fix(responses): preserve streamed tool call history

[asuan-dev]

Summary

• recover complete function-call arguments and custom-tool input from *.done, response.output_item.done, and response.completed.output[]
• preserve Responses item.id separately from call_id, including custom-tool namespaces and protocol-valid generated item IDs
• make raw SSE aggregation follow the same terminal reconciliation rules so persisted history is not saved with empty tool payloads
• reject ambiguous output_index fallback instead of merging or misrouting multiple tool calls

Root cause

The Responses stream transformer only tracked tool calls by call_id and ignored several authoritative terminal payload sources. Some providers send the complete tool payload only in terminal events, so AxonHub could emit and persist a tool call with empty arguments/input. Replaying that malformed item in the next Codex turn could make the conversation fail permanently.

Verification

• cd llm && go test ./transformer/openai/responses -count=10
• cd llm && go test ./... -count=1
• gofmt -d on all changed Go files
• git diff --check
• independent code/protocol/architecture review: PASS

[greptile-apps]

Greptile Summary

This PR fixes a correctness bug in the Responses API stream transformer where tool calls could be persisted with empty arguments/input, causing downstream Codex turns to fail permanently. The fix recovers complete payloads from terminal events (output_item.done and response.completed), introduces ToolPayloadDone/OutputItemDone state flags, and adds ResponseItemID to disambiguate the Responses item.id from call_id.

• aggregator.go: Adds finalItem() (shared terminal-event reconciliation), toolItem() (payload-validated item projection), orderedItems(), and completedToolCalls(). Ambiguous output_index fallback now errors rather than silently misrouting.
• outbound_stream.go: Tool call chunks are now batched and emitted at response.completed via enqueueAggregatedToolCalls(); reconcileToolPayload() handles providers whose final payload diverges from the streamed bytes.
• inbound_stream.go / inbound.go / outbound_convert.go: ResponseItemID and Namespace propagated through all tool call conversion paths; item ID generation gains type-based prefixes (fc_, ctc_, rs_, msg_).

Confidence Score: 4/5

Safe to merge with awareness of one gap in aggregator.go: a function_call arriving solely via output_item.done with empty arguments cannot be rescued by response.completed.

The fix is well-structured and well-tested for the primary scenarios. One gap exists: finalItem's unconditional early-return for OutputItemDone items prevents response.completed from filling in a function_call whose output_item.done carried empty arguments. If a provider exhibits this pattern the call is silently dropped from history, which is the exact class of bug this PR aims to fix. All other paths — terminal-only payloads, ambiguous index rejection, namespace propagation, and reconciliation of diverging final payloads — are verified by the new tests.

aggregator.go — specifically the else-if OutputItemDone early-return guard in finalItem and the length-based heuristic in applyDoneText.

Important Files Changed

Filename	Overview
llm/transformer/openai/responses/aggregator.go	Core refactor introducing finalItem(), toolItem(), orderedItems(), completedToolCalls() and ToolPayloadDone/OutputItemDone flags. OutputItemDone early-return unconditionally blocks response.completed from patching empty-argument function_calls already marked done by output_item.done.
llm/transformer/openai/responses/outbound_stream.go	Replaces immediate delta emission with batched enqueueAggregatedToolCalls() at response.completed; adds reconcileToolPayload() to gracefully handle final-payload divergence.
llm/transformer/openai/responses/inbound_stream.go	Adds toolCallItemIDs map for stable Responses item-ID tracking and propagates ResponseItemID/Namespace through tool call lifecycle. Error propagation from aggregator now surfaced via enqueueEvent.
llm/tools.go	Adds ResponseItemID to ToolCall and Namespace to ResponseCustomToolCall to preserve protocol-level identities separately from call_id.
llm/transformer/openai/responses/outbound_convert.go	generateItemID gains a prefix parameter; convertToResponsesAPIResponse now uses ResponseItemID as the item ID for function_call/custom_tool_call when available and propagates Namespace.
llm/transformer/openai/responses/aggregator_test.go	Seven new tests covering terminal-only payloads, ambiguous output_index rejection, empty done-input semantics, and non-completed response filtering.
llm/transformer/openai/responses/inbound.go	Populates ResponseItemID and Namespace in all convertItemToMessage call sites; adds item.ID to function_call_output and custom_tool_call_output messages.
llm/transformer/openai/responses/stream_tool_roundtrip_test.go	Roundtrip tests for the full inbound/outbound stream path validating consistent item IDs and payload integrity end-to-end.
llm/transformer/openai/responses/stream_tool_cross_protocol_test.go	Cross-protocol tests verifying tool call fidelity across the Responses/unified-LLM boundary.
llm/transformer/openai/responses/inbound_integration_test.go	Updates existing assertions to reflect type-based item ID prefixes (fc_, ctc_) rather than reusing call_id.

