures: `triage.mjs "<error signature>"`.
+1. Check the index is fresh (`.cache/index/`); stale/missing → `node .claude/ukit/index/refresh-index.mjs`. Query: `query-index.mjs "<error|symbol|path>"`; bugs: `triage.mjs "<error>"`.
 <!-- RULE: IDX-02 -->
-3. Open only the **top 1-3 suspect files first**, then widen if needed — the `outline:` block lets you jump straight to `Read(file, offset=<line>)`.
+2. Open only the **top 1-3 suspect files first**, then widen — `outline:` jumps to `Read(file, offset=<line>)`.
 <!-- RULE: IDX-03 -->
-   The outline locates code; it does not describe behaviour — **any code you are about to change must still be read**.
-4. For analog/reuse patterns, check `resolve-context`. Non-code lanes (docs-only, status, task queue) skip the source-code index.
+   The outline locates code, not behaviour — **code you are about to change must still be read**. Analog/reuse → `resolve-context`; non-code lanes skip.
 
 ## Automatic Skill Activation (mandatory)
 <!-- RULE: SKILL-01 -->
-- On every non-trivial task — and again after the first relevant tool calls — inspect installed project-local skills and **auto-activate the matching skill immediately**; end users should not need skill names. Match from prompt wording and tool/file evidence.
+- On every non-trivial task — and again after the first relevant tool calls — **auto-activate the matching skill immediately**; users never name skills. Match from prompt wording and tool/file evidence.
 <!-- RULES: SKILL-02 SKILL-03 -->
-- Use the smallest effective set (usually 1-2 skills); if evidence sharpens, upgrade the active skill choice immediately.
-- Prefer routed context/verification over ad-hoc broad reading; reuse `.claude/ukit/skill-router-state.json` compact route memory.
+- Smallest effective set (1-2 skills); upgrade as evidence sharpens. Prefer routed context/verification over broad reading; reuse `.claude/ukit/skill-router-state.json` route memory.
 
 ### Common skill triggers
-
-- review / audit / diff / PR feedback → `code-review` · bug / error / crash / triage → `debugging-toolkit` · test / spec / coverage / fixture → `testing-quality`
-- docs / README / changelog / handoff / `docs/` edits → `docs-quality` (`.claude/skills/docs-quality/SKILL.md`) · open-ended / status / continue / queued task → `next-step` · wrap up / `docs/STATUS.md` update → `update-status`
-- auth / security / token / permission / risky work → `discover-security` · stale workspace / reinstall / cleanup → `repo-maintenance` (all under `.claude/skills/<name>/SKILL.md`)
+- review/audit/diff → `code-review` · bug/error/crash → `debugging-toolkit` · test/coverage → `testing-quality` · docs/README/`docs/` → `docs-quality` (`.claude/skills/docs-quality/SKILL.md`) · open-ended/continue → `next-step` · wrap-up/`docs/STATUS.md` → `update-status` · auth/security/risky → `discover-security` · stale workspace/cleanup → `repo-maintenance`
 
 ## Internal Helper Policy
 <!-- RULES: HELP-01 HELP-02 -->
-- Prefer the internal index helpers (`node .claude/ukit/index/route-task.mjs`, `resolve-context.mjs`, `verify-context.mjs`) for routing, related-file context, and verification.
+- Prefer the internal index helpers (`node .claude/ukit/index/route-task.mjs`, `resolve-context.mjs`, `verify-context.mjs`) for routing, context, verification.
 - **Do not ask normal contributors to run internal helper commands** or memorize maintainer commands (`ukit doctor`, `ukit diff`, `ukit uninstall`) — run them yourself.
 <!-- RULE: FALLBACK-01 -->
-- Missing/corrupt runtime files or stale workspace → tell maintainers to rerun `ukit install`. Detail: `docs/UKIT_INTERNALS.md`.
+- Missing/corrupt runtime or stale workspace → rerun `ukit install`.
 
 ## Skill Quality (maintainer-only)
-- When editing a template skill/agent under `templates/.claude/`, read `.claude/skills/skill-quality/SKILL.md` before shipping.
+- Editing `templates/.claude/` skills/agents → read `.claude/skills/skill-quality/SKILL.md` first.
 
