Releases: caty-ai/context-kit
Release list
v0.4.0: brief-validator accepts Japanese or English section tokens by default (#8)
v0.4.0: brief-validator accepts Japanese or English section tokens by default (#8)
- Each brief layer is satisfied by either the handbook heading or its English
layer name: ## 実装仕様 / ## Goal, ## 実装チェック / ## Self-verification,レビュー基準 / ## Reviewer criteria. Mixing is allowed; matching stays plain
case-sensitive substring. - Block message names both accepted tokens per missing layer and lists the
English tokens; the skeleton keeps the canonical Japanese headings. - CK_BRIEF_REQUIRED_SECTIONS keeps its single-token contract (no aliases).
- Also since v0.3.3: make lint is a real shellcheck gate with no SC2155
carve-out (#38, #59); the test-lint caller requires suite reconciliation (#38).
PR #62 (merge ff51119). Seats GLM 5.3 / Opus 5 / Kimi K3, cumulative GO x3.
repo-state caller v0.12.1: app-authenticated stamp pushes (family-os#140)
repo-state caller v0.12.1: app-authenticated stamp pushes (family-os#140)
v0.3.2
repo-state caller + generation stamp (family-os#127 rollout; closes #48)
repo-state caller + generation stamp (family-os#127 rollout; closes #48)
v0.3.0 — npm one-shot install: @caty-ai/context-kit installer + provenance publish path (#41, PR #44)
v0.2.1 — wt-snapshot works on stock bash 3.2 and prune expires refs on Linux
shojikumaru
Sho Jikumaru
v0.2.1 — wt-snapshot works on stock bash 3.2 and prune expires refs on Linux
Fixes shipped in this release (user-facing behavior):
- bash-3.2-safe empty-array expansions: wt-snapshot no longer dies with
'unbound variable' on macOS system bash (GitHub macos-latest runners,
stock Macs without Homebrew bash). Issue #30, merge 30b2442. - timestamp_to_epoch GNU fallback now parses the tool's own compact
timestamp format: prune actually expires old snapshot refs on Linux
(previously it silently expired nothing). Issue #30, merge 30b2442. - Test-suite honesty fixes (#24/#33/#31) and family CI gates B5/B6/B7
(#17/#20/#21, merges a234ff7/067af02/d08a997).
Evidence: both-OS green CI on the release state —
ubuntu: https://github.com/caty-ai/context-kit/actions/runs/32357716364/job/96390424263
macOS: https://github.com/caty-ai/context-kit/actions/runs/32357716364/job/96390424505
Fulfils the 'release: deferred' recorded on issue #30 (trigger: #20 close).
v0.2.0 — wt-snapshot: capture a worktree before anyone deletes it
shojikumaru
Sho Jikumaru
v0.2.0 — wt-snapshot: capture a worktree before anyone deletes it
The weekly cleanup report kept finding worktree remnants because "may I delete
this?" had no safe answer. bin/wt-snapshot gives it one: it captures a dirty or
not-reconstructible worktree into a commit object under
refs/worktree-snapshots// in the main repository, so
deletion becomes reversible. The tool never deletes anything itself.
- captures tracked dirty changes, untracked files, and the configured ignored
path set (e.g. .omc/) through a temporary index; the user's HEAD, index and
config are never touched and identity is passed explicitly per git call - a clean worktree at a reachable HEAD is an explicit no-op, so noise is never
snapshotted - CK_WTSNAP_SECRET_SCAN_CMD, when set, sees every captured byte before anything
is written; a hit aborts with the scanner's status and leaves refs and the
object store untouched - restore rebuilds the worktree including the ignored set; prune lists expired
refs and deletes nothing
Evidence for the two properties that matter most, measured on this head:
extended-attribute content cannot smuggle bytes past the scanner (payload holds
0 LIBARCHIVE.xattr records and 0 plain hits for a probe secret placed in an
xattr), and a staged submodule pointer is captured rather than reported as a
no-op (restored index gitlink oid matches the source exactly). A plain-file
secret aborts with exit 17 leaving loose-object count unchanged at 3 and no ref
written. Payload verification is pinned from both directions: an injected
unlisted member aborts with exit 70, and neutralising the member-type
comparison turns exactly one case red.
Suite on this head: 29 passed in tests/test_wt_snapshot.sh, all seven suites
green. Review: 4 rounds, 3 seats each round (Grok 4.6, Kimi K3, Muse Spark 1.2)
— two MAJORs found and closed in rounds 3 and 4, final round GO / GO-WITH-MINOR
/ GO.
v0.1.0 — 公開初版
公開初版。現行 main を v0.1.0 として打刻（2026-08-08・遡及）。単体 AI エージェントのためのコンテキスト衛生キット — プルーニング・スクラッチ退避・委譲ブリーフ検査・セーフティフック・記憶リコール。
EN: First public release, stamping the current main as v0.1.0 (retroactive, 2026-08-08). Context hygiene kit for a single AI agent — pruning, scratch persistence, delegation-brief validation, safety hooks, memory recall.