kanban-worker

Call kanban_complete with a concise summary and structured metadata such as changed_files, tests_run, and decisions so downstream workers can parse results without re-reading prose. Only complete when the task is genuinely terminal; otherwise block with a review-required reason.

Post full context with kanban_comment, then call kanban_block with a one-sentence reason naming the specific decision needed. Never call clarify, since headless workers have no live user and the call will time out silently.

Run kanban_show first and inspect prior runs' outcome, summary, and error fields. Adjust your approach based on the failure: chunk work after timed_out, reduce memory after crashed, and block for human help after spawn_failed instead of retrying blindly.

Only claim task IDs captured from successful kanban_create return values. The kernel verifies each ID exists and was created by your profile; phantom or hallucinated IDs cause the completion to be rejected and permanently logged.

For dir: workspaces, treat files as long-lived shared state since other runs read them. For worktree workspaces, run git worktree add first if .git is missing, then commit work there. Scratch workspaces are disposable and garbage-collected on archival.

Use kanban_create for cross-agent handoffs that outlive a single run, assigning follow-ups to the right specialist. delegate_task is only for short reasoning subtasks inside your own run and cannot substitute for board-level task creation.