Repository navigation
OpenAIResponsesLlm drops the keys beside "result" in a tool's dict return #7836
Description
🔴 Required Information
Describe the Bug:
OpenAIResponsesLlm drops every key except result from a tool's dict return. _serialize_json_value (integrations/openai/_openai_responses_llm.py) unwraps any mapping that has a non-None result key and sends only that value as the function_call_output. ADK wraps a non-dict tool return as exactly {"result": value}, but a tool's own dict can also have a result key next to other data. For example, {"result": "found 2 contracts", "contracts": [...], "next_page_token": "abc"} reaches the model as just found 2 contracts.
AnthropicLlm had the same bug, fixed in #7073: it now unwraps only when the keys are exactly {"result"}.
Steps to Reproduce:
- Give an
LlmAgentusingOpenAIResponsesLlm(model="gpt-5")a function tool that returns{"result": "found 2 contracts", "contracts": ["NDA-12", "MSA-7"], "next_page_token": "abc"}. - Let the model call it and inspect the
function_call_outputitem in the nextresponses.createinput.
Expected Behavior:
The whole dict is sent as JSON. Only the bare {"result": value} wrapper is unwrapped.
Observed Behavior:
function_call_output sent to OpenAI: found 2 contracts
Environment Details:
- ADK Library Version (pip show google-adk): 2.13.0 (also on
mainat 75fb788) - Desktop OS: macOS
- Python Version (python -V): 3.11.16
Model Information:
- Are you using LiteLLM: No
- Which model is being used: gpt-5 via
OpenAIResponsesLlm
🟡 Optional Information
Additional Context:
if 'result' in value and value['result'] is not None:
result = value['result']
if isinstance(result, str):
return result
return json.dumps(result, default=str)How often has this issue occurred?:
- Always (100%), for tool dicts that have a
resultkey beside other keys
Activity
surajksharma07 commentedon Oct 30, 2026
@adityasingh2400 Reproduced this on 2.13.0 and main: only “found 2 contracts” reaches the model while contracts/next_page_token get dropped.
A workaround that worked locally: unwrap only when value.keys() == {'result'} in _serialize_json_value (the same rule AnthropicLlm got in #7073) so any other dict goes out as full JSON. Could you give that a try on your side and check that the bare {"result": x} returns and the MCP content outputs still behave the same?
Once you’ve confirmed it holds up #7837 looks like the right place for it. Will flag it for review. #7847 looks like it covers the same change so it’s worth settling on one of the two.
adityasingh2400 commentedon Oct 31, 2026
Thanks for reproducing it. #7837 already uses exactly that rule (
value.keys() == {'result'}), and I re-checked both cases you mentioned on the PR branch:test_function_response_unwraps_adk_result_wrapperchecks{"result": "sunny"}→sunnyand{"result": [1, 2]}→[1, 2].contentbranch returns before theresultcheck is reached, and the existingtest_function_response_serializes_mcp_content_as_textpasses unmodified.tests/unittests/integrations/openaipasses (227 tests). The PR also has an end-to-end run with a realRunnershowing the full dict reachingfunction_call_output.On the duplicate: #7837 was opened with this issue, about three hours before #7847 (19:05 vs. 22:09 UTC), so I'd suggest keeping #7837. Happy to go with whichever you prefer.
surajksharma07 commentedon Nov 3, 2026
Thanks @adityasingh2400 for re-checking both cases. Went through #7837 and it looks good: it’s the same keys() == {'result'} rule as #7073, and since the content branch returns first, the MCP outputs aren’t affected. Testing through _content_to_response_input_items is a nice touch too, and it still merges cleanly on current main.
Checked main and v2.13.0 again and the old 'result' in value check is still there (it’s the last place in the model connectors with that pattern), so this isn’t fixed in any release yet.
Agreed on keeping #7837 since it came first. @ninglinLiu the change in #7847 is identical, so could you close it in favour of #7837? @xuanyang15 #7837 has 2 workflows waiting on approval whenever you get a chance.
fix: keep the keys beside result in a Responses tool output