Configure max_tokens for Anthropic provider to prevent early streaming response truncation

Author: reichaves

lib/agent/agent_loop.py

@@ -138,6 +138,7 @@ def _get_database_id(indicator_code: str) -> str:
         "model": "claude-sonnet-4-6",
         "extra_headers": {"anthropic-version": "2023-06-01"},
         "thinking": {"type": "enabled", "budget_tokens": 8000},
+        "max_tokens": 16000,
         "temperature": 1,
     },
     "gemini-user": {
@@ -1229,6 +1230,8 @@ def _create_chat_completion(client, config, messages):
     extra_kwargs: dict = {}
     if "thinking" in config:
         extra_kwargs["extra_body"] = {"thinking": config["thinking"]}
+    if "max_tokens" in config:
+        extra_kwargs["max_tokens"] = config["max_tokens"]
     return client.chat.completions.create(
         model=config["model"],
         messages=messages,
@@ -1923,6 +1926,8 @@ def worker():
     extra_kwargs: dict = {}
     if "thinking" in config:
         extra_kwargs["extra_body"] = {"thinking": config["thinking"]}
+    if "max_tokens" in config:
+        extra_kwargs["max_tokens"] = config["max_tokens"]
 
     # Strip source_seal from tool results before streaming re-fetch.
     # The LLM does not need seal hashes to write its response; including them

tests/test_agent_loop.py

@@ -178,12 +178,13 @@ def test_run_agent_anthropic_thinking(mock_openai_class):
 
     run_agent("test", [], "anthropic", user_api_key="sk-ant-key")
 
-    # The completions.create call should include extra_body with thinking
+    # The completions.create call should include extra_body with thinking and max_tokens
     create_call = mock_client.chat.completions.create.call_args
     assert create_call.kwargs.get("temperature") == 1
     extra_body = create_call.kwargs.get("extra_body")
     assert extra_body is not None
     assert extra_body.get("thinking", {}).get("type") == "enabled"
+    assert create_call.kwargs.get("max_tokens") == 16000
 
 
 # ── UserKeyError exception propagation ───────────────────────────────────