@ngockhoale/ukit 3.4.77 → 3.4.79
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 +12 -471
- package/manifests/documentation.yaml +12 -0
- package/package.json +1 -1
- package/template_project/.claude/commands/ukit/handoff-create.md +7 -0
- package/template_project/.claude/commands/ukit/handoff-fullstack.md +5 -0
- package/template_project/.claude/hooks/skill-router.sh +104 -21
- package/template_project/.claude/skills/advisor-plan/REFERENCE.md +2 -0
- package/template_project/.claude/ukit/index/publication-intent.mjs +61 -5
- package/template_project/.claude/ukit/runtime/resumable-run.mjs +12 -0
- package/template_project/.claude/ukit/runtime/stop-coordinator.mjs +424 -26
- package/template_project/docs/AI_HANDOFF/queued/README.md +32 -0
package/CHANGELOG.md
CHANGED
|
@@ -2,6 +2,17 @@
|
|
|
2
2
|
|
|
3
3
|
All notable changes to UKit are documented here.
|
|
4
4
|
|
|
5
|
+
## 3.4.79 - 2026-11-07
|
|
6
|
+
|
|
7
|
+
- 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 `QueueIntent: superseded|stale` + `QueueSupersededBy: archive:cycle-NN[, …] | queue: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 A frozen `ScopeQueue` naming a superseded blueprint warns instead of consuming it silently; a sibling `queue/` directory is reported, never migrated.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`.
|
|
8
|
+
- 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.
|
|
9
|
+
- 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).
|
|
10
|
+
- 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.
|
|
11
|
+
|
|
12
|
+
## 3.4.78 - 2026-11-06
|
|
13
|
+
|
|
14
|
+
- 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.
|
|
15
|
+
|
|
5
16
|
## 3.4.77 - 2026-11-06
|
|
6
17
|
|
|
7
18
|
- 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.
|
|
@@ -2782,381 +2793,6 @@ Verified: `yarn test` 1277/1277 (74 files), `test:release-core` 1270, `test:arti
|
|
|
2782
2793
|
`release:verify` all checks passed, 13/13 handoff node suites, sensitive-data-guard
|
|
2783
2794
|
live-fire 5/5 as designed.
|
|
2784
2795
|
|
|
2785
|
-
## 2.2.16 - 2026-09-06
|
|
2786
|
-
|
|
2787
|
-
The vision gate is gone — image work is advisory-routed, never blocked. Since TASK-008, a
|
|
2788
|
-
pasted image on the UNIC gateway armed `vision-gate.sh`, which refused every Edit/Write until
|
|
2789
|
-
`ukit-vision-analyst` had produced receipts for ALL pending images. In practice this
|
|
2790
|
-
repeatedly stranded sessions (receipts that never landed, one stale marker blocking
|
|
2791
|
-
unrelated writes), so the maintainer retired the hard block: `vision-gate.sh` and its
|
|
2792
|
-
manifest entry `hook-vision-gate` are removed from both the live tree and templates, the omp
|
|
2793
|
-
bridge and `hook-chain-runner.mjs` no longer list it, and `vision-router.sh` now says "never
|
|
2794
|
-
guess at image contents" instead of claiming edits are GATED. Routing to
|
|
2795
|
-
`ukit-vision-analyst` still happens — on this gateway the main tier genuinely cannot read
|
|
2796
|
-
images — but nothing blocks while it runs. Supersedes the 2.2.15 escape hatch, which patched
|
|
2797
|
-
a message on a gate that no longer exists. Verified: vision-hook, route-vision,
|
|
2798
|
-
vision-agent, install-wiring, provision-worktree, and ompHookBridge (52/52) all green.
|
|
2799
|
-
(An interim 2026-09-05 patch that added an escape hatch to the block message is included
|
|
2800
|
-
here and immediately superseded — the gate itself no longer exists.)
|
|
2801
|
-
|
|
2802
|
-
## 2.2.14 - 2026-09-05
|
|
2803
|
-
|
|
2804
|
-
The context hard-cap gate no longer bricks healthy sessions with a phantom token count. The
|
|
2805
|
-
tracker behind `compact.hardCapTokens` accumulates every prompt and tool output across a
|
|
2806
|
-
session, and it only reset on the PreCompact hook — which fires for a manual `/compact` but not
|
|
2807
|
-
for every harness auto-compact path (SDK/VSCode being the known gap). A session whose live
|
|
2808
|
-
context had been compacted down to almost nothing could therefore reach a tracked 528k against
|
|
2809
|
-
the 500k cap: Read/Grep stayed open but every Edit/Write/Bash was refused, and the user saw the
|
|
2810
|
-
assistant freeze with no error. Every real compaction — manual or automatic — now also leaves a
|
|
2811
|
-
`compact_boundary` entry in the session transcript, so `compact-threshold.mjs` rebase the
|
|
2812
|
-
counter from that record (`syncCompactPressureStateWithTranscript`, stamped
|
|
2813
|
-
`lastCompactBoundaryAt` so each boundary resets exactly once). Both the prompt-time hook and
|
|
2814
|
-
`context-hardcap-gate.sh` sync before acting, and the reset is persisted so the gate re-arms
|
|
2815
|
-
instead of disabling itself.
|
|
2816
|
-
|
|
2817
|
-
When the cap is genuinely reached, the model is now ordered to tell the user instead of going
|
|
2818
|
-
quiet. The hard-phase prompt warning and the gate's block message both lead with an explicit
|
|
2819
|
-
directive — tell the user "gõ /compact ngay" as the first sentence of the reply, since typing
|
|
2820
|
-
"continue" alone hits the same block — and the advisory soft phase tells the model to surface
|
|
2821
|
-
`/compact` at the next natural pause. `context-window-guard.sh`'s overflow directive likewise
|
|
2822
|
-
demands the compact instruction as the first sentence rather than a buried suggestion.
|
|
2823
|
-
|
|
2824
|
-
Vietnamese concrete bug reports route to the debug lane again. Signals such as `lỗi`, `sửa`,
|
|
2825
|
-
`đăng nhập`, `đang ở đâu` were written inside `\b(...)\b`, but JavaScript `\b` is ASCII-only and
|
|
2826
|
-
those words have diacritic edges, so the patterns could never match — "app bị lỗi khi đăng nhập,
|
|
2827
|
-
sửa giúp tôi" fell through to informational lanes while the identical English prompt routed
|
|
2828
|
-
correctly. All Vietnamese intent patterns in the router and catalog now use explicit
|
|
2829
|
-
`(?<![A-Za-z0-9_])` / `(?![A-Za-z0-9_])` edges; regression tests cover the debug-lane prompt,
|
|
2830
|
-
the open-ended status question, and hook-level state.
|
|
2831
|
-
|
|
2832
|
-
## 2.2.13 - 2026-09-05
|
|
2833
|
-
|
|
2834
|
-
Question-shaped prompts no longer get blocked by the completion gate. "Đây là lỗi gì?",
|
|
2835
|
-
"model bạn đang dùng là model gì", or "what is this error?" route to the informational lane and
|
|
2836
|
-
answer freely; only prompts that actually ask for a change (fix/sửa/build/…) still owe write and
|
|
2837
|
-
verification evidence before claiming done. Two layered causes: the Vietnamese question pattern
|
|
2838
|
-
only matched adjacent `là gì` (missing "là model gì" / "là lỗi gì"), and the `\blỗi` veto never
|
|
2839
|
-
fired at all — JavaScript `\b` word boundaries don't exist around diacritic characters. A
|
|
2840
|
-
`strongQuestion` signal (an explicit `?`, a leading interrogative, or the Vietnamese pattern) now
|
|
2841
|
-
overrides error-wording vetoes, while "error when I run tests" and "why is build failing?" still
|
|
2842
|
-
route to investigation.
