fix(backend): downgrade user-caused LLM API errors to warning level (#12516)

Requested by @majdyz Follow-up to #12513. Anthropic/OpenAI 401, 403, and 429 errors are user-caused (bad API keys, forbidden, rate limits) and should not hit Sentry as exceptions. ### Changes **Changes in `blocks/llm.py`:** - Anthropic `APIError` handler (line ~950): check `status_code` — use `logger.warning()` for 401/403/429, keep `logger.error()` for server errors - Generic `Exception` handler in LLM block `run()` (line ~1467): same pattern — `logger.warning()` for user-caused status codes, `logger.exception()` for everything else - Extracted `USER_ERROR_STATUS_CODES = (401, 403, 429)` module-level constant - Added `break` to short-circuit retry loop for user-caused errors - Removed double-logging from inner Anthropic handler **Changes in `blocks/test/test_llm.py`:** - Added 8 regression tests covering 401/403/429 fast-exit and 500 retry behavior **Sentry issues addressed:** - AUTOGPT-SERVER-8B6, 8B7, 8B8 — `[LLM-Block] Anthropic API error: Error code: 401 - invalid x-api-key` - Any OpenAI 401/403/429 errors hitting the generic exception handler Part of SECRT-2166 ### Checklist 📋 #### For code changes: - [x] I have clearly listed my changes in the PR description - [x] I have made a test plan - [x] I have tested my changes according to the test plan #### Test plan: - [x] Unit tests for 401/403/429 Anthropic errors → warning log, no retry - [x] Unit tests for 500 Anthropic errors → error log, retry - [x] Unit tests for 401/403/429 OpenAI errors → warning log, no retry - [x] Unit tests for 500 OpenAI errors → error log, retry - [x] Verified USER_ERROR_STATUS_CODES constant is used consistently - [x] Verified no double-logging in Anthropic handler path --- Co-authored-by: Zamil Majdy (@majdyz) <zamil.majdy@agpt.co> --------- Co-authored-by: Zamil Majdy (@majdyz) <zamil.majdy@agpt.co>
2026-04-08 03:00:28 -04:00 · 2026-03-24 10:59:04 +00:00
parent ee5382a064
commit f21a36ca37
2 changed files with 215 additions and 54 deletions
--- a/autogpt_platform/backend/backend/blocks/llm.py
+++ b/autogpt_platform/backend/backend/blocks/llm.py
@@ -49,6 +49,9 @@ settings = Settings()
 logger = TruncatedLogger(logging.getLogger(__name__), "[LLM-Block]")
 fmt = TextFormatter(autoescape=False)

+# HTTP status codes for user-caused errors that should not be reported to Sentry.
+USER_ERROR_STATUS_CODES = (401, 403, 429)
+
 LLMProviderName = Literal[
    ProviderName.AIML_API,
    ProviderName.ANTHROPIC,
@@ -891,65 +894,60 @@ async def llm_call(
        client = anthropic.AsyncAnthropic(
            api_key=credentials.api_key.get_secret_value()
        )
-        try:
-            resp = await client.messages.create(
-                model=llm_model.value,
-                system=sysprompt,
-                messages=messages,
-                max_tokens=max_tokens,
-                tools=an_tools,
-                timeout=600,
-            )
+        resp = await client.messages.create(
+            model=llm_model.value,
+            system=sysprompt,
+            messages=messages,
+            max_tokens=max_tokens,
+            tools=an_tools,
+            timeout=600,
+        )

-            if not resp.content:
-                raise ValueError("No content returned from Anthropic.")
+        if not resp.content:
+            raise ValueError("No content returned from Anthropic.")

-            tool_calls = None
-            for content_block in resp.content:
-                # Antropic is different to openai, need to iterate through
-                # the content blocks to find the tool calls
-                if content_block.type == "tool_use":
-                    if tool_calls is None:
-                        tool_calls = []
-                    tool_calls.append(
-                        ToolContentBlock(
-                            id=content_block.id,
-                            type=content_block.type,
-                            function=ToolCall(
-                                name=content_block.name,
-                                arguments=json.dumps(content_block.input),
-                            ),
-                        )
+        tool_calls = None
+        for content_block in resp.content:
+            # Antropic is different to openai, need to iterate through
+            # the content blocks to find the tool calls
+            if content_block.type == "tool_use":
+                if tool_calls is None:
+                    tool_calls = []
+                tool_calls.append(
+                    ToolContentBlock(
+                        id=content_block.id,
+                        type=content_block.type,
+                        function=ToolCall(
+                            name=content_block.name,
+                            arguments=json.dumps(content_block.input),
+                        ),
                    )
-
-            if not tool_calls and resp.stop_reason == "tool_use":
-                logger.warning(
-                    f"Tool use stop reason but no tool calls found in content. {resp}"
                )

