Implementing Chopin plans

Chopin's MCP initialize instructions and current tool descriptions are authoritative. This skill adds local work practices; it does not restate the lifecycle.

Claim the canonical graph

1. Pass the supplied canonical document URL directly as the id input to read_implementation; the read tool accepts either that URL or a document UUID.
2. Read the returned document UUID, graph, acceptance criteria, dependencies, repository context, revisions, lifecycle state, MCP initialize instructions, and current tool descriptions. Copied plans and remembered command sequences are not substitutes.
3. Compare the graph's repository, base branch, and base commit with the local checkout. Inspect repository identity, branch, remotes, commit, and working tree before claiming or editing.
4. If the checkout does not match, surface the mismatch and stop.
5. Call start_implementation with the returned document UUID, revisions, and current checkout context. Begin only when the claim succeeds, and retain the returned run identity for later calls.

Execute ready work

Re-read the implementation before every lifecycle action, then follow the service instructions for the current state. Reads may use the canonical URL or UUID, but start_implementation and every later lifecycle call must use the UUID returned as document.id.

• Work only on tasks whose dependencies are complete. Independent ready roots may be delegated when the runtime supports it; the owning agent remains responsible for MCP reporting, review, verification evidence, and one pull request per task.
• If read_implementation returns unapprovedDecisions, those decisions are still open. Do not wait for one, block a task on one, or answer or close one: use your best judgment, and put the returned pullRequestSection, unchanged, at the end of every pull request's body, adding a few words after each item on what you assumed.
• Make only the required changes. Do not edit Chopin plan or graph content or start another top-level agent CLI session.
• Stop code changes whenever the service requires a blocker or graph release. Explain what must change and wait for the next valid lifecycle state.
• Perform a separate review pass after implementation. Prefer a fresh reviewer sub-agent; otherwise review independently after stepping away. Resolve in-scope findings and run focused checks.
• Create exactly one pull request per task. Update that pull request if later verification returns the task to work. A task that needs no code change gets no pull request: complete it with noChange, giving the reason and the evidence you checked, rather than opening an empty one.

When the document needs revision

update_document is for document authoring, not an implementation lifecycle shortcut. An active implementation locks the document and returns document-locked for a new update. If scope, acceptance criteria, or dependencies must change, stop code work and follow the current request_revision contract to release the graph. A task blocker alone does not release that lock.

After release, hand the required document changes back to the authoring workflow. It should read the latest source and revision, preserve server-owned projections, and use a new revision-guarded update rather than recreate the document. Do not silently switch from implementing to rewriting or approving the plan. Resume implementation only after the service exposes an approved graph and a new claim succeeds against its current revisions and checkout context. Approval currently has no production UI or route. If no operator-provided approval path is available, report that blocker rather than promising an automatic resume.

Verify the graph

After every task is complete, perform an independent whole-graph verification. Run all relevant checks and capture command names, outcomes, and limitations as evidence for every task. Submit that evidence through the current service contract and follow the returned lifecycle state; only an accepted passing result releases a successful implementation.

Boundaries

Keep this provider-neutral Agent Skills directory intact when installing it in the shared skill location supported by the local runtime.

Use prompt.md as a starter only after Chopin exposes an approved graph through read_implementation.