Commitsmith

Forge Conventional Commits messages from your git diff with one click — using any OpenAI-compatible API, a local Ollama model, or llama.cpp. No telemetry, no server, no code indexing.

Works in VS Code and Cursor.

Features

• ✨ One click in the Source Control title bar generates a message from your diff.
• 🔒 Local-first. Point it at Ollama or llama.cpp and nothing leaves your machine. No API key required for local backends.
• 🧠 Grounded & truthful. The prompt is hardened against hallucination and treats the diff as data, not instructions (prompt-injection resistant).
• 🪙 Token-budgeted. Diffs are measured in tokens (tiktoken), with more budget given to larger files and whole hunks kept intact. Oversized diffs auto-fall back to a two-pass summarize-then-synthesize strategy.
• 🤔 Reasoning-model aware. Handles "thinking" models (qwen3, gemma-thinking, deepseek-r1, gpt-oss, …): disables chain-of-thought where possible, retries when a model spends its whole budget reasoning, and strips any stray <think>/harmony output.
• 🧹 Noise filtering. Lockfiles, generated files, maps and binaries are dropped from the prompt (still listed in the change summary).
• 🔐 Secure key storage. API keys are stored in VS Code SecretStorage (encrypted at rest on your machine). Never sent to third parties.
• 📊 Smart error handling. Clear, actionable error messages for rate limits, auth failures, network issues, and server errors.

Usage

1. Stage your changes in the Source Control view (or leave them unstaged).
2. Click the ✨ button in the Source Control title bar (or run Commitsmith: Generate Commit Message from the Command Palette).
3. Review the generated Conventional Commits message in the commit message box.
4. Edit if needed, then commit as usual.

That's it. No configuration required for basic use — the default is OpenAI with gpt-4o-mini.

How it works

Click the ✨ button in the Source Control title bar. Commitsmith:

1. Reads your diff (staged first, falling back to unstaged).
2. Builds grounded context — a change-summary table (files, status, +/- counts) placed before and after the code, so grounding survives any backend truncation.
3. Budgets the diff in tokens, allocating more to larger files and keeping whole hunks.
4. Sends one request to your configured model and writes the result into the commit message box for you to review and commit.

For large diffs (above commitsmith.largeDiffThreshold tokens, default 7000), Commitsmith automatically uses a two-pass strategy:

• Pass 1: Summarize each changed file into a single line.
• Pass 2: Synthesize those summaries into one commit message.

This ensures the generated message captures the overall change even when the diff is too large for a single request.

It is a pure, stateless client — no server, no vector DB, no code indexing. (That decision is backed by research; see docs/superpowers/specs/.)

Install

From a packaged .vsix (replace 0.1.2 with the latest version):

code --install-extension commitsmith-0.1.2.vsix      # VS Code
cursor --install-extension commitsmith-0.1.2.vsix    # Cursor

Or search for Commitsmith in the VS Code/Cursor extensions marketplace.

Then reload the window.

Setup

Run Commitsmith: Select API Provider from the Command Palette (or the Source Control ⋯ menu) to choose between Ollama, llama.cpp, and an OpenAI-compatible endpoint — or set it in Settings directly.

Ollama (local, private — no API key)

// settings.json
"commitsmith.api": "ollama",
"commitsmith.baseUrl": "http://localhost:11434",
"commitsmith.model": "qwen2.5-coder:3b"   // or any model you've pulled

Commitsmith calls Ollama's /api/chat with num_ctx (default 8192) so large diffs are not silently truncated — a common failure of the OpenAI-compat /v1 path on Ollama.

llama.cpp (local, private — optional API key)

Run llama-server (from llama.cpp) with your model:

llama-server -m path/to/model.gguf --port 8080

// settings.json
"commitsmith.api": "llama.cpp",
"commitsmith.baseUrl": "http://localhost:8080/v1",
"commitsmith.model": "your-model-name"

llama.cpp exposes an OpenAI-compatible /v1/chat/completions endpoint — no API key required by default. For authenticated setups, run llama-server with --api-key and configure it via Commitsmith: Set API Key.

OpenAI-compatible (OpenAI, OpenRouter, LM Studio, Open WebUI, …)

"commitsmith.api": "openai",
"commitsmith.baseUrl": "https://api.openai.com/v1",
"commitsmith.model": "gpt-4o-mini"

Run Commitsmith: Set API Key to store your key securely in VS Code SecretStorage. For Open WebUI the base URL is https://your-host/api. Local OpenAI-compatible servers (LM Studio, etc.) need no key.

