fix-bug-issues

name	fix-bug-issues
description	Use when asked to clear, triage, fix, or resolve the open `bug`-labeled GitHub issue backlog for this repo (SubsetJuliaVM / AtelierArith/ailujsoi) — collect all open bug issues, fix each in parallel git worktrees, open PRs, and merge.

Fix Bug Issues (collect → parallel fix → PR → merge)

Overview

Clears the open bug backlog end to end. A multi-agent workflow does the heavy lifting (cluster → fix in isolated git worktrees → adversarially verify → merge each confirmed PR independently, without waiting for siblings). This skill wraps it with the two steps the workflow cannot do itself: collecting the live issue list (gh) and the final full-suite gate (cargo nextest).

The workflow is opt-in multi-agent orchestration; invoking this skill IS the user's opt-in.

Steps

1. Collect the LIVE list (never a hardcoded snapshot — stale snapshots have fixed already-closed issues):

   gh issue list --repo AtelierArith/ailujsoi --label bug --state open --limit 100 --json number --jq '[.[].number]'
   

   If empty, report "no open bug issues" and stop.

2. Run the workflow, passing the collected numbers with merge: false so implementation PRs remain draft for lead certification:

   Workflow({ name: "fix-bug-issues", args: { issues: [<numbers>], merge: false } })
   

   It triages into work-items, fixes each in an isolated worktree with a fixture verified against upstream julia, opens each implementation PR as draft, and adversarially verifies it. Implementation agents never mark ready or merge (Issue #11056).

3. Lead-certify confirmed PRs sequentially. For every confirmed result, review the exact diff, update it to current origin/main, run the required guarded gate (full suite for VM/compiler/dispatch/inference changes), then land it with bash scripts/premerge_gate.sh --pr <N>. Keep concerns/rejected PRs draft for fix-forward or close them; never merge them manually around the gate.

4. Final full-suite gate (REQUIRED after parallel merges — the promote-fallback / dispatch-order OOM trap only shows in the full run; AGENTS.md):

   # from the repository root
   git checkout main && git pull --no-edit
   timeout 1800 cargo nextest run --release --no-fail-fast   # never pipe to `tail` alone — note which test failed
   

   Treat a failure as a regression ONLY if it does not also fail on the pre-merge main; pre-existing failures (e.g. a separately-tracked one) are out of scope — say so explicitly.

5. Report: issues closed, PRs merged, any concerns/rejected left for review, and the full-suite result.

6. Post-mortem (sjulia-postmortem): record run-level insights in ./memory/ (clusters that recurred, verification gaps), and file follow-up Issues for anything deferred.

Notes

• The repo is under heavy parallel development; the workflow reproduces each bug on up-to-date main first and skips ones already fixed (status fixed-no-pr) — no duplicate PRs.
• New bug issues can appear during a run from other work. This skill handles the snapshot taken in step 1; re-run to pick up newcomers.
• Incidental discovery: if you hit a construct that works in upstream julia but not sjulia, file an unsupported-feature issue (AGENTS.md Discovery Rule), don't silently route around it.
• Engine: .claude/workflows/fix-bug-issues.js (args.issues required; returns verified draft PRs for the lead to certify sequentially).
• Requires the Claude Code Workflow tool (step 2). Agents without it (Codex, opencode, Cursor, …) fall back to a sequential loop: for each issue, reproduce on up-to-date main first (skip if already fixed), fix it in its own worktree/branch following sjulia-dev, verify the fixture against upstream julia, then PR + merge via sjulia-finish-branch — and still run step 4's full-suite gate at the end.