GitHub Code Review Bot

Scheduled AI review of pull requests. A reusable workflow selects the open, human-authored pull requests of an organisation, runs one Copilot CLI review session per pull request against an offline packet and an unprivileged checkout, and lets a trusted job approve the changes that are trivial or that a code review finds low risk. The bot posts nothing else: a pull request it cannot approve appears in the run report for a human, and the pull request itself stays untouched.

How a run works

select (trusted)
  | ledger, scan, skip rules, per-PR packet
  | commit SHA, evidence ID, digests
  v
review (untrusted, one runner per pull request)
  | offline packet + head checkout; no App key or installation token
  v
apply (trusted, one runner per pull request)
  | verify evidence, corroborate the verdict, re-read live state
  | approve with a per-repository token, or record for a human
  v
report (trusted)
  | report + the next run's ledger

Tiers

Tier	Meaning	Outcome
trivial	Pins, lockfiles, documentation, formatting; corroborated by a deterministic classifier	Approved
low-risk	Logic changed, full review found no security or CI/CD concern and no feedback	Approved
needs-human	Anything else, including any uncertainty	Reported, nothing posted

Where to go next

• Setup: App permissions, secrets, validation, going live, targeted reviews and recovery.
• Design: the architecture, trust model, selection rules, ledger, inputs and data contracts.