|
|
2843
|
-
|
|
2844
|
-
The source-code index understands SQL from the canonical source, not just the installed mirror.
|
|
2845
|
-
`CREATE TABLE/VIEW/FUNCTION/PROCEDURE/INDEX/TYPE` and `ALTER TABLE` statements are extracted as
|
|
2846
|
-
symbols with schema-qualified names captured by their bare object name (`qas.fn_rpt_billing` →
|
|
2847
|
-
`fn_rpt_billing`), and `.sql` files are tracked and code-indexed. The schema bump had previously
|
|
2848
|
-
landed only in the runtime mirror; `src/index/` now carries the same v8 schema so the mirror-parity
|
|
2849
|
-
test holds and the `scripts/index` lane benefits too.
|
|
2850
|
-
|
|
2851
|
-
The handoff model-tier guard closes an under-block found while scenario-testing it. Once a real
|
|
2852
|
-
`EXECUTOR_MODEL` or `REVIEWER_MODEL` was on disk, an in-place rewrite could silently demote the
|
|
2853
|
-
executor to a lite tier or set executor == reviewer; validation now also fires when an
|
|
2854
|
-
already-real model line changes value (same-model rewrites and upward escalation stay free). The
|
|
2855
|
-
push-time backstop was unaffected — this moves the rejection to write time, where it belongs.
|
|
2856
|
-
|
|
2857
|
-
`handoff.maxParallelAgents` template default is now 2, matching the operator setting; the config
|
|
2858
|
-
help prose no longer claims a default of 12 one line above a configured 2.
|
|
2859
|
-
|
|
2860
|
-
## 2.2.12 - 2026-09-04
|
|
2861
|
-
|
|
2862
|
-
`ukit install` and `ukit update` now clean up broken Claude Code hook registrations automatically.
|
|
2863
|
-
|
|
2864
|
-
A hook whose stdout contains a `hookSpecificOutput` object without the required `hookEventName`
|
|
2865
|
-
field fails Claude Code's schema validation on **every** tool call, surfacing as a repeating
|
|
2866
|
-
`PreToolUse:<Tool> hook error — Hook JSON output validation failed`. Such a hook has no effect
|
|
2867
|
-
either way, since its output is discarded — it produces nothing but noise.
|
|
2868
|
-
|
|
2869
|
-
New `src/core/repairBrokenHooks.js` detects that shape generically (not keyed to any one vendor)
|
|
2870
|
-
across user settings, project settings, and `settings.local.json`, expanding
|
|
2871
|
-
`$CLAUDE_PROJECT_DIR`/`$HOME`/`~` in hook commands and also catching hooks inlined into
|
|
2872
|
-
`settings.json` instead of a script file. Offending registrations are removed after a
|
|
2873
|
-
`.ukit-backup` restore point is written next to each settings file. Detection matches only real
|
|
2874
|
-
key assignments, so prose mentions of the field in a hook's own comments or docstrings never
|
|
2875
|
-
trigger a removal.
|
|
2876
|
-
|
|
2877
|
-
Wired into the install pipeline; since `ukit update` re-runs `ukit install`, both commands are
|
|
2878
|
-
covered with no new flags or commands to learn — teammates still only need `ukit install`.
|
|
2879
|
-
|
|
2880
|
-
Scope is deliberate: UKit removes the *registration* only. An unregistered hook script never
|
|
2881
|
-
executes, so that is sufficient, and UKit does not delete the script file or uninstall the tool
|
|
2882
|
-
that wrote it. When the owning tool can be identified from the hook source, the install report
|
|
2883
|
-
names it and notes that it may re-register on activation, leaving permanent removal as a human
|
|
2884
|
-
decision.
|
|
2885
|
-
|
|
2886
|
-
Known real-world instance: VSCode extension `tjcg.auto-accept-claude-code` v0.5.0 hardcodes two
|
|
2887
|
-
such hook templates and rewrites its hook file on activation.
|
|
2888
|
-
|
|
2889
|
-
## 2.2.11 - 2026-08-29
|
|
2890
|
-
|
|
2891
|
-
Same-day follow-up to 2.2.10. The two premises that release shipped as unverified were checked
|
|
2892
|
-
against omp's own installed TypeScript source (`@oh-my-pi/pi-coding-agent` v18.0.10, not a live
|
|
2893
|
-
session — omp still was not run) and both hold:
|
|
2894
|
-
|
|
2895
|
-
- **`steer` after a compaction is accepted, not dropped.** `session-maintenance.ts` fires
|
|
2896
|
-
`session_compact` only after `agent.replaceMessages()`/`rebaseAfterCompaction()` already ran, so
|
|
2897
|
-
the session is idle at that point, and `agent-session.ts`'s `sendCustomMessage()` non-streaming
|
|
2898
|
-
branch appends the message to the live context regardless of `deliverAs`.
|
|
2899
|
-
- **`display: true` does trigger a visible re-render.** `extension-ui-controller.ts`'s
|
|
2900
|
-
`applyCustomMessageDisplay()` calls `rebuildChatFromMessages()` for the non-streaming case (the
|
|
2901
|
-
streaming case renders via the `message_end` event instead).
|
|
2902
|
-
|
|
2903
|
-
`templates/.omp/hooks/pre/ukit-bridge.js`'s two `TODO(verify)` comments (PLAN.md §3 D3, D4) are
|
|
2904
|
-
updated to record this — from "unverified, additive, worst-case-unchanged" to "source-confirmed,
|
|
2905
|
-
one rung below a live repro." No runtime behaviour changes; both delivery paths (`steer` +
|
|
2906
|
-
`nextTurn`, and the `finalNotice` fallback) stay in place as defense in depth either way. A stale
|
|
2907
|
-
local `.omp/hooks/pre/ukit-bridge.js` install artifact in this repo — untracked, regenerated by
|
|
2908
|
-
`ukit install`, and out of sync with `templates/.omp/` since before 2.2.10 shipped — was also found
|
|
2909
|
-
and resynced; that file is never part of the published package.
|
|
2910
|
-
|
|
2911
|
-
## 2.2.10 - 2026-08-29
|
|
2912
|
-
|
|
2913
|
-
User report: long omp (Oh My Pi) sessions go idle roughly 15 minutes into a task — no output, no
|
|
2914
|
-
error, no notification, indistinguishable from a hang. Typing `continue` makes the session resume
|
|
2915
|
-
and finish correctly (the user's observation, not a measurement — no field reproduction of the
|
|
2916
|
-
stall exists on this machine to time it). Three real root causes were found by reading the source,
|
|
2917
|
-
not guessed:
|
|
2918
|
-
|
|
2919
|
-
### Fixed
|
|
2920
|
-
|
|
2921
|
-
- **`map-impact` — the mode wide implementation work actually routes to — was never gated by the
|
|
2922
|
-
completion loop.** `templates/.claude/ukit/runtime/execution-ledger.mjs:12`'s `IMPLEMENT_MODES`
|
|
2923
|
-
set omitted it, while `src/index/taskRouting.js:414` routes exactly this kind of task to
|
|
2924
|
-
`map-impact` and `src/index/taskRouting.js:610` gives it the strictest contract
|
|
2925
|
-
(`impact-evidence`, `write-evidence`, `verification-evidence`). So the longest tasks declared the
|
|
2926
|
-
most evidence and got no gate at all — the session could stop with nothing written and nothing
|
|
2927
|
-
verified, silently. `map-impact` is now in `IMPLEMENT_MODES`.
