Repository navigation
ci: bestaxbot signs off the terminal states in surfer dialect - #249
Conversation
…recipe On PR #247 the fix agent ran clean (opus, 24 turns, success) but hit 15 permission denials and produced ZERO output — no commits, no thread replies, no recap — so the loop escalated with the findings unaddressed. Same class as the original #238 thrash: the agent reached for a tool that is not allowlisted and gave up. The execution-output artifact that would name the tool is defeated by the proxy blocking its blob-storage host, and the job log hides per-turn detail by default. Two fixes: - show_full_output: true on the fix AND verify jobs, so every tool call and denial lands in the readable job log (no artifact needed). - Give both agents the explicit gh-api thread-reply recipe (addPullRequestReviewThreadReply) and state plainly that there is NO MCP 'reply' tool here — only mcp__github_inline_comment__create_inline_comment for a NEW comment. Reaching for a nonexistent MCP reply tool is the most likely denial source; this makes the reply path deterministic. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NVR5yWevceZEFbTJmpviiM
The loop's human-facing terminal comments (converged / contested / paused / wipeout) were posted by github-actions[bot] in flat prose. Give them a personality: bestaxbot now authors them (via AI_LOOP_PAT; the label + reviewer edits stay on the job token), in a Californian/Hawaiian surfer voice with an AI-superiority streak — 'clean set, total convergence, need a meat sack to hit merge.' Covers all five terminal comments: converged handoff, contested findings, halt (cap/protected/review-failed/cr-stalled), fix-run wipeout, and verify no-progress. Only the comment author + wording change; labels, reviewers, and the machine-parsed ai-loop-state comment are untouched. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NVR5yWevceZEFbTJmpviiM
|
Caution Review failedThe pull request is closed. ℹ️ Recent review info⚙️ Run configurationConfiguration used: Path: .coderabbit.yaml Review profile: CHILL Plan: Pro Plus Run ID: 📒 Files selected for processing (1)
WalkthroughThis PR updates the Claude PR loop GitHub Actions workflow. It enables full output streaming for fix and verify Claude steps, changes prompt instructions to reply in review threads via gh api graphql (with resolveReviewThread for verify), adds explicit BOT_TOKEN sourcing from secrets.AI_LOOP_PAT for comment operations, and rewrites several pause/halt/handoff comment bodies. ChangesClaude PR loop workflow updates
Estimated code review effort: 2 (Simple) | ~12 minutes Possibly related PRs
Suggested labels: ✨ Finishing Touches🧪 Generate unit tests (beta)
Comment |
Preview DeploymentPreview URL: https://b1044e64.bestax.pages.dev |
|
🎉 This PR is included in version 5.3.0 🎉 The release is available on: Your semantic-release bot 📦🚀 |
|
🎉 This PR is included in version 3.2.0 🎉 The release is available on: Your semantic-release bot 📦🚀 |
|
🎉 This PR is included in version 1.0.0 🎉 The release is available on: Your semantic-release bot 📦🚀 |
|
🎉 This PR is included in version 1.0.0 🎉 The release is available on: Your semantic-release bot 📦🚀 |
What
The loop's five human-facing terminal comments are now authored by bestaxbot (via
AI_LOOP_PAT) and written in a Californian/Hawaiian surfer voice with an AI-superiority streak — instead of flat prose from github-actions[bot].The voice
Mechanics
<!-- ai-loop-state -->comment are untouched (still github-actions[bot]) — the gate's parsing is unaffected.GH_TOKEN="$BOT_TOKEN"(AI_LOOP_PAT); the label/reviewergh pr editcalls stay on the job'sGITHUB_TOKEN.Note
These are comments — no CI/rebuild cost. YAML validated.
Generated by Claude Code
Summary by CodeRabbit