@ngockhoale/ukit 3.5.1 → 3.5.3
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/CHANGELOG.md +35 -297
- package/README.md +1 -0
- package/manifests/platform.full.yaml +23 -0
- package/package.json +8 -1
- package/src/cli/commands/doctor.js +14 -2
- package/src/cli/commands/memory.js +30 -2
- package/src/core/memory/memoryHit.js +2 -0
- package/src/core/memory/mutateMemory.js +9 -0
- package/src/core/memory/projectIdentity.js +9 -2
- package/src/core/memory/recallEngine.js +86 -4
- package/src/core/memory/registryPrune.js +230 -0
- package/src/core/memory/retrieval.js +12 -3
- package/src/core/runtimeConfig.js +1 -0
- package/src/index/buildIndex.js +22 -0
- package/src/index/impactContext.js +11 -1
- package/src/index/queryIndex.js +151 -20
- package/template_project/.claude/hooks/edit-hint.mjs +119 -0
- package/template_project/.claude/settings.json +2 -2
- package/template_project/.claude/ukit/index/edit-graph.mjs +404 -0
- package/template_project/.claude/ukit/index/lib/index-core.mjs +185 -20
- package/template_project/.claude/ukit/index/post-edit-verify.mjs +8 -5
- package/template_project/.claude/ukit/index/resolve-context.mjs +17 -0
- package/template_project/.claude/ukit/index/route-task.mjs +4 -1
- package/template_project/.claude/ukit/runtime/hook-chain-runner.mjs +62 -5
- package/template_project/.claude/ukit/runtime/memory-lanes.mjs +259 -0
- package/template_project/.claude/ukit/runtime/memory-recall.mjs +86 -4
- package/template_project/.claude/ukit/runtime/memory-validity.mjs +232 -0
- package/template_project/.claude/ukit/runtime/reinject-context.mjs +4 -2
- package/template_project/.omp/hooks/pre/ukit-bridge.js +2 -1
- package/template_project/AGENTS.md +1 -1
- package/template_project/CLAUDE.md +1 -1
- package/template_project/instructions/core.md +1 -1
- package/src/core/observability/evaluation/experimentPlan.js +0 -95
- package/src/core/observability/evaluation/perturbation.js +0 -273
- package/src/diagnostics/classifyHang.js +0 -246
- package/src/diagnostics/ompStallSnapshot.js +0 -199
- package/src/render/instructionRenderer.js +0 -290
- package/src/skill/auditSkill.js +0 -140
- package/template_project/instructions/overlays/repo.md +0 -9
package/CHANGELOG.md
CHANGED
|
@@ -2,6 +2,40 @@
|
|
|
2
2
|
|
|
3
3
|
All notable changes to UKit are documented here.
|
|
4
4
|
|
|
5
|
+
## 3.5.3 - 2026-11-15
|
|
6
|
+
|
|
7
|
+
Source index ranking, and memory that knows when the code it describes has changed. Every number below was measured on this repository only.
|
|
8
|
+
|
|
9
|
+
- **Index search ranks the right file first far more often.** `query-index` (and everything that uses it: routing, triage, `resolve-context`) has a new scoring formula. An exact symbol name dominates; one file can no longer fill the list with many weak symbol hits (the best symbol counts in full, further ones are discounted); rare query words weigh more than common ones; substring matching only applies to words of 3+ characters; a path query matches the exact path, then a path suffix, then the basename; an original file ranks above its mirror copy (`template_project/`, `.claude/`, `.codex/`, `.omp/`); import-graph, hotspot and archetype boosts are capped so they cannot outrank a direct match.
|
|
10
|
+
- **Measured on a new 67-query gold set** (`scripts/bench/index-gold.mjs`, `scripts/bench/indexGold.json`; before = 3.5.2 scoring, frozen in `scripts/bench/baselines/index-gold.json`): right file first 16.4% → 61.2%, in the top 3 32.8% → 65.7%. Per group, first / top 3: exported symbol 41.7 → 100 / 66.7 → 100; path 8.3 → 100 / 8.3 → 100; English sentence 18.2 → 63.6 / 45.5 → 72.7; Vietnamese sentence 0 → 50 / 40 → 50; error string 20 → 40 / 30 → 50; unexported symbol 8.3 → 8.3 / 8.3 → 16.7. No group got worse. Median result size 7,001 → 3,788 bytes; median query time 2.9 → 3.8 ms. A test holds the new numbers as a floor.
|
|
11
|
+
- **Still weak, not fixed here:** unexported symbols and error strings, because the index only knows 32.7% of the function declarations under `src/` (752 of 2,299) and has no content search. A mirror still outranks its original in 3 of 12 queries that have one. Both are scheduled for 3.5.4.
|
|
12
|
+
- **Notes whose source changed are labelled.** When a memory note names a file, UKit records that file's state at write time (and for existing notes, the first time the CLI recalls them). If the file later changes or disappears, recall shows the note with `[source changed since noted — verify]` and ranks it slightly below equally relevant notes that are still current; a note whose file is gone ranks last. `ukit memory v2 confirm <record-id | path:line>` marks a note as re-checked. State lives in `.ukit/storage/memory/v2/anchors.json`; hooks only read it. Off: `UKIT_MEMORY_VALIDITY=0`. On the 48-question memory gold set the label and the mild ranking keep 37/38 answers in the top 3 and 10/10 abstentions when 6 of the 12 notes that name a file (of 68 notes) are marked changed; a strict "changed notes last" ranking was tried, lost an answer (36/38) and was not kept as default.
|
|
13
|
+
- **A record that another record supersedes is no longer returned** (`supersedes` on memory records; markdown bullets have no such field yet).
|
|
14
|
+
- **Test hints after an edit use git history.** The edit graph now mines which files changed together (last 500 commits, at index build, about 70 ms) and uses it to rank the suggested tests. Replay of 148 commits, top 3: per commit recall 0.39 → 0.43, precision 0.18 → 0.21, at least one right test in 97 → 116 commits. Per source file, recall is 0.06 → 0.07 (the best possible with three suggestions is 0.24) — the target of a clear rise was not met; test hints remain a starting point.
|
|
15
|
+
- **"Often changed together" line: shipped off.** `memory.editHints.coChange: true` adds one line of files that historically changed with the edited file. Top-3 precision was 0.51 overall but 0.27 on the older half of the history and 0.60 on the newer half, and no threshold kept both halves above 0.5, so it is not on by default.
|
|
16
|
+
- **Tried and removed:** a recall lane that resolves file and symbol names in the question and fuses the notes attached to those files with the text ranking. On 12 new file-naming questions it raised answers in the top 3 from 9 to 11, but 2 of 4 new no-answer questions still returned unrelated notes (with or without the lane) and the hook got about 8–10 ms slower at the median and about 35 ms at p95. The code is not in the package; the questions stay in the gold file as a separate, non-gating block.
|
|
17
|
+
- **The user-level project registry prunes itself.** `~/.ukit/storage/projects/registry.json` kept an entry for every directory UKit was ever installed into, including deleted temp directories; on the maintainer's machine 48,075 of 48,101 entries were dead (19.7 MB; counted by the maintainer, not by the independent review). On a synthetic registry of 48,001 entries a project registration took 272 ms. An entry is now removed only when every root and git directory is provably gone, its drive or volume is mounted, and no user-level memory record points at it. It runs at most once a day once the registry has 500+ entries, within 1.5 s, or on demand: `ukit doctor --prune-registry [--dry-run]`. After pruning the synthetic registry, registration took 2 ms.
|
|
18
|
+
- Engines: the index and memory changes are shared runtime and CLI code, used by Claude Code, omp (through the bridge) and Codex (through the CLI helpers). None of the three hosts was run live for this release; coverage is the tests and benches above.
|
|
19
|
+
- Repository only: `CHANGELOG.md` entries 3.4.76–3.4.85 moved verbatim to `docs/archive/CHANGELOG-3.4.76-3.4.85.md`, and six modules used only by repository scripts and tests plus the repo-only instruction overlay are no longer packed, to keep the package inside its size budget (limits unchanged).
|
|
20
|
+
|
|
21
|
+
## 3.5.2 - 2026-11-15
|
|
22
|
+
|
|
23
|
+
Edit hints. After the agent edits a file, UKit now tells it what else usually has to change with that file, instead of leaving the agent to find the mirror copy or the test on its own.
|
|
24
|
+
|
|
25
|
+
- **A hint after each successful Edit or Write, once per file per task.** It lists the file's mirror copies, the tests most likely to cover it, and project notes (from `docs/MEMORY.md` and memory records) that name that file, its mirror, its tests or its directory. Mirrors are byte-identical twins found at index time, the pairs UKit already declares, and any pair you add under `memory.editHints.mirrors` in `.ukit/storage/config.json`. A pair that was a mirror and has since drifted is still listed, labelled as drifted. List a pair under `memory.editHints.ignoreMirrors` to silence it.
|
|
26
|
+
- **Delivery per engine.** Claude Code: PostToolUse `additionalContext`, which reaches the model with the tool result. omp: the bridge appends the hint as text to the tool result. Codex has no hooks, so `node .claude/ukit/index/resolve-context.mjs "<intent>" --target <file>` prints the same block, and the instructions tell the agent to run it before editing.
|
|
27
|
+
- **New index artifact `.cache/index/edit-graph.json`**, rebuilt with the index only when the index changed. On this repository it is 289 KB; the graph step adds about 218 ms to a one-file refresh, and a full index build with no graph took about 1.2 s (both figures are for this repository only). A per-file token cache was tried to cut the 218 ms and gave no measurable gain in total refresh time, so it was not kept.
|
|
28
|
+
- **`resolve-context` impact output now includes the graph's mirrors**, next to the mirror pairs it already declared.
|
|
29
|
+
- **Off switches:** `UKIT_EDIT_HINTS=0` in the environment, or `memory.editHints.enabled: false`.
|
|
30
|
+
- **Hook runner:** a context-only document (PostToolUse `hookSpecificOutput` with no `permissionDecision` and no top-level `decision`, `continue: false` or `stopReason`) no longer ends the chain's output; it is merged with the other steps' output, and plain-text lines printed alongside it are folded into its `additionalContext` so Claude Code still receives one valid JSON document. Everything else behaves as before.
|
|
31
|
+
- **Safe Patch OK line:** `post-edit-verify` now prints its `[ukit-safe-patch] OK: ...` line as a `systemMessage` JSON document on Claude Code (it was plain text, which made the merged hook output invalid JSON and lost the hint on Safe Patch-backed files). omp still gets the plain line.
|
|
32
|
+
- Measured by replaying this repository's last 400 commits against each commit's parent tree (`scripts/bench/edit-hint-history.mjs`; repository only): mirrors recall 34/34 and precision 34/35. Tests, top 3 per commit: recall 0.39, precision 0.18, and at least one right test in 97 of 148 commits; per source file, recall is 0.06. Test hints are a starting point, not a complete list. The replay does not use importing-test evidence, which the shipped ranking also uses.
|
|
33
|
+
- Memory attached to code is thin so far: in this repository 13 of 66 `docs/MEMORY.md` bullets name a file that exists, and 0 of 9 memory records do, so most hints carry mirrors and tests only.
