fix(responses): preserve function calls completed in done events

[MrLiuGangQiang]

Summary

Fixes an OpenAI Responses streaming conversion issue where a function call completed by response.function_call_arguments.done can lose its arguments or tool identity before it reaches the client.

This affects upstreams that provide arguments, name, namespace, or call_id only in a later event. In Codex, that can leave collaboration.spawn_agent without a dispatchable identity even when the downstream response is valid.

Root cause

The outbound transformer previously treated response.function_call_arguments.done as state-only. The inbound transformer also emitted response.output_item.added immediately, even when the function name was absent. A later terminal event could repair the final item, but it could not repair the already-emitted initial item that clients use to route the call.

Changes

Outbound Responses transformer

• Resolves calls from call_id, with an item_id to call_id fallback.
• Treats non-empty terminal arguments as authoritative.
• Emits the missing argument suffix after partial deltas, without duplicating complete payloads.
• Emits an identity-only unified delta when name or namespace arrives after all arguments have already streamed.
• Preserves id, type, name, and namespace in emitted tool-call deltas.
• Returns an explicit mismatch error for divergent terminal arguments.

Inbound Responses transformer

• Defers output_item.added for unnamed function calls until a tool name is available.
• Buffers arguments received before tool identity and replays them only after emitting the correctly identified output_item.added.
• Merges late tool identity into tracked state.
• Includes the final call_id, name, namespace, and arguments in terminal response.function_call_arguments.done and response.output_item.done events.

JSON comparison safety

Semantically equivalent JSON that differs only in whitespace or object-key order is accepted without a duplicate delta. Comparisons use json.Decoder.UseNumber(), preserving numeric lexemes so distinct large integers cannot collapse through float64 conversion.

Regression coverage

• Complete arguments provided only in function_call_arguments.done.
• Reformatted but equivalent terminal JSON.
• Distinct large integers (9007199254740992 and 9007199254740993) rejected as a mismatch.
• Complete outbound to unified to inbound round trip when terminal events provide spawn_agent, collaboration, call ID, and JSON arguments.
• Arguments that arrive before late tool identity are buffered until a correctly identified output_item.added is emitted.

Verification

cd llm
go test ./transformer/openai/responses -run 'TestOutboundTransformer_TransformStream_(EmitsArgumentsProvidedOnlyInDone|AcceptsEquivalentFinalArguments|RejectsDistinctLargeJSONNumbers)|TestResponsesStream_RoundTrip_(PreservesToolIdentityProvidedOnlyInDone|DefersFunctionCallUntilLateIdentityArrives)' -count=1
git diff --check

[greptile-apps]

Greptile Summary

This PR preserves completed function calls during OpenAI Responses stream conversion. The main changes are:

• Reconciles arguments and tool identity from terminal function-call events.
• Defers unnamed inbound function calls until their identity arrives.
• Buffers and replays arguments received before tool identity.
• Compares reformatted JSON without losing large-number precision.
• Adds tests for terminal arguments, late identity, and JSON comparison.

Confidence Score: 5/5

This looks safe to merge.

• The updated JSON comparison preserves large-number precision.
• Terminal arguments and late tool identity are retained across the stream conversion.
• No remaining blocking issue qualifies for this follow-up review.

Important Files Changed

Filename	Overview
llm/transformer/openai/responses/outbound_stream.go	Reconciles terminal function-call arguments and identity while preserving JSON number precision.
llm/transformer/openai/responses/inbound_stream.go	Defers unnamed calls, buffers early arguments, and includes final identity in completion events.
llm/transformer/openai/responses/outbound_stream_test.go	Adds coverage for terminal arguments, late identity, equivalent JSON, and distinct large numbers.

_{Reviews (6): Last reviewed commit: "test(responses): preserve tool identity ..." | Re-trigger Greptile}

[MrLiuGangQiang]

主要是为了解决codex使用子agent报错的问题