 ## UKit v{{ukit.version}} Shared Runtime
-- Runtime state lives in `.ukit/storage/`; `.ukit/storage/config.json` holds runtime toggles (compact, token pipeline, router, memory, validation, Safe Patch).
-- Reuse `.ukit/storage/memory/` + `ukit memory recall "<current task>"` before asking users to restate decisions; `ukit memory learn` surfaces pending proposals, `ukit memory promote` writes approved rules to MEMORY.md, `ukit memory episode` records session episodes; inspect via `ukit status` / `ukit memory export`.
-- Route memory: `.claude/ukit/skill-router-state.json` — reuse compact `previous-context`/`recent-output` first. Cache state: `.ukit/storage/cache/output-history.json`, `retriever-lanes.jsonl`, `tee/`. Lifecycle hooks all degrade to exit 0 (SessionEnd `session-episode.sh` auto-writes episodes; full map: `docs/UKIT_INTERNALS.md`). Missing/corrupt runtime or old `ukit/` root → rerun `ukit install`. Detail: `docs/UKIT_INTERNALS.md`.
-
+- Runtime state: `.ukit/storage/`; `config.json` holds toggles (compact, token pipeline, router, memory, validation, Safe Patch). `ukit memory recall "<task>"` before asking users to restate decisions; `learn`/`promote`/`episode` manage memory; inspect via `ukit status` / `ukit memory export`. Route memory: `.claude/ukit/skill-router-state.json` (`previous-context`/`recent-output` first); cache: `.ukit/storage/cache/output-history.json` + `retriever-lanes.jsonl` + `tee/`. Hooks degrade to exit 0; missing/corrupt runtime → rerun `ukit install`.
 ## Prompt Caching
 <!-- RULES: CTX-01 CTX-02 CTX-03 CTX-04 CTX-05 CTX-06 CTX-07 CTX-08 CTX-09 CTX-10 -->
-- Full ruleset: `docs/PROMPT_CACHING.md` (read on demand; not loaded every session).
-- CTX-01 deterministic segment bytes · CTX-02 keep roles and order · CTX-03 keep tool IDs and continuation state · CTX-04 no clock/random IDs in static blocks · CTX-05 compaction starts a new epoch · CTX-06 never change data to match a cache · CTX-07 no unconfirmed cache fields · CTX-08 tool-result reuse needs valid freshness · CTX-09 missing usage is unknown, not zero · CTX-10 never cut a required check to reduce calls.
+- Full ruleset: `docs/PROMPT_CACHING.md` (read on demand).
 
 ## Safe Patch Protocol
 <!-- RULES: SAFE-02 SAFE-03 SAFE-01 -->
-- Risky/shared/large edits: prefer unique current-file anchors over line numbers or stale pasted blocks; if `old_string` is missing/ambiguous, re-read current source and ask whether to apply as-is, adapt, or skip.
-- Preserve UTF-8 BOM/no-BOM and LF/CRLF for existing multilingual/user-authored files. Helper: `node .claude/ukit/index/safe-patch.mjs`; detail: `docs/UKIT_INTERNALS.md`.
+- Risky/shared/large edits: prefer unique current-file anchors over line numbers or stale pasted blocks; missing/ambiguous `old_string` → re-read source, then apply, adapt, or skip.
+- Preserve UTF-8 BOM/no-BOM and LF/CRLF on existing files. Helper: `node .claude/ukit/index/safe-patch.mjs`.
 
 ## Handoff Quality Gate — OPT-IN
 <!-- RULE: HAND-01 -->
-CHỈ kích hoạt khi task đi qua `docs/AI_HANDOFF/` (user nói "execute task TASK-xxx" hoặc target là `docs/AI_HANDOFF/tasks/*.md`). Daily prompt → KHÔNG đụng, flow cũ giữ nguyên. Khi Handoff mode: đọc `docs/AI_HANDOFF/RULES.md` để biết 4 phase (Idea+Plan → Create Tasks → Implement+Test → Review+Test) + state machine + self-report model. Config: `.ukit/storage/config.json` → `handoff.*`.
+Chỉ active khi task đi qua `docs/AI_HANDOFF/` ("execute task TASK-xxx" hoặc `docs/AI_HANDOFF/tasks/*.md`); daily prompt giữ flow cũ. Khi active: đọc `docs/AI_HANDOFF/RULES.md` (4 phase + state machine + self-report). Config: `.ukit/storage/config.json` → `handoff.*`.
 
