fix(skills): make /fixer loop until the backlog is drained and park on stall, not fixed counts

[carlos-alm]

Summary

• Loop until actually done. /fixer previously processed one batch of --count issues and stopped, even if more qualifying open issues remained. It now loops across batches by default (new Phase 3 — Continue the Batch Loop, or Proceed to Drain), re-fetching a fresh queue and continuing until no qualifying open issues remain or a 15-batch safety cap is hit. --once restores the old single-batch behavior.
• Repo scope + do-nothing. /fixer only ever acts on the repo it's invoked in. An empty issue queue is now reported and exits cleanly (exit 0, not an error) — with an explicit rule against searching other repos or inventing unrelated work to fill a batch.
• Stall-based parking, not fixed round/pass counts. Convergence rounds (2f) and drain passes (Phase 4) were both capped at a flat 3, parking a PR/stopping drain regardless of whether real progress was still happening. Both now track a per-round/per-pass signature and park only after 3 consecutive rounds/passes with zero measurable change (genuinely blocked) — backed by a 15-round / 15-pass safety cap for runaway cases.

This replaces #2165, reverted in #2167. Fixes the 2 issues Greptile found on #2165 (Confidence Score 2/5):

1. Drain progress never tracked. The drain phase's stall detection relied on parked.txt shrinking, but nothing removed a PR's number from it after merging — so a pass that successfully merged a PR still looked stalled, and the drain could wrongly report the remaining PRs as blocked after 3 passes. Fixed: the "merge lowest-ready PR" step now explicitly removes that PR from parked.txt.
2. Batch state leaked across runs. Phase 0 only checked whether state.json existed to decide "resume", so a fresh invocation after a fully-completed prior run inherited its stale batches-done count and empty queue.json — causing new issues to be silently skipped or the 15-batch cap to trigger early. Fixed: resume now requires queue.json to be non-empty (real unfinished work); otherwise every run-scoped artifact is reset before starting.

Test plan

• Syntax/logic-checked the stall-detection and parked.txt-removal bash snippets standalone
• Verified the Phase 0 fresh-run-vs-resume branching with both queue states (empty → reset, non-empty → resume)
• npm run lint clean (skill file is outside src//tests/ scope)
• Dry-run /fixer --dry-run on a future run to confirm batch/round bookkeeping behaves as designed

[greptile-apps]

Drained merges leave stale state

When Phase 4 merges a parked PR, this block removes it only from parked.txt; its state.json entry remains parked, causing Phase 5 to undercount merged issues and report completed work as unfinished.

[carlos-alm]

Fixed in 66125b6 — the merge-lowest-ready-PR step in Phase 4 now also flips that issue's state.json entry from parked to merged (matched by PR number), in the same step that removes the PR from parked.txt. Traced through the logic to confirm: previously only parked.txt shrank, so Phase 5's final report kept listing a successfully-merged issue as parked and undercounted the merged total, exactly as described. (Note: this comment's diff anchor landed on the drain-stall stopping-condition block a few lines below the actual removal snippet — the concern described in the prose is what's fixed.)

[greptile-apps]

Progress signature misses active changes

When new commits, review revisions, or check runs retain the same aggregate score, unanswered count, ancestry, mergeability, and failing check names, the signature remains unchanged and the stall counter parks an actively progressing PR.

[carlos-alm]

Fixed in b56330b — the gate signature now also folds in the current commit SHA (git rev-parse HEAD), a hash of the full concatenated Greptile review/comment text (not just the extracted N/5 digit), and the specific set of currently-unanswered comment ids (not just their count). Traced through it: a push landing before the score/checks catch up, a re-review that restates the same score for different underlying reasons, or a CI rerun that fails the same named check for a new cause will now all change the signature, so genuine activity between rounds is never misread as zero measurable progress.

[greptile-apps]

Backlog scan stops at 400

This completion check filters only the first 400 open issues returned by gh issue list. In repositories with more open issues, qualifying entries outside that window remain uncounted while the skill reports the backlog as drained.

[carlos-alm]

Fixed in af322d0 — both the queue-build (Phase 1) and the batch-completion check (Phase 3) now pass --author to 'gh issue list' so the author filter runs server-side instead of being applied client-side after a flat --limit 400 fetch, and the limit is raised to 1000 as a generous safety bound on one author's own backlog rather than a cap on the repo's total open-issue count. Confirmed this was a real gap: with the old client-side filter, issues by other authors interspersed in the 400-issue window could crowd out this author's own qualifying issues, letting REMAINING read 0 while qualifying issues still existed.

[greptile-apps]

Greptile Summary

The fixer skill now processes multiple issue batches and uses progress-based convergence and drain stopping conditions.

• Adds fresh-run versus interrupted-run state handling and a single-batch --once option.
• Adds cumulative batch processing with repository- and author-scoped backlog checks.
• Reconciles drained PRs in both parked.txt and state.json.
• Expands convergence signatures and replaces fixed round/pass limits with stall detection and safety caps.

Confidence Score: 3/5

The PR is not yet safe to merge because active reruns of the same named CI check can still be classified as repeated zero-progress rounds and cause a PR to be parked.

The convergence signature stores failing checks only by name, while the parking logic treats three identical signatures as a stall; a rerun without a new commit or review event therefore remains indistinguishable from inactivity.

Files Needing Attention: .claude/skills/fixer/SKILL.md

Important Files Changed

Filename	Overview
.claude/skills/fixer/SKILL.md	Adds backlog-wide batching, fresh-run state resets, richer convergence tracking, progress-based drain behavior, and drained-PR state reconciliation; same-named CI reruns remain invisible to stall detection.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
  Start["Initialize or resume fixer state"] --> Queue["Build current issue batch"]
  Queue --> Empty{"Queue empty?"}
  Empty -->|Yes| Done["Exit successfully"]
  Empty -->|No| Solve["Solve issues and converge PRs"]
  Solve --> BatchCheck{"More qualifying issues?"}
  BatchCheck -->|Yes, below cap| Queue
  BatchCheck -->|No or --once| Drain["Drain parked PRs"]
  Drain --> Report["Reconcile state and report outcomes"]

Loading

_{Reviews (2): Last reviewed commit: "fix: filter gh issue list by author serv..." | Re-trigger Greptile}

[carlos-alm]

Addressed all three findings from the Greptile review (Confidence Score 2/5):

1. Drained merges leave stale state — fixed in 66125b6. Phase 4's merge-lowest-ready-PR step now flips the issue's state.json entry from parked to merged (matched by PR number) in the same step that removes it from parked.txt, so Phase 5's final report no longer undercounts merged work.
2. Progress signature misses active changes — fixed in b56330b. The gate signature now includes the current commit SHA, a hash of the full Greptile review/comment text, and the specific set of unanswered comment ids, not just aggregates that can lag a push or restate an unchanged score for different reasons.
3. Backlog scan stops at 400 — fixed in af322d0. Both the queue-build and batch-completion checks now filter by author server-side via --author, with the --limit raised to a generous 1000 as a safety bound on one author's own backlog rather than the repo's total open-issue count.

All three were traced through standalone before fixing, per the review guidance. Lint and the full test suite (260 files / 4178 tests) pass locally. Ready for re-review.

[carlos-alm]

@greptileai