ship
Release-gate sequence: validate → review → push → tag. Wraps kernel:git mechanics with the pre-ship safety chain. Triggers: ship, release, push to main, ready to merge, deploy.
- License 许可
- Apache-2.0
- 最近更新
- 2026-11-14
- Stars
- ★ 1.6k
- Forks
- 491
01 前言 Front Matter
- name
- ship
- kernel
- { "kind": "operator", "version": 1, "confirmation": "on_side_effect", "side_effects": "writes_remote" }
- description
- Release-gate sequence: validate → review → push → tag. Wraps kernel:git mechanics with the pre-ship safety chain. Triggers: ship, release, push to main, ready to merge, deploy.
- allowed-tools
- Read, Bash, Task, Edit
02 资源文件 1
- 参考文档reference/ship-research.md2.7 KB
03 SKILL.md 正文
<skill id="ship">
Sequence
-
Preflight
git status --porcelain— clean? If not: ask user to commit, stash, or abandon.git branch --show-current— matches intent? Wrong branch → stop.git log --oneline {main}..HEAD— commit range matches what you think you're shipping?- (gate: any mismatch → AskUserQuestion before continuing)
-
Validate
- Invoke
/kernel:validate(spawns validator agent, full 9-gate safety chain). - If unavailable: run project's nearest configured command (see reference/ship-research.md for equivalents).
- (gate: any FAIL → stop; report which gate; do NOT push)
- For plugins that support multiple loaders, validate each loader's schema independently and exercise one installed payload per loader. A shared file parsing successfully is not proof that its armed behavior is equivalent.
- Run install, supported upgrade, and documented recovery commands in a disposable plugin/cache copy from outside the source checkout. Assert user data and user-owned files are unchanged.
- Put an explicit resource ceiling around heavyweight suites when available; a release gate must fail loudly instead of exhausting the host machine.
- Invoke
-
Review
- Tier 1 (1–2 file changes, low risk): self-review via Big 5 from skills/quality/SKILL.md.
- Tier 2+: invoke
/kernel:review(spawns reviewer agent, >80% confidence threshold). - (gate: REQUEST CHANGES → stop; address feedback; restart from step 2)
- (gate: COMMENT → AskUserQuestion: "Address now, ship anyway, or hold?")
- (gate: APPROVE → continue)
-
Push
- Feature branch (
feat/*,fix/*,chore/*, etc.):git push(orgit push -u origin {branch}if upstream not set). main/master: STOP. AskUserQuestion required. (NEXUS I0.8)- Detached HEAD or unexpected state: stop; investigate.
- (gate: push rejected / non-fast-forward → surface to user; do NOT force-push)
- Feature branch (
-
Version + Tag (on a release)
- Semver: patch=fix, minor=feature or behavior-preserving refactor, major=breaking. Confirm the number with the user.
- Bump ALL canonical declarations in one shot:
scripts/bump-version.sh X.Y.Zupdates.claude-plugin/plugin.json,.claude-plugin/marketplace.json,governance/kernel.md.tmpl, regeneratesAGENTS.mdandCLAUDE.md, and updatesskills/help/SKILL.md. NEVER hand-edit one location — drift failstest_version_sync_all. - Human-author the release prose the script does NOT touch: the plugin/marketplace
descriptionhighlight + aCHANGELOG.mdentry (## [X.Y.Z] - DATE+ Added/Changed/Fixed). - (gate:
bash tests/run-tests.shgreen —test_version_sync_allconfirms no stale version anywhere.) - If a native manifest validator rejects required safety metadata, do not weaken the safety metadata or hand-author an unvalidated manifest. Keep the proven compatibility loader, record the limitation, and defer the native manifest until both schemas can be satisfied.
- Tag (only if user requested a tagged release):
git tag -lto avoid clobber →git tag -a v{X.Y.Z} -m "{summary}"→git push origin v{X.Y.Z}. - Before upgrading the installed plugin on this machine:
pgrep -fl codex. The upgrade deletes the old cache directory under a live session's feet and every hook in that session exits 127 until it restarts. Name the live sessions in the handoff and say they need a restart; never report "upgraded" as if it covered them (docs/upgrading.md, 2026-08-27). - After publishing the GitHub release, install it into Claude's own cache and verify a
fresh headless Claude process:
scripts/install-release.sh X.Y.Z. The script updates only Claude's marketplace/cache, atomically advances Claude'scurrentselector, and fails unless both the installed manifest and the fresh process reportX.Y.Z. Codex installation remains a separate operation.