|
|
34
|
+
- Limits: `MultiEdit` and `NotebookEdit` get no hint. A hint dropped by context compaction is not re-sent in the same task. The step has a 400 ms budget (250 ms of it for memory) and stays silent rather than delay the edit; measured 164-199 ms through the real hook runner on the maintainer's machine, with the hint delivered together with memory lines.
|
|
35
|
+
- Engines: observed live on Claude Code for Edit: editing `docs/STATUS.md` and `docs/MEMORY.md` in this repository put the hint into the model's context as PostToolUse additional context, once per file (a second edit of the same file produced none). Codex and omp were not run live; omp is covered through the real bridge function with the real runner, not the omp host. Coverage is otherwise the shared runtime's tests.
|
|
36
|
+
- The step keeps its per-task state small: once there are more than 64 state files it removes those older than 7 days, at most 32 per run.
|
|
37
|
+
- Repository only: `CHANGELOG.md` entries 3.4.51-3.4.75 moved verbatim to `docs/archive/CHANGELOG-3.4.51-3.4.75.md` to keep the package inside its size budget (limits unchanged).
|
|
38
|
+
|
|
5
39
|
## 3.5.1 - 2026-11-14
|
|
6
40
|
|
|
7
41
|
First release of the memory line (3.5.x). Recall now returns the project's written decisions instead of content-free session stubs. Design and evidence: `docs/research/2026-11-14_memory-system-research.md` (repository only).
|
|
@@ -16,302 +50,6 @@ First release of the memory line (3.5.x). Recall now returns the project's writt
|
|
|
16
50
|
- Engines: Claude Code and omp use all lanes (omp through the Claude runtime bridge). Codex has no UKit hooks, so it gets the new recall through `ukit memory recall`, and its episodes are recorded by the idle-ledger backfill rather than at session end. Codex and omp were not run live for this release; coverage is the shared runtime's tests.
|
|
17
51
|
- Repository only: `CHANGELOG.md` entries 2.7.0–2.8.0 and 3.4.0–3.4.50 (except 3.4.14 and 3.4.22) moved verbatim to `docs/archive/` to keep the package inside its size budget (limits unchanged). The memory module inventory moved to `docs/design/memory-io-inventory.md`.
|
|
18
52
|
|
|
19
|
-
## 3.4.85 - 2026-11-14
|
|
20
|
-
|
|
21
|
-
- Release gate (repository only): **the publish guard no longer has an owner-authorization prerequisite.** Since 3.4.72 `npm publish` also required a fresh owner chat message naming the exact version, so finished, verified and reviewed work stopped to ask for a sentence. The project owner did not want that step: a release is a new version number for users to check. The per-version message check and the standing-grant form added in 3.4.84 are both removed; a receipt needs no `authorization` block and a leftover one is ignored. Still mandatory: an official `release:verify` PASS on the exact candidate (unchanged), and an independent review whose final verdict is exactly APPROVED, derived from a native reviewer session. One review rule replaces a side effect of the removed step: the owner-session comparison was also what kept the reviewer apart from the working session, so the cited review message must now come from a separate reviewer agent's own transcript (`<session>/subagents/agent-<id>.jsonl`); a message from a main session is refused. `PROJECT_IMPORTANT.md` §4.4 now records the rule (finished work ships without asking; the two quality gates stay).
|
|
22
|
-
- Tests: `tests/release/publishGuard.test.js` drops the owner-message and standing-grant cases; the session-store provenance cases (unknown uuid, hash mismatch, foreign store, cwd binding, session metadata, text blocks) now run against the reviewer message. New cases: a main-session reviewer message refuses; a reviewer agent working in a linked worktree passes.
|
|
23
|
-
|
|
24
|
-
## 3.4.84 - 2026-11-14
|
|
25
|
-
|
|
26
|
-
Follow-up to the 3.4.83 hook audit: the items that release left open.
|
|
27
|
-
|
|
28
|
-
- Fix: **a command that only mentions a test runner no longer counts as verification.** The completion gate's classifier matched runner names anywhere in the command text, so `grep -n pytest file`, `ls node_modules/.bin/vitest`, `echo yarn test`, `yarn add -D vitest`, `npm install jest` or `which vitest`, run with exit 0, were recorded as a verification receipt with a satisfying `test-verified` verdict — a task could be declared verified with no test run. Each simple command is now judged by the word that executes: direct runners (`vitest`, `jest`, `mocha`, `ava`, `pytest`, `tsc`, `eslint`, `biome`, `ruff`, `mypy`), package-manager scripts (`test`, `test:*`, `lint`, `typecheck`, `check`, `build`, `release:verify`, via `run`/`exec`/`dlx`/`workspace` and `npx`), `python -m pytest`, `uv`/`poetry run …`, `node --test` / `node --check`. Leading assignments, the env/time/timeout/sudo/nohup/corepack wrappers and value-taking options (`yarn --cwd DIR …`, `pnpm -C DIR …`, `npm --prefix DIR …`) are skipped; arguments, paths, comments, quoted strings and heredoc bodies are never evidence. Newly recognised: `pnpm -C pkg test`, `npm --prefix pkg test`, `yarn workspace foo test`, `node node_modules/vitest/vitest.mjs run`. `[ vitest = vitest ]` / `test vitest = vitest` now produce no verification receipt at all.
|
|
29
|
-
- Fix: **`reset-auto-permissions.mjs --help` no longer performs a reset.** Arguments were checked with `includes`, so `--help` or any typo (`--dry-runn`) fell through to the real reset. `--help` / `-h` print usage and exit 0; any unknown argument exits 2; both return before a file is read.
|
|
30
|
-
- Fix: **decision-plane shadow calls stop timing out on a healthy endpoint.** The detached hook-path shadow child blocks no hook, yet shared the inline path's 3 s cap; an 8-question batch to the local `unic-decision` endpoint measured about 2 s typical with 3–6 s tails, so whole batches timed out and then opened the circuit breaker. The detached child now honours `decisionPlane.timeoutMs` up to 8 s; the inline path keeps 3 s and shadow answers still never change a route. The "five keys never answered" note in 3.4.83 was not reproduced: all eight asked keys returned valid answers in the audited week (five have no baseline to compare against and log `agree=unknown`).
|
|
31
|
-
- Fix: self-improve classifies `timeout-hang` from host evidence. A Bash receipt carries `timedOut` when Claude Code reports the call as a failure whose own error reads "Command timed out after …"; mined patterns carry the marker and `timeout-hang` accepts it. Exit 137/143 alone stay unclassified (137 is also the OOM-kill status). Codex and omp expose no equivalent field that could be confirmed, so their timeouts still rely on exit 124 or the text markers.
|
|
32
|
-
- Release gate (repository only): **the publish guard accepts a standing owner authorization.** Since 3.4.72 every publish needed a fresh owner chat message naming the exact version, so a finished, verified and reviewed fix still stopped to ask. The receipt may now cite `authorization: { by: 'user', standing: true }`; the guard then reads the grant from the "Standing publish authorization" section of the git-tracked `PROJECT_IMPORTANT.md` (regular file, committed and unmodified at HEAD, names the package, covers every version, one line that authorizes publishing, states the verification + review conditions, no denial, pending or negated grant). The verified run and the independent APPROVED review are unchanged and still mandatory; a receipt that cites a message is judged on that message only. The template `PROJECT_IMPORTANT.md` shipped to installed projects carries no grant.
|
|
33
|
-
- Tests: new `tests/hooks/verificationCommandWord.test.js`; argument-safety cases in `tests/hooks/c91ResetAutoPermissions.test.js`; timeout-budget cases in `tests/index/unicDecisionCircuit.test.js`; `timedOut` cases in the ledger, pattern and proposal tests; standing-authorization cases in `tests/release/publishGuard.test.js`. `tests/hooks/nativePublicationRouting.test.js` now reports router and gate output when its foreign-session assertion fails.
|
|
34
|
-
- Not changed: that test failed once under full-suite load in 3.4.83 and could not be reproduced (load average 60, slowed git, paused coordinator), so no product change was made for it. A related gap was found and is not fixed here: when the router is killed by its deadline before its first route write, or its session lock is busy, the following Stop evaluates no route and releases without a message.
|
|
35
|
-
|
|
36
|
-
## 3.4.83 - 2026-11-13
|
|
37
|
-
|
|
38
|
-
Hook fixes from a one-week audit of this repository's own Claude Code transcripts, hook-latency rows, decision log and permission audit log (6,685 hook runs, no hook timeouts; Stop p95 about 1.6s).
|
|
39
|
-
|
|
40
|
-
- Fix: **`auto-allow-bash` no longer mints junk permission rules.** The binary was picked by whitespace splitting, so a leading assignment with a quoted or substituted value was cut in half and the next fragment became the "binary": `UKIT_EVIDENCE='… HEAD …' yarn test` minted `Bash(HEAD:*)`, `d=$(ls -td …)` minted `Bash(-td:*)`. The hook now scans real shell words (single/double quotes, `$(…)` with nesting, backticks, backslash escapes), skips whole `NAME=value` / `NAME+=value` words, and mints nothing when the first command word cannot be determined (unbalanced quote or substitution, leading operator, redirection or subshell) or starts with `-`. Behaviour change: `command -v node` and `env -i FOO=1 ls` now mint nothing (they minted `-v` / `-i`). The denylist block, charset check, interpreter list and locked atomic write are unchanged. omp runs this same script through its bridge; Codex has no auto-allow lane.
|
|
41
|
-
- Fix: `reset-auto-permissions.mjs --sanitize` also removes rules whose token starts with `-` (option fragments left by the old tokenizer), behind the same proven-provenance gate. Other junk words (`HEAD`, `status`, `view` …) cannot be told apart from real binaries and are left in place; the full reset removes every UKit-managed rule.
|
|
42
|
-
- Fix: **`task-watchdog` no longer announces "payload was truncated or stalled" on large Edit/Write calls.** Its PostToolUse path reads only the event name from the payload (tasks, state and config come from disk), yet any payload over the 64 KiB staging cap was reported as incomplete input. A payload cut only by the size cap, whose event name is provably complete and exactly `PostToolUse`, now runs normally. Stalled staging, an unrecoverable event name, Stop (which forwards the raw payload to the coordinator) and the pre-install fallback branch keep the announced degrade.
|
|
43
|
-
- Fix: **self-improve no longer mints handoff tasks for failures that never happened.** The `timeout-hang` class matched `timeout|hang|hung` anywhere in a command's text, so "c*hang*elog", "unc*hang*ed", a test file named `verifyReleaseTimeout.test.js` or an env var `UKIT_VERIFY_STEP_TIMEOUT_MS` each produced an `AUTO-<n>.md` task telling the executor to raise a timeout. Mined patterns now carry the failed receipts' exit codes (`exitCodes`); `timeout-hang` needs exit 124 or a whole-word marker ("timed out", "timeout exceeded/expired", "hangs", "hanging", "hung") on the command's invocation lines — heredoc bodies and comment lines are not read. The other three classes use whole-word matches on the same stripped text. Known limit: a timeout the host kills with exit 1 or 137 is not classified (a missed task, not a wrong one).
