Faut-il créer votre propre Greptile ?

Build a self-hosted AI pull request reviewer for a single GitHub repository. No web UI, no accounts, no telemetry.

Stack, no substitutions: Python 3.12, FastAPI, uvicorn, SQLite for state, sqlite-vec for vector search, httpx for GitHub REST calls, OpenAI API for embeddings and review generation. Package with uv. Everything runs in one process plus one CLI.

Secrets in .env, loaded with python-dotenv: GITHUB_APP_ID, GITHUB_PRIVATE_KEY_PATH, GITHUB_WEBHOOK_SECRET, OPENAI_API_KEY, REPO_FULL_NAME.

Part 1, indexer CLI (index.py):
- Walk the local clone of the repo, respect .gitignore, skip binaries, lockfiles, and anything over 400 KB.
- Chunk source files by function or class using tree-sitter for Python, TypeScript, JavaScript, and Go; fall back to 60-line sliding windows with 10-line overlap for everything else.
- Embed each chunk, store text, path, start line, end line, git blob sha, and vector in SQLite.
- Support incremental reindex: given two git revisions, only re-embed chunks whose file blob sha changed. Print counts of added, updated, deleted chunks.

Part 2, webhook service (server.py):
- POST /webhook, verify the X-Hub-Signature-256 HMAC against GITHUB_WEBHOOK_SECRET, reject on mismatch.
- Handle pull_request opened and synchronize events only. Enqueue work in a SQLite-backed job table and return 202 immediately.
- A background worker fetches the PR diff, splits it per file hunk, and for each hunk retrieves the top 8 related chunks by vector similarity plus the full current version of the changed file if it is under 800 lines.
- Send one LLM call per changed file with the hunk, retrieved context, and a review prompt that demands: only comment on correctness, security, or clear API misuse; no style nits; no praise; no summaries; return an empty array when the change is fine.
- Model output must be strict JSON: a list of objects with path, line, severity, body. Validate with Pydantic and drop anything whose line is not inside the diff.

Part 3, dedupe and posting:
- Store a hash of path plus normalized comment body per PR. Never post the same finding twice across force pushes.
- Post surviving comments as a single GitHub review with inline comments, authenticated as the GitHub App via a short-lived installation token.
- Add a CLI command 'review-local PR_NUMBER' that prints what it would post without calling GitHub, for prompt tuning.

Out of scope: GitLab and Bitbucket, multi-repo support, a dashboard, learning from dismissed comments, autofix suggestions, chat replies to review threads.

Deliver a README with GitHub App setup steps, the exact permissions needed (contents read, pull requests write, metadata read), and how to expose the webhook locally with a tunnel. Include pytest coverage for signature verification, diff line mapping, and dedupe.