 ## Context + Verification Budget
 <!-- RULE: BUDGET-01 -->
-- **Trivial**: no docs, no index query unless the file target is unclear. **Simple**: `docs/MEMORY.md` only + resolver-selected files/tests. **Non-trivial**: `docs/MEMORY.md` + `docs/PROJECT.md` + `docs/CODE_MAP.md`.
+- **Trivial**: no docs, no index query unless the target is unclear. **Simple**: `docs/MEMORY.md` + resolver-selected files/tests. **Non-trivial**: + `docs/PROJECT.md` + `docs/CODE_MAP.md`.
 <!-- RULES: BUDGET-02 BUDGET-03 -->
-- `docs/STATUS.md` for open-ended/continue prompts; `docs/TASKS.md` only for queued-task prompts; `docs/WORKLOG.md` recent entries only (archive overflow). Verification: targeted first, widen only on risk/shared scope, ask before blanket broad runs.
+- `docs/STATUS.md` for open-ended/continue; `docs/TASKS.md` only for queued-task prompts; `docs/WORKLOG.md` recent entries only. Verification: targeted first, widen on risk/shared scope, ask before broad runs.
 
 ## Living Status Workflow
 <!-- RULE: STATUS-01 -->
-- `docs/STATUS.md` captures compact current state — not source truth, never replaces source/index-first investigation.
-- "What next?"/"continue" → `next-step` with a freshness cue; after meaningful work → `update-status`. `docs/TASKS.md` is the local AI task queue — prefer `Ready for AI`. Detail: `docs/UKIT_INTERNALS.md`.
+- `docs/STATUS.md` = compact current state — not source truth, never replaces index-first investigation.
+- "What next?"/"continue" → `next-step`; after meaningful work → `update-status`. `docs/TASKS.md` = local AI task queue — prefer `Ready for AI`.
 
 ## Subagent Lanes (internal)
 <!-- RULE: SUBAG-02 -->
-- The `ukit-small-task-maintainer` subagent (`subagents.smallTaskModel`, default `unic-lite`) handles safe/reversible UKit chores as a sidecar lane — never block or slow the user task; risky work hands back to the main model. Detail: `docs/UKIT_INTERNALS.md`.
-- When routed state's `routeSummary.line` carries `review=code-reviewer(diff)`, launch `code-reviewer` in background (`smart` tier) **only after** write + verification evidence; findings advisory — never block the reported completion. Detail: `docs/UKIT_INTERNALS.md`.
+- `ukit-small-task-maintainer` (`subagents.smallTaskModel`, default `unic-lite`) = sidecar lane for safe/reversible UKit chores — never block the user task; risky work hands back.
 <!-- RULE: SUBAG-01 -->
-- Direct execution is default for trivial/simple work; delegate only on meaningful context shrink or parallel gains (noisy side lanes, 3+ independent failures, batch plans). Never ask end users to name agents or remember agent commands.
+- Direct execution is default for trivial/simple work; delegate only on real context shrink or parallel gains. Never ask users to name agents or agent commands.
 
 ## Adaptive Autonomy
 <!-- RULE: AUTO-01 -->
-- `autonomy.level` in `.ukit/storage/config.json` controls how much UKit acts without asking: `conservative` (ask more), `balanced` (default), `free-run` (auto-run more), `vibecode` (one prompt to a finished result; the gate stops only on completion evidence, a genuine blocker, or a dangerous-command decision). End users should not need to change it.
+- `autonomy.level` in `.ukit/storage/config.json`: `conservative` | `balanced` (default) | `free-run` | `vibecode` (one prompt to a finished result; stops only on completion evidence, a genuine blocker, or a dangerous-command decision).
 
 ## Unattended Completion Loop
 <!-- RULE: UNATTENDED-01 -->
-- Under `orchestration.permissionMode: unattended` there is no human in the loop — the loop is UNDERSTAND → PLAN → IMPLEMENT → VERIFY → FIX → REVERIFY → COMPLETE. A turn ends only on COMPLETE evidence, a real external blocker, or a deny decision.
+- `orchestration.permissionMode: unattended` = no human in the loop — UNDERSTAND → PLAN → IMPLEMENT → VERIFY → FIX → REVERIFY → COMPLETE. A turn ends only on COMPLETE evidence, a real external blocker, or a deny.
 <!-- RULE: UNATTENDED-02 -->
-- `maxAttemptsPerFailure: 5` per failing verification, `maxRecoveryStrategies: 3` distinct fix strategies before reassessing — retry the same edit, fix the cause differently, then reduce scope/approach; do not loop one strategy.
+- `maxAttemptsPerFailure: 2`, `maxRecoveryStrategies: 2` distinct fix strategies before reassessing — retry the same edit, then fix the cause differently; never loop one strategy.
 <!-- RULE: UNATTENDED-03 -->
-- BLOCKED only on true external blockers: missing credentials, dead service, unreachable network/API. Test failures, lint errors, review findings, and ambiguity are NOT blockers — they feed FIX.
+- BLOCKED only on true external blockers: missing credentials, dead service, unreachable network/API. Test failures, lint errors, review findings, ambiguity are NOT blockers — feed FIX.
 <!-- RULE: UNATTENDED-04 -->
-- A deny (hook/gate/permission) never becomes a question: pick the next safer alternative that satisfies the intent and record the substitution. Never emit "do you want me to…" — decide and proceed.
+- A deny (hook/gate/permission) never becomes a question: pick the next safer alternative satisfying the intent, record the substitution. Never emit "do you want me to…" — decide and proceed.
 