|
|
2928
|
-
- **Reaching the continuation cap ended the session with a log line, not a message.** At
|
|
2929
|
-
`MAX_CONTINUATIONS` (6, unchanged), `evaluateCompletion` returned `{ capped: true }` and the only
|
|
2930
|
-
reaction was `pi.logger.warn` — a log sink, not the transcript. This is also the mechanical reason
|
|
2931
|
-
a new prompt (`continue`) appeared to "fix" the stall: a new prompt writes a fresh `requestKey`,
|
|
2932
|
-
which resets the per-request continuation count back to zero. The gate now spends exactly one
|
|
2933
|
-
final continuation (`finalNotice`) instructing the model to state what is unfinished and stop, and
|
|
2934
|
-
persists a one-shot `notified` flag on the ledger so the notice cannot refire.
|
|
2935
|
-
- **The post-compact resume was delivered only to the next user turn, and every UKit→omp message was
|
|
2936
|
-
`display: false`.** After an omp auto-compaction, the `docs/AI_HANDOFF/RUN.md` resume cursor was
|
|
2937
|
-
sent with `deliverAs: 'nextTurn'` only — if compaction ended the agent loop, the cursor waited for
|
|
2938
|
-
a human to type something. Separately, every message the bridge built hardcoded `display: false`,
|
|
2939
|
-
so UKit had no user-visible channel in omp at all, for context or for stop notices.
|
|
2940
|
-
`runSessionCompact` now sends the resume as `steer` (reaches a still-running turn) *and*
|
|
2941
|
-
`nextTurn` (unchanged fallback), and `runSessionStop` now sends one `display: true` notice
|
|
2942
|
-
whenever it stops with evidence still missing — including the previously-silent
|
|
2943
|
-
`review-release` / ungated-mode case.
|
|
2944
|
-
|
|
2945
|
-
**Two premises behind the third fix are unverified — omp is not installed in this development
|
|
2946
|
-
environment.** Whether omp still accepts a `steer` message after a compaction, and whether omp
|
|
2947
|
-
honours `display: true` on a bridge message, were not and could not be reproduced or measured here.
|
|
2948
|
-
Both changes are additive: if either premise is false, the `nextTurn` delivery and the model's own
|
|
2949
|
-
`finalNotice` message (from the second fix) still reach the user, so the worst case is unchanged
|
|
2950
|
-
from today's behaviour, never worse. See `docs/STATUS.md` for the open thread and the evidence that
|
|
2951
|
-
would confirm the continuation-cap path directly.
|
|
2952
|
-
|
|
2953
|
-
- **`templates/.omp/RULES.md` gains one sticky rule**: ending a turn with no output is a defect, not
|
|
2954
|
-
a pause — a model-side backstop for whatever a code gate cannot reach.
|
|
2955
|
-
|
|
2956
|
-
## 2.2.9 - 2026-08-28
|
|
2957
|
-
|
|
2958
|
-
Retuning UKit's context budget used to mean editing five scattered places and hand-syncing two
|
|
2959
|
-
of them. It is now **three numbers in one file**, and everything else is derived at install time.
|
|
2960
|
-
|
|
2961
|
-
The shipped defaults also move up for a 1M context window: compaction is recommended at 150,000
|
|
2962
|
-
estimated tokens (was 50,000) and the hard cap sits at 500,000 — half the window (was 220,000).
|
|
2963
|
-
|
|
2964
|
-
### Changed
|
|
2965
|
-
|
|
2966
|
-
- **`compact.tokenThreshold` / `compact.hardCapTokens` / `compact.autoCompactWindowRatio` in
|
|
2967
|
-
`.ukit/storage/config.json` are now the only knobs.** New `src/core/compact/contextBudget.js`
|
|
2968
|
-
reads them and derives the rest, so the defaults in `runtimeConfig.js` and
|
|
2969
|
-
`compact/threshold.js` no longer carry their own copies of the number. Edit the three values,
|
|
2970
|
-
rerun `ukit install`, and the whole budget follows. Documented in README →
|
|
2971
|
-
*Retuning The Context Budget (maintainers)*, including a window→cap sizing table, because the
|
|
2972
|
-
new 500,000 default sits **above** a 200k or 256k model's window, where the gate could never fire.
|
|
2973
|
-
- **New defaults:** `tokenThreshold` 50,000 → **150,000**; `hardCapTokens` 220,000 → **500,000**;
|
|
2974
|
-
auto-compact fires at **70%** of the cap → 350,000.
|
|
2975
|
-
|
|
2976
|
-
### Fixed
|
|
2977
|
-
|
|
2978
|
-
- **omp auto-compact had silently drifted out of sync.** `.omp/config.yml`'s
|
|
2979
|
-
`compaction.thresholdTokens` was still 180,000 while Claude Code's
|
|
2980
|
-
`CLAUDE_CODE_AUTO_COMPACT_WINDOW` had moved on — the two were separate hand-maintained numbers
|
|
2981
|
-
with nothing pinning them together. Both are now rendered from the same derived value at
|
|
2982
|
-
install time, so **omp and Claude Code auto-compact at the same point** and cannot drift apart
|
|
2983
|
-
again.
|
|
2984
|
-
- **`ompDocsSync > historical records are untouched by this cycle` failed on every new plan doc.**
|
|
2985
|
-
The guard compared *all* changes under `docs/plans/` and friends against an old cycle's base
|
|
2986
|
-
commit, so the first plan added afterwards (2.2.8's) tripped it permanently. Narrowed to
|
|
2987
|
-
`--diff-filter=MDR`: rewriting, deleting or renaming a historical record still fails; adding a
|
|
2988
|
-
new plan does not.
|
|
2989
|
-
|
|
2990
|
-
## 2.2.8 - 2026-08-28
|
|
2991
|
-
|
|
2992
|
-
UKit already picked the right files to open, but the step *after* that — actually entering the
|
|
2993
|
-
file — was unoptimised: the agent read the whole thing to find the one function it needed. This
|
|
2994
|
-
release adds a signature outline so it can jump straight to a line instead.
|
|
2995
|
-
|
|
2996
|
-
The outline is a map for **locating** code, not a substitute for reading it. Any code that is
|
|
2997
|
-
about to be changed must still be read; the saving comes from narrower reads, not skipped ones.
|
|
2998
|
-
That rule is stated in the shipped `CLAUDE.md` / `AGENTS.md` guidance next to the feature itself.
|
|
2999
|
-
|
|
3000
|
-
### Added
|
|
3001
|
-
|
|
3002
|
-
- **Indexed symbols now carry `line` and `signature`.** `query-index` and `resolve-context` print
|
|
3003
|
-
an `outline:` block (`line: signature`) for the top suspect files, so an agent can go straight
|
|
3004
|
-
to `Read(file, offset=<line>)`. On this repo, locating the top-3 suspects for a representative
|
|
3005
|
-
query costs 2,082 bytes / 25 lines instead of 131,702 bytes / 4,064 lines of whole-file reads.
|
|
3006
|
-
On by default; `--no-outline` turns it off. Capped at the top 3 files and 12 symbols each.
|
|
3007
|
-
- **`getFileOutline()`** in the index core, returning line-sorted, name-deduped rows. Missing
|
|
3008
|
-
file, missing artifact, or a stale schema all resolve to `[]`, so callers can render
|
|
3009
|
-
unconditionally.
|
|
3010
|
-
- **First behavioural parity test for the hand-maintained index mirror**
|
|
3011
|
-
(`tests/index/indexCoreParity.test.js`). `templates/.claude/ukit/index/lib/index-core.mjs`
|
|
3012
|
-
duplicates `src/index/*` by hand and was flagged in `MIRROR_PAIRS` but never actually compared;
|
|
3013
|
-
drift could only be caught by reading both files. It is now asserted.
