This pull request introduces a major refactor and expansion of the action modeling and candidate generation pipeline, centralizing action semantics and profiles, improving action risk handling, and simplifying/focusing the LLM client logic. It also removes legacy fallback code and consolidates action schema logic into a new core module. The changes are grouped into three main themes:

1. Action Modeling Refactor and Centralization

• Introduced core/actions.py to define ActionSpec, ModeProfile, and ActionProfile dataclasses, along with canonical action definitions, aliases, and motivational profiles. This centralizes all action semantics, planning/execution instructions, and motivational mappings in a single location for consistent use across the codebase.
• Removed the legacy llm/action_schema.py file, moving all action schema logic to core/actions.py.

2. Candidate Generation and Risk Handling Improvements

• Replaced the old LLM candidate action generation with a new pipeline: the LLM now only returns contextual action risk overrides (get_action_risks_from_text), and candidate actions are constructed using the new build_candidate_actions function (from magus.candidates), which leverages the centralized action profiles and risk overrides. [1] [2] [3] [4]
• Removed all local heuristic fallback logic for candidate generation and stimulus appraisal from llm/client.py, enforcing a single, consistent LLM-driven pipeline.

3. Codebase Simplification and Consistency

• Updated the Action and Stimulus dataclasses in core/state.py to remove unused fields (e.g., delta_g from Action), and clarified type annotations. [1] [2] [3]
• Refactored the chat assistant logic in llm/conversation.py for improved reliability and retry handling, and to use the new action schema interface from core/actions.py. [1] [2] [3]

These changes collectively make action handling more robust, maintainable, and extensible, and pave the way for more advanced motivational reasoning and action selection.