qubie-org/qi

Entry point for working on the qi repository — a local-first text river running IBM Granite 4.1 via llama.cpp, with inline-only rendering, a citation-verifying research process, and a standalone macOS app. Load this first for any qi task, then follow it to the specific skill. Use when the user mentions qi, qiui, @qi-ui/cli, the text river, sigils, Granite intrinsics, aLoRA adapters, answerability, the research note or deck, or asks to set up, run, edit, evaluate, convert models for, or ship this project.

Edit the qi codebase correctly — house style, the docblock convention, the test suite, and the verify-against-the-real-app loop. Use when adding or changing a command, an app, a grounding source, a pack, a verb, a sigil, the inline renderer, the engine (colour, glyphs, key, sound), the agent loop, or the store; and before writing any commit that touches src/, cli/ or native/.

Every model qi can run, and everything recorded about it — the two offered Granite sizes with install size, RAM, hashes and which activated adapters exist; the ten packs in the catalogue; the five RAG intrinsics and the aLoRA distinction. Use when choosing or comparing a model size, adding or editing a pack, looking up a checksum, repo id, port, byte count or licence, explaining why the activated gate matters, or answering any question about what weights qi uses and where they come from.

The research process in src/ground/research.ts — the eight stages, the six named failure exits, the guards that do not depend on a prompt, and the 22-question ten-domain evaluation that decides whether a change helped. Use when changing anything about how qi searches, reads, gates, extracts, verifies or composes a note or a deck; when a note comes back empty or with bad citations; or when asked to evaluate or measure the research process.

Install, run and verify qi — either from npm via @qi-ui/cli or from source with bun, vite and llama-server. Use when setting up qi for the first time, choosing a model size, starting or stopping the model server, checking that generation, tool calling, grammars and the adapter shelf actually work, or diagnosing an install that produces no answers, no sound, no images, a blank page or a connection error.

Build, sign, notarise, package and release qi — the Swift app bundle, the dylib closure rewrite, Sparkle updates and the appcast, the weights release, and publishing @qi-ui/cli to npm. Use when asked to make a build, produce a DMG, sign or notarise the Mac app, cut a release, publish converted adapters, ship an update, or explain the licence and the notices the app carries.

Convert, fine-tune or train weights for qi — chiefly converting IBM's activated LoRA safetensors to GGUF so the answerability gate costs milliseconds instead of seconds, and the pre-Granite research lane under train/. Use when asked to convert an adapter, produce activated adapters for a new model size, fine-tune anything, run a training job, or touch train/, and read it before running any GPU work for this project.