|
|
3014
|
-
- **Both harnesses covered.** omp runs the same Node helpers and reads `AGENTS.md`, so the
|
|
3015
|
-
guidance ships identically in `templates/CLAUDE.md` and `templates/AGENTS.md` (byte-identical
|
|
3016
|
-
section bodies, enforced by `tests/consistency/contextDocsParity.test.js`). No separate omp
|
|
3017
|
-
code path was needed.
|
|
3018
|
-
|
|
3019
|
-
### Fixed
|
|
3020
|
-
|
|
3021
|
-
- **`.vue` line numbers would have been silently wrong.** `extractScriptContent` joined the
|
|
3022
|
-
`<script>` blocks together, so a symbol in the second block reported a line number offset by the
|
|
3023
|
-
whole `<template>` above it. Script regions are now masked in place, preserving byte offsets, so
|
|
3024
|
-
reported lines match the file on disk.
|
|
3025
|
-
|
|
3026
|
-
### Changed
|
|
3027
|
-
|
|
3028
|
-
- `INDEX_SCHEMA_VERSION` 6 → 7. Existing `.cache/index/` artifacts are invalidated and rebuilt
|
|
3029
|
-
automatically on first use; no migration step and no user action.
|
|
3030
|
-
|
|
3031
|
-
## 2.2.7 - 2026-08-28
|
|
3032
|
-
|
|
3033
|
-
Follow-up to 2.2.6: that release only made the wrong-runtime failure message *explain* the
|
|
3034
|
-
`UKIT_NODE_PATH` workaround better, but a user still had to diagnose the error and export the
|
|
3035
|
-
variable by hand. This release removes the manual step for the common case.
|
|
3036
|
-
|
|
3037
|
-
### Fixed
|
|
3038
|
-
|
|
3039
|
-
- **The omp bridge now auto-detects a working Node.js binary instead of requiring
|
|
3040
|
-
`UKIT_NODE_PATH` to be set by hand.** [`ukit-bridge.js`](.omp/hooks/pre/ukit-bridge.js) probes
|
|
3041
|
-
`process.execPath` with `--version` before trusting it; if that fails (e.g. a Bun-hosted or
|
|
3042
|
-
self-compiled omp reporting its own binary as `process.execPath`), it falls back to the bare
|
|
3043
|
-
`node` command name and lets normal OS `PATH` resolution find a real Node.js binary — this
|
|
3044
|
-
works the same way on macOS, Linux, and Windows, since Node's `child_process` spawn does `PATH`
|
|
3045
|
-
lookup for bare executable names on all three. An explicit `UKIT_NODE_PATH` still always wins
|
|
3046
|
-
and skips the probe. The probe result is cached for the life of the omp process, so it only
|
|
3047
|
-
runs once, not on every tool call. The transport-failure message's `UKIT_NODE_PATH` hint now
|
|
3048
|
-
only appears when both the probe and the `node` PATH fallback have failed.
|
|
3049
|
-
|
|
3050
|
-
### Added
|
|
3051
|
-
|
|
3052
|
-
- **1 regression test** in `tests/hooks/ompHookBridge.test.js` — with `UKIT_NODE_PATH` unset and
|
|
3053
|
-
`process.execPath` mocked to fail the `--version` probe, the bridge falls back to invoking the
|
|
3054
|
-
bare `node` command instead of blocking the tool call.
|
|
3055
|
-
|
|
3056
|
-
## 2.2.6 - 2026-08-27
|
|
3057
|
-
|
|
3058
|
-
Follow-up to a real-world report: a user on a Bun-hosted/self-compiled omp build kept getting
|
|
3059
|
-
every Edit/Write blocked with `UKit OMP hook runner failed before producing a valid result
|
|
3060
|
-
(killed=true, ..., runtime=/Users/.../bin/omp)`. The diagnostic added in 2.2.4 was already
|
|
3061
|
-
honest about the failure, but it did not say *why* — `runtime=` pointed straight at omp's own
|
|
3062
|
-
binary (proof that `process.execPath` inside that omp build resolves to omp itself, not a real
|
|
3063
|
-
Node.js binary), yet the message never told the user the fix was to set `UKIT_NODE_PATH`.
|
|
3064
|
-
|
|
3065
|
-
### Fixed
|
|
3066
|
-
|
|
3067
|
-
- **The omp bridge's transport-failure message named the wrong-runtime symptom but never the
|
|
3068
|
-
fix.** [`ukit-bridge.js`](.omp/hooks/pre/ukit-bridge.js) already resolved
|
|
3069
|
-
`process.env.UKIT_NODE_PATH || process.execPath` and reported the resolved `runtime` on failure
|
|
3070
|
-
(2.2.4), but a user seeing `runtime=/Users/lenk/.local/bin/omp` had no way to know that string
|
|
3071
|
-
meant "wrong binary" or what to do about it. When `UKIT_NODE_PATH` is unset, the failure
|
|
3072
|
-
message now appends: the reported runtime may not be a real Node.js binary, and the exact fix —
|
|
3073
|
-
`export UKIT_NODE_PATH="$(command -v node)"`. No behavior change when `UKIT_NODE_PATH` is
|
|
3074
|
-
already set.
|
|
3075
|
-
|
|
3076
|
-
### Added
|
|
3077
|
-
|
|
3078
|
-
- **2 regression tests** in `tests/hooks/ompHookBridge.test.js` (case 13) — the wrong-runtime
|
|
3079
|
-
hint appears when `UKIT_NODE_PATH` is unset, and is omitted when it is already set.
|
|
3080
|
-
|
|
3081
|
-
## 2.2.5 - 2026-08-27
|
|
3082
|
-
|
|
3083
|
-
Follow-up to the user's report that UKit had become too heavy-handed: too many hook scripts firing per tool call, too much blocking, and a concrete false-positive in the destructive-command gate itself. This wave trims default hook wiring and fixes the false positive, while explicitly preserving `rm`/destructive-command protection as instructed.
|
|
3084
|
-
|
|
3085
|
-
### Changed
|
|
3086
|
-
|
|
3087
|
-
- **`PreToolUse:Bash` now runs 4 hook scripts instead of 6.** `verification-guard.sh` was removed from the default Bash chain (both `.claude/settings.json` copies, the omp bridge's `HOOK_EVENT_MAP`/`FAIL_CLOSED_SCRIPTS`, and `hook-chain-runner.mjs`'s own classification). Its hard-block path is gated behind `UKIT_VERIFICATION_GUARD_ENFORCE=1`, unset by default, so it was pure per-call overhead (a full `node` spawn) for zero present safety value. The script and its standalone tests are untouched — still shippable, opt-in.
|
|
3088
|
-
- **`PreToolUse:Read|Grep|Glob` no longer runs `skill-router.sh`.** Automatic skill activation still runs on `UserPromptSubmit` and `Edit|Write`, where a skill choice actually matters;
|
|
3089
|
-
|
|
3090
|
-
### Fixed
|
|
3091
|
-
|
|
3092
|
-
- **`block-dangerous.sh` misidentified dangerous *text* as a dangerous *command*.** Both its per-pattern denylist and its `rm` handling matched against the entire raw command string with no awareness of shell quoting, so `grep -rn "rm -rf" .` or `echo "rm -rf /tmp/foo"` — which merely mention the text — were blocked as if they executed it. Quoted spans are now scrubbed before matching, but only when the command's own invoked program is a pure text tool (`grep|egrep|fgrep|rg|ag|awk|sed|echo|printf|cat`) and it does not pipe into a shell/eval sink (`| bash|sh|zsh|eval|xargs`), since `echo "rm -rf /" | bash` really does execute the quoted text. Every other command shape is scanned exactly as strictly as before.