 ## 3-Tier Model Routing
 <!-- RULES: TIER-01 TIER-02 -->
-**Internal orchestration only — end users still just use natural language. No new commands.**
+**Internal orchestration only — end users still use natural language. No new commands.**
 
-| Tier | Generic alias | Claude model | Typical tasks |
-|------|--------------|--------------|---------------|
-| lite | `unic-lite` | claude-haiku | Reads, git queries, bash summaries, small doc edits |
-| code | `unic-code` | claude-sonnet | Normal coding, local fixes, shared edits, builds, debugging, impact mapping |
-| smart | `unic-smart` | claude-opus | Release review/audit, escalated deep reasoning after repeated failure |
+| Tier | Alias | Claude model |
+|------|-------|--------------|
+| lite | `unic-lite` | claude-haiku |
+| code | `unic-code` | claude-sonnet |
+| smart | `unic-smart` | claude-opus |
 
-- A tier takes effect only when work is handed to an agent whose definition binds that model (`model:` frontmatter in `.claude/agents/*.md`; `model:` `@smol`/`@default`/`@slow`/`@vision` in `.omp/agents/*.md` via `modelRoles`) — the main session model never changes mid-turn.
-- Contract map: `tiny-fix` → lite · `local-fix`, `local-build`, `shared-edit`, `find-cause`, `map-impact` → code · `review-release` → smart. Escalation: same file/symbol failing `debugLoopThreshold` (default 2) times routes the next attempt one tier higher, capped at `smart`.
-- `unic-vision` is a capability lane, not a cost tier — unverified vision must never guess image contents; route images to `ukit-vision-analyst`. Harness table: `docs/UKIT_INTERNALS.md`.
+- A tier binds only via an agent's `model:` field (`.claude/agents/*.md`; `.omp/agents/*.md` `@smol`/`@default`/`@slow`/`@vision` via `modelRoles`) — the session model never changes mid-turn. Contract map: `tiny-fix` → lite · `local-fix`/`local-build`/`shared-edit`/`find-cause`/`map-impact` → code · `review-release` → smart; escalation: same file/symbol failing `debugLoopThreshold` (2) → one tier up, cap `smart`. `unic-vision` = capability lane (not cost) — never guess images; route to `ukit-vision-analyst`.
 
 ## Skills
-- Canonical skills live in `.claude/skills/`; adapter mirrors may exist (`.codex/skills/` → symlink). **omp** reads `.claude/skills/` via its `claude` discovery provider.
+- Canonical skills: `.claude/skills/`; adapter mirrors may exist (`.codex/skills/` → symlink). **omp** reads `.claude/skills/` via its `claude` provider.
 
 ## Project Snapshot
-- Project: {{project.name}} | Root: {{project.root}}
-- Packs: {{project.stack}} | Frontend: {{stack.frontend}} | Backend API: {{stack.backendApi}} | PostgreSQL: {{stack.postgres}}
-- Package manager: {{runtime.packageManager}} | OS: {{runtime.os}} | Node: {{runtime.nodeVersion}} | Provider: {{providers.unic}}
+- Project: {{project.name}} | Root: {{project.root}} | Packs: {{project.stack}} | PM: {{runtime.packageManager}} | OS: {{runtime.os}} | Node: {{runtime.nodeVersion}} | Provider: {{providers.unic}}
 
 ## Working Rules
-- Keep scope tight, prefer the smallest correct change set, and reuse existing code.
-- Update `docs/WORKLOG.md` after significant work; if source contradicts docs, update docs immediately.
-- Use `{{runtime.packageManager}}`.
+- Keep scope tight; smallest correct change set; reuse existing code. Update `docs/WORKLOG.md` after significant work; source contradicts docs → update docs. Use `{{runtime.packageManager}}`.
 
