Task Agent

Each invocation does exactly one task: the next pending item from the tasks file. The same file (or a sibling state file, in legacy mode) tracks completion, so reruns always pick up where the previous run left off.

Arguments (all optional, key=value form):

• tasks=<path> — path to the tasks file. Default: agent-tasks.yml or agent-tasks.yaml in cwd.
• state=<path> — path to a separate state file. Forces legacy mode. Default: auto-derived (<stem>-state.yml) and only used if the tasks file is in legacy format.

Two file formats are supported:

• Unified (preferred): one file holds tasks and their completion status. Each task has an id, a description, and a status (pending | done | failed | skipped). On completion the same entry gains branch, pr_url, date. Unknown keys (e.g. external_ref, labels, priority) are preserved verbatim across rewrites so external sync tools can attach metadata.
• Legacy: two files — a tasks file with bare-string tasks and a separate *-state.yml listing completed entries. Auto-detected when the tasks file has no status: field. Existing setups keep working unchanged.

See ../../references/format.md for the full schema, the status enum, and the passthrough-metadata contract.

Prereqs: python3, git.