|
|
3093
|
-
- **`rm -rf ./dist` (and other allowlisted safe-cleanup targets with a `./` prefix) was blocked as an "unsafe delete target."** The unsafe-target check ran before the safe-cleanup allowlist check and its pattern misfired on the literal `./` prefix. Reordered so the allowlist is checked first; genuinely risky targets like `rm -rf ../important` still block.
|
|
3094
|
-
|
|
3095
|
-
### Added
|
|
3096
|
-
|
|
3097
|
-
- **11 regression tests** in `tests/hooks/blockDangerous.test.js` covering the false-positive fix, the closed shell-pipe escape hatch, the `./dist` allowlist fix, and all pre-existing destructive-command blocks (`rm -rf /`, `rm -rf ~`, `rm -rf .`, `rm -rf ../important`, `git push --force`, `git reset --hard`).
|
|
3098
|
-
|
|
3099
|
-
## 2.2.4 - 2026-08-27
|
|
3100
|
-
|
|
3101
|
-
Follow-up to 2.2.3's hook-chain collapse: a real omp session on a CloudMounter/FUSE-mounted project reported every tool call — including harmless `echo`, `pwd`, `ls`, and `Edit` — failing with `invalid hook-chain output`, and in one case `protect-files.sh hook chain was killed before safety gates completed`, despite `protect-files.sh` itself passing trivially. The bridge was misattributing any aggregate-runner transport failure to the first fail-closed script in the chain, whether or not that script ever ran.
|
|
3102
|
-
|
|
3103
|
-
### Fixed
|
|
3104
|
-
|
|
3105
|
-
- **Outer transport failures no longer impersonate a specific safety gate.** `runScriptChain` in `ukit-bridge.js` treated a killed/timed-out `pi.exec`, empty stdout, or malformed JSON from `hook-chain-runner.mjs` as proof that the chain's first fail-closed script (e.g. `protect-files.sh`) had failed — without verifying that script ever ran. It now distinguishes "the runner returned a verifiable per-script result" from "the runner itself failed before producing one," and on the latter reports the real failure (killed/code/elapsedMs/runtime) instead of a fabricated gate name. Diagnostics are appended to `.ukit/storage/cache/hook-errors/<session>.jsonl`.
|
|
3106
|
-
- **Bash and Read/Grep/Glob no longer fail closed on a bridge/runner outage.** Only `Edit`/`Write` chains still block when the runner can't produce a verdict, since those guard file mutation.
|
|
3107
|
-
- **`process.execPath` is no longer trusted blindly for the runner subprocess.** The bridge resolves `process.env.UKIT_NODE_PATH || process.execPath` and records the resolved runtime in the failure diagnostic, so a Bun-hosted omp launching the Node-only runner under the wrong binary is now visible instead of silently assumed.
|
|
3108
|
-
|
|
3109
|
-
### Added
|
|
3110
|
-
|
|
3111
|
-
- **6 regression tests** in `tests/hooks/ompHookBridge.test.js` (`case 13`) covering a killed outer exec, malformed stdout, a non-zero exit with empty stdout before any script ran, fail-open behavior for Bash and Read chains under the same failures, and `UKIT_NODE_PATH` override.
|
|
3112
|
-
|
|
3113
|
-
## 2.2.3 - 2026-08-26
|
|
3114
|
-
|
|
3115
|
-
Sessions were stalling mid-task — the model would read source, then simply stop, with the work half done and nothing wrong upstream. This wave stops treating that as a prompting problem. Instructions asking a model to "continue until the edit is made" are advisory by nature: every hidden backend behind `unic-lite` / `unic-code` / `unic-smart` / `unic-vision` reads them slightly differently, and the ones that read them loosely stall. UKit now records what actually happened — real Edit/Write receipts, real verification exit codes — and enforces completion mechanically at the point of stopping, identically on Claude Code and omp. Nothing in the enforcement path branches on provider branding.
|
|
3116
|
-
|
|
3117
|
-
### Fixed
|
|
3118
|
-
|
|
3119
|
-
- **Silent stalls were unenforceable, because nothing measured them.** A stop after a read-only pass was indistinguishable from a stop after finished work:
|
|
3120
|
-
- **The omp bridge could exceed omp's own handler budget and freeze the session.** A single `Edit` fired 7 hook scripts as 7 serial `pi.exec` subprocesses at up to 8s each — worst case well past omp's ~30s outer timeout, which surfaces to the user as the session simply hanging. They now run in one bounded `hook-chain-runner.mjs` process: 4s per script, 10s total budget, under a 12s outer timeout. Per-event latency is appended to `.ukit/storage/cache/hook-latency/<session>.jsonl` so hook cost stops being invisible.
|
|
3121
|
-
- **omp writes bypassed the safety gates entirely.** omp's `normalizeToolEventInput` emits `path` / `paths`;
|
|
3122
|
-
- **An unmapped mutation-capable tool failed open and silently.** `mapToolName` returned `null` for any tool not in its table, and the bridge ran an empty script chain — so a host tool that writes files but is not yet mapped skipped every guard with no signal. Such tools are now blocked with a message naming the tool; unmapped read-only tools still pass through.
|
|
3123
|
-
- **The ledger CLI silently did nothing under a symlinked project root.** Its main-guard compared `fileURLToPath(import.meta.url)` against `path.resolve(process.argv[1])`.
|
|
3124
|
-
- **`.omp/config.yml` documented the model aliases backwards**, calling the UNIC names "LITERAL model names, not aliases".
|
|
3125
|
-
|
|
3126
|
-
### Added
|
|
3127
|
-
|
|
3128
|
-
- **`tests/hooks/executionLedgerCli.test.js`** (4 tests) — drives the real CLI through a symlinked root:
|
|
3129
|
-
- **Manifest entries and dev-mirror parity for the four new files** — `record-execution.sh`, `completion-gate.sh`, `execution-ledger.mjs`, `hook-chain-runner.mjs`.
|
|
3130
|
-
|
|
3131
|
-
## 2.2.2 - 2026-08-22
|
|
3132
|
-
|
|
3133
|
-
Running Claude Code and omp side by side exposed that they were not, in fact, running the same UKit. `CLAUDE.md` and `AGENTS.md` are hand-maintained forks of one instruction set read by different harnesses, and `AGENTS.md` had fallen three minor versions behind — so omp was never told the model-tier table existed. This wave re-unifies them and adds a test that makes the drift impossible to repeat.
|
|
3134
|
-
|
|
3135
|
-
It also settles the three `TODO(verify)` items 2.2.1 shipped with. Those were plan-time guesses about omp's API, unresolvable at the time because omp was not installed on any machine that could answer them. omp v17.4.2 is now installed and was inspected directly — and two of the three guesses were wrong, in the quiet way where nothing errors and the feature simply never engages.
|
|
3136
|
-
|
|
3137
|
-
### Fixed
|
|
3138
|
-
|
|
3139
|
-
- **Auto-compact was never actually configured on omp.** `.omp/config.yml` set `compact.autoCompactWindow:
|
|
3140
|
-
- **Post-compact hooks never ran on omp.** The bridge assumed omp re-emitted `session_start` after a compaction, the way Claude Code's matcher-less `SessionStart` hook fires with `source:
|
|
3141
|
-
- **Two UNIC gateway probes could never have fired.** `unic-gateway.mjs` probed an `OMP_BASE_URL` env var and a `baseUrl` key in `.omp/config.yml`.
|
|
3142
|
-
- **omp behaved as if UKit only had one model, because it was never told otherwise.** `templates/AGENTS.md` — the context doc omp, Codex and OpenCode read — was missing the entire `3-Tier Model Routing` block, plus `Post-Edit Sidecar Review` and `Skill Quality`, and still announced itself as `UKit v1.5.1`. `.omp/config.yml` declared `modelRoles` for all four lanes, but nothing in omp's context ever said *when* to use a tier, so everything ran inline on the session model.
