Agentic.Loop.Stages.LLMCall (agentic v0.2.2)

Makes the LLM API call and stores the response in context.

Uses the llm_chat callback from ctx.callbacks to make the actual API call. The response is stored in ctx.last_response for the next stage (typically ModeRouter) to process.

Model Routing

Resolves the best model route via Agentic.ModelRouter before each call. The resolved route is passed to the callback under "_route" key. If routing fails, falls back to direct callback invocation.

Cache Awareness (V1.2)

Separates params into a stable prefix (system prompt, workspace snapshot, tool definitions) and volatile suffix (recent transcript). Computes a stable_prefix_hash so the host can detect when the prefix changed and pass cache boundary hints to the LLM provider.

The params map sent to llm_chat includes a "cache_control" key with:

• "stable_hash" — hash of the stable prefix content
• "prefix_changed" — boolean, true when prefix differs from last call