iterate-pr: Claude reviews post as github-actions[bot] and are skipped as info-bot noise #514

fetch_pr_feedback.cjs never treats this repo's Claude review as a review. The in-progress placeholder and the finished review are both filed in the bot bucket, which is skipped silently. So summary.review_in_progress stays 0 while a review is still running, and a finished review's findings never reach needs_attention.

Observed on #510, after an @claude /review comment:

• The placeholder opened with ### Review in progress, yet review_in_progress read 0 and the comment was bucketed as bot.
• The iterating agent only waited correctly because it polled the comment body itself. An agent trusting the script would have treated "no review in progress, nothing needing attention" as a clean review and exited before the review landed.

Cause

• The review comment's author is github-actions[bot], not claude[bot]. claude-code-review-on-demand.yml passes github_token on purpose (see the comment near line 109 about avoiding the broader claude App token), so the action posts under the Actions identity.
• In .agents/skills/iterate-pr/scripts/fetch_pr_feedback.cjs:
  • REVIEW_BOT_PATTERNS (line ~90) includes /^claude/i, which expects claude[bot].
  • INFO_BOT_PATTERNS (line ~104) includes /^github-actions/i.
  • bucketByAuthor only runs the in-progress check, and only flags review-bot findings, when isReviewBot(author) is true, so a github-actions[bot] comment drops straight into bot.
• The placeholder regex (IN_PROGRESS_MARKERS, line ~198) is fine. It is never reached.

Evidence

gh api repos/taskless/cli/issues/510/comments --jq '.[] | "\(.user.login) | \(.body[0:50])"'
github-actions[bot] | **Claude finished @theCodeDrift's task in 6m 57s** —— [View

Possible directions, not decided

• Recognise a github-actions[bot] comment as the Claude review by its body: it opens with Claude finished, Claude Code is working, or Review in progress. Route only those to the review-bot path, so real Actions status comments stay bot.
• Or match on the action's own markers (the [View job] link) rather than the login.

Whichever fix lands needs a test with a github-actions[bot] placeholder and a github-actions[bot] finished review. The existing tests use claude[bot], which is why this passed.