|
|
3143
|
-
- **Five section pairs had been renamed apart** — `Core Rule`/`Core UKit Rule`, `Fast Classification`/`Fast Task Classification`, `Internal Helper Policy`/`Internal Helper Routing`, `Selective Subagent Policy`/`Subagent Policy`, `Project Snapshot`/`Environment Snapshot`. This is what let the drift hide: once the headings differ, no diff lines up and the gap is invisible. Headings are now shared, and the two docs carry byte-identical bodies for every shared section.
|
|
3144
|
-
- **Every session opened by being told to resume a finished cycle.** `docs/AI_HANDOFF/RUN.md` was left at `Phase:
|
|
3145
|
-
- **`/ukit:*` commands re-pasted the user's whole problem statement up to seven times per file.** Only one `$ARGUMENTS` site per command is a real interpolation (the `## Problem / feature` body);
|
|
3146
|
-
- **Eight fallback lines told the model to do something only a human can do.** `> Other tools without subagent support:
|
|
3147
|
-
|
|
3148
|
-
### Added
|
|
3149
|
-
|
|
3150
|
-
- **`### How a tier is actually bound`** in both context docs. The main session model never changes mid-turn;
|
|
3151
|
-
- **`tests/consistency/contextDocsParity.test.js`** (29 tests) — every `## ` section must exist in both `templates/CLAUDE.md` and `templates/AGENTS.md` with byte-identical bodies, unless allowlisted with a stated reason (currently one entry: `Session Start — OpenCode`, which exists because OpenCode does not auto-load `.claude/skills/`). Also asserts neither doc hardcodes a version, and that both manifest items declare the same variable set.
|
|
3152
|
-
- **`.omp/RULES.md` grew from 4 rules to 7**, targeting direct chat specifically — direct chat is steered only by static context, so what is missing there is simply never applied.
|
|
3153
|
-
- **The three `/ukit:*` sites that named only Claude Code now name omp too** — `handoff-fullstack.md`'s Model Split note, and the small-task-maintainer lines in `handoff-clear.md` and `handoff-status.md`.
|
|
3154
|
-
|
|
3155
|
-
### Changed
|
|
3156
|
-
|
|
3157
|
-
- **Four more expired one-shot guards re-anchored, same class as the two below.** `ompDocsSync` cases 3 and 6 validated "the top CHANGELOG entry" — true on the day 2.2.1 shipped, false at the next release; they now find the omp entry by version. `ompContextLayer` case 10 guarded the compact key with a `TODO(verify)` marker instead of a value, and now asserts the verified key plus its provenance. `tests/handoff/cycle4/unic-gateway.test.mjs` case 13b had never actually run (it took its skip branch because the file it checked did not exist yet) and was matching a key spelling — `@vision:` — that belongs to the agent side, not the config side.
|
|
3158
|
-
- **`tests/consistency/ompCommandParity.test.js` cases 4 and 5 no longer diff against `git show HEAD:`.** They were scope guards for one in-flight cycle ("TASK-005 must not modify these"), and that premise expired when the cycle shipped: against a committed baseline they pass vacuously, and against a dirty tree they block every later legitimate edit instead — which is exactly what they did. Both are re-expressed as the durable properties they stood in for (no `subagent_type` site in the two non-delegation commands; pinned per-file `Do NOT` and tier-justification-clause counts). No protection was dropped, only re-anchored.
|
|
3159
|
-
|
|
3160
2796
|
## 2.2.1 - 2026-08-22
|
|
3161
2797
|
|
|
3162
2798
|
UKit gains a fourth runtime and loses one. omp (Oh My Pi) reaches full parity with the Claude Code workspace — same agents, same skills, same hooks, same `/ukit:*` commands — and the Antigravity adapter, unused and unmaintained, is gone. The parity is deliberate rather than incidental: the port reuses the existing `.claude/` assets wherever omp's discovery can already reach them, so there is one source of truth per capability, not two that drift.
|
|
@@ -3188,99 +2824,4 @@ UKit gains a fourth runtime and loses one. omp (Oh My Pi) reaches full parity wi
|
|
|
3188
2824
|
|
|
3189
2825
|
**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`.
|
|
3190
2826
|
|
|
3191
|
-
|
|
3192
|
-
|
|
3193
|
-
`local-build` was the one execution contract in the daily (non-handoff) flow that let the executor claim done off write evidence alone — and it is also the busiest lane a weaker `unic-code` backend hits. This closes that gap and adds a non-blocking second opinion, without adding wait time to the main task.
|
|
3194
|
-
|
|
3195
|
-
### Changed
|
|
3196
|
-
|
|
3197
|
-
- **`local-build` now requires verification evidence, not just write evidence, before claiming done.** `orchestration.contracts['local-build'].completionRule` was `require-write`;
|
|
3198
|
-
|
|
3199
|
-
### Added
|
|
3200
|
-
|
|
3201
|
-
- **Non-blocking sidecar diff review for `local-build` and `shared-edit`.** After write + verification evidence exist for one of these tasks, routed state now carries a `review=code-reviewer(diff)` segment (`routeSummary.postEditReview`) instructing the agent to launch `code-reviewer` in the background (`run_in_background: true`, `REVIEW_TARGET_TYPE=diff`, model `unic-smart` per the new `subagents.diffReviewModel`) — reusing the reviewer-model-differs-from-executor precedent from `handoff.reviewer`, since a background pass can afford a stronger model without slowing the main task down. Findings are advisory only; the main task's completion is never blocked or reopened on their account. New `code-reviewer.md` mode: `REVIEW_TARGET_TYPE=diff`, documented in `templates/.claude/agents/code-reviewer.md` (mirrored byte-identically to `.claude/agents/code-reviewer.md`) and wired into `templates/CLAUDE.md` / this repo's own `CLAUDE.md` under "Post-Edit Sidecar Review (internal)".
|
|
3202
|
-
- **`subagents.diffReviewEnabled` / `diffReviewAgent` / `diffReviewModel`** — new config fields in both `.ukit/storage/config.json` and `templates/ukit/storage/config.json`, defaulting to `true` / `code-reviewer` / `unic-smart`.
|
|
3203
|
-
|
|
3204
|
-
## 2.1.4 - 2026-08-19
|
|
3205
|
-
|
|
3206
|
-
More room before compaction, the handoff hooks stop taxing every call they were meant to speed up, and the shipped handoff instructions stop contradicting the config they tell agents to read.
|
|
3207
|
-
|
|
3208
|
-
### Changed
|
|
3209
|
-
|
|
3210
|
-
- **`compact.hardCapTokens` 160,000 → 220,000 and `env.CLAUDE_CODE_AUTO_COMPACT_WINDOW` 150,000 → 180,000.** 2.1.3 fixed the deadlock but left the pair sized for a 200k window, which meant compacting far more often than necessary on a 256k model. The new pair leaves roughly 40k of working room after auto-compact fires, and the ordering invariant `autoCompactWindow < hardCapTokens` still holds — `tests/core/autoCompactWindow.test.js` fails if either number moves in isolation. **On a 200k model, lower `hardCapTokens` back to 160,000 or below**; the `_help` text in `config.json` says so, and the gate cannot protect you if the cap sits above your real context window.
|
|
3211
|
-
- **Config-missing fallbacks follow the default**: `compact-threshold.mjs` and `context-window-guard.sh` now fall back to `220_000` rather than `160_000`, so a corrupt or absent `config.json` behaves like a fresh install instead of silently reverting to the old cap.
