Pass instruction depth config to llm-connect

src/activity_core/rules/executor.py

@@ -144,15 +144,16 @@ def _execute(
     # Step 2 — render prompt (raises UntrustedFieldError on policy violation)
     rendered = _render_prompt(instr.prompt, instr.trusted_fields, event, context)
     prompt_hash = hashlib.sha256(rendered.encode()).hexdigest()
+    llm_config = _llm_run_config(instr)
 
     # Step 3 — call LLM
-    raw_output = llm_client.complete(rendered, model=instr.model)
+    raw_output = llm_client.complete(rendered, model=instr.model, config=llm_config)
 
     # Step 4 — validate and optionally retry
     task_specs, report, error = _validate_output(raw_output, instr)
     if error:
         retry_prompt = rendered + f"\n\nPrevious output was invalid: {error}\nPlease fix."
-        raw_output = llm_client.complete(retry_prompt, model=instr.model)
+        raw_output = llm_client.complete(retry_prompt, model=instr.model, config=llm_config)
         task_specs, report, error = _validate_output(raw_output, instr)
         if error:
             logger.warning(
@@ -172,6 +173,19 @@ def _execute(
     )
 
 
+def _llm_run_config(instr: Any) -> dict[str, Any]:
+    """Build the llm-connect RunConfig payload from instruction metadata."""
+    config: dict[str, Any] = {"model_name": instr.model}
+    for field in ("temperature", "max_tokens", "max_depth"):
+        value = getattr(instr, field, None)
+        if value is not None:
+            config[field] = value
+    model_params = getattr(instr, "model_params", None)
+    if model_params:
+        config["model_params"] = model_params
+    return config
+
+
 def _empty_result(instr: Any, prompt_hash: str | None = None) -> InstructionResult:
     return InstructionResult(
         tasks=[],