watch-reviews

Invoke the skill with a PR number and it polls the inline comment, conversation, and review feeds on a backoff schedule: every 5 minutes for 2 hours, then every 30 minutes. Each round reports new activity and returns, carrying the seen-set and watermark forward.

It authorizes each comment author in two steps: verifying ethereum-optimism org membership, then confirming write, maintain, or admin permission on the repository via the collaborators permission API. Only after both checks pass is the body fetched, one comment at a time, by id.

No. Accounts with user.type Bot are never authorized, and comments from anyone without verified write access are quarantined as metadata only. Their bodies are never read, quoted, summarized, or acted upon.

Edited bodies are treated as unattributable because the REST API does not expose editor identity. The skill compares updated_at against created_at, discards the body, reports it as edited metadata, and escalates to the operator instead of acting.

Only changes to the PR's own diff plus its tests and docs. Everything else, including merging, force-pushing, editing CI configuration, adding dependencies, or running scripts, is escalated to the human operator regardless of the requester's authority.

That command fetches every review body wholesale, including unauthorized ones, and a forgotten jq projection dumps them all into context at once. The skill instead reads metadata-only feeds and fetches bodies individually by id after authorization.