https://github.com/artttj/llama-review
Multi-model code review. Different models, different strengths, one report, reasonable actions.
https://github.com/artttj/llama-review
claude-code claude-code-plugin code-review developer-tools multi-model ollama
Last synced: 6 days ago
JSON representation
Multi-model code review. Different models, different strengths, one report, reasonable actions.
- Host: GitHub
- URL: https://github.com/artttj/llama-review
- Owner: artttj
- License: mit
- Created: 2026-05-25T04:06:43.000Z (about 2 months ago)
- Default Branch: main
- Last Pushed: 2026-06-05T23:54:46.000Z (about 1 month ago)
- Last Synced: 2026-06-06T01:20:03.655Z (about 1 month ago)
- Topics: claude-code, claude-code-plugin, code-review, developer-tools, multi-model, ollama
- Language: JavaScript
- Homepage:
- Size: 3.38 MB
- Stars: 1
- Watchers: 0
- Forks: 0
- Open Issues: 1
-
Metadata Files:
- Readme: README.md
- License: LICENSE
- Agents: AGENTS.md
Awesome Lists containing this project
README
# Llama Review
[](LICENSE)  
**[Live demo → sonto.space/llama-review](https://sonto.space/llama-review/)**
Different models have different strengths. Llama Review assigns the right model to the right task.
Five review lanes, each running on a model chosen for that domain:
- **Qwen 3.5**: frontend review with vision, layout checks, state management, accessibility
- **GLM-5.1**: backend bugs, N+1 queries, race conditions, unhandled exceptions, architecture
- **Kimi K2.6**: security across large diffs using a 262K context window
- **DeepSeek V4-Flash**: test gaps, broken assertions, coverage holes
- **MiniMax M2.7**: dead code, duplicate logic, over-engineering
Results merge, deduplicate, and rank into a single report.
## What you get
- **Critical**: security holes, data loss risks, auth bypasses
- **Needs attention**: real bugs, edge cases, performance regressions, unnecessary complexity
- **Noted**: things that passed, lanes that found nothing, low-confidence flags
- **Models Used**: which model ran each lane, effort level, and result
- **Suggested test commands**: what to run to verify the changes
- **One-paragraph summary**: ready to paste into a PR or Slack
- **Structured JSON output**: `--json` flag writes machine-readable findings to a file
- **Next Steps**: concrete subagent commands to fix each finding
Every finding has the file, line, what's broken, and how to fix it. No generic filler.
## Verdict
Every run ends with one verdict, derived from the ranked findings:
| Verdict | When | Exit code |
|---------|------|-----------|
| **BLOCK** | any critical finding | `2` |
| **REVIEW** | only needs-attention items | `0` |
| **CLEAN** | nothing actionable | `0` |
```
## Verdict: BLOCK
1 critical, 4 to review · fix before merge (exit 2)
```
Pass `--strict` to also block on needs-attention items (REVIEW becomes BLOCK, exit `2`) — useful as a CI gate. The verdict is included in `--json` output under the `verdict` key.
## Install
```
/plugin marketplace add artttj/llama-review
/plugin install llama-review
/reload-plugins
```
Then run:
```
/llama-review
```
Requires the `ollama` CLI on PATH and Node.js 18+ (bundled with Claude Code). On first run without a config file, llama-review offers to create `.llama-review.yml` from defaults.
## Usage
```
/llama-review # defaults: all lanes, origin/main, normal effort
/llama-review target=origin/staging # diff against a branch
/llama-review lanes=frontend,security # only specific lanes
/llama-review --local # use local ollama models
/llama-review --init # create .llama-review.yml from defaults
/llama-review --effort deep # 64k tokens per lane
/llama-review --jira # append a Jira comment block
/llama-review --json # write structured findings to JSON file
```
The `llama-review.mjs` script handles the full pipeline: detects changed files with `git diff`, applies exclude patterns, auto-assigns files to lanes by pattern, scales token budgets by diff size, dispatches parallel Ollama API calls with per-lane timeout and retry, handles thinking model output (falls back to `message.thinking` when `message.content` is empty), parses structured JSON output with text fallback, then merges and ranks the findings.
