Contextual, Capability-Driven Primers

[ashenfad]

Implement a system for dynamically injecting primer instructions based on registered capabilities.

Core Principle: The framework should intelligently coach the agent on how to use its tools, but only when those tools are relevant.

Architecture:

• Modular primer library (e.g., with_statements.md, view_image.md)
• Registry of primer triggers based on capability registration
• Primer assembler that runs before each LLM call

Initial Use Cases:

1. Auto-inject guidance for stateful/unpickleable objects when agent.module(instance) is called
2. Auto-inject view_image primer when agent has image-producing capabilities

Benefits:

• Dramatically reduces barrier to entry
• Makes custom primers like db_primer.py obsolete
• Prevents context pollution by only showing relevant guidance