|
|
44
|
-
- Fix (tests only): hook latency telemetry from test runs no longer lands in the repository's own `.ukit/storage/cache/hook-latency/` (seven test files spawned hooks against the real root; about 4,200 rows over five days). `appendTelemetryRow` drops rows for the root named by `UKIT_TELEMETRY_DENY_ROOT`, which only this repository's `vitest.config.js` sets; unset, it is inert.
|
|
45
|
-
- Fix (pre-release review): the new scanner gives up — mints nothing — on constructs whose end it cannot prove. Two review rounds each found a command resolving to the wrong binary (`X=$'\' curl ' ls` and `A=${x:- curl } ls` both minted `Bash(curl:*)` while bash runs `ls`). Every `$` now goes through one helper that accepts only `$(…)`, `$NAME` and a simple `${NAME}`; ANSI-C/locale quoting (`$'…'`, `$"…"`), any other `${…}`, `$[…]`, and a `#`, heredoc or `case` inside `$(…)` mint nothing. The scan is bounded to the first 4,096 characters of the command (a 100 KB assignment value stalled the hook for more than 45 seconds; a first command word not resolved inside the window mints nothing), and variable names are matched against explicit ASCII classes (`É=a ls` no longer counts as an assignment).
|
|
46
|
-
- Tests: `tests/hooks/autoAllowBinaryResolution.test.js` (41), four size-truncation cases in `tests/hooks/taskWatchdog.test.js`, classifier and `exitCodes` cases in `tests/core/codeProposals.test.js` / `tests/core/failurePatterns.test.js`.
|
|
47
|
-
- Packaging: the 2.5.1–2.6.11 changelog entries moved verbatim to `docs/archive/CHANGELOG-2.5.1-2.6.11.md` (repository only); size limits unchanged.
|
|
48
|
-
- Not changed: decision-plane shadow calls to the local `unic-decision` endpoint timed out 184 times in the audited week and five decision keys never returned a valid answer — shadow-only, not root-caused in this release.
|
|
49
|
-
|
|
50
|
-
## 3.4.82 - 2026-11-10
|
|
51
|
-
|
|
52
|
-
Completion-gate fixes from a field-telemetry audit of the projects running UKit (Stop/gate logs, hook latency, route state, session transcripts).
|
|
53
|
-
|
|
54
|
-
- Fix: **a stale legacy handoff cursor no longer marks every stop BLOCKED.** A `docs/AI_HANDOFF/RUN.md` written before `RunId` existed has no authority record, so nobody can own it until `/ukit:handoff-fullstack --adopt-legacy-run=<RunId>`. A session whose published route holds no handoff binding is now `NOT_OWNED` (an ordinary session beside an old cursor) instead of `UNKNOWN`; telemetry showed ~2,500 "handoff ownership UNKNOWN … report BLOCKED" advisories, 2,000 of them in one project in one month. A route that carries a `handoffRunId`, a duplicated `RunId`, or a missing/foreign/incomplete route still fails closed as `UNKNOWN`; the cursor is never touched and adoption is unchanged.
|
|
55
|
-
- Fix: **questions and acknowledgements are no longer routed as change orders.** Tone-folding turned "cái" into the verb *cài* (install) — likewise chính/chỉnh, Việt/viết, đợi·đối·đội/đổi, cháy·chảy/chạy, sữa/sửa, thèm/thêm — so "Cái này là gì, tại sao có chế độ này" owed write + verification evidence. Also released: "chưa được cấp quyền cập nhật …" / "no permission to update" (a read-only banner, not an order), "còn cần làm gì thêm nữa không" / "có thêm test nữa không" ("anything else", not "add") — only in the closed "còn/có … gì thêm [nữa] [không]" / clause-final "thêm nữa không" shape, so "thêm gì vào README.md" stays an order, a what-is question followed by punctuation, a leading "giải thích …" / "explain …", and bare acknowledgements after a task ("ok", "có", "được", "ừ", "được rồi") which now park the previous contract instead of re-arming it. Every release keeps its order control: "Cái này là gì, sửa giúp tôi", "thêm test nữa đi", "ok sửa đi", "giải thích rồi sửa luôn", "cấp quyền cập nhật cho user admin trong file roles.js" still block an empty stop. A bare "yes / ok / có / được / ừ / đồng ý" that answers a question the agent itself just asked (the last assistant text in the host transcript ends with a question, with no other owner turn in between) is the go-ahead for that same task and keeps its contract armed; "đóng nhé" (close it) is no longer read as "đồng ý". The homograph / banner / question rules are applied to `src/index/taskRouting.js`, `route-task.mjs` and the `skill-router.sh` hook; the acknowledgement rules live in `route-task.mjs` and the hook (the `src/index` copy has no greeting lane).
|
|
56
|
-
- Improved: **an unproven verification now names the command that proves it.** Most verification-evidence blocks were a runner piped into `tail`/`grep`/`tee` or followed by `; echo exit=$?`, and sessions repeated the same masked shape. The PostToolUse notice and the gate instruction now add "Run exactly this once: …" (`set -o pipefail; <same command>` for a pipe, the command up to and including the runner for a status echo, keeping a leading `cd`/env prefix) — only when the same classifier accepts that command and no second runner would be dropped. Guidance only: what counts as proof is unchanged, and loop/background shapes keep the generic wording.
|
|
57
|
-
- Fix: **"semantic contract invariant violation … contradictory contracts" no longer blocks a healthy session.** Telemetry: 172 such Stop blocks. The router re-scores the execution mode on every tool event while the contract stays the task's own, and both the router and the gate re-derived the contract from the *current* mode — so a contract built under `find-cause`/`map-impact` compared unequal to itself once the mode moved to `local-fix`. The comparison now accepts a stored contract that any execution mode reproduces for the same prompt and engine; a contract no mode reproduces (foreign evidence tag, different task kind) still blocks. Two feeders are closed as well: a new task no longer inherits the previous task's evidence tags, operation goals or state targets (a read-only review's `impact-evidence` leaking into the next change order), and a system-injected turn (compact summary, `/loop` wakeup) after an answer-only turn no longer mints a contract bound to the owner's prompt.
|
|
58
|
-
- Fix: **a slow Stop evaluation is no longer reported as a broken install.** One evaluation repeated the same git probes (16 spawns → 10; about 950ms → 550ms measured on this repo); they are now memoized for that single evaluation only. The evaluator's time slice scales with the hook deadline (1500ms → up to 3000ms; `completion-gate.sh` default deadline 3000 → 5000ms, inside the hook's slot, see below), and an evaluator that still runs out of time blocks fail-closed with its own fixed text ("ran out of time … Stop again to re-evaluate") instead of "Run: ukit install".
|
|
59
|
-
- Fix: the Stop message for an unproven verification cut the command at 80 characters — on a real `cd <root> && PYTHONPATH=. venv/bin/python3.12 -m pytest … | grep … | tail -40` that ended exactly at `python3.12 -`, naming a command that never ran and hiding the pipe that made it unproven. The label now keeps the whole command up to 240 characters (both ends beyond that). The hook-deadline block now says to stop again first and reach for `ukit install` only if it repeats; it stays fail-closed.
|
|
60
|
-
- Fix: a Stop on which the completion evaluator could not finish (killed at its time slice, hook deadline, evaluator crash) blocked every time with no cap — one field session took 170+ such blocks in a row on a loaded machine, each costing a turn the agent could do nothing with. These blocks now have their own bounded breaker in `completion-gate.sh`: counted per session (one session never releases another), released only after more than `stopGateMaxCrashStreaks` (default 3) consecutive failures (no more than 10 minutes apart) that also span at least 120 seconds — stopping again a few times in a row never releases — and the release is announced as a blocked, non-complete outcome ("completion was NOT verified … not a pass"). Any real verdict resets the count; evidence blocks, and any decision carrying a handoff-cursor or task-watchdog block, are never counted. The crash breaker for a coordinator that exits non-zero is unchanged and separate. A block verdict the evaluator had already written when its time-slice kill landed is now honoured as that verdict instead of being reported as an evaluator failure (a cut-short release is still never trusted).
|
|
61
|
-
- Fix: the Stop hook's chain slot is 12s (was 8s) and its host timeout 22s (was 18s), keeping the chain-budget rule host ≥ slot + 10s. With the 5000ms coordinator deadline plus the 1500ms killer the wrapper could need 6.5s before bash start-up and staging, and a chain-runner timeout of this wrapper would have been a silent release.
|
|
62
|
-
- Tests: `tests/hooks/semanticContractModeChurn.test.js`, `tests/hooks/stopEvaluatorBudget.test.js`, `tests/hooks/telemetryAuditRouting.test.js` (real hook + gate, fresh and after-task, with order controls), a legacy-cursor bystander case with fail-closed controls in `handoffProvenance.test.js`, and rerun-guidance cases in `stopGateReceiptCapture.test.js`.
|
|
63
|
-
- Not changed (by design, per the audit): pipe/`; echo $?` shapes stay unproven; a carried compact-summary contract still owes its evidence; the teammate-message lane has no fix in this release.
|
|
64
|
-
|
|
65
|
-
## 3.4.81 - 2026-11-08
|
|
66
|
-
|
|
67
|
-
- New: **automatic Claude Code session discovery for UKit Advisor.** `ukit:advisor` / `$ukit-advisor` with nothing after the trigger (or only a short remark such as "lỗi rồi") reads the latest main-agent Claude Code session of the current project — no paste needed; `session <id>` and `project <path>` name one explicitly; anything else is still pasted output, as in 3.4.80. A bundled read-only Node reader (`scripts/claude-session-reader.mjs` inside the skill, also `ukit advisor context [--last|--session <id>|--project <path>] [--json]`) matches the project by the session's recorded cwd (realpath, git root, linked worktrees; nested repos and sibling folders excluded), honours `CLAUDE_CONFIG_DIR`, skips subagent/sidechain transcripts and symlinks, refuses to guess between concurrent sessions, streams huge/partial/concurrently-written JSONL with bounded secret redaction, and renders a bounded, redacted, untrusted digest that keeps owner messages, tool results, assistant claims and Stop-hook/gate refusals apart (owner STOP ≠ hook refusal). It never writes, spawns, resumes or messages Claude Code; no session, no permission or a Codex model that cannot run shell commands yields one safe status-request prompt.
|
|
68
|
-
- Tests: 41 session-reader tests on synthetic transcripts with an isolated `CLAUDE_CONFIG_DIR`; the repo-only live runner `scripts/dev/advisor-live-eval.mjs` gains `--mode=auto` and counts only real `command_execution` events as a reader run; two load-sensitive hook tests get timing headroom (assertions unchanged).
|
|
69
|
-
- Release note: 3.4.80 was tagged and GitHub-released but never published to npm; 3.4.81 is the first npm release carrying UKit Advisor (both the paste mode and session discovery).
|
|
70
|
-
- Packaging: the 2.3.0–2.4.3 changelog entries moved verbatim to `docs/archive/CHANGELOG-2.3.0-2.4.3.md` (repository only); size limits unchanged.