-            reasoning = None
-            for content_block in resp.content:
-                if hasattr(content_block, "type") and content_block.type == "thinking":
-                    reasoning = content_block.thinking
-                    break
-
-            return LLMResponse(
-                raw_response=resp,
-                prompt=prompt,
-                response=(
-                    resp.content[0].name
-                    if isinstance(resp.content[0], anthropic.types.ToolUseBlock)
-                    else getattr(resp.content[0], "text", "")
-                ),
-                tool_calls=tool_calls,
-                prompt_tokens=resp.usage.input_tokens,
-                completion_tokens=resp.usage.output_tokens,
-                reasoning=reasoning,
+        if not tool_calls and resp.stop_reason == "tool_use":
+            logger.warning(
+                f"Tool use stop reason but no tool calls found in content. {resp}"
            )
-        except anthropic.APIError as e:
-            error_message = f"Anthropic API error: {str(e)}"
-            logger.error(error_message)
-            raise ValueError(error_message)
+
+        reasoning = None
+        for content_block in resp.content:
+            if hasattr(content_block, "type") and content_block.type == "thinking":
+                reasoning = content_block.thinking
+                break
+
+        return LLMResponse(
+            raw_response=resp,
+            prompt=prompt,
+            response=(
+                resp.content[0].name
+                if isinstance(resp.content[0], anthropic.types.ToolUseBlock)
+                else getattr(resp.content[0], "text", "")
+            ),
+            tool_calls=tool_calls,
+            prompt_tokens=resp.usage.input_tokens,
+            completion_tokens=resp.usage.output_tokens,
+            reasoning=reasoning,
+        )
    elif provider == "groq":
        if tools:
            raise ValueError("Groq does not support tools.")
@@ -1462,7 +1460,16 @@ class AIStructuredResponseGeneratorBlock(AIBlockBase):
                    yield "prompt", self.prompt
                    return
            except Exception as e:
-                logger.exception(f"Error calling LLM: {e}")
+                is_user_error = (
+                    isinstance(e, (anthropic.APIStatusError, openai.APIStatusError))
+                    and e.status_code in USER_ERROR_STATUS_CODES
+                )
+                if is_user_error:
+                    logger.warning(f"Error calling LLM: {e}")
+                    error_feedback_message = f"Error calling LLM: {e}"
+                    break
+                else:
+                    logger.exception(f"Error calling LLM: {e}")
                if (
                    "maximum context length" in str(e).lower()
                    or "token limit" in str(e).lower()
--- a/autogpt_platform/backend/backend/blocks/test/test_llm.py
+++ b/autogpt_platform/backend/backend/blocks/test/test_llm.py
@@ -1,9 +1,18 @@
+from typing import cast
 from unittest.mock import AsyncMock, MagicMock, patch

+import anthropic
+import httpx
+import openai
 import pytest

+import backend.blocks.llm as llm
 from backend.data.model import NodeExecutionStats

+# TEST_CREDENTIALS_INPUT is a plain dict that satisfies AICredentials at runtime
+# but not at the type level. Cast once here to avoid per-test suppressors.
+_TEST_AI_CREDENTIALS = cast(llm.AICredentials, llm.TEST_CREDENTIALS_INPUT)
+