 ## DuraOne Skill — Conditional Activation
 <!-- RULE: DURA-01 -->
-DuraOne skill chỉ active khi pack `duraone` được cài hoặc `.claude/skills/duraone/SKILL.md` tồn tại — khi active, luôn đọc SKILL.md + references trước khi code; khi không, dùng generic standards + index patterns. Chi tiết: `docs/UKIT_INTERNALS.md`.
+DuraOne skill chỉ active khi pack `duraone` được cài hoặc `.claude/skills/duraone/SKILL.md` tồn tại — khi active, đọc SKILL.md + references trước khi code; khi không, dùng generic standards + index patterns.
 
 ## Completion Checklist
-- Requirements implemented · No unrelated changes · Verification executed and reported · Docs updated when source truth changed.
+- Requirements implemented · No unrelated changes · Verification executed + reported · Docs updated when source truth changed.
 
 ## Handoff Fullstack Rules
 <!-- RULE: HAND-02 -->
-- `docs/AI_HANDOFF/RUN.md` là run cursor có thẩm quyền; `Phase:` ≠ `done`/`blocked` nghĩa là run còn sống — Stop gate từ chối stop và trả về `Next:` step.
-- Recap/checkpoint không bao giờ là completion — chỉ `HANDOFF FULLSTACK COMPLETE` (sau `Phase: done`) hoặc `HANDOFF FULLSTACK BLOCKED` (sau `Phase: blocked`) mới kết thúc run. Resume tự động mọi task chưa xong (current, legacy, pending, interrupted, recovery `-R<n>`).
-- Kết thúc cycle: docs sync → archive `docs/AI_HANDOFF/archive/cycle-NN/` → `Phase: done`. `handoff-clear` bắt buộc đóng RUN.md. Full rules: `docs/AI_HANDOFF/RULES.md`.
+- `docs/AI_HANDOFF/RUN.md` là run cursor có thẩm quyền; `Phase:` ≠ `done`/`blocked` = run còn sống — Stop gate từ chối stop, trả về `Next:`. Chỉ `HANDOFF FULLSTACK COMPLETE`/`BLOCKED` kết thúc run — recap/checkpoint không phải completion; resume tự động mọi task chưa xong. Kết thúc cycle: docs sync → archive `archive/cycle-NN/` → `Phase: done`; `handoff-clear` bắt buộc đóng RUN.md. Full rules: `docs/AI_HANDOFF/RULES.md`.
 
 ## Compact Instructions
-Khi compact giữa một handoff run: giữ lại goal, RUN.md path + phase hiện tại, task inventory, task đang làm, `Next:` step, blockers, verification evidence, commits, worktree/copy-back state, và quy tắc "compact không phải completion". Sau compact: đọc lại RUN.md + INDEX.md rồi chạy tiếp `Next:` ngay.
+Compact giữa handoff run: giữ goal, RUN.md path + phase, task inventory + task đang làm, `Next:` step, blockers, verification evidence, commits, worktree/copy-back state, "compact không phải completion". Sau compact: đọc lại RUN.md + INDEX.md rồi chạy `Next:`.

