Required lifecycle for non-trivial work. Create a task, claim it, do the work, mark done. Include the task name in your reply.
Note: For idchain agents, these rules are also injected via
defaults.claudeMdinconfigs/idchain.yaml, so the lifecycle is always in context without invoking this skill. This file is kept as reference documentation and for non-idchain agents that load the skill explicitly.
You treat every non-trivial request as a first-class task in the manager's /tasks system. The task lifecycle is mandatory for any multi-step work or work that produces an artifact.
POST $MANAGER_URL/tasks with {title, name, from: <your-name> }POST $MANAGER_URL/tasks/<name>/claim with {agent_id: <your-name> } (status flips to doing)./output/ in your working directory.POST $MANAGER_URL/tasks/<name>/done with {agent_id: <your-name> } (status flips to done)Done. Task: implement-x. Output: ./output/report.mdMark the task done with a failure note in the reply. Do not leave it in doing.
Other agents reading the task stream need to see the terminal state, even if it is failure.
Use kebab-case for task names: implement-feature-x, audit-contracts-sept, review-pr-42.
Avoid reserved command verbs (delete, deploy, sync, etc.) which will be rejected by the validator.
A verifier agent walking the task stream can see every unit of work, every artifact, every completion or failure, but only if every agent uses the system. Your discipline is what makes the team auditable.