## Configuration
Drop a `.llama-review.yml` in your project root:
```yaml
# Global exclude patterns — strip from diff before dispatch
exclude:
- "packages/exercises/src/data/exercises/**"
- "**/seed.sql"
- "**/messages.js"
- "**/messages.po"
models:
frontend: "qwen3.5:cloud"
backend: "glm-5.1:cloud"
security: "kimi-k2.6:cloud"
tests: "deepseek-v4-flash:cloud"
simplify: "minimax-m2.7:cloud"
effort:
quick: 8000
normal: 32000
deep: 64000
local: false
# Per-lane overrides
lane_config:
backend:
timeout: 240 # seconds
retries: 1
thinking: true # increase num_predict for reasoning models
security:
timeout: 180
retries: 1
thinking: true
# Custom lanes
lanes:
infra:
files: "terraform/**, docker/**, .github/**, k8s/**"
focus: "misconfigured resources, missing secrets, unsafe defaults"
model: "kimi-k2.6:cloud"
timeout: 180
retries: 1
```
Set a lane's model to `false` to disable it. Custom lanes extend the built-in ones.
## Custom prompt templates
Override the review prompt for any lane by creating a markdown file:
- **User-level** (all projects): `~/.claude/skills/llama-review/prompts/.md`
- **Project-level**: `/.llama-review/prompts/.md`
For example, `.llama-review/prompts/security.md` replaces the built-in security prompt. User-level files take priority over project-level. If neither exists, the built-in default is used.
## Review lanes
| Lane | Files | Default model | Type | Why this model |
|------|-------|---------------|------|----------------|
| frontend | `*.tsx, *.jsx, *.vue, *.svelte, *.astro, *.css, *.scss, *.less, *.html, *.mdx, *.d.ts, *.j2, *.twig, *.blade.php, templates/` | qwen3.5:cloud | cloud | Vision + thinking + tools for UI review |
| backend | `*.php, *.py, *.rb, *.go, *.java, *.rs, *.kt, *.ts, *.js, *.cs, *.scala, *.c, *.cpp, *.h, *.hpp, *.sql, *.graphql, *.proto, *.tf` (excludes test and frontend files) | glm-5.1:cloud | cloud | Strongest code reasoning, 9.5/10 |
| security | all files | kimi-k2.6:cloud | cloud | 262K context for full attack surface review |
| tests | `*.test.*, *_test.*, *.spec.*, *_spec.*, *.phpunit.*, *.cy.*, *.e2e.*, *.integration.*, *.stories.*, tests/, __tests__/, spec/` | deepseek-v4-flash:cloud | cloud | Fast structured analysis |
| simplify | all files | minimax-m2.7:cloud | cloud | Cheap pattern matching for dead code and over-engineering |
Each lane only gets files matching its patterns. Security and simplify always get the full diff. Empty lanes are skipped. Files are auto-assigned by the script — no manual categorization needed.
## Model setup
### Cloud models (default)
Models with the `:cloud` suffix are dispatched through the standard Ollama HTTP API (`/api/chat`). The suffix is a naming convention — there is no separate cloud endpoint or proxy. Both cloud and local models hit the same `OLLAMA_HOST`.
This means cloud models only work if your Ollama server actually serves a model with that name. Plain `ollama serve` won't have them unless you've set up a proxy or custom model names. If a cloud model is unavailable, the lane fails and reports the error.
To make cloud models work, point `OLLAMA_HOST` to an Ollama-compatible server that serves them:
```bash
OLLAMA_HOST=https://your-proxy.example.com /llama-review
```
Or run locally with `--local` instead.
### Local models
Pass `--local` to strip `:cloud` suffixes and dispatch to your local Ollama instance. If a local model isn't pulled, the lane fails and reports the error.
```bash
ollama pull qwen3:8b # example
/llama-review --local
```
Recommended local models:
- `qwen3:8b`: fits in 8GB VRAM, good for quick reviews
- `deepseek-r1:14b`: reasoning-focused, good for security lanes
- `devstral:24b`: agentic coding, good for backend lanes
## License
MIT. See [LICENSE](LICENSE).