kagura-code-reviewer

kagura-code-reviewer

Pro-grade code review on your git diff with zero Anthropic billing. A local (or cloud) Ollama model fans out multi-angle finders, adversarially majority-vote verifies each one, dedups, and returns a ranked Markdown/JSON report with a green / yellow / red verdict you can gate CI on.

Cost-free, Ollama-powered code review for Claude Code. The review "brain" runs on an Ollama model (cloud or local), so reviews consume zero Anthropic billing. A companion slash command (/kagura-code-reviewer) integrates with Kagura Memory: the outer Claude session retrieves past conventions and findings, passes them as context to the CLI, then writes durable knowledge back after the review. The CLI itself is a self-contained tool — it does not call Kagura Memory directly.

How it works

Claude Code (outer session)
   │  /kagura-code-reviewer slash command
   │  1. Recalls past findings from Kagura Memory (trust_tier: trusted filter)
   │  2. Writes assembled context to /tmp/kcr-ctx.md
   │  3. Invokes CLI ──────────────────────────────────────────────────────┐
   │                                                                        │
   └── Presents report to user ←── 5. Writes durable knowledge back        │
                                                                            ▼
                                                         kagura-code-reviewer CLI
                                                            │  git diff (base...HEAD)
                                                            │  sandboxed repo tools
                                                            │    read_file / grep / git
                                                            ▼
                                                         Ollama (local or cloud)
                                                            │  agentic review loop
                                                            ▼
                                                         Markdown / JSON report
                                                         exit 0 = clean
                                                         exit 1 = blocking issues

Install

pip install kagura-code-reviewer

System prerequisites

These are not installed by pip — you must set them up separately:

• Ollama daemon running with at least one model pulled.
  Default cloud alias uses qwen3-coder:480b-cloud; default local alias uses qwen2.5-coder:7b.
  Pull with: ollama pull qwen2.5-coder:7b
• claude CLI — required only for the /kagura-code-reviewer slash-command workflow.
  Install via: npm install -g @anthropic-ai/claude-code

Quickstart

# Review current branch vs main (Markdown to stdout)
kagura-code-reviewer --base main

# Write report as JSON to a file
kagura-code-reviewer --base main --format json --out report.json

# Use the local model alias (faster, smaller)
kagura-code-reviewer --local

# Limit review to specific paths
kagura-code-reviewer --base main --paths src/foo.py --paths tests/test_foo.py

# Review a GitHub PR by URL (open or closed/merged) — run inside a clone of the PR's repo
kagura-code-reviewer --pr https://github.com/owner/repo/pull/123

# Stream review progress to stderr (so a slow local run doesn't look frozen)
kagura-code-reviewer --base main -v

Progress (-v/--verbose): streams phase progress (finders → dedup → per-candidate verify i/N → final count) to stderr so a slow local-model run doesn't look frozen. stdout / --out (the md/json report) stays clean and pipe-safe, so --format json --out is unaffected.