package/templates/ukit/storage/config.json

@@ -75,19 +75,17 @@
   "tokenPipeline": {
     "inputCompression": true,
     "outputCompression": true,
-    "promptCache": true
+    "promptCache": true,
+    "outputMaxTokens": 600,
+    "outputMaxLines": 40
   },
   "router": {
     "enabled": true,
-    "defaultModel": "claude-sonnet-5",
-    "advisorModel": "claude-opus-5",
-    "advisorEnabled": true,
-    "maxAdvisorCalls": 3
+    "defaultModel": "claude-sonnet-5"
   },
   "orchestration": {
     "enabled": true,
     "orchestratorModel": "claude-sonnet-5",
-    "advisorEnabled": true,
     "permissionMode": "unattended",
     "allowUnsafeEval": false,
     "contracts": {
@@ -110,8 +108,7 @@
         "maxContextPulls": 1,
         "verificationPolicy": "targeted-if-covered",
         "completionRule": "require-write-and-verification",
-        "delegationPolicy": "disallow-by-default",
-        "postEditReviewPolicy": "sidecar-non-blocking"
+        "delegationPolicy": "disallow-by-default"
       },
       "find-cause": {
         "maxReadPassesBeforeReassess": 3,
@@ -124,8 +121,7 @@
         "maxContextPulls": 2,
         "verificationPolicy": "targeted-then-widen-on-risk",
         "completionRule": "require-write-and-verification",
-        "delegationPolicy": "allow-qualified-sidecar",
-        "postEditReviewPolicy": "sidecar-non-blocking"
+        "delegationPolicy": "allow-qualified-sidecar"
       },
       "map-impact": {
         "maxReadPasses": 3,
@@ -213,9 +209,8 @@
   },
   "memory": {
     "enabled": true,
-    "autoCapture": true,
     "progressiveRetrieval": true,
-    "maxInjectionTokens": 1000,
+    "maxInjectionTokens": 320,
     "archiveAfterDays": 30,
     "maxSessions": 20,
     "redactSecrets": true
@@ -326,9 +321,6 @@
     "enabled": true,
     "smallTaskModel": "unic-lite",
     "smallTaskAgent": "ukit-small-task-maintainer",
-    "diffReviewEnabled": true,
-    "diffReviewAgent": "code-reviewer",
-    "diffReviewModel": "unic-smart",
     "visionEnabled": true,
     "visionModel": "unic-vision",
     "visionAgent": "ukit-vision-analyst",
@@ -537,28 +529,25 @@
     "tokenPipeline": {
       "inputCompression": "Nén/tái sử dụng input context khi an toàn.",
       "outputCompression": "Tóm tắt output command ồn nhưng vẫn giữ recovery hint.",
-      "promptCache": "Dùng cache prompt/context compact để giảm lặp lại."
+      "promptCache": "Dùng cache prompt/context compact để giảm lặp lại.",
+      "outputMaxTokens": "Ngân sách token tối đa cho output command sau khi nén. Output nhỏ hơn ngưỡng này được bỏ qua nén hoàn toàn. Mặc định 600.",
+      "outputMaxLines": "Số dòng tối đa trong bản tóm tắt output đã nén. Mặc định 40."
     },
     "router": {
       "enabled": "Bật task routing nội bộ.",
-      "defaultModel": "Model hint cân bằng cho công việc bình thường. Adapter/provider có thể map khác.",
-      "advisorModel": "Model hint mạnh hơn cho planning/reasoning khó khi được bật.",
-      "advisorEnabled": "Cho phép advisor-style planning cho task phức tạp.",
-      "maxAdvisorCalls": "Giới hạn số lần gọi advisor để không lãng phí."
+      "defaultModel": "Model hint cân bằng cho công việc bình thường. Adapter/provider có thể map khác."
     },
     "orchestration": {
       "enabled": "Bật ladder điều phối nội bộ của UKit. Ladder này thay cách nghĩ LITE/FULL cứng nhắc bằng các tầng tiny-fix → review-release an toàn hơn.",
       "orchestratorModel": "Model điều phối chất lượng cao để chọn execution layer. Đây là model quan trọng cho chất lượng completion, không nên dùng model quá yếu.",
-      "advisorEnabled": "Cho phép UKit dùng advisor-style reasoning khi chọn layer trong tình huống mơ hồ.",
       "permissionMode": "Chế độ quyền của orchestration. Ba giá trị dành riêng: interactive, safe-auto, unattended. Chu kỳ này chỉ triển khai unattended.",
       "allowUnsafeEval": "Mặc định false. Chỉ khi user tự bật true, .omp/config.yml mới render tools.approval.eval: allow; mọi trường hợp khác render deny (fail-closed, không bao giờ prompt). Residual risk: eval chạy payload subprocess (os.remove, shutil.rmtree) mà bash.patterns text deny không chặn được — chỉ dựa vào bridge eval->Bash của TASK-004.",
       "contracts": "Bộ contract cho 7 tầng điều phối: tiny-fix, local-fix, local-build, find-cause, shared-edit, map-impact, review-release. Khi mơ hồ, UKit nên lệch lên tầng an toàn hơn dù tốn token hơn."
     },
     "memory": {
       "enabled": "Bật memory local của UKit.",
-      "autoCapture": "Cho phép UKit lưu decision/rule bền vững khi an toàn.",
       "progressiveRetrieval": "Lấy memory liên quan nhỏ trước, chỉ mở rộng khi cần.",
-      "maxInjectionTokens": "Số token memory tối đa được inject vào context.",
+      "maxInjectionTokens": "Số token memory tối đa được inject vào context (clamp thực tế trong khoảng 160–640).",
       "archiveAfterDays": "Sau bao nhiêu ngày memory có thể được archive/compact.",
       "maxSessions": "Số session memory gần đây giữ active.",
       "redactSecrets": "Cố tránh lưu secret vào memory."