PM Agent Strategy Authoring

⚠️ **Trading warning / Avertissement trading** : informational only, not investment advice. Prediction-market contracts can go to zero, a strategy that behaved in simulation can lose real money, and no backtest or paper run promises a return. Curated by Skill Harbor: the strategy-authoring runbook of oracle3, an open-source prediction-market trading engine and MCP server for Polymarket, Kalshi and other venues (pip install oracle3). This skill covers the LLM-driven flavour: it scaffolds an AgentStrategy subclass with the strategy create command, has the model or an external API make decisions inside process_event, and requires the reasoning to be written down on every decision with record_decision so the run can be monitored and reviewed afterwards. The discipline is explicit: check the control plane's pause flag before every decision, give every external call a timeout so it can never block the event loop, record the strategy file, class name and command for reproducibility, and never use future information. Validation is a short dry run on a handful of events; evaluation is paper trading with the monitor on, because agent strategies cannot be grid-searched or auto-tuned, a limit the skill states plainly. Mid-run you can pause, resume, and even hot-swap the strategy file. From the YichengYang-Ethan/oracle3-prediction-market-agent repository (Apache-2.0). Honest caveats: an LLM-driven strategy is non-deterministic by design, so identical code can decide differently tomorrow, and a calm paper run is evidence, not a promise; taking a strategy live afterwards is a separate, gated step (see the live trade ops skill), not a formality. Skill Harbor never reviews the code, review it yourself before use.