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. Only complete when the task is genuinely terminal; otherwise block with a review-required reason.

Post the structured metadata (changed files, test counts, diff path) as a kanban_comment first, then call kanban_block with a reason prefixed 'review-required:'. A reviewer approves via hermes kanban unblock, which re-spawns the worker with the comment thread.

The kernel verifies every id in created_cards against cards actually created by your profile. Phantom or invented ids cause the completion to be rejected with an error, and the failed attempt is permanently recorded on the task's event log.

Run kanban_show first and inspect prior runs' outcome, summary, and error fields. Outcomes like timed_out, crashed, spawn_failed, reclaimed, or blocked each indicate different fixes, such as chunking work, reducing memory, or asking a human instead of retrying blindly.

No, the CLI is not installed in containerized backends like Docker or Modal, so it will fail there. Use the kanban_* tools, which work across all terminal backends; the CLI exists only for human operators at a terminal.

Skip heartbeats entirely for tasks under about two minutes, and never send empty notes or 'still working' messages. Good heartbeats name concrete progress like 'epoch 12/50' and are sent every few minutes at most.