Sequence Diagram

%%{init: {'theme': 'neutral'}}%%
sequenceDiagram
    participant Provider as Provider SSE
    participant Agg as streamAggregator
    participant OB as responsesOutboundStream
    participant IB as responsesInboundStream

    Provider->>Agg: response.output_item.added (function_call)
    Agg->>Agg: "newAggregatedItem(), ToolPayloadDone=false"

    Provider->>Agg: function_call_arguments.delta
    Agg->>Agg: "item.Arguments += delta, ToolPayloadDeltas append"

    Provider->>Agg: response.output_item.done (complete args)
    Agg->>Agg: "finalItem(fromCompleted=false), ToolPayloadDone=true, OutputItemDone=true"

    Provider->>Agg: response.completed (args in snapshot)
    Agg->>Agg: "finalItem(fromCompleted=true), OutputItemDone=true EARLY RETURN"

    Agg->>OB: completedToolCalls(), toolItem() validates payload
    OB->>OB: enqueueAggregatedToolCalls(), reconcileToolPayload(deltas, finalPayload)
    OB->>OB: emit identity chunk then reconciled delta chunks

    OB->>IB: ToolCall identity + argument deltas (unified LLM format)
    IB->>IB: resolveToolCallItemID (ResponseItemID or generated fc_/ctc_)
    IB->>IB: output_item.added, function_call_arguments.delta, done events
    IB->>IB: buildResponse() for response.completed snapshot

Loading

%%{init: {'theme': 'base', 'themeVariables': {"darkMode": true, "background": "#0d1117", "primaryColor": "#21262d", "primaryTextColor": "#e6edf3", "primaryBorderColor": "#8b949e", "lineColor": "#8b949e", "textColor": "#e6edf3", "edgeLabelBackground": "#161b22", "actorBkg": "#21262d", "actorBorder": "#8b949e", "actorTextColor": "#e6edf3", "actorLineColor": "#8b949e", "signalColor": "#8b949e", "signalTextColor": "#e6edf3", "noteBkgColor": "#373320", "noteBorderColor": "#d4a72c", "noteTextColor": "#f0e6c0", "labelBoxBkgColor": "#21262d", "labelBoxBorderColor": "#8b949e", "labelTextColor": "#e6edf3", "loopTextColor": "#e6edf3", "activationBkgColor": "#30363d", "activationBorderColor": "#8b949e"}}}%%
sequenceDiagram
    participant Provider as Provider SSE
    participant Agg as streamAggregator
    participant OB as responsesOutboundStream
    participant IB as responsesInboundStream

    Provider->>Agg: response.output_item.added (function_call)
    Agg->>Agg: "newAggregatedItem(), ToolPayloadDone=false"

    Provider->>Agg: function_call_arguments.delta
    Agg->>Agg: "item.Arguments += delta, ToolPayloadDeltas append"

    Provider->>Agg: response.output_item.done (complete args)
    Agg->>Agg: "finalItem(fromCompleted=false), ToolPayloadDone=true, OutputItemDone=true"

    Provider->>Agg: response.completed (args in snapshot)
    Agg->>Agg: "finalItem(fromCompleted=true), OutputItemDone=true EARLY RETURN"

    Agg->>OB: completedToolCalls(), toolItem() validates payload
    OB->>OB: enqueueAggregatedToolCalls(), reconcileToolPayload(deltas, finalPayload)
    OB->>OB: emit identity chunk then reconciled delta chunks

    OB->>IB: ToolCall identity + argument deltas (unified LLM format)
    IB->>IB: resolveToolCallItemID (ResponseItemID or generated fc_/ctc_)
    IB->>IB: output_item.added, function_call_arguments.delta, done events
    IB->>IB: buildResponse() for response.completed snapshot

Loading

_{Reviews (3): Last reviewed commit: "refactor(responses): centralize streamed..." | Re-trigger Greptile}

[greptile-apps]

response.completed cannot rescue a function_call that output_item.done marked done with empty arguments

The early return fires for every caller, including applyResponseSnapshot (fromCompletedResponse=true). When a provider sends output_item.done with arguments:"" and there are no prior argument deltas, finalItem marks the item OutputItemDone=true but leaves item.Arguments.Len() == 0. A subsequent response.completed snapshot carrying the full arguments hits this early return before applying them, so toolItem() — which requires Arguments.Len() > 0 for function_call — silently drops the call.

The fromCompletedResponse parameter was introduced to distinguish the two callers, but the guard does not consult it. Guarding with && !fromCompletedResponse would allow response.completed to fill in missing payload while still letting output_item.done win whenever it carries a non-empty value (since if src.Arguments != "" { ... } only overwrites when the snapshot has real data).