|
|
3212
|
-
|
|
3213
|
-
### Fixed
|
|
3214
|
-
|
|
3215
|
-
- **The shipped handoff instructions no longer restate a config value that had gone stale.** Eight lines across `handoff-fullstack.md`, `handoff-implement.md`, `handoff-review.md` and `handoff-planner.md` told agents `handoff.maxParallelAgents` defaults to **10**, while `config.json` in the same release ships **12** — so an installed project got instructions that contradicted its own runtime config. Every one of those lines already said "read from `.ukit/storage/config.json`", so the restated number was redundant as well as wrong; it is now gone and the value has exactly one home. This was the second occurrence (the first was the `3 → 10` bump), and the cause both times was hand-copying a config value into prose. `tests/consistency/configDocsSync.test.js` now fails the suite if any doc restates a numeric `handoff.*` default that contradicts config, in English or Vietnamese, and if a `templates/.claude` file without `{{ }}` variables drifts from its active copy — the "keep the two mirrors identical" rule was written in two places but had never been enforced by anything except a maintainer running `diff` by hand.
|
|
3216
|
-
- **The worktree early-exit guard no longer costs more than it saves.** `skill-router.sh` and `stale-spec-guard.sh` skip routing for tool calls scoped to a disposable `.worktrees/task-*` tree — but the first cut spawned `node -e` on every call to decide that, measured at 53ms → 124ms per call in the main tree, where the skip never applies. A pure-bash `case "$INPUT" in *".worktrees/"*)` pre-check now gates the node spawn; it is a strict superset of what the node guard can match, so the main tree pays nothing. Both hooks still fail **open** — any parse error, missing field or node failure falls through to the normal path.
|
|
3217
|
-
|
|
3218
|
-
### Added
|
|
3219
|
-
|
|
3220
|
-
- **`.claude/ukit/index/provision-worktree.mjs`** — deterministic worktree provisioning for handoff waves, packaged via `manifests/platform.full.yaml`.
|
|
3221
|
-
- **Test-selection policy in `docs/AI_HANDOFF/RULES.md`** — per-task related tests are the floor, and a full suite at the wave/cycle boundary is mandatory, not optional.
|
|
3222
|
-
- **`scripts/bench/parallel-agents.mjs`** — harness for measuring concurrent test-process contention at several parallelism levels, recording `slowdownFactor` / `perRunCost` / a `recommended` value plus the load averages the run happened under. It measures test-process contention, not LLM agents.
|
|
3223
|
-
- **`handoff.maxParallelAgents`: 10 → 12.** Chosen, not measured — the benchmark above was never run on a quiet enough machine, so this is the middle of the ~10-15 safe band the config's own `_help` has always documented. Lower it to 3-5 if your tasks produce long verification output or you see compaction firing repeatedly.
|
|
3224
|
-
|
|
3225
|
-
## 2.1.3 - 2026-08-16
|
|
3226
|
-
|
|
3227
|
-
The reason auto-compact never fired in the terminal was UKit itself. This makes it fire.
|
|
3228
|
-
|
|
3229
|
-
### Fixed
|
|
3230
|
-
|
|
3231
|
-
- **Auto-compact now triggers below the hard cap instead of never.** `context-hardcap-gate.sh` refuses Edit/Write/Bash at `compact.hardCapTokens` (160,000 — 80% of a 200k window), while Claude Code's own auto-compact fires near the real window limit. The gate froze the session first, so the transcript stopped growing, so the auto-compact threshold was never reached: a deadlock whose only exit was typing `/compact` by hand. Default settings now ship `env.CLAUDE_CODE_AUTO_COMPACT_WINDOW = 150000`, which lands 10,000 tokens *before* the gate — compaction happens on its own, the run keeps going, and the grace window added in 2.1.2 stays as the backstop rather than the normal path.
|
|
3232
|
-
- **`autoCompactEnabled: true` is pinned in default settings**, so a stale toggle in `/config` cannot silently reinstate the deadlock.
|
|
3233
|
-
|
|
3234
|
-
### Added
|
|
3235
|
-
|
|
3236
|
-
- **`tests/core/autoCompactWindow.test.js`** — locks the ordering invariant `env.CLAUDE_CODE_AUTO_COMPACT_WINDOW < compact.hardCapTokens`, the 100,000–1,000,000 range Claude Code accepts, `autoCompactEnabled`, template/mirror parity, and the `overwrite_with_backup` merge strategy that lets existing installs pick the fix up. Change either number in isolation and the suite fails instead of the deadlock returning silently.
|
|
3237
|
-
|
|
3238
|
-
### Notes
|
|
3239
|
-
|
|
3240
|
-
- `env` in `settings.json` applies to **new sessions**. After `npm i -g @ngockhoale/ukit@latest && ukit install`, restart the session (or `/compact` once) before the change takes effect.
|
|
3241
|
-
- Still true, and unchanged: no hook can invoke `/compact` — it is client-only. 2.1.3 does not call it; it moves the client's own threshold below UKit's gate so the client compacts on its own.
|
|
3242
|
-
|
|
3243
|
-
## 2.1.2 - 2026-08-16
|
|
3244
|
-
|
|
3245
|
-
Follow-up to 2.1.0, which fixed the wrong hook. Runs were still freezing mid-task because the gate that actually blocks tools was never touched.
|
|
3246
|
-
|
|
3247
|
-
### Fixed
|
|
3248
|
-
|
|
3249
|
-
- **`context-hardcap-gate.sh` gets a grace window for unfinished runs.** 2.1.0 made the *advisory* `context-window-guard.sh` (UserPromptSubmit) run-cursor aware, but the halt users hit came from `context-hardcap-gate.sh` — a separate PreToolUse hook that hard-refuses Edit/Write/Bash with `exit 2` once `estimatedTotalTokens >= compact.hardCapTokens`. It knew nothing about `docs/AI_HANDOFF/RUN.md`, so it froze runs mid-edit exactly as before 2.1.0. It now reads the run cursor and, while a run is unfinished, allows a bounded `compact.hardCapGraceCalls` (default 10) further calls with an explicit "spend this on landing, not new work" directive — finish the edit, commit, write the cursor, push — then blocks hard as it always did.
|
|
3250
|
-
|
|
3251
|
-
### Added
|
|
3252
|
-
|
|
3253
|
-
- **`compact.hardCapGraceCalls`** (default `10`). Budget is per over-cap episode, not per wave:
|
|
3254
|
-
- **`tests/hooks/contextHardcapGate.test.js`** — 12 cases covering under-cap pass-through, blocking with no run in flight, `Phase:
|
|
3255
|
-
|
|
3256
|
-
### Notes
|
|
3257
|
-
|
|
3258
|
-
- **No hook can call `/compact`** — it is a client-only command, so UKit cannot auto-compact on your behalf.
|
|
3259
|
-
|
|
3260
|
-
## 2.1.0 - 2026-08-16
|
|
3261
|
-
|
|
3262
|
-
Handoff autonomy wave: `/ukit:handoff-fullstack` runs a whole goal end-to-end without stopping to ask. Questions are collected once, up front, in the planning phase; everything that used to be a mid-run STOP now resolves automatically and keeps going.
|
|
3263
|
-
|
|
3264
|
-
### Added
|
|
3265
|
-
|
|
3266
|
-
- **Autonomy Contract in all four handoff commands.** `handoff-create` owns the single question window — one batched `AskUserQuestion` (max 4, recommended option first) before planning starts, covering scope, conflicting interpretations, subsystem choice and acceptance criteria only. `handoff-fullstack`, `handoff-implement` and `handoff-review` ask nothing between that window and the Final Report; each command carries a table mapping every former STOP to its required automatic behavior.
|
|
3267
|
-
- **Run cursor + resume protocol (`docs/AI_HANDOFF/RUN.md`).** Every phase transition writes `Phase`/`Cursor`/`Next`.