Settings

Commands

• Commitsmith: Generate Commit Message — also the ✨ Source Control button.
• Commitsmith: Select API Provider — pick Ollama, llama.cpp, or OpenAI-compatible (dropdown).
• Commitsmith: Set API Key / Clear API Key.

Development

npm install
npm run lint         # ESLint (type-checked, strict)
npm run typecheck    # tsc --noEmit
npm test             # unit tests (vitest)
npm run build        # bundle with esbuild → dist/extension.js
npm run package      # produce a .vsix

Press F5 in VS Code to launch the Extension Development Host.

Live integration check (optional)

npm run integration drives the real pipeline against live endpoints. Everything is env-driven, and each test is skipped (not failed) when its endpoint is unset or unreachable:

# Ollama
OLLAMA_HOST=http://localhost:11434 OLLAMA_MODEL=llama3.2 npm run integration

# llama.cpp (test runs only when LLAMACPP_HOST is set)
LLAMACPP_HOST=http://localhost:8080/v1 LLAMACPP_MODEL=default npm run integration

# OpenAI-compatible (test runs only when OPENAI_KEY is set)
OPENAI_BASE=https://api.openai.com/v1 OPENAI_MODEL=gpt-4o-mini OPENAI_KEY=sk-... npm run integration

# Reasoning models (opt-in)
REASONING=1 REASONING_MODELS=qwen3:latest,gpt-oss:20b npm run integration

On Windows PowerShell, set the variables with $env:NAME = "value" before the command.

Troubleshooting

"No API key set for a remote OpenAI-compatible endpoint"

Run Commitsmith: Set API Key from the Command Palette, or configure a local provider (Ollama or llama.cpp) that doesn't require a key.

"Model returned an empty response"

The model may have spent its entire token budget reasoning. Try:

• Using a non-reasoning model (e.g., gpt-4o-mini instead of gpt-oss)
• Increasing commitsmith.maxTokens (default 200)
• Reducing commitsmith.maxDiffTokens to shrink the prompt

"Rate limit exceeded (HTTP 429)"

Your API provider is throttling requests. Wait a moment and try again, or reduce request size by lowering commitsmith.maxDiffTokens.

"Could not resolve hostname" or "Connection refused"

Check that commitsmith.baseUrl points to a running server. For Ollama, verify it's started with ollama serve. For llama.cpp, run llama-server -m model.gguf --port 8080.

"Server error (HTTP 5xx)"

The remote endpoint is experiencing issues. Check its logs or try again later.

Two-pass strategy triggered unexpectedly

The diff exceeded commitsmith.largeDiffThreshold tokens (default 7000). To avoid this, lower commitsmith.maxDiffTokens or split large changes into smaller commits.

Limitations

• Single-commit scope. Commitsmith generates a message for one commit at a time. For multi-commit workflows, run it once per commit.
• English-first. The built-in prompt is optimized for English commit messages. For other languages, override commitsmith.systemPrompt with a multilingual variant.
• No code understanding. Commitsmith analyzes the diff, not the codebase. It doesn't understand semantics beyond what's visible in the diff.
• Token counting is approximate. Uses cl100k_base (tiktoken) as a general-purpose tokenizer. For non-OpenAI models, this is an approximation — far better than byte counting but not perfect.

Contributing

See CONTRIBUTING.md.

Development Workflow

npm install
npm run watch          # Watch mode — rebuilds on save
# Press F5 in VS Code to launch Extension Development Host

The extension is bundled with esbuild for fast builds. Tests run with vitest and lint with ESLint.

Running Integration Tests

npm run integration drives the real pipeline against live endpoints. Each test is skipped (not failed) when its endpoint is unset or unreachable:

# Ollama
OLLAMA_HOST=http://localhost:11434 OLLAMA_MODEL=llama3.2 npm run integration

# llama.cpp
LLAMACPP_HOST=http://localhost:8080/v1 LLAMACPP_MODEL=default npm run integration

# OpenAI-compatible
OPENAI_BASE=https://api.openai.com/v1 OPENAI_MODEL=gpt-4o-mini OPENAI_KEY=sk-... npm run integration

License

Apache License 2.0 © 2026 Anatoly Lazarev. See NOTICE.