Git remembers what changed. This remembers why.
Semantic version control plus living architecture docs. Git tells you what changed — gitmind tells you why, what system structure exists today, and where the risks are.
Every commit is automatically analyzed by a local LLM (Ollama + qwen2.5-coder:7b), then merged into structured metadata and an incremental architecture model stored alongside your code. No API costs. No external services. Works with any editor.
| Feature | Status |
|---|---|
| Semantic commit summaries with local LLM analysis | ✅ Available |
| Feature-level metadata and commit history tracking | ✅ Available |
| Staleness detection for inactive features | ✅ Available |
| Incremental architecture model generation | ✅ Available |
| Auto-generated architecture, contracts, and quality findings docs | ✅ Available |
| Local dashboard for feature health and architecture findings | ✅ Available |
| Architecture Decision Record generation for new features | ✅ Available |
| Full-system professional diagrams and review dashboard expansion | Coming Soon |
| On-demand docs rebuild CLI | Coming Soon |
| Editor integrations and richer repo analysis workflows | Coming Soon |
git log --oneline
a3f92b1 fix auth
9c21d44 update middleware
3b891ef refactorThis tells you nothing useful. Which feature does middleware.py belong to? Is it still active? What's safe to delete after 6 months?
gitmind answers all of that.
git commit
↓
post-commit hook (bash)
↓
diff_reader.py → reads diff + commit message
↓
llm.py → Ollama (qwen2.5-coder:7b) analyzes the diff
↓
metadata.json → semantic commit/feature history
↓
system_model.json / findings.json / dashboard.json
↓
docs/*.md → architecture, contracts, quality findings
↓
dashboard.py → local architecture and staleness dashboard
# 1. Install Ollama and pull the model
curl -fsSL https://ollama.ai/install.sh | sh
ollama serve
ollama pull qwen2.5-coder:7b
# 2. Clone gitmind and install into your repo
git clone https://github.com/chanakyavasantha/gitmind
bash gitmind/hooks/install.sh /path/to/your/repo
# 3. Make a commit — gitmind runs automatically
cd /path/to/your/repo
git commit -m "add rate limiting to login"Output:
[gitmind] Analyzing commit...
[gitmind] Feature tracked: auth_rate_limiting
[gitmind] Added rate limiting to the login endpoint
# List all tracked features
python3 cli/query.py features
# Files touched by a feature
python3 cli/query.py files auth_middleware
# Full semantic commit history
python3 cli/query.py history
# Dead code candidates — no activity in 90 days
python3 cli/query.py stale --days 90gitmind now keeps both human-readable docs and machine-readable architecture state in the repo.
Generated markdown:
docs/generated/architecture.md— current system structure, component graph, and public API referencedocs/generated/contracts.md— module contracts for changed source filesdocs/generated/quality-findings.md— architecture risks, quality findings, and strengths
Generated JSON:
metadata/system_model.json— incremental architecture modelmetadata/findings.json— structured findings for the current codebasemetadata/dashboard.json— precomputed dashboard payload
To inspect them:
# Open the local dashboard
python3 cli/dashboard.py
# Or serve the docs site
mkdocs serveThe dashboard runs at http://localhost:4242 by default.
{
"features": {
"auth_middleware": {
"introduced": "2026-04-10T20:00:00",
"files_touched": ["src/auth.py", "src/middleware.py"],
"still_active": true,
"last_built_upon": "2026-04-10T20:00:00",
"commit_count": 7
}
},
"history": [
{
"commit_hash": "4f2a91c",
"timestamp": "2026-04-10T20:00:00",
"what_changed": "Added JWT expiry validation",
"why_it_likely_changed": "Security requirement to reject expired tokens",
"feature_name": "auth_middleware",
"impact": "All protected routes now reject expired tokens",
"files_touched": ["src/auth.py"]
}
]
}Tested on the gitmind repo itself during initial development:
| Commit | Hook result | Finding |
|---|---|---|
| Initial scaffold | No diff found — skipped |
Root commit has no HEAD~1 — fixed with git show HEAD fallback |
| First-commit fix | JSON parse error | deepseek-coder outputs // comments — fixed with format: "json" |
| Format enforcement | Parse error still | Field type mismatches — fixed with _extract_json() + _coerce() |
| Robust parser | Feature tracked: gitmind | First clean end-to-end run |
| Model upgrade | Clean names on all commits | Switched to qwen2.5-coder:7b — no hallucinations, clean JSON |
Current model: qwen2.5-coder:7b — significantly better feature naming and JSON reliability than deepseek-coder.
gitmind/
├── core/
│ ├── architecture/ # extraction, analyzers, contracts, ADRs, renderers
│ ├── diff_reader.py # reads git diffs, handles first-commit edge case
│ ├── llm.py # Ollama interface with JSON enforcement + coercion
│ ├── metadata.py # read/write metadata.json
│ ├── system_model.py # compatibility wrapper for architecture model
│ ├── doc_generator.py # compatibility wrapper for doc generation
│ └── engine.py # orchestrates commit-time updates
├── hooks/
│ ├── post-commit # bash hook installed into target repo
│ └── install.sh # copies gitmind into any repo, installs hook
├── cli/
│ ├── query.py # feature/history query tool
│ ├── dashboard.py # local dashboard server
│ └── dashboard.html # dashboard UI
├── metadata/ # generated architecture model + dashboard payload
├── docs/ # generated and hand-written docs
├── tests/
│ └── test_core.py # pytest suite
└── requirements.txt
| Phase | Goal | Status |
|---|---|---|
| 1 | Core hook — diff → LLM → metadata.json | ✅ Done |
| 2 | Feature grouping across commits | Planned |
| 3 | Staleness detection + removal planner | Planned |
| 4 | Team support — Husky, large diff chunking | Planned |
| 5 | SQLite + ChromaDB semantic search | Planned |
| 6 | VS Code extension + Claude Code integration | Planned |
| 7 | Open source launch | Planned |
Full roadmap: docs/roadmap.md
- Free — no API costs per commit
- Private — your code never leaves your machine
- Offline — works without internet after model pull
See CONTRIBUTING.md.
pip install pytest black
pytest tests/ -v
black core/ cli/MIT