📝 Pull Request Template

1. Related Issue

Closes # (issue number)

Type of Change (select one)

Bug Fix, New Feature

3. Description

• Composer — robust LLM parsing:

  • _call_llm now accepts agent results that are already a LlmPlanProposal, a dict, or a JSON/text string. It attempts to validate/parse into LlmPlanProposal and falls back to a minimal proposal that preserves the raw model output in rationale.
  • Why: many LLM providers sometimes return raw JSON strings or unwrapped dicts; this change avoids crashes and preserves the model output for debugging/retry logic.

• Composer — safety for missing prices & conservative BP accounting:

  • _normalize_quantity now blocks exposure-increasing trades when no valid price is available, only allowing reduce-only (flattening) actions.
  • Consumed buying-power is computed using the same slippage-adjusted effective price used for the clamp (more conservative).
  • Why: prevents creating “free” leverage/large positions when price data is missing and avoids underestimating buying-power consumption that previously led to inflated portfolio values.

• Portfolio settlement correctness:

  • Trade settlement and cash updates now prefer avg_exec_price, then exit_price, then entry_price (fallback to market snapshot) when computing notional and mark price.
  • Why: ensures cash reflects realized PnL at the actual execution price so equity (total_value) equals account_balance + net_exposure after closes; fixes the observed mismatch after full liquidation.

• Logs & diagnostics: composer and portfolio code now emit clear warnings/errors when encountering missing/invalid model or price data and preserve raw LLM outputs in rationale for auditing.

  • Why: improves observability for debugging model/parsing and market data issues in production.

Summary: these changes make decision-time sizing and portfolio accounting safe and deterministic: they prevent exposure creation when market prices are unavailable, ensure buying-power is conservatively estimated, and make cash/realized PnL settle correctly — plus they add tests and better logging so the behaviors are observable and reproducible.

4. Testing

• I have tested this locally.
• I have updated or added relevant tests.

5. Checklist

• I have read the Code of Conduct
• I have followed the Contributing Guidelines
• My changes follow the project's coding style