Skip to content

Review bot wastes CI compute with unbounded polling and monitors self-authored PRs #1186

@worktrunk-bot

Description

@worktrunk-bot

Problem

The review bot's Step 5 ("Monitor CI") has two issues that waste CI compute across every review session:

1. Unbounded polling loop

Step 5 says "poll until complete (sleep 30–60s between checks)" with no cap. In practice, the bot enters a sleep/poll cycle that consumes 6–13 minutes per review session. Across 5 review sessions in the past 24 hours, roughly 30–40 minutes of CI compute was spent on sleep commands.

Evidence from the past 24 hours:

2. CI monitoring on self-authored PRs is pointless

The bot monitors CI for self-authored PRs even though it has no approval to dismiss. Step 5 says "After approving, check whether CI has finished" — but for self-authored PRs, the bot cannot approve (GitHub rejects self-approvals), so there's nothing to dismiss on failure. Run 22295767857 spent 6 poll cycles monitoring CI on PR #1183 with no possible action on the outcome.

This was also identified by yesterday's nightly review subagent (run 22294949725, agent-a748832) as a "Moderate Severity" finding, but the main agent dropped it from the final report.

Suggested fix

In the pr-review skill's Step 5:

  1. Cap polling at ~5 iterations to limit waste
  2. Explicitly skip CI monitoring for self-authored PRs where no approval was posted
  3. Use gh run watch --exit-status where possible instead of manual sleep/poll loops

Source

Nightly reviewer analysis of runs from 2026-02-23.

Metadata

Metadata

Assignees

No one assigned

    Labels

    claude-behaviorIssues with Claude CI bot behavior

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions