implement-spec
Build a whole spec's tickets in parallel subagents, onto one integration branch.
- License 许可
- Apache-2.0
- 最近更新
- 2026-11-13
- Stars
- ★ 0
- Forks
- 1
01 前言 Front Matter
- name
- implement-spec
- description
- Build a whole spec's tickets in parallel subagents, onto one integration branch.
- allowed-tools
- Read, Grep, Glob, Bash(gh issue *), Bash(gh pr create*), Bash(gh pr ready*), Bash(gh pr view*), Bash(git switch*), Bash(git branch*), Bash(git merge*), Bash(git push*), Bash(git worktree*), Bash(git status*), Bash(git log*), Bash(<shared-scripts-dir>/sub-issue-frontier.sh*)
- argument-hint
- <spec issue>
02 资源文件 1
- 参考文档agents/openai.yaml174 B
03 SKILL.md 正文
Implement Spec
Build every ticket of a spec, as filed by the to-tickets skill, onto one integration branch. The tickets aren't a list but a task graph: each names the tickets that block it, so there is always a frontier of tickets ready to start. Run implementers across the frontier in parallel; as each lands, the frontier moves.
This needs subagents that can work in their own git worktrees (Claude Code's Agent tool with worktree isolation, or Codex's spawned agents). Without them, work the tickets one at a time with the implement skill instead.
Keep traffic to and from subagents small. Point at the spec, the ticket, notes, and commits rather than pasting their contents.
1. Read the graph
Read the spec and list its tickets with
<shared-scripts-dir>/sub-issue-frontier.sh <spec>. On Claude Code, call the
Skill tool with "budget" before a large fan-out. Done when you know the
frontier and how many implementers to run at once.
2. Explore once (optional)
When several tickets need the same reading (code areas, external docs), have one exploration subagent write its notes to a directory outside the repo that every later subagent can read. Done when the notes' path is ready to hand to the implementers.
3. Create the integration branch
Branch from the base and push it. Done when the branch exists on the remote.
4. Run implementers across the frontier
For each frontier ticket, claim it (gh issue edit <n> --add-assignee @me),
then re-read its assignees: assigning never fails when someone else is already
on it, so if anyone else is, leave it and take the next. Start an implementer
subagent for each claimed ticket in its own worktree and branch, in the
background where the host allows. Each implementer:
- checks that its worktree is based on the integration branch, and resets onto it if not;
- calls the Skill tool with "tdd" to build the ticket, and calls it again with "minimal-change" before adding a dependency or an abstraction;
- runs the project's verification;
- merges the integration branch's latest tip into its own branch before reporting done.
Done when every frontier ticket has an implementer running or reported done (or blocked, with the reason).
5. Merge as they land
When an implementer reports done, merge its branch into the integration branch with a merger subagent, resolving conflicts there, and run verification on the result. After the first merge, if the repo closes work through pull requests (or the user wants one), open a draft PR from the integration branch that closes the spec and its tickets, with a body in the pr-body skill's shape. Otherwise call the Skill tool with "issue-close" for each ticket as it lands. If a merge unblocks tickets, start implementers for the new frontier. Done when every ticket is merged and either linked from the draft PR or closed.
6. Review and finish
With a draft PR, mark it ready, then call the Skill tool with "review-pr",
asking for the agent review (its standards and spec passes against the whole
spec), and fix their findings
in a single implementer subagent. Without one, report the integration branch.
Done when the review gate is met (or the branch is reported), CI is green, and
every implementer worktree is removed (git worktree remove).
04 同仓库更多 Skills
teach
Teach yourself a topic over several sessions, with lessons, references, and a record of what you've learned.
svg-logo-designer
Generate SVG logos with multiple concepts, layouts, and color variations.
documentation-writer
Writes and revises human-facing documentation — READMEs, contributor and setup guides, reference pages, architecture notes, CLAUDE.md, and changelog entries. Covers picking the right document type for the question being asked, structuring for skimming, writing examples that actually run, and keeping docs from going stale. Use when creating or rewriting docs, when a README has drifted from the code, or when deciding where a piece of information belongs. For SKILL.md and agent files, use /skill-writer instead.
adhd-mode
Shapes every response for a reader with ADHD: the next action first, numbered steps, progress restated each turn, tangents held back, concrete time estimates, wins made visible. Stays on for the rest of the session until the reader says "stop adhd mode". On Claude Code, the harness-kit:adhd output style does the same without invoking this skill.
wait-what
That last message didn't land: re-explain it with the missing context.
to-questionnaire
Turn a decision you can't make alone into a questionnaire for the one person who can answer it, filled in async or together in a meeting.