|
|
3268
|
-
- **`handoff-resume.sh` SessionStart hook** (manifest id `hook-handoff-resume`). Reads the run cursor and prints a resume banner with Goal/Base/Phase/Cursor/Next, plus two extra lines when the session started from `source: compact`. Advisory only, always exits 0, silent when there is no unfinished run.
|
|
3269
|
-
- **Auto-fix loop after review (max 2 rounds).** Non-approved tasks are re-waved and re-implemented in parallel — agents read their own `## Reviewer Verdict` from disk — then committed and re-reviewed.
|
|
3270
|
-
- **Planner Grounding + 12-point Self-Audit.** `handoff-planner` must verify that every Target File exists (or is marked `(new)`), that every Verification Command is a real `package.json` script, and that test file paths match the neighbouring test layout; then run a 12-point self-audit (coverage / correctness / test quality, including "could every listed test pass against an empty implementation?") and report the result. Compensates for dropping plan review from 3 rounds to 2.
|
|
3271
|
-
- **Wave-boundary commit + context checkpoint.** Each wave ends with `git add -A && git commit -m "handoff:
|
|
3272
|
-
|
|
3273
|
-
### Changed
|
|
3274
|
-
|
|
3275
|
-
- **`git push` moved from the `ask` list to `allow`** in default settings, alongside `git add`, `git commit` and `git worktree`.
|
|
3276
|
-
- **Subagent return contracts are now capped.** Executors return ≤10 lines (`TASK`/`STATUS`/`EXECUTOR_MODEL`/`FILES`/`RED`/`VERIFY`/`NOTE`), reviewers ≤6.
|
|
3277
|
-
- **`context-window-guard.sh` is now run-cursor aware.** With an unfinished run it emits shed-context-and-continue instructions instead of the stop-and-`/compact` directive.
|
|
3278
|
-
- **Same-wave file conflicts auto-serialize** instead of marking both tasks `needs_breakdown`:
|
|
3279
|
-
- **A dirty working tree no longer blocks Phase 3.** It is committed as `handoff: checkpoint before implement` and the run continues.
|
|
3280
|
-
- **Plan review loop cap: 3 rounds → 2**, then findings are applied and the run proceeds (logged as `### Round <N> — findings applied without re-review`). Config: `handoff.autonomy.planReviewRounds`.
|
|
3281
|
-
- **Review diff range is now anchored to the plan commit** (`git log --grep='^handoff: plan' -n 1`) rather than the working tree, since waves are committed as they finish.
|
|
3282
|
-
- **`handoff-planner` maximizes wave width.** `Dependencies: none` unless task B genuinely cannot be written without A's output — ordering preference, same code area and review order are explicitly not dependencies.
|
|
3283
|
-
- **`feature-implementer` and `code-reviewer` run unattended in handoff mode.** Ambiguity resolves through task file → PLAN.md → surrounding code → recommended choice, recorded in `## Discussion`;
|
|
3284
|
-
- **`handoff-model-guard.sh` gained a bash fast path** that skips the node spawn when the payload contains neither `AI_HANDOFF` nor `git push`.
|
|
3285
|
-
|
|
3286
|
-
Entries for 1.0.0 through 2.0.7 are archived in the repository at `docs/archive/CHANGELOG-1.0.0-2.0.7.md` (not shipped in the npm package).
|
|
2827
|
+
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).
|
|
@@ -1257,6 +1257,18 @@ entries:
|
|
|
1257
1257
|
validation: [manual]
|
|
1258
1258
|
archive_policy: never
|
|
1259
1259
|
notes: Team upgrade notice for 3.4.28 (local RC); maintainer-only, not shipped to user projects
|
|
1260
|
+
|
|
1261
|
+
- id: docs-release-3-4-79-queue-lifecycle-migration
|
|
1262
|
+
path: docs/release/3.4.79-QUEUE-LIFECYCLE-MIGRATION.md
|
|
1263
|
+
class: canonical
|
|
1264
|
+
audience: [maintainer]
|
|
1265
|
+
owner: product
|
|
1266
|
+
source_of_truth: docs/release/3.4.79-QUEUE-LIFECYCLE-MIGRATION.md
|
|
1267
|
+
merge_strategy: none
|
|
1268
|
+
load_policy: on-demand
|
|
1269
|
+
validation: [manual]
|
|
1270
|
+
archive_policy: never
|
|
1271
|
+
notes: Consumer migration notes for the 3.4.79 handoff queue lifecycle; maintainer-only, not shipped to user projects
|
|
1260
1272
|
# Declared complexity→docs mapping consumed by the router's `docs=` summary segment
|
|
1261
1273
|
# (src/index/taskRouting.js CONTEXT_LAYER_DOCS mirrors this block; DOC-201 FR-001/002).
|
|
1262
1274
|
# queued-task intent omitted v1: docs/TASKS.md is optional per-repo (SPEC §14).
|
package/package.json
CHANGED
|
@@ -173,6 +173,13 @@ the document into a LEDGER, a bounded task set and a queue, and never implements
|
|
|
173
173
|
first), `QueueDrain: separate-backlog` and `QueueSource: <R-IDs/epics>`; `/ukit:handoff-fullstack`
|
|
174
174
|
reports such a blueprint as `Backlog: <slug>` and never folds it into the current cycle; it drains
|
|
175
175
|
only when the human names the slug (one blueprint per invocation).
|
|
176
|
+
- **Queue root:** always write blueprints under `docs/AI_HANDOFF/queued/` (canonical). If a legacy
|
|
177
|
+
`docs/AI_HANDOFF/queue/` directory also exists, report it as a diagnostic in the planner report;
|
|
178
|
+
never move, merge or migrate it automatically.
|
|
179
|
+
- **Retiring a blueprint:** a blueprint that is already delivered or replaced is marked, not deleted:
|
|
180
|
+
`QueueIntent: superseded|stale` + `QueueSupersededBy: archive:cycle-NN[, …] | queue:<slug>`. The
|
|
181
|
+
runtime verifies that provenance and fails closed (the blueprint keeps counting) when it cannot;
|
|
182
|
+
see `docs/AI_HANDOFF/queued/README.md`.
|
|
176
183
|
|
|
177
184
|
Model split on Claude Code: haiku/`unic-lite` gathers, opus/`unic-smart` decides, and
|
|
178
185
|
sonnet/`unic-code` implements later; omp maps these to `@smol`/`@slow`/`@default`.
|
|
@@ -358,6 +358,11 @@ work, not only the current plan:
|
|
|
358
358
|
whose LEDGER rows are `deferred` carry that header. Mandatory overflow (`feasible`/`adapted`
|
|
359
359
|
rows queued by create because the plan exceeded 10 tasks) is queued WITHOUT `QueueIntent`,
|
|
360
360
|
so it is folded in here like any legacy blueprint and is never silently excluded.
|
|
361
|
+
A blueprint marked `QueueIntent: superseded|stale` with a verified `QueueSupersededBy:` provenance
|
|
362
|
+
(`archive:cycle-NN` covering every task, or `queue:<slug>`) is excluded as `Superseded: <slug>`;
|
|
363
|
+
unprovable metadata fails closed and the blueprint still counts — reconcile it (or set
|
|
364
|
+
`Phase: blocked` for an owner decision), never archive or delete it to silence the gate. Inspect
|
|
365
|
+
read-only with `node .claude/ukit/runtime/resumable-run.mjs queue-status <projectRoot>`.
|
|
361
366
|
Backlog items do not count toward `index:all-done` / `ExitPredicate`, so they never block completion. Only
|
|
362
367
|
an explicit slug in the invocation drains one (one blueprint per invocation). The
|
|
363
368
|
create phase of this command applies the same advisor intake as `/ukit:handoff-create`
|