The commit message generator that actually understands your code.

Most tools in this space pipe raw git diff to an LLM and hope for the best. CommitBee parses your code with tree-sitter, maps diff hunks to symbol spans, and gives the LLM structured semantic context — producing fundamentally better commit messages, especially for complex multi-file changes.

CommitBee uses tree-sitter to parse both the staged and HEAD versions of every changed file — in parallel across CPU cores. It extracts 10 symbol types (functions, methods, structs, enums, traits, impls, classes, interfaces, constants, type aliases) with full signatures — the LLM sees pub fn connect(host: &str, timeout: Duration) -> Result<Connection>, not just "Function connect." Modified symbols show old → new signature diffs so the LLM understands exactly what changed. Cross-file relationships are detected automatically: if validator.rs calls parse() and both changed, the prompt says so. Symbols are tracked in three states: added, removed, and modified-signature.

Supported languages: Rust, TypeScript, JavaScript, Python, Go, Java, C, C++, Ruby, C# — all enabled by default, individually toggleable via Cargo feature flags. Files in other languages still get full diff context — just without symbol extraction.

Before the LLM generates anything, CommitBee computes deterministic evidence from your code and encodes it as hard constraints in the prompt:

• Bug-fix evidence in the diff → fix. No bug evidence → the LLM can't call it a fix.
• Formatting-only changes (whitespace, import reordering) → style. Detected both heuristically and via per-symbol whitespace classification.
• Dependency-only changes → chore. Always.
• Public API removed → breaking change flagged automatically.
• MSRV bumps, engines.node, requires-python changes → metadata-aware breaking detection.

Commit types are driven by code analysis, not LLM guesswork. The prompt includes computed EVIDENCE flags, CONSTRAINTS the model must follow, the primary change for subject anchoring, a character budget for the subject line, and anti-hallucination rules with negative examples.

Every generated message passes through a 7-rule validation pipeline:

1. Fix requires evidence — no bug comments, no fix type
2. Breaking change detection — removed public APIs must be flagged
3. Anti-hallucination — breaking change text can't copy internal field names
4. Mechanical changes must use style
5. Dependency-only changes must use chore
6. Subject specificity — rejects generic messages like "update code" or "improve things"
7. Subject length — enforces the 72-character first line limit

If any rule fails, CommitBee appends targeted correction instructions and re-prompts the LLM — up to 3 attempts, re-validating after each. The final output goes through a sanitizer that strips thinking blocks, extracts JSON from code fences, removes conversational preambles, and wraps the body at 72 characters. You get a clean, spec-compliant conventional commit or a clear error — never a silently mangled message.

When your staged changes mix independent work (a bugfix in one module + a refactor in another), CommitBee detects it and offers to split them into separate, well-typed commits. The splitter uses diff-shape fingerprinting combined with Jaccard similarity on content vocabulary — files are grouped by the actual shape and language of their changes, not just by directory. Symbol dependency merging keeps related files together even when their diff shapes differ: if foo() is removed from one file and added in another, they stay in the same commit.

⚡ Commit split suggested — 2 logical change groups detected:

  Group 1: feat(llm)  [2 files]
    [M] src/services/llm/anthropic.rs (+20 -5)
    [M] src/services/llm/openai.rs (+8 -3)

  Group 2: fix(sanitizer)  [1 file]
    [M] src/services/sanitizer.rs (+3 -1)

? Split into separate commits? (Y/n)

┌─────────┐    ┌──────────┐    ┌────────────┐    ┌──────────┐    ┌───────────┐    ┌─────────┐
│  Stage  │ →  │   Git    │ →  │ Tree-sitter│ →  │  Split   │ →  │  Context  │ →  │   LLM   │
│ Changes │    │  Service │    │  Analyzer  │    │ Detector │    │  Builder  │    │Provider │
└─────────┘    └──────────┘    └────────────┘    └──────────┘    └───────────┘    └─────────┘
                    │                │                 │                │               │
               Staged diff      Symbol spans     Group files      Budget-aware     Commit message
               + file list      (functions,      by module,       prompt with      (conventional
                                classes, etc.)   suggest split    semantic context    format)

