cairn-autonomous-round

cairn: Autonomous-round skill

Drive exactly one bounded autonomous cycle. One task, one commit, one journal entry, explicit handoff. No re-scheduling, no chaining — when the cycle is done, control goes back to the user.

For repeating cycles, see the sibling cairn-autonomous-loop skill.

When to use this skill

• User says "do one task", "pick one from the punch list and run it", "finish the Kanban refactor autonomously" (scoped to one item), "take this one to done then stop".
• A scheduled wakeup fired (loop mode) but the user has asked to convert the remaining work to a single round.

Do NOT use for

• Interactive sessions where the user is actively typing — use your normal flow; the round protocol is overkill mid-dialog.
• Multi-step plans that need user checkpoints between steps. If the "task" is actually "build feature X" with three design decisions along the way, stop and ask before starting.
• Open-ended "keep going while I sleep" — that's cairn-autonomous-loop, not this skill.

Protocol

The substance — autonomy-level menu, task-pick rules, gate requirements, hard rules, mid-cycle ask-anyway cases — lives in <project>/docs/workflow/autonomous-protocol.md. Read that before proceeding; this skill is a thin orchestration wrapper over the protocol, specialised for one cycle only.

Summary of the round-specific shape:

1. Check for new user input; abandon if present.
2. Calibrate autonomy level (menu; L2 default). Log the selected level in the session journal.
3. Verify prerequisites (clean tree, journal exists or is creatable, docs/todo.md exists).
4. Pick one bounded task. If cairn's autonomous-planner sub-agent is available, dispatch it for a recommendation; otherwise scan docs/todo.md directly per the autonomy level's rules.
5. Announce the pick in the journal: Picked X because Y. Autonomy level: LN.
6. Implement.
7. Gate. If the review-phase skill is available, invoke it for quality + security passes.
8. Commit locally. No push.
9. Write the handoff summary and stop. Do NOT schedule a next wakeup — this skill is one-cycle-only.

Handoff format

The final action in the session journal (written by this skill's last step) must match the protocol's handoff block:

## Handoff — <timestamp>

**Shipped this round:** <SHA> — <commit subject>.

**Autonomy level used:** <L0/L1/L2/L3/L4>.

**Stopped because:** round complete (single-cycle scope).

**Queued if you want more:** <next bounded tasks, or "nothing
obvious">.

**For your review:** <numbered yes/no questions>.

Relationship to the loop skill

• autonomous-round = this skill. One cycle. Stop.
• autonomous-loop = sibling skill. Multiple cycles with checkpoint pauses and wakeup scheduling.

Both dispatch the same underlying protocol. Pick the skill that matches what the user asked for; don't try to run a loop through this skill by invoking it repeatedly.