spec-agent 0.3.0

Auto-generate an Obsidian knowledge wiki from every git push, powered by Claude

spec-agent

Auto-generate an Obsidian knowledge wiki from every git push — powered by Claude, Ollama, or Gemini.

Every time you push code, spec-agent reads the diff, classifies the change, and writes a structured spec document into your Obsidian vault. Over time, your vault becomes a living, visually-navigable graph of everything you've ever built.

Inspired by Andrej Karpathy's LLM Wiki pattern: the LLM acts as a compiler (raw commits → structured wiki), not a retrieval engine. No RAG, no embeddings — just well-organized markdown with [[wikilinks]] that Obsidian renders as a knowledge graph.

---

How It Works

git push
  └── ~/.git-hooks/pre-push         ← global hook, fires on every repo
        └── spec-agent run           ← Python CLI
              └── LLM (tool-use)     ← agentic loop (Anthropic / Ollama / Gemini)
                    ├── search_wiki       ← finds related existing pages
                    ├── read_wiki_file    ← reads context before updating
                    ├── write_wiki_file   ← creates or updates spec
                    └── update_index      ← appends row to index.md

The agent:

1. Classifies the commit type from the message prefix (feat→feature, fix→bug, refactor→refactor, chore→skip)
2. Searches your vault for related pages to link to (as many queries as needed)
3. Writes an adaptive spec using the appropriate template
4. Updates index.md — the master log that Claude reads at session start

Chore commits are skipped. Bot branches (dependabot/*, renovate/*) are skipped. Tiny commits (below a configurable character threshold) are skipped.

---

Features

• Zero friction — fires automatically on every git push, no developer action needed
• Cold-start KB bootstrap — spec-agent init-repo scans an existing codebase and writes a structured knowledge base into the vault before any pushes have occurred
• Pluggable LLM backends — use Anthropic (cloud), Ollama (local/free), or Gemini (free tier)
• Adaptive templates — selects the right format based on commit type
  • feature → full spec (summary, problem, implementation, files, open questions)
  • bug → short report (root cause, fix applied)
  • refactor → brief note (what changed, why, before/after)
  • arch → ADR format (context, decision, consequences, alternatives)
• Accurate [[wikilinks]] — searches vault for existing pages before writing, so links are real
• Living index — index.md is a table of every spec ever written; share it with Claude at session start to give it full project memory
• Obsidian graph — concepts referenced by many specs become visual hubs after 10+ specs
• Works on every repo — one global hook installation covers all your projects
• Per-repo opt-out — exclude specific repos from the global hook with a single command

---

Prerequisites

• Python 3.11+
• Obsidian — download here (free)
• Git 2.9+ (for core.hooksPath support)
• One of the following LLM backends (you choose during setup):
  • Anthropic — API key required
  • Ollama — free, runs locally; download here
  • Gemini — free tier available; API key required

---

Installation

From PyPI (recommended)

pip install spec-agent

From source

git clone https://github.com/visheshchaitanya/spec-agent.git
cd spec-agent
pip install -e .

For Gemini backend support, install the extra:

pip install "spec-agent[gemini]"

---

Setup

1. Initialize your vault

spec-agent init --vault ~/Documents/dev-wiki

This creates the vault directory structure and writes a default ~/.spec-agent/config.yaml.

Then open ~/Documents/dev-wiki as a vault in Obsidian (File → Open vault as folder).

2. Choose your LLM backend

spec-agent configure

This interactive command walks you through picking a backend and saves your settings — no YAML editing needed:

spec-agent configure

Choose an LLM backend:

  anthropic  Cloud — best quality, requires ANTHROPIC_API_KEY
  ollama     Local — free, runs on your machine (no API key)
  gemini     Cloud — free tier available, requires GEMINI_API_KEY

Backend [anthropic]: ollama

Popular Ollama models:
  qwen2.5:7b   — fast, good reasoning, ~4 GB
  qwen2.5:14b  — better quality, ~8 GB
  gemma3       — Google Gemma 3 (12B), ~7 GB
  llama3.2     — Meta Llama 3.2, ~2 GB

Ollama server URL [http://localhost:11434]:
Model name [qwen2.5:7b]:

✓ Config saved → backend: ollama, model: qwen2.5:7b

To install Ollama and pull the model:
  # 1. Install Ollama: https://ollama.com/download
  # 2. Pull your chosen model:
  ollama pull qwen2.5:7b

Setting API keys (Anthropic / Gemini)

API keys are read from environment variables — they are never written to disk.

Anthropic:

export ANTHROPIC_API_KEY="sk-ant-..."
# Add to ~/.zshrc or ~/.bashrc to make permanent

Gemini (free tier at aistudio.google.com):

export GEMINI_API_KEY="AIza..."
# Add to ~/.zshrc or ~/.bashrc to make permanent

3. Install the global git hook

spec-agent install-hook

This creates ~/.git-hooks/pre-push and sets git config --global core.hooksPath ~/.git-hooks. The hook fires on every push in every repository on your machine.

