Snif

Context-Aware Code Review Agent

Snif is a code review agent that understands your codebase. It indexes the repository structure, generates semantic summaries, and uses that context to review pull requests with specific, evidenced, and actionable findings.

It ships as a single Rust binary for CI pipelines. On each commit, it builds a structural graph, generates LLM summaries, and embeds them for search. On each pull request, it retrieves related code using structural, semantic, and keyword methods, then runs a single review call and posts only findings that survive aggressive filtering.

---

Getting Started

Prerequisites

Rust toolchain (1.70+). An OpenAI-compatible LLM provider endpoint and API key.

Build

cargo build --release

The binary is at target/release/snif.

Configure

Copy the example configuration to your repository root:

cp .snif.json.example .snif.json

Edit .snif.json to set your LLM provider. At minimum you need the endpoint and model names:

{
  "model": {
    "review_model": "gpt-4o",
    "summary_model": "gpt-4o-mini",
    "endpoint": "https://api.openai.com/v1"
  }
}

Set the API key as an environment variable:

export SNIF_API_KEY=your-api-key

See Configuration for all options, provider examples, and environment variables.

Index

Build the repository index. This parses source files, builds the structural graph, generates LLM summaries, and computes vector embeddings.

snif index --path /path/to/repo

Use --full for a clean rebuild. Without it, indexing is incremental.

Review

Review a code change using a local diff file:

git diff main > /tmp/change.diff
snif review --path /path/to/repo --diff-file /tmp/change.diff

Review a GitHub pull request directly:

GITHUB_TOKEN=your-token snif review --repo owner/repo --pr 123

Findings are printed to stdout as JSON. When using --repo and --pr, findings are also posted as inline review comments on the pull request.

Evaluate

Run the evaluation harness against benchmark fixtures:

snif eval --fixtures ./fixtures/

The command exits with code 0 if quality gates pass and code 1 if they fail.

Clean

Remove all local runtime data (index database, embedding cache, feedback store):

snif clean

This does not touch your source code or configuration — only Snif's generated data.

Examples

Documentation

• Configuration setup, config format, CLI reference
• CI Integration GitHub Actions, GitLab CI, Docker, generic CI
• Bot Identity GitHub App and GitLab bot user setup
• Testing local testing and evaluation
• Product scope, delivery plan, success criteria
• Architecture system design, modules, data flows