fix(investigation): actually use thinking_budget parameter for per-specialist max_tokens

The _run_specialist_session() function accepted thinking_budget as a parameter
but completely ignored it, always deriving max_thinking_tokens from thinking_level
instead. This made the per-specialist max_tokens configuration (e.g., 128000 for
root_cause agents) non-functional.

Updated the thinking_kwargs logic to prioritize explicit thinking_budget when
provided, with fallback to thinking_level-based derivation for backward
compatibility.

This fix ensures both investigation and PR review specialists correctly use
their configured max_tokens budgets.

Co-Authored-By: Claude Sonnet 4.5 <[email protected]>

apps/backend/runners/github/services/parallel_agent_base.py

@@ -194,11 +194,19 @@ class ParallelAgentOrchestrator:
             # Create SDK client for this specialist
             # Use per-specialist model for betas (not the global config model)
             betas = get_model_betas(model or self.config.model or "sonnet")
-            # Use per-specialist thinking level when provided
-            effective_thinking = thinking_level or self.config.thinking_level or "medium"
-            thinking_kwargs = get_thinking_kwargs_for_model(
-                model, effective_thinking
-            )
+
+            # Get thinking budget - use explicit budget if provided, otherwise derive from thinking level
+            if thinking_budget is not None:
+                thinking_kwargs = {
+                    "max_thinking_tokens": thinking_budget
+                }
+            else:
+                # Use per-specialist thinking level when provided
+                effective_thinking = thinking_level or self.config.thinking_level or "medium"
+                thinking_kwargs = get_thinking_kwargs_for_model(
+                    model, effective_thinking
+                )
+
             # Override effort_level if explicitly provided (e.g., investigation
             # agents always use "high" effort regardless of thinking level).
             # Only applies to adaptive models (Opus 4.6+) where thinking_kwargs