fix(openai): v5 - skip unpaired reasoning item_reference in Responses API #10827
Conversation
When using the OpenAI Responses API with store: true and reasoning models, the API returns an error if a reasoning item_reference is sent without its required paired item (text or function_call with item_reference). This fix adds look-ahead logic when processing reasoning parts to check if the following part (after skipping consecutive reasoning parts) will also emit an item_reference. A reasoning item_reference is only included when: 1. The following part is a text or tool-call 2. The following part has a provider-assigned itemId (meaning it will emit an item_reference, not a full message/function_call object) This prevents errors like: - "Item 'rs_xxx' of type 'reasoning' was provided without its required following item" - "Item 'fc_xxx' of type 'function_call' was provided without its required 'reasoning' item"
![vercel[bot]](https://avatars.githubusercontent.com/in/8329?s=60&v=4){kind=small}
|
tagging @aayush-kapoor for visibility since I note you are assigned to the issue that this resolves 😄 |
|
cc @lgrammel |
|
can you provide an example under examples/ai-core with a tool that throws an error which leads to this problem? i want to make sure we fix the root cause and i am not convinced yet that this isn't patching the symptom |
|
also please do all work in a pr against main; we have tooling for backports |
|
@lgrammel This one got a lot harder to reproduce, however, eventually I managed to track down the root cause of the issue itself - which was resolved a couple of days ago in #10733 So we're good to close this, and related issues, as resolved! The chain was: |
3 participants
Background
When using the OpenAI Responses API with
store: trueand reasoning models (e.g., o3, o4-mini), the API returns errors whenitem_referenceitems are unpaired:Item 'rs_xxx' of type 'reasoning' was provided without its required following itemItem 'fc_xxx' of type 'function_call' was provided without its required 'reasoning' itemThis occurs in multi-turn conversations when:
The typical scenario is that the agent reasons, calls a tool, the tool returns an error (e.g., validation failure), and on retry the SDK rebuilds the conversation context. The reasoning has an
itemIdfrom the previous turn, but the failed tool call does not, causing the mismatch.Summary
Added look-ahead logic when processing reasoning parts with
store: true. Before including a reasoningitem_reference, the code now scans past consecutive reasoning parts to find the first non-reasoning part. The reference is only included if that part has a valid itemId that it will then include in it's iteration. Failed tool calls do not have theitemIdand thus do not include theitem_reference, causing the issue.Manual Verification
Tested with a reasoning model that calls a custom tool with invalid input. Previously this caused the API error; now the reasoning
item_referenceis correctly skipped and the conversation continues.Checklist
pnpm changesetin the project root)Future Work
Port into v6 main branch. #10829
Related Issues
#8379