implement-piece

/implement-piece <repo> <n> implements the Piece described by the spec or sub-issue <repo>#<n>.

Fetch it with npm run issue:read -- <repo> <n> --body --comments and state its title before starting. If the reference is ambiguous, ask. If the issue is Research, read .claude/skills/research-piece/SKILL.md and follow it on the same issue instead; it is user-invoked, so the Skill tool can't reach it.

Work in the Piece's issue workspace. If this session did not start at its root, run npm run issue:start -- <repo> <n> from the primary root and tell the developer to start /implement-piece <repo> <n> at the path it prints; stop there.

Draft by the rules under "Implementing a piece spec" in docs/pieces/rules.md: its text, its supporting material, when it may write without asking, its style and the Outline. On a sub-issue, write only the Sections it names. Save each passage you read once, as a text file in the passages folder, .runner/passages/<repo>-<n> under the workspace root, named for its source and range; the folder is gitignored.

Save the Agent draft, every time, without offering, before review. Its path, its shape and how a save replaces its blocks are in docs/pieces/agent-draft.md.

1. get_doc_as_markdown the Deliverable as it is now.
2. Save the Sections this issue drafted into the Agent draft, each marked as the agent's. A later revision round saves the Sections it changed the same way.

Then read .claude/skills/review-piece/SKILL.md and follow it on the same issue; it reads the Agent draft and the passages folder, never the Doc. It is user-invoked, so the Skill tool can't reach it.

Then:

• Fix every finding suggested as Fix in the Deliverable, then save the Agent draft again as above. After fixing copied wording, rerun review-piece's Overlap command on the Agent draft as it is now, and judge any new hit the same way.
• Report every other finding to the developer, unfixed, with its suggested verdict and reason. The developer's own edits are what /to-writing-style learns from.

Done when no finding suggested as Fix is left, or each one left is reported with why.

Last, commit only the Agent draft, in <repo>'s clone in the workspace on its issue-<n>-<slug> branch: git -C <clone> add writing/agent-drafts/<spec n>.md, then git -C <clone> commit -m "Agent draft of <repo>#<n> before review.", "review" being the developer's.

End with the landing, which waits for the developer's review. Ask which model and effort pair ran this and each implementing session's peak /context. Then write the wrap-up facts per "Landing a spec" in docs/agents/git-workflow.md: the Piece's branch as the only branch, the Deliverable's link in the closing comment, and their answers as pair and peakContext. Tell them to run npm run wrap-up -- <repo> <n> once they've finished editing; it lands the Agent draft, records the outcome row and closes the issue.