|
|
71
|
-
|
|
72
|
-
## 3.4.80 - 2026-11-08
|
|
73
|
-
|
|
74
|
-
- New: **UKit Advisor for Codex.** Paste Claude Code output (status, errors, stalled transcripts, STOP/gate/release messages — VN/EN/mixed) into Codex after `$ukit-advisor` or `ukit:advisor` and get back exactly ONE copy-ready Vietnamese prompt for Claude Code (PROMPT ONLY: no analysis section, no code fence, no alternatives). The advisor is read-only, treats the paste as untrusted data, never lifts an owner STOP, never forges ownership/receipts, never starts duplicate jobs, keeps release states distinct (never republishes a verified npm version) and redacts secrets.
|
|
75
|
-
- Surfaces: `$ukit-advisor` is a native Codex skill seeded into `~/.agents/skills/ukit-advisor` (`allow_implicit_invocation: false`). `ukit:advisor` is instruction-routed — a marker-delimited block in the active global Codex instructions file (`$CODEX_HOME/AGENTS.override.md` when present, else `AGENTS.md`) points the prefix at the same SKILL.md; it is not a native Codex command.
|
|
76
|
-
- Lifecycle: installed/refreshed by `ukit install` when the Codex adapter is selected, and by `ukit update` (with `--no-install` or outside a UKit project it only refreshes an advisor UKit already installed); idempotent; ownership via `.ukit-managed.json` hashes, so an unrelated same-name skill is never overwritten and user-edited files are kept; the user's AGENTS.md bytes, BOM and line endings are preserved; a damaged BEGIN/END marker pair or a symlinked AGENTS.md is reported and never rewritten. New `ukit advisor status|install|uninstall [--dry-run]`; opt out with `UKIT_CODEX_ADVISOR=0`.
|
|
77
|
-
- Tests: lifecycle/ownership/cross-platform unit tests, a deterministic PROMPT-ONLY contract checker with 13 eval fixtures, test isolation now pins `CODEX_HOME`, and a repo-only live runner `scripts/dev/advisor-live-eval.mjs`.
|
|
78
|
-
- Packaging: the 3.0.0–3.3.3 changelog entries moved verbatim to `docs/archive/CHANGELOG-3.0.0-3.3.3.md` (repository only); size limits unchanged.
|
|
79
|
-
|
|
80
|
-
## 3.4.79 - 2026-11-07
|
|
81
|
-
|
|
82
|
-
- P0 handoff queue lifecycle: one read-only classifier (`classifyQueuedBlueprints`) now decides which `docs/AI_HANDOFF/queued/` blueprints still count as unfinished work. `QueueIntent: superseded|stale` + `QueueSupersededBy: archive:cycle-NN[, …] | queue:<slug>` excludes a blueprint only when the provenance is proven (every package task archived in an 8-column cycle snapshot as delivered/grounded-cancelled, or a live non-cyclic replacement inside the project); self-references, cycles, paths, symlink escapes, partial archives and malformed headers fail closed with diagnostics. Legacy blueprints and `deliberate-future` keep their behaviour; the array API stays compatible. A frozen `ScopeQueue` naming a superseded blueprint warns instead of consuming it silently; a sibling `queue/` directory is reported, never migrated. A `queue:<slug>` replacement that is `deliberate-future` backlog cannot carry the obligation (fails closed). New read-only `resumable-run.mjs queue-status`.
|
|
83
|
-
- P0 completion contract: plan/doc-only prompts (Vietnamese "tài liệu/kế hoạch", guide/plan/notes nouns, `update docs/x.md`) owe only the written `docs/…md` artifact instead of code write+test evidence, and a filename such as `docs/plans/deploy.md` is no longer read as a deploy order. Code prompts keep write+verification; deploy prompts keep operation verification. Every clause must be doc-only (whitelist): an inspection lead, a bare `write`/`update` with no doc noun, or a vague order ("do it", "proceed", "làm tiếp") keeps the code contract; inside a docs-path clause free prose is allowed only after a describe/explain lead, and any order behind plus/once/as well as/parentheses/colon keeps the code contract.
|
|
84
|
-
- Acceptance: 6-column INDEX/snapshot layouts fail closed and are never rewritten; the consumer migration procedure (backup → dry-run on a copy → 8-column re-snapshot → idempotent → rollback, collision, 3-digit cycle, crash recovery) is covered by `tests/handoff/queueLifecycleAcceptance.test.mjs`. Consumer guidance: `docs/release/3.4.79-QUEUE-LIFECYCLE-MIGRATION.md` (repository only).
|
|
85
|
-
- Packaging: the 2.1.0–2.2.16 changelog entries moved verbatim to `docs/archive/CHANGELOG-2.1.0-2.2.16.md` (repository only); size limits unchanged.
|
|
86
|
-
|
|
87
|
-
## 3.4.78 - 2026-11-06
|
|
88
|
-
|
|
89
|
-
- P0 RC4 (delayed wakeup): a `/loop` fire queued behind a busy turn is delivered only after that turn ends, so its `scheduled_task_fire` record sat behind the turn's assistant entries and beyond the router's 128KB tail (real incident: 177KB, 7 minutes) and the wakeup minted a fresh CODE_IMPLEMENTATION task. An UNCONSUMED same-session fire (no later entry carries its `scheduledFireId`), matching the prompt, at most 6h old, found in a bounded 8MB tail is now host provenance: the previous task and contract carry, a read-only wakeup stays a coordination turn. Consumed, foreign-session, stale, timestamp-less or different-prompt fires never exempt an owner prompt. Replay of the real transcript: HEAD minted a new task, fixed router carries.
|
|
90
|
-
|
|
91
|
-
## 3.4.77 - 2026-11-06
|
|
92
|
-
|
|
93
|
-
- Packaging: the 1.0.0–2.0.7 changelog entries moved verbatim to `docs/archive/CHANGELOG-1.0.0-2.0.7.md` (repository only) so the npm artifact stays inside the unchanged size budget; no runtime or safety code was removed and the size limits were not raised.
|
|
94
|
-
- P0 acceptance: a Claude Code scheduled wakeup (`scheduled_task_fire` in the current delivery window) is recognized as host provenance, never owner authorization; commit/push + script deploy keeps a separate git goal; a milestone with no attempted operation hands back instead of demanding operation verification; an unchanged-state Stop gives a single finite handback.
|
|
95
|
-
- P0 continuation: a pending host background job yields a WAIT Stop (no budget, no edit demand, one release per job); an explicit owner continue/authorization adopts the predecessor task's content-current evidence via a single-hop link; scoped (conditional/purpose/other-subject/quoted) bans no longer prohibit an authorized deploy; non-npm native deploys are proven by deploy + post-deploy receipts; unproven-exit messages name the runner. Replay/unit evidence only.
|
|
96
|
-
|
|
97
|
-
## 3.4.76 - 2026-11-03
|
|
98
|
-
|
|
99
|
-
**P0 follow-up: UNKNOWN-stop marker store, explicit STOP publication and negation-aware state classification. Isolated process/replay evidence only; not live-host certification.**
|
|
100
|
-
|
|
101
|
-
- `ukit marker-store migrate`: plan by default; `--apply --writers-stopped` refuses unless the owner attests every OMP/bridge session is closed, no process uses the project, no writer temp file exists and the store is unchanged across an observation window. It never kills processes, keeps releases byte-for-byte, turns blocks into sticky releases under the same hashed names, quarantines (never deletes) the old store, and resumes after interruption.
|
|
102
|
-
- Bridge: a fenced or unavailable store fails open (`degraded:storage`) without crash or repeated calls; release-only saturation keeps every release.
|
|
103
|
-
- Explicit owner STOP survives pending route publication and never expires by TTL.
|
|
104
|
-
- Classifier: a negated verb ("do not stop the job") no longer mints a state-verification debt; `khong sua ledger/gate/receipts` stays READ_ONLY.
|
|
105
|
-
- Consumers: close related OMP/bridge sessions, `ukit update`, `ukit marker-store migrate`, then `--apply --writers-stopped`, verify, reopen. STOP stays in effect; legacy cursors are adopted only via explicit `--adopt-legacy-run`.
|
|
106
|
-
- Artifact budget: packed < 3,500,000 and unpacked < 11,300,000 bytes (limits unchanged). To stay under the unpacked limit, about 92 verbose bullets in historical entries (mostly older than 2.7) were shortened in this file, saving about 22 KB; no release was removed, and the full original text is in git history at 46b48d11.
|
|
107
|
-
|
|
108
|
-
## 3.4.75 - 2026-11-01
|
|
109
|
-
|
|
110
|
-
- Publish receipts bind to artifact integrity.
|
|
111
|
-
|
|
112
|
-
## 3.4.74 - 2026-10-30
|
|
113
|
-
|
|
114
|
-
**Route publication protocol, handoff ownership authority and explicit legacy adoption; no live-host certification beyond the isolated native OMP A/B/C acceptance recorded in `docs/STATUS.md`.**
|
|
115
|
-
|
|
116
|
-
- Route publication: pending sidecar → shared write with UUID revision → final sidecar commit across all route writers/readers (`docs/ROUTE_PUBLICATION.md`); Stop coherence tests and a fault-injection matrix cover torn/concurrent publication.
|
|
117
|
-
- Handoff ownership: immutable run binding plus per-run authority record (generation/history) in `handoff-provenance.mjs`; `resolveCursorApplicability` yields OWNED / NOT_OWNED / UNKNOWN and UNKNOWN is never treated as ownership. Not described as fail-safe beyond what the tests prove.
|
|
118
|
-
- OMP case A: a Stop with unavailable route identity is released as NOT_OWNED only when a bounded transcript intent scan proves the session never targeted the handoff; torn or non-JSON-shaped transcript lines stay unknown. Intent only releases a non-claimant and never grants ownership. Claude/OMP Stop differences are documented and tested (`stopHostSemantics.test.js`); advisory or loop-breaker release is not completion evidence.
|
|
119
|
-
- Legacy adoption: owner-invoked `--adopt-legacy-run=<sha256>` only (never at session start or Stop), validated under lock against the cursor snapshot hash; preserves content and state and never declares a run done.
|
|
120
|
-
- Publication intent: `publicationScope` treats "do not publish until/unless …" as a temporary restriction rather than a task-wide prohibition (it never grants authorization); `evidenceStatement` masks the repo's own release-policy sentence so it is not read as an execution order; `SEMANTIC_ENGINE_VERSION` bumped so cached intent scans are invalidated.
|
|
121
|
-
- Release completion: `releasePublicationMissing` accepts only a task-owned, tool-use-attributed, non-masked publish receipt with a parsed `+ pkg@version` line followed by a later same-scope registry view (registry state alone is not proof); `cron-operation-adapter` reports publish receipts without that line as unproven; background completions apply the masked-exit check.
|
|
122
|
-
- Artifact budget: packed < 3,500,000 and unpacked < 11,300,000 bytes.
|
|
123
|
-
|
|
124
|
-
## 3.4.73 - 2026-10-28
|
|
125
|
-
|
|
126
|
-
- Git history proof tolerates unrelated untracked files; tracked/staged drift and explicit whole-tree clean obligations remain strict. No filename exemptions.
|
|
127
|
-
- Native ledger-to-Stop fixtures cover Claude Code/Codex/omp, repeated Stop and untracked shipped-source publication drift; not live-host certification.
