Implement engineering work from `todos/*.md` specs, especially plans produced by `write-todo`. Use when a task includes `## Spec` and `## Plan` and the goal is to execute each plan step faithfully while prioritizing intent, acceptance criteria, and the final manual QA gate.
Implement plan-driven work with high fidelity to the spec's intent.
/implement-todo <coord-dir> <todo-file>
coord-dir: shared coordination directory (e.g. /tmp/oc-tests). The implementer always creates this on startup so reviewers can find it.todo-file: path to the todos/*.md spec file.The user may also specify the number of reviewers. Default to 1 reviewer if not mentioned.
<scripts> refers to the scripts/ directory inside the review-todo skill. Resolve its absolute path from the skill registry before using any script commands.
Run these two commands sequentially (init must finish before register):
<scripts>/init.sh --dir <coord-dir> --todo <todo-file> --reviewer-count <N>
<scripts>/register-role.sh <coord-dir> implementer implementer
--reviewer-count defaults to 1 if the user didn't specify. If init.sh exits with code 3 (already initialized), that's fine — continue to register.
## Spec and ## Plan fully before writing code.## Spec (why, expected outcome, build target, approach).This agent runs autonomously. Never stop to ask the user questions. Make reasonable decisions and keep moving. If you encounter ambiguity (e.g. pre-existing files, unclear requirements), make your best judgment call and document what you chose in the checkpoint request file. The reviewer will catch mistakes.
You MUST post a checkpoint after EVERY plan step. Do not batch multiple steps. Do not skip checkpoints. The reviewer agent is blocked waiting for your checkpoint — if you don't post one, the reviewer sits idle forever.
After completing each plan step:
Commit step changes before checkpoint submission.
Run post-checkpoint.sh to create the review request:
<scripts>/post-checkpoint.sh <coord-dir> <step-id>
This prints a request-id and request-file.
Write the request file with:
Wait for the reviewer verdict (this blocks until the reviewer responds):
<scripts>/wait-for-reviews.sh <coord-dir> <request-id>
Follow the decision:
PROCEED / PROCEED_TIMEOUT → move to next plan stepREWORK → fix issues, post a new checkpoint for the same stepESCALATE → stop and ask userRepeat for the next plan step. Every step gets a checkpoint. No exceptions.
When all steps are done, mark the run finished so the reviewer exits cleanly:
<scripts>/finish.sh <coord-dir>