• 🏠 Local-first — Ollama by default. Your code never leaves your machine. No API keys needed.
• 🔒 Secret scanning — 24 built-in patterns across 13 categories (cloud keys, AI/ML tokens, payment, database, crypto). Add custom patterns or disable built-ins via config.
• ⚡ Streaming — Real-time token display from all 3 providers (Ollama, OpenAI, Anthropic) with Ctrl+C cancellation.
• 📊 Token budget — Smart truncation that prioritizes the most important files within ~6K tokens.
• 🎯 Multi-candidate — Generate up to 5 messages and pick the best one interactively.
• 🪝 Git hooks — prepare-commit-msg hook with TTY detection for safe non-interactive fallback.
• 🔍 Prompt debug — --show-prompt shows exactly what the LLM sees. Full transparency.
• 🩺 Doctor — commitbee doctor checks config, connectivity, and model availability.
• 🐚 Shell completions — bash, zsh, fish, powershell via commitbee completions.
• ⚙️ 5-level config — Defaults → project .commitbee.toml → user config → env vars → CLI flags.
• 🦀 Single binary — ~18K lines of Rust. Compiles to one static binary with LTO. No runtime dependencies.
• 🧪 367 tests — Unit, snapshot, property (proptest for never-panic guarantees), and integration (wiremock).

cargo install commitbee

git clone https://github.com/sephyi/commitbee.git
cd commitbee
cargo build --release

The binary will be at ./target/release/commitbee.

• Rust 1.94+ (edition 2024)
• Ollama running locally (default provider) — Install Ollama
• A model pulled in Ollama (recommended: qwen3.5:4b)

ollama pull qwen3.5:4b

# Stage your changes
git add src/feature.rs

# Generate and commit interactively
commitbee

# Preview without committing
commitbee --dry-run

# Auto-confirm and commit
commitbee --yes

# See what the LLM sees
commitbee --show-prompt

That's it. CommitBee works with zero configuration if Ollama is running locally.

If CommitBee saves you time, consider sponsoring the project 💛

• Full Guide — configuration, providers, splitting, validation, troubleshooting
• PRD & Roadmap — product requirements and future plans

Run commitbee init to create a config file. Works out of the box with zero config if Ollama is running locally.

See Configuration for the full config reference, environment variables, and layering priority.

commitbee [OPTIONS] [COMMAND]

CommitBee scans all content before it's sent to any LLM provider with 24 built-in patterns across 13 categories:

• ☁️ Cloud providers — AWS access/secret keys, GCP service accounts & API keys, Azure storage keys
• 🤖 AI/ML — OpenAI, Anthropic, HuggingFace tokens
• 🔧 Source control — GitHub (PAT, fine-grained, OAuth), GitLab tokens
• 💬 Communication — Slack tokens & webhooks, Discord webhooks
• 💳 Payment & SaaS — Stripe, Twilio, SendGrid, Mailgun keys
• 🗄️ Database — MongoDB, PostgreSQL, MySQL, Redis, AMQP connection strings
• 🔐 Cryptographic — PEM private keys, JWT tokens
• 🔑 Generic — API key assignments, quoted/unquoted secrets
• ⚠️ Merge conflict detection — Prevents committing unresolved conflicts

Add custom patterns or disable built-ins in your config:

custom_secret_patterns = ["CUSTOM_KEY_[a-zA-Z0-9]{32}"]
disabled_secret_patterns = ["Generic Secret (unquoted)"]

The default provider (Ollama) runs entirely on your machine. No data leaves your network unless you explicitly configure a cloud provider.

cargo test   # 367 tests — unit, snapshot (insta), property (proptest), integration (wiremock)

See Testing Strategy for the full breakdown.

See CHANGELOG.md for the full version history.

Current: v0.5.0 Beyond the Diff — Full signature extraction, semantic change classification, cross-file connections, security hardening, and 36-fixture eval harness.

Contributions are welcome! By contributing, you agree to the Contributor License Agreement — you'll be asked to sign it when you open your first pull request.

If you find CommitBee useful, consider sponsoring my work — it helps keep the project going.

This project is dual-licensed under AGPL-3.0-only and a commercial license. See LICENSE for details.

REUSE compliant — every file carries SPDX headers.

Copyright 2026 Sephyi