Claude Pr Review

claude-pr-review

Multi-axis PR review orchestration for Claude Code — one /pr-review  command that runs up to five independent review perspectives against the same pull request, cross-verifies every finding between axes, and compiles a single comparison report with copy-paste-ready inline comments.

Built and battle-tested over months of daily production PR review. Extracted from the author's personal setup; see Prerequisites honestly before expecting it to run as-is.

Why this design — the methodology behind the command, told through real catches and misses (a dead Save button four reviewers missed, red-team refutation rates, why consensus still gets verified): 一個模型不夠：五軸交叉審的 code review 工作流 (zh-TW). The article states the three design philosophies; this repo is the implementation, and has kept evolving since it was written (formal-spec gate, provenance discipline, Codex presets came later) — where they differ, the command file is current.

Why multi-axis

A single reviewer — human or model — has a single blind spot profile. This command deliberately combines perspectives with *different* blind spots:

The design principle borrowed from security auditing: the agent that finds an issue never verifies it. Context-aware findings are verified by the diff-only axis and vice versa (symmetric cross-verification), consensus findings still get a convention-baseline check, and no finding is ever dropped — refuted ones ship in the report with both sides' evidence so the human makes the final call.

What you get

- A per-run selection matrix — Step 2.98 lays the five axes (plus an experimental web GPT Pro path) out as a model×angle matrix, marks each cell recommended or optional, and waits for your choice. A recommendation is never consent: an unanswered prompt selects nothing,…