work-management

Work Management

1. Inspect relevant instructions and existing work. A small edit uses the conversation and appropriate checks; create no process file automatically.
2. For larger work, reuse or create one .ai/work/<slug>.md per outcome using the template or CLI. Sessions and agents share it.
3. Define the goal, constraints, observable acceptance criteria and a short ordered plan. Use stable task IDs with states open, in_progress, blocked and done.
4. Update evidence as work happens. Implementation progress and acceptance verification are separate: a changed file alone does not prove resulting behavior.
5. Preserve human feedback verbatim with stable IDs and an explicit disposition. Record only consequential decisions and their reasons.
6. Archive in .ai/archive/ when the outcome is verified and unresolved feedback is reviewed. Link reusable knowledge into existing documentation or the private vault.

Safe updates and handoffs

• Read the current record and relevant files; load historical work only when needed. Before rewriting, compare against the version read and preserve intervening edits. Retry or resolve conflicts; never replace a record from a stale session snapshot.
• The dashboard edits the same Markdown file using ETag checks. Filesystem updates require the same reread and comparison discipline.
• A completed task needs evidence. If a required check cannot run, record why, the affected criterion and the next action. Keep verification outstanding unless other adequate evidence establishes it; never archive an unverified outcome.
• Keep shared rules in the toolkit and local constraints in project AGENTS.md. Keep docs/ for its existing purpose. CSV is generated, not maintained separately.
• Follow repository privacy rules. Keep private vault notes and configuration out of public records. Report the outcome, actual checks and material limits concisely.