GitHub Copilot

Topic file for running pr-reviewer from GitHub Copilot (CLI agent mode, IDE chat, Copilot Workspace). Not yet exercised in staged mode.

Fetch tooling

Metadata can come from the GitHub MCP server (get_pull_request gives head.sha and base.sha in one response, which is what Step 2 needs) or from gh in a shell. The clone and pin in Step 4 need a shell regardless; git is the universal path.

Models

Copilot does not expose a per-subagent model choice at the time of writing. Record copilot as the model for every pass, and record the effort you asked for in prose. If your Copilot does offer model selection for sub-tasks, map Hard to the strongest model available, Medium to a mid-cost one, Easy to the cheapest, and record what you chose.

Fan-out primitive

Copilot CLI's agent mode can run a task end to end but, at the time of writing, has no primitive to spawn three isolated sub-agents that cannot see each other. Without one, run single mode and record mode: single (fallback: no isolated sub-agent primitive) in findings.md so the report says so. If a future version adds isolated sub-tasks, each reviewer prompt from references/passes.md is one sub-task; the isolation requirement is that no reviewer's context contains another's output.

Looping primitive

No /loop or /schedule yet (copilot-cli#2056). Autopilot completes one task; it does not repeat a prompt. Use:

• Manual re-invocation after each round until converged: true or --rounds=N.
• Autopilot task phrasing: "Run /pr-reviewer <num> --rounds=3 repeatedly until findings.md records converged: true for the current revision."
• External shell driver that breaks on the convergence marker (see the awk check in generic.md).

Posting

Step 11 of the skill. Through the GitHub MCP server's review-creation resource if connected, else gh api … /reviews --input. Always after the report, always with explicit confirmation.