To disable at any time:

spec-agent uninstall-hook

4. Push anything to test

cd ~/any-repo
git push
# → spec-agent fires in the background
# → spec appears in ~/Documents/dev-wiki

---

LLM Backends

Backend	Cost	Quality	Setup
anthropic	~$0.003–0.015 / push	Best	ANTHROPIC_API_KEY
ollama	Free (local compute)	Good (model-dependent)	Install Ollama + pull model
gemini	Free tier / ~$0.001	Good	GEMINI_API_KEY

Running with Ollama (local, zero cost)

1. Install Ollama
2. Pull a model:

   ollama pull qwen2.5:7b   # fast, good reasoning
   # or
   ollama pull gemma3        # Google Gemma 3
   

3. Configure spec-agent:

   spec-agent configure     # choose "ollama", set model name
   

Ollama runs the model locally on your GPU/CPU — no data leaves your machine.

Running with Gemini (cheap cloud alternative)

1. Get a free API key at aistudio.google.com
2. Install the extra dependency:

   pip install "spec-agent[gemini]"
   

3. Configure:

   export GEMINI_API_KEY="AIza..."
   spec-agent configure     # choose "gemini"
   

---

Configuration

Config lives at ~/.spec-agent/config.yaml. Run spec-agent configure to update it interactively, or edit it directly:

vault_path: ~/Documents/dev-wiki
model: claude-sonnet-4-6
ignored_repos: []
ignored_branches:
  - dependabot/*
  - renovate/*
min_commit_chars: 50
# LLM backend
llm_backend: anthropic        # anthropic | ollama | gemini
ollama_url: http://localhost:11434
ollama_model: qwen2.5:7b
gemini_model: gemini-2.0-flash

Key	Default	Description
vault_path	~/Documents/dev-wiki	Absolute path to your Obsidian vault
model	claude-sonnet-4-6	Anthropic model (used when llm_backend: anthropic)
ignored_repos	[]	Exact repo names to never process
ignored_branches	[dependabot/*, renovate/*]	Glob patterns — matching branches are skipped
min_commit_chars	50	Skip pushes where total commit message length is below this
llm_backend	anthropic	Which backend to use: anthropic, ollama, or gemini
ollama_url	http://localhost:11434	Ollama server URL (change if running on a remote machine)
ollama_model	qwen2.5:7b	Ollama model name (must be pulled first via ollama pull)
gemini_model	gemini-2.0-flash	Gemini model name

---

Vault Structure

~/Documents/dev-wiki/
├── index.md                    ← master log — give this to Claude at session start
├── features/
│   ├── alert-ingestion.md
│   └── auth-system.md
├── bugs/
│   ├── fix-status-migration.md
│   └── fix-null-pointer.md
├── refactors/
│   └── extract-auth-middleware.md
├── concepts/                   ← auto-created graph hubs when first referenced
│   ├── clickhouse.md
│   └── jwt.md
└── projects/
    └── my-app/
        ├── overview.md         ← created by init-repo
        ├── AuthService.md
        └── ApiGateway.md

index.md — your project memory

index.md is a markdown table of every spec ever written:

# Dev Wiki — Index

| Date | Type | Title | Project | Link |
|------|------|-------|---------|------|
| 2026-04-07 | bug | Fix status migration | my-app | [[bugs/fix-status-migration]] |
| 2026-04-05 | feature | Alert ingestion pipeline | my-app | [[features/alert-ingestion]] |

Paste the contents of index.md at the start of any Claude session — Claude immediately knows everything you've built across all projects.

---

Giving Claude context at session start

Paste this into your Claude session to give it full project memory:

Here is my dev wiki index — everything I've built:

<paste contents of ~/Documents/dev-wiki/index.md>

I'm working on <task>. Based on the index, what related specs should I look at?

Claude can then ask you to paste specific spec files for deeper context.

---

CLI Reference

spec-agent [COMMAND] [OPTIONS]

Commands:
  configure        Interactively set up your LLM backend (recommended)
  init             Initialize vault directory and write config
  init-repo        Bootstrap a knowledge base for the current repo
  install-hook     Install global git pre-push hook
  uninstall-hook   Remove the global git pre-push hook
  opt-out          Exclude current repo from the global hook
  opt-in           Re-include current repo in the global hook
  run              Run the agent (called automatically by git hook)
  config-get       Read a config value (used internally by hook)

spec-agent configure

Interactive setup wizard. Prompts for backend choice and backend-specific settings, then saves to ~/.spec-agent/config.yaml. No options required.

spec-agent init

Options:
  --vault TEXT  Path to Obsidian vault directory (required)

spec-agent init-repo

Options:
  --deep    Full breadth-first scan (reads up to 40 files) [default: shallow]
  --force   Update existing KB without prompting

