feat(loop): comment-fix — automated bounded fix attempts for unresolved PR review threads (issue #96, part 2) by robercano-ghbot · Pull Request #159 · robercano/reCode

robercano-ghbot · 2026-07-18T07:51:56Z

What this does

Part 2 of #96: gives the autonomous PR loop a fourth reaction to a
loop-authored, open PR — an unresolved, qualifying inline review-comment
thread now gets an automated fix attempt instead of sitting open forever
waiting on a human to notice.

pr-comment-fix.sh (new): detects open, bot-authored PRs with an
UNRESOLVED review-comment thread (GraphQL reviewThreads, since REST has
no "resolved" concept for these) that has at least one comment from the
repo owner or an allowlisted bot commenter, excluding: PRs that are ALSO a
pr-feedback.sh candidate (precedence); PRs labeled needs-human or
claude-comment-fixing (in-flight guard, mirroring pr-ci-fix.sh's
claude-ci-fixing); and threads whose fix was already posted for the
thread's CURRENT state (a 
marker cursor, keyed to the thread id + attempt number, so a genuinely NEW
comment reopening the same thread still re-triggers).
Anti-livelock: fix-attempt count is tracked PER THREAD, derived
directly from the marker's own embedded attempt number (no new state
file). A thread that reopens after 2 completed fix attempts is escalated
to needs-human instead of retried again.
loop-census.sh: surfaces the count as comment_fix_prs=N, folded
into the existing cadence=FAST trigger alongside feedback_prs/ci_fix_prs.
loop-tick.sh: wires pr-comment-fix.sh in as step 5/6 (between
pr-feedback.sh and pr-ci-fix.sh), parses its TSV, and emits
action=comment-fix pr=N at the documented precedence.
loop-event.sh: builds the COMMENT-FIX driver prompt — label
claude-comment-fixing first (in-flight guard), fix on the same branch
(no new branch), push, resolve ONLY the threads actually addressed, then
post the claude-comment-addressed marker(s) with the real thread id(s)
and attempt number(s). Never merges, never force-pushes.
loop-daemon.sh: gives comment-fix drivers their own transient
systemd unit shape (pr-loop-driver-commentfix-pr<N>, distinct from both
pr-loop-driver-pr<N> for feedback and pr-loop-driver-cifix-pr<N> for
ci-fix) so drivers for the same PR number can never collide in naming or
reattach.
pr-ci-fix.sh: now ALSO excludes comment-fix candidates from its own
detection (mirroring its existing feedback exclusion), enforcing the
ci-fix < comment-fix precedence twice (belt and suspenders).
gates.json / self/gates.json: new commentFix.botAllowlist config
surface (empty array = disabled, same empty-means-skip convention as
notify) — the repo owner's own review threads always qualify regardless
of this list; fill it in to let a specific bot reviewer's threads
auto-trigger comment-fix too.
.claude/commands/pr-loop.md / .claude/self/pr-loop-self.md: STEP 0
tick contract updated to mention the new step and verdict.

Decision: verdict precedence

feedback > comment-fix > ci-fix > advance. Owner CHANGES_REQUESTED
always outranks everything else. An unresolved qualifying review-comment
thread outranks a CI fix (the reopened conversation is addressed before
chasing a possibly-unrelated CI failure) and a fresh advance dispatch. Merge
still requires a fresh owner approval regardless of any of this —
merge-ready.sh is untouched.

Bounded attempts

Comment-fix dispatches reuse the same per-issue attempt-budget counter
introduced in #95 (loop-issue-attempts.json) for the overall
dispatch-count ceiling — no new state file for that. The PER-THREAD retry
budget (2 attempts before escalating that specific thread) is a
separate, finer-grained bound, derived entirely from the marker comments
already posted to the PR (see pr-comment-fix.sh's header doc).

Out of scope

Part 3 of #96 (conflict-after-sibling-merge rebase) is not included
here — still open, tracked separately.

🤖 Generated with Claude Code

…ed PR review threads (issue #96, part 2) Adds pr-comment-fix.sh (issue #96 capability #2): detects open, bot-authored PRs with an unresolved, qualifying (owner or allowlisted-bot) review-comment thread, excluding feedback candidates, needs-human/claude-comment-fixing labeled PRs, and threads already addressed for their current state. Bounds retries to 2 attempts per thread (derived from the claude-comment-addressed marker itself, no new state file) before escalating to needs-human. Wires the new comment-fix reaction into the loop at the documented precedence (feedback > comment-fix > ci-fix > advance > resume): census surfaces comment_fix_prs=N, loop-tick.sh adds it as step 5/6 and emits action=comment-fix pr=N, loop-event.sh builds its driver prompt, and loop-daemon.sh gives comment-fix drivers their own transient unit naming (pr-loop-driver-commentfix-pr<N>). pr-ci-fix.sh now also excludes comment-fix candidates from its own detection (mirroring its existing feedback exclusion). Adds the commentFix.botAllowlist config surface to gates.json (empty by default — only the owner's threads qualify). Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

…t empty allowlist TESTS reviewer rejection on PR #159 (issue #96 part 2): the anti-livelock escalation scenario had no positive assertion that PR 17's budget-exhausted thread actually triggers `gh pr edit 17 --add-label needs-human` -- "PR 17 not emitted" alone can't distinguish correct escalation from a silent drop. Mirror pr-feedback.test.sh's gh-call-log convention: the fixture's fake bot-gh.sh now appends every invocation to a log file, and the test greps it directly (PR 17 labeled, PR 16 -- attempt 2, not yet exhausted -- is not). Also add a fixture2 scenario with commentFix.botAllowlist: [] (the shipped DEFAULT), previously never exercised since fixture1 hardcodes a non-empty allowlist everywhere: a bot-only commenter does not qualify, while an owner comment in the same run still does, proving the owner always qualifies independent of the allowlist. Test-only change; no production logic touched.

robercano-ghbot · 2026-07-18T08:14:42Z