kronos-agent

# Kronos Agent — Financial Candle Forecasting Generic inference wrapper around [Kronos](https://github.com/shiyu-coder/Kronos), a decoder-only foundation model pre-trained on K-line (candlestick) sequences from 45+ global exchanges. Feed it OHLC DataFrames, get forecasted OHLC DataFrames back. **This skill is infrastructure, not a personality.** It has no identity, no daily routine, no opinions. Trader agents call it to get predictions; scheduled scripts call it to pre-compute signals. All personalization (which symbols, what account size, what risk limits) lives in the caller's config. ## Requirements - Python 3.10+ - [`uv`](https://docs.astral.sh/uv/) (both scripts use PEP 723 inline dependencies) - Firebase Application Default Credentials for Firestore writes (only needed for non-dry-run use; see "Firebase auth" below) First model download pulls weights from HuggingFace (`NeoQuasar/Kronos-*`). This takes 1–3 minutes depending on model size and network. Subsequent runs use the local HF cache. ## Quickstart ### 1. Verify the submodule is initialized ```bash cd /path/to/Jarvis git submodule status claude/skills/kronos-agent/vendor/kronos # Should print: d5ffd46... claude/skills/kronos-agent/vendor/kronos (heads/master) ``` If the hash is missing or prefixed with `-`, run: ```bash git submodule update --init claude/skills/kronos-agent/vendor/kronos ``` ### 2. Smoke-test the inference wrapper ```bash cd claude/skills/kronos-agent ./inference.py --smoke-test --size ...