bug-batch

name	bug-batch
description	[Tier 2 · moderate autonomy · full review gate · reproduce-first required] Close a batch of bugs — from a list, an issue tracker, or a described set — one PR per bug, each gated by a FAILING TEST written first that reproduces the bug, then a fix that turns it green. Use for a backlog of bugs, a tracker label, or a known cluster of defects. Bug-fixes are among the categories agents are WEAKEST at (they need exact, not approximate, changes), so this mission forces exactness: prove the bug with a red test before fixing. Does not add features. Runs via the autonomous-fleet-core engine. Trigger on: "fix these bugs", "work through the bug backlog", "close the bugs labelled X", "fix this list of issues", "batch bug fixing".
license	MIT
compatibility	Requires git and gh CLI in the target repository
metadata	{"author":"ravidsrk","version":"1.0.0","tier":"2","fleet-component":"mission"}
status	exploratory

Status: exploratory. This mission is documented but has not yet been run end-to-end on an external repo with an archived run-archive. It is preserved for future promotion (see docs/exploratory/missions/README.md § Promotion criteria). Do not invoke this skill in a production fleet until it has been promoted.

Mission: bug-batch

Required skills

Before executing, activate these skills and read their full instructions:

1. autonomous-fleet-core — read references/engine.md and references/composition.md when coordinating
2. One runtime adapter: autonomous-fleet-adapter-orca, autonomous-fleet-adapter-claude-code, autonomous-fleet-adapter-grok, or autonomous-fleet-adapter-codex

Follow the core and your adapter in full, then apply the mission parameters below.

Do not load a second mission skill in the same run. For chained missions, use fleet-program.

Optional skills

Skill	Activate when	If unavailable
—	—	—

Worker skills

Role	Skills	If unavailable
@codex (repro test, fix)	—	Repo test conventions
@claude (review, integrator)	—	Mission review gate only

Deferred missions

Record in <LEDGER_DIR>/bug-batch-readiness.md under Recommended next missions and in DECISIONS.md.

Finding type	Route to
Fix reveals missing feature (not a bug)	take-product-to-completion or user-scoped mission
Root cause is architectural	adversarial-review-and-fix
Area lacks regression tests	test-coverage

Empirical note: bug-fix is among the two WEAKEST task categories for AI agents — the cross-agent fix-PR merge rate is materially lower than docs/build work (arXiv 2601.15195, Ehsani et al., MSR 2026) — because success depends on EXACT (not approximate) code changes and on correctly localizing the defect. (There is no published category-level merge rate for this mission as scoped; per-task merge rates vary by agent, so treat the tier ordering as qualitative.) The reproduce-first gate below is the specific countermeasure: a failing test that pins the bug forces an exact fix and proves it landed. This gate is MANDATORY — it is the reason this mission works.

GOAL

Close every bug in the batch with a verified fix. The batch source is whatever the user supplies: a list in the prompt, an issue tracker query/label, or a described cluster. For EACH bug: reproduce it with a failing test, fix the root cause, prove the test now passes, and ensure no regression elsewhere. One PR per bug. No feature work — if a "bug" is actually a feature request, record it and skip.

ROLE PIPELINE

• @claude triages the batch into a frozen BUG INDEX.
• @codex (per bug) writes the reproducing test + the fix.
• A fresh build-blind @claude REVIEWS each PR: the test genuinely reproduces the bug (fails WITHOUT the fix), the fix addresses root cause (not a symptom mask), no regression, scoped.
• @claude is the INTEGRATOR: opens PR, merges (conflict-aware), cleans worktree.

LEDGER

<LEDGER_DIR>/bug-batch-progress.md. Per-task flags: REPRO=<t/f> FIXED=<t/f> PR_OPEN=<t/f> REVIEWED=<t/f> MERGED=<t/f>. Plus a BUG INDEX: each bug ID/description, OPEN | FIXED via PR#n | SKIPPED (reason).

TASK STRUCTURE

• T-TRIAGE [@claude] — assemble the batch into a BUG INDEX (if the source is a tracker, pull the matching issues; cite IDs). For each: a one-line repro hypothesis and the suspected area. Separate real bugs from feature-requests-mislabelled-as-bugs (skip the latter, record why). Output docs/bug-batch-triage.md. Freeze the batch, then fix.
• T-FIX… [per bug, loop — REPRODUCE-FIRST GATE] — each bug is one PR, in strict order:
  1. REPRODUCE: write a test that FAILS because of the bug (red). Confirm it red. Set REPRO. If the bug can't be reproduced, mark the bug NEEDS-INFO in the index and move on — do NOT guess-fix an unreproduced bug.
  2. FIX: fix the root cause so the test goes green; run the full suite (no regression). Set FIXED.
  3. A fresh build-blind @claude reviews the PR — independently confirms the test fails without the fix and passes with it, the fix is root-cause not symptom, no regression → @claude merges. Parallelize bugs in non-overlapping files; serialize same-file (hot-file rule). Update the BUG INDEX.
• T-FINAL [@claude] — full suite green incl. every new reproducing test; walk the BUG INDEX, every bug FIXED or SKIPPED/NEEDS-INFO with reason. Output <LEDGER_DIR>/bug-batch-readiness.md with fleet-outcome YAML (bugs_open, bugs_skipped), bug index resolution, the reproducing test per fix, skips/needs-info, all PRs). Ship as the final PR.

Runtime goal

After ledger init, SET_GOAL per autonomous-fleet-core/references/runtime-goals.md. Record ## Runtime goal in <LEDGER_DIR>/bug-batch-progress.md. GOAL_COMPLETE only after ## DONE below.

Mission bug-batch DONE: <LEDGER_DIR>/bug-batch-progress.md all task flags true,
<LEDGER_DIR>/bug-batch-readiness.md with fleet-outcome.status done and mission metrics satisfied,
./scripts/validate-fleet-outcome.sh passes, all PRs merged into BASE.

DONE

Every BUG-INDEX item FIXED (with its reproducing test) or explicitly SKIPPED/NEEDS-INFO, every fixed task terminal, suite green, <LEDGER_DIR>/bug-batch-readiness.md exists. Then send the FINAL report.

DECISION DEFAULTS (mission-specific)

• REPRODUCE-FIRST IS MANDATORY. No fix without a test that first fails because of the bug. An unreproduced bug is marked NEEDS-INFO, never guess-fixed.
• Fix ROOT CAUSE, not the symptom. A change that hides the symptom but leaves the cause is rejected.
• A "bug" that is actually a feature request → SKIP, record why; don't build features here.
• The reproducing test stays in the suite as a permanent regression guard.
• One PR per bug — never batch multiple bugs into one PR (keeps localization and review clean).
• Any ambiguity → reproduce it precisely first; let the failing test define exactly what "fixed" means.