 class TestLLMStatsTracking:
    """Test that LLM blocks correctly track token usage statistics."""
@@ -655,3 +664,148 @@ class TestAITextSummarizerValidation:
        error_message = str(exc_info.value)
        assert "Expected a string summary" in error_message
        assert "received dict" in error_message
+
+
+def _make_anthropic_status_error(status_code: int) -> anthropic.APIStatusError:
+    """Create an anthropic.APIStatusError with the given status code."""
+    request = httpx.Request("POST", "https://api.anthropic.com/v1/messages")
+    response = httpx.Response(status_code, request=request)
+    return anthropic.APIStatusError(
+        f"Error code: {status_code}", response=response, body=None
+    )
+
+
+def _make_openai_status_error(status_code: int) -> openai.APIStatusError:
+    """Create an openai.APIStatusError with the given status code."""
+    response = httpx.Response(
+        status_code, request=httpx.Request("POST", "https://api.openai.com/v1/chat")
+    )
+    return openai.APIStatusError(
+        f"Error code: {status_code}", response=response, body=None
+    )
+
+
+class TestUserErrorStatusCodeHandling:
+    """Test that user-caused LLM API errors (401/403/429) break the retry loop
+    and are logged as warnings, while server errors (500) trigger retries."""
+
+    @pytest.mark.asyncio
+    @pytest.mark.parametrize("status_code", [401, 403, 429])
+    async def test_anthropic_user_error_breaks_retry_loop(self, status_code: int):
+        """401/403/429 Anthropic errors should break immediately, not retry."""
+        import backend.blocks.llm as llm
+
+        block = llm.AIStructuredResponseGeneratorBlock()
+        call_count = 0
+
+        async def mock_llm_call(*args, **kwargs):
+            nonlocal call_count
+            call_count += 1
+            raise _make_anthropic_status_error(status_code)
+
+        with patch.object(block, "llm_call", new=AsyncMock(side_effect=mock_llm_call)):
+            input_data = llm.AIStructuredResponseGeneratorBlock.Input(
+                prompt="Test",
+                expected_format={"key": "desc"},
+                model=llm.DEFAULT_LLM_MODEL,
+                credentials=_TEST_AI_CREDENTIALS,
+                retry=3,
+            )
+
+            with pytest.raises(RuntimeError):
+                async for _ in block.run(input_data, credentials=llm.TEST_CREDENTIALS):
+                    pass
+
+        assert (
+            call_count == 1
+        ), f"Expected exactly 1 call for status {status_code}, got {call_count}"
+
+    @pytest.mark.asyncio
+    @pytest.mark.parametrize("status_code", [401, 403, 429])
+    async def test_openai_user_error_breaks_retry_loop(self, status_code: int):
+        """401/403/429 OpenAI errors should break immediately, not retry."""
+        import backend.blocks.llm as llm
+
+        block = llm.AIStructuredResponseGeneratorBlock()
+        call_count = 0
+
+        async def mock_llm_call(*args, **kwargs):
+            nonlocal call_count
+            call_count += 1
+            raise _make_openai_status_error(status_code)
+
+        with patch.object(block, "llm_call", new=AsyncMock(side_effect=mock_llm_call)):
+            input_data = llm.AIStructuredResponseGeneratorBlock.Input(
+                prompt="Test",
+                expected_format={"key": "desc"},
+                model=llm.DEFAULT_LLM_MODEL,
+                credentials=_TEST_AI_CREDENTIALS,
+                retry=3,
+            )
+
+            with pytest.raises(RuntimeError):
+                async for _ in block.run(input_data, credentials=llm.TEST_CREDENTIALS):
+                    pass
+
+        assert (
+            call_count == 1
+        ), f"Expected exactly 1 call for status {status_code}, got {call_count}"
+
+    @pytest.mark.asyncio
+    async def test_server_error_retries(self):
+        """500 errors should be retried (not break immediately)."""
+        import backend.blocks.llm as llm
+
+        block = llm.AIStructuredResponseGeneratorBlock()
+        call_count = 0
+
+        async def mock_llm_call(*args, **kwargs):
+            nonlocal call_count
+            call_count += 1
+            raise _make_anthropic_status_error(500)
+
+        with patch.object(block, "llm_call", new=AsyncMock(side_effect=mock_llm_call)):
+            input_data = llm.AIStructuredResponseGeneratorBlock.Input(
+                prompt="Test",
+                expected_format={"key": "desc"},
+                model=llm.DEFAULT_LLM_MODEL,
+                credentials=_TEST_AI_CREDENTIALS,
+                retry=3,
+            )
+
+            with pytest.raises(RuntimeError):
+                async for _ in block.run(input_data, credentials=llm.TEST_CREDENTIALS):
+                    pass
+
+        assert (
+            call_count > 1
+        ), f"Expected multiple retry attempts for 500, got {call_count}"
+
+    @pytest.mark.asyncio
+    async def test_user_error_logs_warning_not_exception(self):
+        """User-caused errors should log with logger.warning, not logger.exception."""
+        import backend.blocks.llm as llm
+
+        block = llm.AIStructuredResponseGeneratorBlock()
+
+        async def mock_llm_call(*args, **kwargs):
+            raise _make_anthropic_status_error(401)
+
+        with patch.object(block, "llm_call", new=AsyncMock(side_effect=mock_llm_call)):
+            input_data = llm.AIStructuredResponseGeneratorBlock.Input(
+                prompt="Test",
+                expected_format={"key": "desc"},
+                model=llm.DEFAULT_LLM_MODEL,
+                credentials=_TEST_AI_CREDENTIALS,
+            )
+
+            with (
+                patch.object(llm.logger, "warning") as mock_warning,
+                patch.object(llm.logger, "exception") as mock_exception,
+                pytest.raises(RuntimeError),
+            ):
+                async for _ in block.run(input_data, credentials=llm.TEST_CREDENTIALS):
+                    pass
+
+        mock_warning.assert_called_once()
+        mock_exception.assert_not_called()