enable thinking in anthropic CUA (#2240)

skyvern/forge/agent.py

@@ -1416,6 +1416,8 @@ class ForgeAgent:
                 "display_width_px": settings.BROWSER_WIDTH,
             }
         ]
+        thinking = {"type": "enabled", "budget_tokens": 1024}
+        betas = ["computer-use-2025-01-24"]
         if not llm_caller.message_history:
             llm_response = await llm_caller.call(
                 prompt=task.navigation_goal,
@@ -1423,7 +1425,8 @@ class ForgeAgent:
                 use_message_history=True,
                 tools=tools,
                 raw_response=True,
-                betas=["computer-use-2025-01-24"],
+                betas=betas,
+                thinking=thinking,
             )
         else:
             llm_response = await llm_caller.call(
@@ -1431,7 +1434,8 @@ class ForgeAgent:
                 use_message_history=True,
                 tools=tools,
                 raw_response=True,
-                betas=["computer-use-2025-01-24"],
+                betas=betas,
+                thinking=thinking,
             )
         assistant_content = llm_response["content"]
         llm_caller.message_history.append({"role": "assistant", "content": assistant_content})

skyvern/forge/sdk/api/llm/api_handler_factory.py

@@ -693,6 +693,7 @@ class LLMCaller:
         max_tokens = active_parameters.get("max_completion_tokens") or active_parameters.get("max_tokens") or 4096
         model_name = self.llm_config.model_name.replace("bedrock/", "").replace("anthropic/", "")
         betas = active_parameters.get("betas", NOT_GIVEN)
+        thinking = active_parameters.get("thinking", NOT_GIVEN)
         LOG.info("Anthropic request", betas=betas, tools=tools, timeout=timeout)
         response = await app.ANTHROPIC_CLIENT.beta.messages.create(
             max_tokens=max_tokens,
@@ -701,6 +702,7 @@ class LLMCaller:
             tools=tools or NOT_GIVEN,
             timeout=timeout,
             betas=betas,
+            thinking=thinking,
         )
         LOG.info("Anthropic response", response=response, betas=betas, tools=tools, timeout=timeout)
         return response