auto-fix-all

---

name: auto-fix-all description: Autonomously runs the full pipeline (new issue → plan → fix → monitor) for a queue of issue IDs, one at a time, forever — waiting for new IDs to be pushed onto the queue whenever it runs dry — with no user interaction except when a PR is closed without merging. Usage: /auto-fix-all ...

You are the coordinator. Your job is to manage the queue and the two things the architect agent cannot do itself (ScheduleWakeup between issues, asking the user what to do about a closed PR) — everything else (implementation, PR review, comments, CI) is delegated to a spawned architect agent, one per issue. Follow the steps below precisely and in order.

The issues folder is always docs/agents/issues and the plans folder is always docs/agents/plans.

Step 1 — Initialize the queue

If skill arguments were provided (space-separated IDs), run:

scripts/queue.sh save <id1> <id2> ...

If no arguments were given, this is a re-invocation after context clearing — the queue already contains the remaining issues. Skip this step and go directly to Step 2.

Step 2 — Process the next issue

Get the next id (blocks until the queue has one — if it's currently empty, it sleeps 5 seconds and checks again, forever, so a run that drains the queue keeps waiting for issues pushed onto it later, e.g. via push-issue-to-queue, instead of exiting):

scripts/queue.sh wait-next

Call this id <id>. Spawn:

Agent(subagent_type: "architect", prompt: "Read steps/process_one_issue.md (resolved relative to the auto-fix-all skill folder) and follow it for issue . Report OUTCOME=merged or OUTCOME=closed PR_NUMBER=.")

Wait for the agent to finish, then parse OUTCOME from its report, and proceed to Step 3.

Step 3 — React to the outcome

OUTCOME=merged

scripts/queue.sh pop

Check whether to clear context:

scripts/config.sh is-enabled clear_context

• Exit 0 (true): call ScheduleWakeup(delaySeconds=60, prompt="/auto-fix-all", reason="clearing context before next issue") and stop. Do not loop back to Step 2. (Requires that auto-fix-all was invoked via /loop /auto-fix-all <ids>; the 60-second wakeup fires a fresh iteration that reads the queue and continues.)
• Exit 1 (false or absent): go back to Step 2 to process the next issue.

OUTCOME=closed PR_NUMBER=<n>

This is the one point in the whole pipeline where you ask the user something — the spawned architect agent cannot, so it stopped and handed this back to you.

PR #<n> for issue <id> was closed without merging. What would you like to do?

1. Reimplement from scratch (start over from a clean main for this issue)
2. Skip this issue and move on to the next one

• Reimplement — the rejected branch must be discarded first, since process_one_issue.md's branch bootstrap now reuses an existing issue-<id> branch instead of always recreating it:

  scripts/github.sh cleanup-branch <id>
  

  Then go back to Step 2 (the id stays at the front of the queue; a fresh architect agent will find no existing issue-<id> branch and create a genuinely clean one from main via process_one_issue.md).
• Skip — scripts/queue.sh pop, then go back to Step 2.

Step 4 — Done

This skill runs forever by design — Step 2 blocks and waits whenever the queue is empty instead of stopping, so issues pushed onto the queue later are still picked up. This step is only reached if the run is stopped externally (e.g. the user interrupts it): report a summary at that point, for each ID processed so far, of the final PR URL and outcome (merged/skipped).

Do not ask for confirmation at any point except the single explicit question above for the closed outcome.