|
|
128
|
-
|
|
129
|
-
## 3.4.72 - 2026-10-28
|
|
130
|
-
|
|
131
|
-
- Native evidence: reject background, masked and test-expression exits as operation/verification proof; retain supported foreground controls across shared Claude Code/Codex/OMP fixtures. No live-host certification.
|
|
132
|
-
- Completion: retained publication-only debt uses task-owned registry proof only without a real operation goal; unrelated Git/service duties remain enforced.
|
|
133
|
-
- Self-improve: serialize deferred AUTO queue replay/merge, retain snapshots/provenance and corrupt/legacy entries, drain only written tasks; dry-run remains read-only.
|
|
134
|
-
- Release recovery: bind mandatory phase exits and independent final review to unchanged candidate bytes; publish lifecycle refuses missing prerequisites. Local evidence files are integrity checks, not cryptographic identity proof. Final acceptance and isolated registry consumers remain pending; 3.4.71 incident evidence is preserved below.
|
|
135
|
-
|
|
136
|
-
## 3.4.71 - 2026-10-28
|
|
137
|
-
|
|
138
|
-
**F-7 intent scope, release-freeze coordination and Claude Code operation-receipt exit proof; no native-host certification.**
|
|
139
|
-
|
|
140
|
-
- Classification (F-7): an inspection verb followed by an interrogative object ("review how the publish and deploy flow works", "explain why update failed") no longer lets its coordinated fragments become orders. The collapse ends at `then` or any coordinator followed by an action verb, so real orders after the question (including bare-verb, Vietnamese and adverb-interposed forms) keep their kind. Independent review probed ~490 prompt pairs against 3.4.70 with no remaining executable-to-READ_ONLY regression except intended compound-noun cases.
|
|
141
|
-
- Self-improve: added a verifier freeze marker and deferred AUTO queue. Later review found repeated frozen passes could overwrite queued proposals; the original no-loss claim was incorrect.
|
|
142
|
-
- Completion evidence: added Claude Code clean-PostToolUse shell-exit inference. Later review found background-child attribution unsafe; shell success alone is not native-operation success. No live-host certification.
|
|
143
|
-
- Verification: full `yarn test` 698 files / 11,371 passed / 3 skipped, exit 0; official verifier PASS. Unpacked budget 11,200,000 owner-approved. Published before final independent CHANGES-REQUESTED review; normal-update testing also changed the owner's global installation instead of staying isolated. 3.4.71 was published, not fully accepted. Wrapper/correlation/stale-obligation and P2 work remained open.
|
|
144
|
-
|
|
145
|
-
## 3.4.70 - 2026-10-26
|
|
146
|
-
|
|
147
|
-
**Task-owned completion evidence and bounded reminder recovery; no native-host certification.**
|
|
148
|
-
|
|
149
|
-
- Classification: configuration comparisons, read-only output specifications and owner-approval reminders do not inherit mutation or verification debt. Explicit verification and publication orders retain their own obligations; private Git scope and later affirmative orders remain distinct. Descriptive acceptance-review prose — subordinator continuation fragments ("…, including source fixes F-3") and non-imperative goal/acceptance lines ("Final product goal: … normal ukit update pass") — no longer routes read-only reviews as implementation orders; any imperative fragment inside such a line still enforces write+verification.
|
|
150
|
-
- Completion: preserve truthful blocked outcomes, scoped exhausted-denial reconciliation and native Git/remote evidence. Launch acknowledgements, missing exit proof, masked commands, foreign receipts and stale parity cannot prove success. Docs-only Handoff creation honors the owner's explicit Git publication scope.
|
|
151
|
-
- Reminder recovery: reapply classified evidence obligations even when the task-kind signature is unchanged. Repeated read-only reminders need no artificial checkpoint verification; owned external blockers remain blocked rather than successful. No permission reset, receipt bypass or safety relaxation.
|
|
152
|
-
- Release verification: bind mandatory phases to candidate bytes and fail on drift; run existing live-documentation checks before expensive gates and retain external per-run result metadata. Mandatory lanes, artifact budgets and timeouts are unchanged.
|
|
153
|
-
- Evidence limits: targeted regressions and independent reviews cover shared Claude/OMP sources and fixtures; Codex remains advisory where no native hook exists. Historical failed verifier runs remain failed. The review integration fixture now has a measured, first-case-only 15s allowance after bounded child-latency reproduction; global/runtime timeouts are unchanged. Discovery deadline testing waits for the intended filesystem phase instead of racing startup I/O. No speculative performance fix or exact historical slowdown attribution is claimed.
|
|
154
|
-
|
|
155
|
-
## 3.4.69 - 2026-10-25
|
|
156
|
-
|
|
157
|
-
**Narrow shared-lock cancellation repair; no new unattended-Handoff certification.**
|
|
158
|
-
|
|
159
|
-
- Fix: cancellation after a successful lock owner stamp now passes through token-checked, best-effort release before returning `aborted`; the callback remains unexecuted. Deterministic regression verifies lock absence and immediate reacquisition.
|
|
160
|
-
- Tests: managed Handoff template cases pin owner-byte preservation, exact sidecar bytes/provenance and no-write aligned behavior. Preflight fixtures now use an independent temporary Git repository with ownership-checked cleanup, avoiding default dirty-worktree/sandbox accumulation without touching retained evidence.
|
|
161
|
-
- Verification: independent original and cleanup-only reviews approved with minor findings; full release verifier passed (core 10,741 passed / three skipped, isolated handoff 351/351, artifact 39/39). Cleanup review retains a non-blocking concurrency-sensitive source-inventory assertion and fail-loud fixture cleanup caveats.
|
|
162
|
-
- Limits unchanged: native isolation dispatch, clean unattended lifecycle/native-host acceptance, historical `ukitcmd25` loss and unidentified `79/1` remain unresolved. This fix is not established as the cause of those incidents. No 3.4.68 live-gate waiver is inherited or live acceptance claimed.
|
|
163
|
-
|
|
164
|
-
## 3.4.68 - 2026-10-25
|
|
165
|
-
|
|
166
|
-
**Stable hardening release; outstanding live acceptance explicitly deferred by the owner for the internal six-person team. This is not certification of unattended Handoff.**
|
|
167
|
-
|
|
168
|
-
- Planning: durable source snapshots, recorded validator/independent-review ready gates, amendment/plan-count/overwrite controls, and realpath-safe installed helper entrypoints.
|
|
169
|
-
- Runtime: authenticated asynchronous delegation completion, bounded recovery and claim fencing; landing coordinated with claims; explicit already-applied adoption and non-completing metadata reconciliation; completed-only archive/resume safeguards and three-engine adapter consistency.
|
|
170
|
-
- Owner preservation: managed Handoff templates refresh only with provenance, retain custom content/sidecars safely through update and uninstall; prevent shell-keyword/state-builtin permission minting, with provenance-only sanitizer.
|
|
171
|
-
- Classification/continuation: read-only output specifications and scheduled permission reminders no longer manufacture implementation work; real edit orders retain evidence obligations.
|
|
172
|
-
- Safety: existing-only delegation guard also honors restrictive runner policy without a RUN marker. Planning instructions omit isolation, but **actual native dispatch is not fixed**: the latest create10 failed at its first isolation-bearing Agent call (0/10 accepted, nine skipped). The guard remains; no payload stripping, trust/permission expansion or provider workaround.
|
|
173
|
-
- Known limits retained: clean unattended multi-plan execution, review-return/replan, blocked-independent scheduling, wait/no-duplicate, interruption/cycle resume, late-create frozen scope, pause/compact, current-candidate OMP lifecycle and missing native-host/platform coverage are not certified. Historical unidentified `79/1` and lost `ukitcmd25` (29/30) remain unexplained; later passing checks do not resolve them. Inspection-command friction also remains.
|
|
174
|
-
- Evidence: reviewed repair groups; full code-candidate suite 673 files / 10,771 tests passed, three skipped. Recovered roadmap 8/8, integration 61/61, completed-only archive and unchanged reinvocation remain recovery evidence, not a clean unattended PASS. The owner explicitly waived these outstanding live gates as release blockers, not technical/package/registry-consumer checks. No FAIL/BLOCKED evidence is relabeled PASS.
|
|
175
|
-
|
|
176
|
-
## 3.4.67 - 2026-10-23
|
|
177
|
-
|
|
178
|
-
- Fix: a wait/monitor/hold-only request ("keep waiting on the verifier job, read output to the footer, do not edit while the candidate is frozen") is now classified verification-only (`READ_ONLY`, `verification-evidence`) instead of the `CODE_IMPLEMENTATION` default that demanded `write-evidence` and forced an Edit/Write. A wait that also carries a real order ("wait, then fix src/app.js") stays an implementation. Vietnamese "rồi" now splits clauses for this check. Both classifier copies (`src/core/executionContracts.js`, installed `route-resolver.mjs`) changed together; regression in `tests/hooks/gateIntentClassification.test.js`.
|
|
179
|
-
- Test-only: `hookProcessTree` escapee regression budget no longer races a cold Node start under full-suite load.
|
|
180
|
-
|
|
181
|
-
## 3.4.66 - 2026-10-23
|
|
182
|
-
|
|
183
|
-
**Handoff hardening: scope-frozen runs, completed-only ARCHIVE, bounded repair, isolated verification.**
|
|
184
|
-
- Stop coordinator: INDEX task rows parse positionally (empty cells keep columns; malformed rows fail closed); an unreadable or vanished archive snapshot no longer falls back to stale task headers; optional `ScopeQueue`/`Cycle` freeze a run to its assigned scope so unrelated queued or archived history cannot prove completion; `Paused: yes` releases continuation without minting completion.
|
|
185
|
-
- Archive writer: canonical `docs/AI_HANDOFF/ARCHIVE/cycle-NN/`; refuses (before any write) symlinked docs/AI_HANDOFF/archive/cycle/INDEX/TASK paths, malformed or unreadable prior snapshots and any task-enumeration error instead of snapshotting a partial cycle; legacy lowercase archives stay read-only.
|
|
186
|
-
- Commands: create queues independent plans instead of overwriting, preserves sources and amendments; fullstack/implement/review commit with `git commit --only` explicit paths (unrelated staged and unstaged owner work preserved, proven against a real Git fixture); failures route to bounded repair or replan instead of blocked; one persisted Recovery exception after escalation; clear pauses and preserves unfinished work and keeps read-back write verification; blocked is reserved for external blockers or genuine owner requirement conflicts.
|
|
187
|
-
- `yarn test:handoff` now runs `tests/handoff/**` in a disposable copy with throwaway HOME, excluded live handoff state (any-case ARCHIVE), vetted symlinks, checked git setup, bounded children and explicit path validation; wired into `release:verify`.
|
|
188
|
-
|
|
189
|
-
## 3.4.65 - 2026-10-22
|
|
190
|
-
|
|
191
|
-
**Configuration capture: Vietnamese comparison is not a mutation order.**
|
|
192
|
-
- Original pilot prompt reproduced CONFIGURATION/external with no operationGoal: “Đối chiếu cấu hình đang hoạt động” normalized to `doi chieu`, then `doi` was mistaken for “đổi” (change). Exclude only `doi chieu` from config mutation verbs in the three shared classifier copies; do not change repository detection, operation receipts or completion gates.