Run from inside a git repository. Writes KB docs to projects/<repo-name>/ in the vault.

spec-agent opt-out

No options. Auto-detects repo name from git rev-parse. Adds the current repo to ignored_repos in config — the global hook will skip it on future pushes.

spec-agent opt-in

No options. Removes the current repo from ignored_repos, re-enabling spec generation on push.

spec-agent install-hook

No options. Installs ~/.git-hooks/pre-push and sets the global git hooks path.

spec-agent uninstall-hook

No options. Removes ~/.git-hooks/pre-push, disabling automatic spec generation on push.

spec-agent run

Options:
  --repo TEXT        Repository name (required)
  --branch TEXT      Branch that was pushed (required)
  --messages TEXT    Newline-separated commit messages (required)
  --diff-file TEXT   Path to temp file containing the git diff (required)
  --config TEXT      Path to config.yaml [default: ~/.spec-agent/config.yaml]

---

Manual run (without pushing)

You can run the agent manually against any repo:

DIFF_FILE=$(mktemp /tmp/spec-agent-diff.XXXXXX)
git diff HEAD~1..HEAD | head -c 50000 > "$DIFF_FILE"

spec-agent run \
  --repo "$(basename $(pwd))" \
  --branch "$(git rev-parse --abbrev-ref HEAD)" \
  --messages "$(git log HEAD~1..HEAD --format='%s')" \
  --diff-file "$DIFF_FILE"

---

Bootstrap an existing repo

If you install spec-agent on a repo that already has history, the vault starts empty and the push agent has no context. Fix this with init-repo:

cd ~/my-project
spec-agent init-repo           # shallow scan — reads README, configs, entry points
spec-agent init-repo --deep    # breadth-first scan of up to 40 files

This writes a projects/<repo-name>/overview.md and individual component docs into the vault. Future pushes immediately find relevant KB pages via search_wiki.

To refresh the KB after significant changes:

spec-agent init-repo --force   # re-runs and updates existing docs; only sends changed files to LLM

---

Per-repo opt-out

To stop spec-agent from firing on a specific repo (useful for personal/dotfiles repos):

cd ~/my-private-repo
spec-agent opt-out    # auto-detects repo name, adds to ignored_repos

To re-enable:

spec-agent opt-in

Or manage ignored_repos directly in ~/.spec-agent/config.yaml:

ignored_repos:
  - my-private-repo
  - dotfiles

---

Development

Requirements

git clone https://github.com/visheshchaitanya/spec-agent.git
cd spec-agent
pip install -e ".[dev]"

Run tests

pytest

Run tests with coverage

pytest --cov=spec_agent --cov-report=term-missing

All tests use temporary directories — no vault or API key required.

---

Architecture

Pluggable LLM backends

The agent loop in agent.py is backend-agnostic. A backend is selected via get_backend(cfg) and implements a common interface. The loop runs until stop_reason == "end_turn", with a hard cap of 20 iterations to prevent runaway execution:

LLMBackend
  ├── AnthropicBackend   uses anthropic SDK, native tool-use
  ├── OllamaBackend      HTTP POST to /api/chat, OpenAI-compatible tool format
  └── GeminiBackend      uses google-genai SDK, function declarations

Each backend normalizes its response to a ChatResponse(stop_reason, text, tool_calls) so the agent loop is identical regardless of which LLM is running.

Tool-using agent loop

backend.chat(system, messages, tools)
  → stop_reason == "tool_use"
      → dispatch tool, collect results
      → backend.make_tool_results_messages(tool_calls, results)
      → append to messages, loop  (max 20 iterations)
  → stop_reason == "end_turn"
      → done
  → API error (429 / 5xx / timeout)
      → retry with backoff (max 3 attempts)
      → abort cleanly if all retries exhausted

Tools

Tool	Purpose
search_wiki	Full-text search across vault using grep -r — finds related pages for wikilinks
read_wiki_file	Reads existing spec file — enables update mode instead of duplicate creation
write_wiki_file	Writes markdown to vault; mode=update appends a dated changelog section
update_index	Appends row to index.md master log

Diff safety

Git diffs larger than 50,000 characters are truncated before being passed to the agent. The diff is passed via a temp file (not shell arguments) to avoid escaping issues with special characters.

---

Roadmap

• Auto-detect Ollama installation and pull model during spec-agent configure (#enhancement)
• GitHub webhook / cloud daemon (replace local hook with HTTP POST)
• Multi-agent parallel processing (Classifier + Writer + Linker)
• Obsidian Dataview dashboards
• Per-project vault paths
• Slack/email notifications on spec creation

---

Contributing

Pull requests welcome. Please:

1. Fork the repo and create a branch from main
2. Add tests for any new behavior
3. Ensure pytest passes
4. Open a pull request

---

License

MIT — © 2026 Vishesh Chaitanya