-
Human pass (any user-facing release: app build, deploy, anything a person will touch)
- Load
skills/human-pass/SKILL.mdand write the guide for THIS build: literal controls, paste-ready inputs, expected outcome per step, worst case first, bundled to one sitting with the time cost stated. - Hand it over and wait. The build is not shipped, it is awaiting verdict.
- (gate: no guide, or no recorded verdict → the release is NOT done. Say "awaiting your pass on <build>", never "shipped". A green pipeline is not a person having used the thing.)
- Findings come back as prose; YOU convert them into issues, not the human.
- Record the verdict as a state-change receipt on the release issue, naming what it unlocks (next build, submission, deploy promotion).
- Load
-
Checkpoint
agentdb learn pattern "ship: {branch} {commit_range} {sha_pushed}" "validate=pass review=pass push=ok"- Profile-gated: github-oss/github-production → post PR or release note via gh CLI.
The verdict is part of done
For anything a person will use, the completion states are ordered: validated, merged, deployed, accepted. CI proves the first three. Only a human who used the build proves the fourth, and only that one predicts whether the release was any good: a 15-minute guided pass found five defects that 500+ green tests had not, because some properties only exist on the far side of a real screen with real data. Report the state you actually reached, by name.
Ask-user gates (mandatory pause points)
| Point | Condition | Question |
|---|---|---|
| Preflight | Branch or commit range is ambiguous | "Shipping {N} commits on {branch}. Confirm?" |
| Review | COMMENT verdict | "Review returned {N} comments. Address now, ship anyway, or hold?" |
| Push | Target is main/master | "About to push to main. NEXUS I0.8 requires explicit say-so. Confirm?" |
Failure modes
validate FAIL→ block; report which gate; do not loop silentlyreview REQUEST_CHANGES→ block; address comments; restart from step 2push rejected (non-fast-forward)→ ask user; never auto-rebase or force-pushtag conflict→ checkgit tag -lfirst; suggest next version; never overwrite existing tag
Anti-patterns
ship_without_validate— every step depends on the previous; skipping validate skips type/test/securitysilent_skip_review— review FAIL with no surfacing to user = trust violationauto_push_to_main— NEXUS I0.8; main requires explicit user confirmationforce_push_on_rejection— rejection means remote has work you don't; investigate before overwritingtag_without_release_intent— tags are durable; casual pushes get no tagskip_checkpoint— shipped-but-unrecorded work breaks retrospective and learning loops
<on_complete> agentdb write-end '{"skill":"ship","branch":"X","commits":N,"validate":"pass","review":"approve|comment|skip","pushed":true,"tagged":"X.Y.Z|none"}' </on_complete>
</skill>
04 同仓库更多 Skills
worktable-review-with-annotations
Create, inspect, reply to, update, or resolve situated Worktable annotations. Use when feedback belongs to a Doc section, block, text range, or HTML Doc, or when the user asks to address existing review feedback and coordinate a requested artifact edit.
worktable-find-and-synthesize-context
Find, assess, and synthesize context from Worktable Docs and Records. Use when the user asks what a workspace contains, wants workspace-backed research, needs current facts reconciled across artifacts, or requests a cited synthesis without changing workspace state.
worktable-create-or-update-docs
Create, edit, rename, or maintain narrative Worktable Docs. Use for plans, decisions, briefs, notes, research syntheses, and other durable prose when the user requests a new artifact or a safe update to an existing one.
worktable-create-or-manage-records
Organize repeated structured information as durable Worktable Records. Use for research datasets, inventories, directories, pipelines, trackers, and other collections whose items need identity, fields, filtering, links, validation, or individual updates.
worktable-create-or-manage-interactive-html
Create or improve durable visual and interactive artifacts in Worktable. Use for dashboards, trackers, calculators, mini apps, slide presentations, and other requests that benefit from a polished HTML Doc.
worktable-collaborate-threads
Read, answer, and hand off messages through durable Worktable threads with users or other authenticated participants. Use when the user asks to check addressed messages, reply to a participant, continue a thread, request another agent's help, wait briefly for a response, or watch for comments and messages sent to the agent.