|
|
193
|
-
- Exact prompt fixture, comparison variants and genuine change commands (including comparison followed by change) retain regression coverage. Existing Stop-boundary reconciliation corrects the stale classified contract without resetting task identity/counters; original rotated live contract is not reconstructed or sealed.
|
|
194
|
-
- Authorized one-pair CLI-worker experiment: cap2 COMPLETE 4/4; cap4 INCOMPLETE 2/4. No four-worker adoption, extra arm, global/BAGuide/default change or native OMP throughput claim. Experiment harness/report are repository evidence, not installed runtime.
|
|
195
|
-
|
|
196
|
-
## 3.4.64 - 2026-10-22
|
|
197
|
-
|
|
198
|
-
**Completion Gate: defect-contingent repair/release is not an unconditional publish order.**
|
|
199
|
-
- Exact approved pilot prompt reproduced false `release-publication`: line/comma splitting detached `release npm` from the preceding wrapped “Nếu gặp bug UKit…” condition. Standing policy/carryover is not needed to reproduce it.
|
|
200
|
-
- Preserve defect-repair sentence scope across line wraps; activate its release branch only after owned runtime writes. Direct publish commands still require registry verification, and foreign-task proofs remain rejected.
|
|
201
|
-
- Do not interpret hashed write-bank keys as paths; bank the runtime-write classification separately across same-task rekeys. Docs/test-only direct writes do not activate repair publication. Legacy pathless Bash deltas remain conservative (may demand release); no new path-attribution architecture added.
|
|
202
|
-
- Pilot preflight narrowed to existing host interfaces and exact condition-change options; no paid arms, no relaxed limits adopted, no BAGuide/default changes.
|
|
203
|
-
|
|
204
|
-
## 3.4.63 - 2026-10-22
|
|
205
|
-
|
|
206
|
-
**Completion recovery: background pending is not a pipeline.**
|
|
207
|
-
- Reproduced task `task-848c14831946f7c359cdf11faf45edd2` / `bk1n4wpks`: bare release verifier exited 0 in host notification; Bash/grep read did not ingest completion. Launch and masked-shell receipts share `exitStatusUnproven`; recovery falsely asserted pipeline and requested reruns.
|
|
208
|
-
- Diagnostic-only fix: background launch -> wait SAME job, no duplicate; host completion -> Read task output through footer, not rerun; masked shell -> wrapper status unproven, actual pipe only a possible cause. Receipt acceptance/freshness unchanged; unsupported hosts remain fail-closed.
|
|
209
|
-
- Regressions: pending, host completion+Bash vs Read, later launch not masking valid completion, failed runner, masked wrapper, analysis-doc freshness vs living STATUS; existing forged/mismatched/duplicate/foreign/stale controls retained.
|
|
210
|
-
- Release-gate fixture repair: TERM-ignoring grandchild now signals readiness after handler installation before the measured runner deadline starts (injected real child); retains ≥700ms grace/escalation and no-survivor assertions. Production hook-process unchanged.
|
|
211
|
-
- Independent four-task fixture pilot and whole-workflow budget proposal recorded; no paid pilot. OMP genuine-service live smoke remains unverified (deterministic RPC bash lacks name/ready); no claim of every-service coverage.
|
|
212
|
-
|
|
213
|
-
## 3.4.62 - 2026-10-22
|
|
214
|
-
|
|
215
|
-
**Handoff commands: state what `handoff.maxParallelAgents` counts (wording only).**
|
|
216
|
-
- The cap is child workers of ONE batch or review set, summed across roles (implementers and reviewers together); the main session is outside it; it is not a session-wide limit; it is prompt-level guidance (no hook counts children) and the host's own concurrency limit is separate. Added to `handoff-implement`, `handoff-fullstack` (implement batch + review set) and `handoff-review`.
|
|
217
|
-
- No default, config key, scope or behavior change. Basis: BAGuide field audit (peak 14 concurrent children while the cap text said 2, never refused), see `docs/analysis/CONCURRENCY-CHECKPOINT-2026-10-22.md`.
|
|
218
|
-
|
|
219
|
-
## 3.4.61 - 2026-10-22
|
|
220
|
-
|
|
221
|
-
**omp bash: stop sending the call shape the host rejects (`name` + `async`/`timeout`).**
|
|
222
|
-
- Field evidence (BAGuide session log, 2026-10-19..22): 1,771 of 6,541 bash calls failed with "Service mode does not accept async or timeout; use ready.timeout for readiness." Every failing call carried `name` with `async`/`timeout`; calls without `name` never failed and 66 immediate retries that dropped `name` succeeded. `name` was used as a label on finite commands.
|
|
223
|
-
- Host contract (omp 18.12.5 bash `execute`, read from the installed binary): `name` selects SERVICE mode, which rejects `async === true` and any `timeout`; readiness is `ready.timeout`. Without `name`, `async`/`timeout` are valid and `ready` is only ignored with a warning.
|
|
224
|
-
- Change: `omp-bash-call-guard.mjs` (block-only, stateless, never rewrites input or changes a command's meaning) wired into the omp bridge `tool_call` for `bash`; the block reason separates the two intents (finite command: drop `name`/`ready`; service: drop `async`/`timeout`, use `ready.timeout`). Same rule added to the Stall & Wait instructions (STALL-02) for all three engines.
|
|
225
|
-
- Live smoke (omp 18.12.5, isolated temp project with the template bridge, model `unic/unic-lite`, BAGuide and the omp binary untouched): `name`+`timeout` was blocked with the UKit preflight text (not the host's "Service mode..." text), so the `tool_call` hook runs before the host validator; a finite command ran; the corrected retry (name dropped) ran and was not blocked. NOT verified live: a genuine service call (the test model kept adding a top-level `timeout`; one attempt hung with no output and was stopped; no process was left running). The service-allowed path is covered only by the bridge-boundary tests. Guidance, not host enforcement.
|
|
226
|
-
- Flaky test root cause + fix: `withTransientFsRetry` compared `Date.now()` against a deadline while sleeping with `setTimeout`; node timers can fire up to ~2% early at small delays, so after a clamped sleep the clock still showed time remaining and an extra attempt ran. It now measures with `performance.now()` (injectable `now`/`sleep`) and stops after a sleep that exhausted the deadline. Regression: `tests/hooks/transientFsRetry.test.js` (controlled clock incl. early-timer skew, plus one real-timer smoke); 30/30 repeated runs.
|
|
227
|
-
- Review freshness through a real Bash mutation (git fixture, real PostToolUse receipt): review approved -> `sed -i` via Bash -> `changeSeq` advances, review is `review-stale`, verification is `edited-after-verification`; a no-op Bash keeps the approval. No defect reproduced; regression `tests/hooks/playbookReviewBashMutation.test.js`.
|
|
228
|
-
- Regression: `tests/hooks/ompBashCallGuard.test.js` (normal finite command with timeout/async allowed; genuine service with `ready.timeout` allowed; `name`+`timeout` / `name`+`async` blocked repeatably; blank `name` is not service mode; input never mutated).
|
|
229
|
-
|
|
230
|
-
## 3.4.60 - 2026-10-22
|
|
231
|
-
|
|
232
|
-
**Completion Gate: a background completion footer is no longer trusted on its own.**
|
|
233
|
-
- Gap: 3.4.59 accepted `[exited with code N]` at the end of `.../tasks/<id>.output` as the runner's exit status. That file holds the runner's own stdout, so a process (or a test fixture) that prints the footer while still running, or before a non-zero exit, could produce a false PASS.
|
|
234
|
-
- Fix: a completion is accepted only when the session transcript holds the HOST's `<task-notification>` for the same task id (queue event or plain user turn; tool-result text, including a Read of the output file, never counts) and its exit code equals the footer. No notification -> `no-host-notification`; disagreement -> `host-footer-mismatch`. A host-reported failure is recorded as failed.
|
|
235
|
-
- Regression: `tests/hooks/backgroundCompletionForgery.test.js` (footer printed while running, imitation notification inside a tool result, matching notification, mismatching exit, host failure, other task id); lifecycle tests now mirror the host notification.
|
|
236
|
-
- Limit: needs `transcript_path` in the hook payload (Claude Code provides it); without it a background completion stays unproven. omp/Codex shapes not covered; replay, not a live host.
|
|
237
|
-
|
|
238
|
-
## 3.4.59 - 2026-10-22
|
|
239
|
-
|
|
240
|
-
**Completion Gate: a host-backgrounded verification can now complete through a supported, provenance-bound result capture.**
|
|
241
|
-
- Gap: 3.4.57 made a background launch acknowledgement unproven, but nothing could later turn it into evidence, so the only route was re-running a long suite in the foreground. Claude Code delivers the completion as a task-notification the hooks never see; the host does write the runner's output to `.../tasks/<id>.output` ending in `[exited with code N]`.
|
|
242
|
-
- Fix: when the agent reads that file back, the ledger records a verification receipt bound to the launch receipt by `backgroundTaskId` (same task ledger, so another task/session finds no launch), exit status from the host footer, judged against the LAUNCH tree fingerprint. Rejected: no footer (still running / "passed" text only), no matching launch, duplicate completion, output file older than the launch. Exit 1 yields a failed verdict.
|
|
243
|
-
- Regression: `tests/hooks/backgroundCompletionLifecycle.test.js` (success, failure, partial, duplicate, other task, other session, late, stale file, unrelated file, Stop-gate outcomes). Also `tests/orchestration/schedulingHarness.test.js`: a deterministic offline scheduling-logic model (not a throughput benchmark).
|
|
244
|
-
- Limit: replay of the host file format observed on Claude Code; omp/Codex background shapes are not covered, and no hook receives the notification itself.
|
|
245
|
-
|
|
246
|
-
## 3.4.58 - 2026-10-22
|
|
247
|
-
|
|
248
|
-
**Completion Gate: plain publish orders create the registry obligation, new tasks do not inherit the previous task's evidence, and a Vietnamese status line about a config is not a config order.**
|
|
249
|
-
- Bug: "Please/Now npm publish the package", "Đưa package lên npm" and "Phát hành bản này lên npm" were not read as publish orders, so code + tests PASS completed the gate with nothing on the registry. `promptOrdersNpmPublish` now accepts polite/sequencing prefixes and the Vietnamese publish verbs; negations, `--dry-run` and prohibitions still create no debt.
|
|
250
|
-
- Bug: a new task inherited the previous task's classifier-owned contract fields and MIXED evidence (a stale `git-integrity push` goal and `operation-verification` under "npm publish the package"), so a correct publish run stayed blocked on an unordered push. `skill-router.sh` now resets them when a new task starts.
|
|
251
|
-
- Bug: "OMP version/config đã load lúc bắt đầu vẫn unknown" classified CONFIGURATION/external (`bat` in "bắt đầu", `tat` in "tất cả" read as toggle verbs), demanding operation proof for a status report. `bat`/`tat` no longer match those phrases; real "Bật/Tắt config" orders stay CONFIGURATION (3 classifier copies + mirrors).
