Code-Review-Loop

A deterministic-first, persona-based review CLI for PR diffs.

Code-Review-Loop runs a structured review pipeline:

1. select personas/lenses,
2. review each lens independently,
3. cross-check findings with discourse (AGREE/CHALLENGE/CONNECT/SURFACE),
4. optionally verify against requirements/conventions, and
5. summarize score/effort/verdict in deterministic tables.

The tool is implemented in Rust. Default LLM backend is Claude Code CLI (claude -p --output-format json); an OpenRouter backend (--backend openrouter + OPENROUTER_API_KEY) is also available and does not require the claude CLI.

What this repository is for

This CLI is designed for PR-level review and post-review remediation checks.

• review : full PR pipeline (primary mode)
• describe : generate PR summary metadata (title, summary, walkthrough, labels)
• improve : produce concrete patch suggestions with before/after code snippets
• ask : free-form Q&A against a diff

It favors traceability and auditability:

• deterministic checks are locally computed,
• LLM output is limited to judgmental parts, and
• fixed schemas keep outputs reviewable by scripts.

This repository is the private implementation counterpart and aligns operationally with the public full-review skill ecosystem where applicable.

Build and requirements

Requirements

• Rust toolchain (for building CLI)
• claude CLI in PATH (for LLM-backed review modes, default backend) — or --backend openrouter with OPENROUTER_API_KEY set, which needs no claude CLI
• optional: semgrep for local deterministic SAST/secrets/semi-static checks

Build

cargo build --release

# Built binary:
# target/release/codereview

If you want to keep a local debug binary:

cargo build

Core usage

All commands expect a diff patch and a spec file.

git diff > diff.patch

review (primary pipeline)

codereview review \
  --spec specs/default.toml \
  --diff diff.patch \
  --requirements requirements.md \
  --conventions conventions.md \
  --deterministic-results deterministic-results.json \
  --human-voice \
  --out runs/pr123

requirements, conventions, and deterministic-results are optional. If omitted, the tool emits explicit “not provided” sections rather than inventing assumptions.

Output (normally under runs/pr123):

• `repo