roborev-review

Quick reference

Consent gate — highly opinionated practice. You MUST ask the user before setting this up in a repository (an AskUserQuestion / explicit confirmation, not an assumption). roborev queues background AI agents on every commit — the cost, quota, and git-hook implications are the user's decision. A repo charter (AGENTS.md/CLAUDE.md) that already mandates roborev counts as standing consent. Never install it silently as a side-effect of another task.

Per-commit AI review with roborev

roborev (https://www.roborev.io/) reviews every commit asynchronously: a post-commit hook queues the commit, a reviewer agent produces a verdict with severity-ranked findings, and refine/fix loops turn findings into follow-up commits. The suite convention is that agents read verdicts between work rounds and address findings before merge — review coverage without a human bottleneck.

Setup

Install via brew install roborev (or the installer in the roborev docs). Then, in the repo, with consent:

roborev init          # writes .roborev.toml
roborev install-hook  # post-commit hook that queues each commit for review

Suite-convention .roborev.toml:

agent = "codex"                  # reviewer agent
model = "gpt-5.6-sol"            # resolve the CURRENT best at setup time — see caveat
security_agent = "claude-code"   # auth / VC / ACL / token surfaces
review_min_severity = "low"

Model pins go stale. Resolve the latest available reviewer model at setup time — today gpt-5.6-sol — do not copy a pin blindly from another repo or from this file. Agents resolve the current default at session start rather than hardcoding.

Reviewer-model diversity. Reviews come from a different model family than the author: Claude-authored commits get a codex/GPT reviewer (and security surfaces get a second, Claude-based pass via security_agent). Same-family review repeats the author's blind spots; cross-family review is the point.

Operating discipline

• Async only, never blocking. Agents run reviews in the background and never block foreground work waiting for a verdict (roborev review --background, or simply let the hook queue it). Sequence verdict-reading after other useful work.
• Read verdicts between rounds: roborev status for the queue, roborev show [<commit>] for findings. A lead session reads verdicts at verify-merge time and routes findings back as fix briefs.
• Address findings — fix, or record why not. roborev refine <commit> iterates on a reviewed commit with the findings as input; roborev fix applies the reviewer's suggested remediation. Findings on security-critical surfaces are merge-blockers.
• Verdicts are advisory input to the human/lead merge decision — a green verdict never auto-merges anything.

Agent persona

When this work is delegated to a sub-agent, the lead runbook in `.claude/agents/solid-app-orchestration.md` owns the review loop — it routes through this skill. Specialist personas run roborev only async/background per their stop-gates.