|
|
252
|
-
- Regression: `tests/hooks/releaseCompletionEvidence.test.js` (resolver → contract → gate), `tests/hooks/configPreferenceNotOrder.test.js` + real-prompt fixture.
|
|
253
|
-
- Limit: the resolver still emits no release-specific contract field; the obligation is enforced gate-side from the prompt. Verified offline on macOS/Node 22; replay, not a live host.
|
|
254
|
-
|
|
255
|
-
## 3.4.57 - 2026-10-22
|
|
256
|
-
|
|
257
|
-
**Completion Gate: background launch acknowledgements are not verification evidence; stale-contract reconciliation covers all obligation fields and the Stop boundary.**
|
|
258
|
-
- Bug: a host-backgrounded Bash runner (`run_in_background`) answers with a launch acknowledgement (`backgroundTaskId` / "Command running in background with ID: …") that carries no exit status, yet the ledger minted a `test-verified` verdict from it. The ack now records an `unproven` verification receipt (`exitStatusUnproven`, not success); a foreground exit status is still required. Completion of the background job is a separate host event and is not accepted as evidence by this change.
|
|
259
|
-
- Stale contract (3.4.56 follow-up): reconciliation now compares kind plus `operationGoal` / `configurationScope` / `releasePolicy` / `reportArtifact` and replaces classifier-owned fields wholesale (no leaked stale `operationGoal`); the Stop evaluator (`execution-ledger.mjs --evaluate-stop`) also reconciles in memory, so read-only / Bash-only tasks (which never trigger the router) no longer stay stuck. Only `contractOrigin: classified` contracts are eligible; system-carried and status-report contracts, and genuine MIXED/OPERATIONAL contracts, are unchanged.
|
|
260
|
-
- Regression: `tests/hooks/backgroundLaunchAckNotEvidence.test.js`, extended `tests/hooks/staleContractRederive.test.js`.
|
|
261
|
-
- Verified offline on macOS/Node 22; replay, not a live host.
|
|
262
|
-
|
|
263
|
-
## 3.4.56 - 2026-10-22
|
|
264
|
-
|
|
265
|
-
**Completion Gate: a stale self-classified contract is re-derived from the task's own prompt on continuation.**
|
|
266
|
-
- Field case: after the 3.4.55 classifier fix, an in-flight task kept its previously stored MIXED contract (continuations reused `routeSummary.executionContract` verbatim), so the gate still demanded `operation-verification` the user's prompt never ordered.
|
|
267
|
-
- Fix (`persistRouteState` in `skill-router.sh`): on a non-system continuation whose stored contract was classified by the router (`contractOrigin` != `carried-system`, not a status report), the contract is re-classified from `lastExplicitUserPromptText`; it is replaced only when the taskKind differs. Genuine contracts re-classify identically and are kept; system-carried contracts are never touched. No gate loosening, no receipts added.
|
|
268
|
-
- Regression: `tests/hooks/staleContractRederive.test.js` (stale MIXED corrected, genuine MIXED retained, carried-system untouched, new task classified on its own).
|
|
269
|
-
- Verified offline on macOS/Node 22; hook replay, not a live host.
|
|
270
|
-
|
|
271
|
-
## 3.4.55 - 2026-10-22
|
|
272
|
-
|
|
273
|
-
**Completion Gate: preference, conditional-future and prohibition-list text is not an order (field false-block).**
|
|
274
|
-
- Field case: an audit prompt saying "Ưu tiên configuration change" (preference) and "Nếu chỉ cần configuration change:" (future branch) was classified `CONFIGURATION`/external, so the Stop gate demanded `operation-verification` for a config edit nobody ordered (task `task-0c7425cb…`, `accepted-count=0`). A follow-up prompt listing what an authorization does NOT allow as bullets ("không cho phép: - restart …") plus "Negative controls: … chỉnh config/restart/deploy" was classified `MIXED`/service-runtime.
|
|
275
|
-
- Fix (`neutralizeNonOrders` in all three classifier copies: `src/core/executionContracts.js`, `route-resolver.mjs` mirror, `skill-router.sh` inline): a config preference, a prohibition-header bullet list (bounded to its own sentence and ended by a blank line) and "negative/positive controls" test-case bullets are neutralized; the audit/inspect/review lead filter now tolerates bullet prefixes. Gate, ledger and loop-breaker logic untouched; genuine config/restart/deploy orders and mixed implementation tasks keep their obligations (negative controls in tests).
|
|
276
|
-
- Regression: `tests/hooks/configPreferenceNotOrder.test.js` (src + installed + hook copy; real prompt fixtures `concurrency-audit-config-preference-prompt.txt`, `authorization-boundary-bullets-prompt.txt`). Independent review found one false negative (header regex swallowed the whole line before it); fixed and pinned.
|
|
277
|
-
- Housekeeping: `docs/STATUS.md` rotated under its 150-line budget (verbatim archive), concurrency audit report registered in `manifests/documentation.yaml`.
|
|
278
|
-
- Verified offline on macOS/Node 22; classifier replay, not a live OMP/Claude Code/Codex session; Windows/Linux not run.
|
|
279
|
-
|
|
280
|
-
## 3.4.54 - 2026-10-22
|
|
281
|
-
|
|
282
|
-
**Completion Gate: a stated prohibition is not an operation order (field false-block).**
|
|
283
|
-
- Fix: `classifyTaskSemantics` (src + shipped `route-resolver.mjs` mirror) treated "Production deploy/restart vẫn bị cấm", "production deploy vẫn bị chặn", "forbidden deploy/publish" and a bare list item "Production restart." as actionable runtime operations, so an audit/report prompt became MIXED + service-runtime and the Stop gate demanded `operation-verification` for a restart nobody ordered.
|
|
284
|
-
- New guard `operationNotOrdered` applies to the git/runtime operation checks only: a prohibition attached to the operation, or a bare production-operation noun item, orders nothing. Lift/past markers ("was prohibited, owner now allows", "even though it is not allowed", "khong bi chan") keep the clause an order. Real orders (restart + verify, deploy to production, fix + restart) are unchanged.
|
|
285
|
-
- Regression: `tests/hooks/prohibitionNotOrder.test.js` (78 cases across src + installed copies) with the real field prompt fixture. Ledger, gates and release logic untouched.
|
|
286
|
-
- Verified offline on macOS/Node 22; OMP/Claude Code coverage is classifier replay, not live host.
|
|
287
|
-
|
|
288
|
-
## 3.4.53 - 2026-10-22
|
|
289
|
-
|
|
290
|
-
**Stop-level closure of the UNPROVEN verdict (follow-up to 3.4.52).**
|
|
291
|
-
- `recoveryInstruction`: a non-current `unproven` verdict no longer yields "verdict is stale, re-run at current head"; it names the real gap (masked pipe, exit status unproven) and never "fix the failure".
|
|
292
|
-
- `stop-coordinator`: the `verify:pass` predicate and `goalVerificationVerdict` treat `unproven` as unmet but report "unproven (exit status not proven)", not a verified failure; `unproven` never counts as a pass.
|
|
293
|
-
- New Stop-level regressions (`tests/hooks/executionLedgerStopLifecycle.test.js`, GC-03e): UNPROVEN neither failure nor PASS; exhausted retry budget ends visibly without a recovery loop or write demand; same-task same-command PASS clears its own blocker, a different command / UNPROVEN rerun / another task does not; a superseded failure stops producing recovery text.
|
|
294
|
-
- Note: `tests/handoff/**` is excluded from `vitest`; run `tests/handoff/goalCompletion.test.js` with `node --test` (vitest exits 1 with "No test files found").
|
|
295
|
-
- Verified offline on macOS/Node 22; no live host proof.
|
|
296
|
-
|
|
297
|
-
## 3.4.52 - 2026-10-22
|
|
298
|
-
|
|
299
|
-
**Verification-loop ledger integrity (post-3.4.51 follow-up).**
|
|
300
|
-
- Fix: the fd number of a redirection (`2>&1`, `2>/dev/null`, `12>&1`) leaked into the verification-loop fingerprint as an argument, so `... .test.js 2>&1 | tail` was reported as `... .test.js 2`. Genuine numeric arguments and quoted text are untouched.
|
|
301
|
-
- A masked-pipe exit (`exitStatusUnproven`) now carries its own non-satisfying `unproven` verdict kind and is no longer counted as a definite failure toward the verification-loop streak.
|
|
302
|
-
- A later valid PASS of the same command on the same task/ledger clears its own sticky `verification-loop` blocker; evidence never crosses tasks or ledgers. Wrapper exit 0 still never becomes PASS.
|
|
303
|
-
- Regressions: `tests/hooks/executionLedger*.test.js`, `verificationExitProof`. Verified offline on macOS/Node 22; live host (Claude Code/OMP) in-memory state unverified, reload/`ukit update` + restart needed to load the fix.
|
|
304
|
-
|
|
305
|
-
## 3.4.51 - 2026-10-21
|
|
306
|
-
|
|
307
|
-
**Handoff orchestration integrity (CKit/OMP 18.12.5 incident audit).**
|
|
308
|
-
- OMP `Task` preflight (block-only): an ordinary agent carrying `sourceSession`/`model` (any value, even empty) or batch-root `sourceSession`/`model` is blocked with exact remove-field guidance; args are never rewritten, seance stays native-owned. Retries are counted per session + task instance + fingerprint (prose/name/status excluded), hashed, bounded and persisted; STOP guidance after two identical blocks.
|
|
309
|
-
- Stop-gate continuation provenance: only positively native/system-origin continuations (OMP branch `session-stop-continuation` match) reuse the task contract; a user paste of the gate text mints a fresh task. Pending-continuation marker is session/task/reason-hash scoped and lock-serialized.
|
|
310
|
-
- Resumable records now require exact stored task-id equality (lossy filename normalization could collide).
|
|
311
|
-
- Classifier: diagnostic "which config changed" reports are READ_ONLY; forward configuration orders and mixed source mutation + report stay CONFIGURATION / CODE_IMPLEMENTATION.
|
|
312
|
-
- Review-panel aggregation: FAIL/NOT_RUN dominate, underscore verdict aliases normalized, executable N/A becomes NOT_RUN; explicit product / integration-evidence / status-only review kinds. Handoff commands carry retry-fingerprint and pending-review continuity rules.
|
|
313
|
-
- Audit: `docs/analysis/HANDOFF-INTENT-DISPATCH-AUDIT.md`. Historical CKit wrapper/contract origin remains unknown; verified offline on macOS/Node 22 only; no live host proof, no CI dispatch.
|
|
314
|
-
|
|
315
53
|
## 3.4.22 - 2026-10-07
|
|
316
54
|
|
|
317
55
|
**Windows: `.sh` hooks no longer deny every Edit/Write/Bash.**
|
|
@@ -432,4 +170,4 @@ UKit gains a fourth runtime and loses one. omp (Oh My Pi) reaches full parity wi
|
|
|
432
170
|
|
|
433
171
|
**Known gap:** omp was not installed on the machine this release was built on. Every omp assertion in the test suite is against the generated `.omp/` tree and the bridge module's shape — the packed-tarball test confirms `templates/.omp/**` actually publishes, but nothing here observes omp's runtime behaviour. The three open verification items (does omp emit `session_start` on a post-compact continuation; omp's real base-URL key; omp's real auto-compact-window key) are recorded in `docs/AI_HANDOFF/tasks/TASK-008.md`.
