conversion from json to POJO error when use tool call in ai stream

[wykeli]

Bug Report: com.fasterxml.jackson.core.JsonParseException: Unexpected close marker ']': expected '}' (for Object starting at [Source: REDACTED (StreamReadFeature.INCLUDE_SOURCE_IN_LOCATION disabled); line: 1, column: 235])
at [Source: REDACTED (StreamReadFeature.INCLUDE_SOURCE_IN_LOCATION disabled); line: 1, column: 289]
Spring AI version: 1.0.1
Spring Boot version: 3.5.5
Java version: 17

Description:
the json content is
{"jobParam": {"clusterType": "xx", "noAddClusterAgain": false, "noQuitCluster": false, "receiveAccount": "xx", "noReturnMachine": true}, "borrowNodeCondition": {"nodeFromLabel": "buffer", "notBorrow": false, "weCubeKeys": [{"field": "idc", "relation": "eq", "value": "x"]}, "creator": "XXX", "rate": 6, "sourceCondition": {"weCubeKeys": [{"field": "idc", "relation": "in", "values": ["s", "a"]}, {"field": "cluster_alias", "relation": "eq", "value": "abc"}, {"field": "node_ip", "relation": "eq", "value": "127.0.0.1"]}, "taskName": ""}, "conversionId": "1"}
that is not an normal json style

---

what the code when occur in
chatClient.prompt().toolCallbacks(ToolCallbacks.from(XXX)).user(xxx).advisors(spec -> spec.param(ChatMemory.CONVERSATION_ID, XXX)).stream().content()
the tool code is
@Tool(description = "发起k8s节点迁移任务") public String startMgrTask( @ToolParam(description = "迁移任务参数") AiMgrJobParam jobParam, @ToolParam(description = "当前会话id") String conversionId) { }
I guess this is a bug when using tool call in ai stream ,beacus no exception orrur when use the sync way

chatClient.prompt().system(SYSTEM_PROMPT).toolCallbacks(ToolCallbacks.from()).user().advisors(spec -> spec.param(ChatMemory.CONVERSATION_ID, conversionKey)).call().content()

[kmw10693]

Hi, thanks for reporting this issue.

I’d like to take this one and work on a fix.

[injae-kim]

Fix PR: #4274 👍

[wykeli]

Thank you. Your feedback is extremely prompt; you are the best

[wykeli]

Fix PR: #4274 👍

may i aks an question, what is the maven config use in you code case, and what is the dependency version ?
because I can not find the OpenAiChatOptions.builder().toolExecutionEligibilityPredicate(),thank you for supporting!!

[dev-jonghoonpark]

@wykeli The PR has been created, but it hasn’t been merged yet.

Once it gets merged and included in the official release, you’ll be able to use it via Maven.
If you need to use it right away, you can also build it yourself and use it directly.

[dev-jonghoonpark]

@wykeli
Looking at the PR, it seems like it would work if you just add the implementation code from the PR into your project. Would you like to try it?

[wykeli]

@wykeli Looking at the PR, it seems like it would work if you just add the implementation code from the PR into your project. Would you like to try it?

@dev-jonghoonpark
ok，thanks
Additionally, I’ve found that when Spring AI performs streaming calls to certain LLMs with “thinking” capabilities, every additional tool call during the process causes one more think tag to appear in the model output. This doesn’t occur when tool calls aren’t used or when the calls are synchronous.

[wykeli]

Fix PR: #4274 👍

I think this PR is not working,ToolExecutionEligibilityPredicate is interface to judge whether to use tool call , and when tool call finish will set finishReason with TOOL_CALLSS or STOP,
however, I guest tool call will not execute in ToolOnFinishPredicate ,because the condition will not be ture for
hasToolCalls && ("tool_calls".equalsIgnoreCase(finishReason) || "stop".equalsIgnoreCase(finishReason))
I think this is impossible to be ture because no any one tool will be execute for this

[wykeli]

The problematic LLM was Qwen3-8B. After debugging the chatCompletionStream in openAiApi class , I found that the chunk returned by the LLM was invalid. I then switched the model to DeepSeek-R1-0528, and the issue was resolved