X Feed Pilot

It needs macOS, Google Chrome, Node.js 20+, Python 3, and Clang with access to Apple's Vision framework. From the project root run npm install, copy .env.example to .env and add both API keys, then npm run build and npm start. Load dist/extension/ as an unpacked extension in chrome://extensions.

TYPESAFE_API_KEY and OPENAI_API_KEY, both required to start a session. They can be placed in .env or supplied through the environment variables of the same names; a nonempty value in .env takes precedence.

Reading history stays on your computer. Records are stored in data/pilot.sqlite, which Git ignores. Image bytes stay in memory only for the decision and follow-ups, and temporary crop files are removed.

Each session runs for up to 10 minutes or 80 posts, with limits of 8 likes and 4 bookmarks. Collection and the session timer pause when the Chrome training tab loses focus. Starting again after stopping creates a new session, and restarting the service ends any unfinished session.

Posts with their own attached images go to OpenAI's gpt-6-luna Decisions API, and posts without images go to Jev. Both use your preferences and evidence from links or mentioned accounts visited as needed.

Each decision makes one request with a 15-second timeout, no retry, and no switching between models. Three consecutive failed decisions stop the session. Posts marked as ads are skipped.

Only the main post's own images are evaluated; images inside quoted posts and video frames are excluded. Up to four full attachments are downloaded, and videos without captions may lack sufficient evidence. Automating likes and bookmarks may violate X's rules and restrict the training account.