|
|
434
172
|
|
|
435
|
-
Entries for 3.4.0 through 3.4.50 (except 3.4.14 and 3.4.22) are archived in the repository at `docs/archive/CHANGELOG-3.4.0-3.4.50.md`. Entries for 2.7.0 through 2.8.0 are archived in the repository at `docs/archive/CHANGELOG-2.7.0-2.8.0.md`. Entries for 2.5.1 through 2.6.11 are archived in the repository at `docs/archive/CHANGELOG-2.5.1-2.6.11.md`. Entries for 3.0.0 through 3.3.3 are archived in the repository at `docs/archive/CHANGELOG-3.0.0-3.3.3.md`. Entries for 2.1.0 through 2.2.16 (except 2.2.1) are archived in the repository at `docs/archive/CHANGELOG-2.1.0-2.2.16.md`; entries for 1.0.0 through 2.0.7 at `docs/archive/CHANGELOG-1.0.0-2.0.7.md` (neither is shipped in the npm package).
|
|
173
|
+
Entries for 3.4.76 through 3.4.85 are archived in the repository at `docs/archive/CHANGELOG-3.4.76-3.4.85.md`. Entries for 3.4.0 through 3.4.50 (except 3.4.14 and 3.4.22) are archived in the repository at `docs/archive/CHANGELOG-3.4.0-3.4.50.md`. Entries for 2.7.0 through 2.8.0 are archived in the repository at `docs/archive/CHANGELOG-2.7.0-2.8.0.md`. Entries for 2.5.1 through 2.6.11 are archived in the repository at `docs/archive/CHANGELOG-2.5.1-2.6.11.md`. Entries for 3.0.0 through 3.3.3 are archived in the repository at `docs/archive/CHANGELOG-3.0.0-3.3.3.md`. Entries for 2.1.0 through 2.2.16 (except 2.2.1) are archived in the repository at `docs/archive/CHANGELOG-2.1.0-2.2.16.md`; entries for 1.0.0 through 2.0.7 at `docs/archive/CHANGELOG-1.0.0-2.0.7.md` (neither is shipped in the npm package).
|
package/README.md
CHANGED
|
@@ -108,6 +108,7 @@ Maintainer/debug commands (advanced workflows, not team onboarding):
|
|
|
108
108
|
|
|
109
109
|
- `ukit diff` — preview file changes before install
|
|
110
110
|
- `ukit doctor` — validate UKit state and docs baseline
|
|
111
|
+
- `ukit doctor --prune-registry [--dry-run]` — drop user-registry entries whose project root provably no longer exists (unmounted volumes, stat errors and entries with user memory are kept; `--dry-run` prints counts only)
|
|
111
112
|
- `ukit uninstall` — remove UKit-managed assets
|
|
112
113
|
- `ukit status` — inspect shared UKit runtime state
|
|
113
114
|
- `ukit memory ...` — inspect/export/forget shared memory items
|
|
@@ -1322,6 +1322,19 @@ items:
|
|
|
1322
1322
|
packs:
|
|
1323
1323
|
- core
|
|
1324
1324
|
|
|
1325
|
+
# 3.5.2 WP-C: advisory PostToolUse Edit|Write module step (edit-graph hint).
|
|
1326
|
+
- id: hook-edit-hint-module
|
|
1327
|
+
type: hook
|
|
1328
|
+
sourceTemplate: .claude/hooks/edit-hint.mjs
|
|
1329
|
+
targetPath: .claude/hooks/edit-hint.mjs
|
|
1330
|
+
requires:
|
|
1331
|
+
- ukit-runtime-scripts
|
|
1332
|
+
mergeStrategy: overwrite_with_backup
|
|
1333
|
+
variables: []
|
|
1334
|
+
enabledByDefault: true
|
|
1335
|
+
packs:
|
|
1336
|
+
- core
|
|
1337
|
+
|
|
1325
1338
|
- id: hook-completion-gate
|
|
1326
1339
|
type: hook
|
|
1327
1340
|
sourceTemplate: .claude/hooks/completion-gate.sh
|
|
@@ -1862,6 +1875,16 @@ items:
|
|
|
1862
1875
|
packs:
|
|
1863
1876
|
- core
|
|
1864
1877
|
|
|
1878
|
+
- id: ukit-index-edit-graph-script
|
|
1879
|
+
type: config
|
|
1880
|
+
sourceTemplate: .claude/ukit/index/edit-graph.mjs
|
|
1881
|
+
targetPath: .claude/ukit/index/edit-graph.mjs
|
|
1882
|
+
mergeStrategy: overwrite_with_backup
|
|
1883
|
+
variables: []
|
|
1884
|
+
enabledByDefault: true
|
|
1885
|
+
packs:
|
|
1886
|
+
- core
|
|
1887
|
+
|
|
1865
1888
|
- id: ukit-index-route-catalog-script
|
|
1866
1889
|
type: config
|
|
1867
1890
|
sourceTemplate: .claude/ukit/index/route-catalog.mjs
|
package/package.json
CHANGED
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
{
|
|
2
2
|
"name": "@ngockhoale/ukit",
|
|
3
|
-
"version": "3.5.
|
|
3
|
+
"version": "3.5.3",
|
|
4
4
|
"description": "Install/update an index-first AI workspace for Claude Code, OpenAI Codex and omp (Oh My Pi).",
|
|
5
5
|
"license": "MIT",
|
|
6
6
|
"type": "module",
|
|
@@ -53,6 +53,12 @@
|
|
|
53
53
|
"!scripts/audit/c100/usage-provenance.mjs",
|
|
54
54
|
"!scripts/audit/decision-coverage.mjs",
|
|
55
55
|
"!scripts/audit/subagent-cost/**",
|
|
56
|
+
"!src/render/instructionRenderer.js",
|
|
57
|
+
"!src/diagnostics/ompStallSnapshot.js",
|
|
58
|
+
"!src/diagnostics/classifyHang.js",
|
|
59
|
+
"!src/skill/auditSkill.js",
|
|
60
|
+
"!src/core/observability/evaluation/perturbation.js",
|
|
61
|
+
"!src/core/observability/evaluation/experimentPlan.js",
|
|
56
62
|
"!scripts/bench/**",
|
|
57
63
|
"scripts/bench/decision-runtime-metrics.mjs",
|
|
58
64
|
"!scripts/perf/**",
|
|
@@ -61,6 +67,7 @@
|
|
|
61
67
|
"!template_project/**/.smbdelete*",
|
|
62
68
|
"!template_project/.smbdelete*",
|
|
63
69
|
"!template_project/.ukit/**",
|
|
70
|
+
"!template_project/instructions/overlays/repo.md",
|
|
64
71
|
"!template_project/**/skill-router-state.json",
|
|
65
72
|
"CHANGELOG.md",
|
|
66
73
|
"template_project/.gitignore",
|
|
@@ -30,6 +30,7 @@ import { inspectPermissions } from '../../core/permissionDoctor.js';
|
|
|
30
30
|
import { inspectHookChainHealth } from '../../core/hookChainDoctor.js';
|
|
31
31
|
import { findOpencodeArtifacts, opencodeSteerMessage } from '../../core/opencodeSteer.js';
|
|
32
32
|
import { buildUserPaths } from '../../core/userPaths.js';
|
|
33
|
+
import { pruneRegistry } from '../../core/memory/registryPrune.js';
|
|
33
34
|
import { loadPlaybooks } from '../../core/userPlaybooks.js';
|
|
34
35
|
import { userMemoryStats } from '../../core/memory/userMemory.js';
|
|
35
36
|
import {
|
|
@@ -54,8 +55,8 @@ import {
|
|
|
54
55
|
} from '../../core/observability/support/schedule.js';
|
|
55
56
|
|
|
56
57
|
export const DOCTOR_HELP_FLAGS = new Set(['--help', '-h']);
|
|
57
|
-
const KNOWN_FLAGS = new Set([...DOCTOR_HELP_FLAGS, '--skills', '--gateway', '--docs', '--permissions', '--conformance', '--json']);
|
|
58
|
-
const SUPPORTED_FLAGS_LIST = '--help, -h, --skills, --gateway, --docs, --permissions, --conformance, --json';
|
|
58
|
+
const KNOWN_FLAGS = new Set([...DOCTOR_HELP_FLAGS, '--skills', '--gateway', '--docs', '--permissions', '--conformance', '--json', '--prune-registry', '--dry-run']);
|
|
59
|
+
const SUPPORTED_FLAGS_LIST = '--help, -h, --skills, --gateway, --docs, --permissions, --conformance, --json, --prune-registry, --dry-run';
|
|
59
60
|
|
|
60
61
|
export function printDoctorHelp() {
|
|
61
62
|
console.log('Usage: ukit doctor [options]');
|
|
@@ -68,6 +69,8 @@ export function printDoctorHelp() {
|
|
|
68
69
|
console.log(' --gateway Live gateway probe (streaming + non-streaming); advisory only');
|
|
69
70
|
console.log(' --docs Run doc-contract checks (manifests/documentation.yaml projects only)');
|
|
70
71
|
console.log(' --permissions Print the full per-host permission report table');
|
|
72
|
+
console.log(' --prune-registry Remove user-registry entries whose project root provably no longer exists');
|
|
73
|
+
console.log(' (unmounted volumes, stat errors and entries with user memory are kept); add --dry-run to print counts only');
|
|
71
74
|
console.log(' --json Emit the permission report as JSON ({permissions:{...}})');
|
|
72
75
|
console.log(' --conformance Print the engine conformance matrix (enforcement class per host)');
|
|
73
76
|
}
|
|
@@ -430,6 +433,15 @@ export async function runDoctor({ packageRoot, projectRoot, argv = [], homeDir =
|
|
|
430
433
|
return;
|
|
431
434
|
}
|
|
432
435
|
|
|
436
|
+
// Standalone maintenance surface: prune stale project-registry entries (counts only).
|
|
437
|
+
if (argv.includes('--prune-registry')) {
|
|
438
|
+
const dryRun = argv.includes('--dry-run');
|
|
439
|
+
const r = await pruneRegistry({ homeDir, dryRun, keepRoots: [projectRoot] });
|
|
440
|
+
console.log(`[UKit] registry prune${dryRun ? ' (dry-run, nothing written)' : ''}: status=${r.status} total=${r.total} stale=${r.stale} removed=${r.removed} keptWithMemory=${r.keptWithMemory} unproven=${r.unproven}`);
|
|
441
|
+
if (r.status !== 'ok' && r.status !== 'no-registry') process.exitCode = 1;
|
|
442
|
+
return;
|
|
443
|
+
}
|
|
444
|
+
|
|
433
445
|
// TASK-006 / FR-006 — `--json` is a machine surface: emit ONLY the permission
|
|
434
446
|
// report object (plus TASK-016's flight-recorder snapshot), exit 1 when any
|
|
435
447
|
// FAIL class fires.
|