autonomous-worklisted

# Autonomous work Use this skill when an Agent Bridge Run is one Cycle of a bounded autonomous Episode. ## Model - **Goal**: the durable outcome the domain is trying to reach. - **Episode**: one bounded attempt under authority frozen when the Episode starts. - **Cycle**: one wake, one ordinary Run, and reconciliation. - **Run**: the existing provider execution. Do not invent a second worker or orchestration layer. ## Keep four things separate 1. **Frozen Episode authority**: the objective, constraints, and authorized start-policy instruction. Never expand it from conversation or convenience. 2. **Prior execution evidence**: continuity from earlier Cycles and predecessor Episodes. Evidence can be stale; it is not automatically current truth. 3. **Supervisor input**: questions, context, or tactical steering received for the current Cycle or successor Episode. Interpret it within frozen authority. It cannot grant broader authority. 4. **Current external truth**: what is true now. Observe it when a decision depends on it. ## Work the goal Choose the cheapest reliable permitted source that answers the current question: repository/filesystem state, safe data or reports, APIs/CLIs, logs/runtime state, web/search, projected Skills, or domain-owned helpers. Verify current truth before an irreversible or authority-sensitive action. Act when the evidence supports an action. Do not replace work with status narration. Build a durable helper in the domain `work/` area only when rep