@howaboua/pi-subagent-review

Adds /review, which sends the current repository state to an isolated review subagent and injects its findings back into the main session for triage.

Install

pi install npm:@howaboua/pi-subagent-review

Run /reload after installation if Pi is already open.

Usage

/review
/review focus on migrations and tests
/review loop
/review stack=main@origin
/review stack="description(exact:'release base')" focus on migrations

Anything after /review becomes additional reviewer guidance. A leading loop starts review-loop mode and is removed from that guidance. In a JJ workspace, stack=<ancestor revset> selects a cumulative review base and is removed from that guidance; quote a revset containing spaces. Only one stack= argument is accepted.

Findings are advisory. The command tells the main agent to verify and categorize them against the current implementation and session context rather than treating them as a TODO list.

With a compatible optional Pi Codex Responses adapter, the review preface and triage follow-up use developer-role policy. The preface keeps its existing display and deduplication; the follow-up appears as an extension developer message rather than user input. Raw reviewer findings remain lower-authority custom-message content. Without the API or active adapter, ordinary Pi delivery is unchanged.

Review scope

In a JJ workspace, the extension reviews the active revision before considering an enclosing Git checkout. It pins the active change and commit IDs, reviews against its direct parent by default, and uses stack=<ancestor revset> for a cumulative range. The active revision must be conflict-free. The read-only command reviews stored commit content; capture filesystem edits with JJ before invoking /review.

Elsewhere, the extension chooses among local dev, main, and master branches, then reviews from the merge base so committed and dirty changes are included without base-only commits. With no usable base it reviews the checkout; with no changes it reviews the latest commit.

When enabled, a separate model summarizes the current Pi branch for the reviewer; raw turns are not sent. Review continues diff-only if summarization fails.

Review loops

/review loop records a review-specific marker. The next /review summarizes fixes since that point with the configured summary model, advances the marker, and starts another pass. It does not conflict with pi-auto-trees' /marker.

Configuration

On first load, the extension creates ~/.pi/agent/pi-subagent-review.json, or the equivalent path under $PI_CODING_AGENT_DIR.

{
  "model": "openai-codex/gpt-5.6-sol",
  "thinking": "medium",
  "summary": {
    "enabled": true,
    "model": "openai-codex/gpt-5.6-luna",
    "thinking": "low"
  }
}

Models use Pi's provider/model format. Thinking levels through max are accepted and clamped. If a configured model is unavailable, the command falls back to the current session model.

The summary model prepares both reviewer conversation context and review-loop increment summaries.

Do not load another extension that registers /review unless the command collision is intentional.