[
https://issues.apache.org/jira/browse/CAMEL-23956?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Federico Mariani updated CAMEL-23956:
-------------------------------------
Attachment: OpenAIAgenticLoopMalformedToolArgumentsTest.java
> camel-openai: malformed tool-call arguments crash the agentic loop instead of
> being returned to the model
> ---------------------------------------------------------------------------------------------------------
>
> Key: CAMEL-23956
> URL: https://issues.apache.org/jira/browse/CAMEL-23956
> Project: Camel
> Issue Type: Bug
> Components: camel-openai
> Affects Versions: 4.21.0
> Reporter: Federico Mariani
> Priority: Major
> Labels: ai
> Attachments: OpenAIAgenticLoopMalformedToolArgumentsTest.java
>
>
> Models occasionally emit syntactically invalid JSON in
> {{tool_calls[].function.arguments}} — a hallucination class every agentic
> runtime must handle. The documented error contract of the agentic loop
> ({{openai-mcp.adoc}}, "Error Handling in the Agentic Loop") is that tool
> execution problems are caught, logged and sent back to the model as tool
> result text so it can recover.
> However, in {{OpenAIProducer.processNonStreamingAgentic}} the argument
> parsing happens *outside* the try/catch that implements that contract:
> {code:java}
> Map<String, Object> argsMap = OBJECT_MAPPER.readValue(argsJson, Map.class);
> // line ~523, not guarded
> String resultContent;
> try {
> McpSchema.CallToolResult toolResult = getEndpoint().callTool(mcpClient,
> toolName, argsMap);
> ...
> } catch (Exception e) {
> resultContent = "Error: Tool execution failed: " + e.getMessage();
> }
> {code}
> A malformed arguments string therefore fails the whole exchange with a raw
> Jackson exception:
> {noformat}
> com.fasterxml.jackson.core.io.JsonEOFException: Unexpected end-of-input:
> expected close marker for Object
> {noformat}
> {{OpenAIToolExecutionProducer.process}} (line ~145) has the identical
> unguarded {{readValue}}.
> h3. Suggested fix
> Move the parse inside the guarded section (or its own try/catch) and feed a
> corrective tool result back to the model, e.g. {{"Error: invalid tool
> arguments: " + e.getMessage()}} — mirroring how MCP execution failures are
> already handled. The attached route-based reproducer
> ({{OpenAIAgenticLoopMalformedToolArgumentsTest}}, using
> {{camel-test-infra-openai-mock}}) asserts the recovery behavior and fails on
> current main.
> Related: CAMEL-23943 documents the same defect family in
> camel-langchain4j-tools. The fix should stay compatible with the error-policy
> model of the unified AI tool abstraction (CAMEL-23382,
> {{AiToolResult.ArgumentError}}), which camel-openai will adopt in roadmap
> step 6.
> _This issue was drafted by Claude Code on behalf of Federico Mariani_
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
