Cross-AI Review
Before a single line of code gets written, PP Core can send your phase plans out to other AI tools for a second opinion. This guide explains how cross-AI plan review works, how to wire it into replanning, and how to configure which reviewers weigh in.
Why review plans, not just code
A flawed plan produces flawed code no matter how cleanly it is executed. Cross-AI review catches design problems at the cheapest possible moment, while the work still lives in a plan file. By routing your plans through reviewers built on different models, you surface blind spots that any single model, including the one that wrote the plan, tends to miss.
Requesting a review
/pp-review --phase N
This dispatches the phase's plans to multiple AI reviewers in parallel and collects their feedback into {phase}-REVIEWS.md, with findings classified by severity. The reviewers read the phase PLAN.md files along with RESEARCH.md as input.
Choosing reviewers
Each reviewer corresponds to an external CLI you must have installed and authenticated. Select them with flags:
| Flag | Reviewer | Requirement |
|---|---|---|
--gemini | Gemini CLI | Installed and authenticated. |
--claude | Claude, in a separate session | claude CLI installed. |
--codex | Codex CLI | Installed. |
--opencode | OpenCode | CLI installed. |
--qwen | Qwen Code | Installed with an API key. |
--cursor | Cursor agent | Agent access configured. |
--agy / --antigravity | Antigravity CLI | Free with Google credentials. |
--all | Every detected reviewer | Each CLI installed and authenticated. |
For example, to get feedback from two reviewers at once:
/pp-review --phase 4 --gemini --codex
Feeding feedback back into planning
Once a review lands, fold its concerns into a fresh plan by replanning with the --reviews flag:
/pp-plan-phase N --reviews
This treats REVIEWS.md as additional context, so the new plan directly addresses what the reviewers raised.
Automating the loop to convergence
Doing plan, review, and replan by hand works, but PP Core can run the cycle for you until the serious concerns are gone:
/pp-plan-review-convergence N
The convergence command repeats the loop: plan, review, count the HIGH-severity concerns, and replan if any remain. It exits when the HIGH count reaches zero or the cycle cap is hit. The default cap is three cycles, which you can extend with --max-cycles. Progress across cycles is tracked in a CYCLE_SUMMARY embedded in REVIEWS.md.
Local models in the loop
If you would rather review with models running on your own machine, the convergence command supports them through --ollama, --lm-studio, and --llama-cpp. Note that local-model support exists only in /pp-plan-review-convergence, not in /pp-review.
/pp-plan-review-convergence 4 --ollama --max-cycles 5
Configuration
Persist your preferences in config.json so reviews run the same way every time:
review.default_reviewers Array of reviewer CLI names to use by default
review.models.<cli> Per-reviewer model specification
workflow.plan_review_convergence Boolean gate, disabled by default
Setting review.default_reviewers means you can run /pp-review --phase N without naming reviewers each time, and turning on workflow.plan_review_convergence makes convergence a standing part of your planning gate rather than something you invoke by hand.