[Quim Muntal (qmuntal)]

Summary

Fix the SDK behavior behind the streaming and cancellation workarounds documented in dkr290/agenticgo#49.

• Assemble complete streamed tool calls and refusals once the OpenAI stream finishes successfully; preserve text, usage, metadata, and the original raw chunks.
• Emit completed and unexecuted tool outcomes before terminal errors, including concurrent batches, approval resumption, and buffered non-streaming runs. Preserve original tool errors.
• Keep non-streaming buffering and ResponseStream.Collect error behavior unchanged. No new public APIs or options.
• Document existing per-call history support for injected images and test typed, serialized external storage without session in-memory history or duplicated messages.

.NET comparison

Checked stable Agent Framework dotnet-1.27.0 (4d1674b78951d6fd127ada532660514637f906e9), which uses Microsoft.Extensions.AI 10.10.0 and its OpenAI adapter 10.10.1.

• The OpenAI adapter accumulates tool fragments by index and emits completed calls after the stream ends.
• Non-streaming invocation returns one completed response. Message-injection retention also relies on per-call history.
• Cancellation exceptions escape the invocation processor before batch-result construction. This change deliberately preserves partial outcomes rather than reproducing that loss in Go.

Validation

• Full suites passed: provider/openaiprovider, agent/harness/toolautocall, and agent.
• Regression coverage includes packed terminal calls, interleaved arguments, terminal refusals, failed streams, completed side effects during cancellation, panics, concurrent dispatch, approvals, middleware reuse, and serialized image history.
• Configured gofumpt applied; compatible golangci-lint reports 0 new issues. Full-package lint has three pre-existing formatting findings in unchanged files.

[github-actions]

API consistency review

Result	Scope
unable to determine	user-visible behavior

Changed Go contract

No exported API changes (per the PR description). Observable behavior changes:

• The OpenAI provider now assembles streamed tool calls and refusals once the stream finishes.
• toolautocall and the agent now emit completed and unexecuted tool outcomes before terminal errors.
• Non-streaming buffering and ResponseStream.Collect error behavior are unchanged.

Upstream evidence reviewed

No upstream sources were independently reviewed in this run. The PR description cites the .NET MEAI OpenAI adapter and FunctionInvokingChatClient, but I did not verify those references.

Assessment

I could not complete a verified comparison with the .NET and Python implementations, so I am reporting no findings. The PR states that Go deliberately preserves partial tool outcomes on cancellation, where .NET lets the exception escape before building the batch result. Maintainers should confirm that this divergence is acceptable and check the Python OpenAI streaming tool-call handling manually. No parity label was added.

Generated by Go API Consistency Review Agent for #1349 · copilot · auto · 14 AIC · ⌖ 0.616 AIC · ⊞ 4.9K · ◷

[Copilot]

🟡 Changes recommended

Unresolved issues allow truncated OpenAI streams to trigger execution of unfinished tool calls and tool-loop failures to label messages with stale conversation IDs.

2 open issues

• Reject incomplete tool calls when the stream ends at EOF · New
• Update conversation history ID on tool-loop failures · New

What changed in this PR

This PR improves streamed OpenAI response assembly and preserves partial tool outcomes when automatic invocation stops, without adding public APIs.

Changes:

• Assemble tool calls and refusals after streaming ends, preserving text, usage, metadata, and raw chunks.
• Emit completed and unexecuted tool outcomes before terminal errors, including buffered runs and approval resumption.
• Document and test injected-image retention through serialized external history.

File	Description
provider/​openaiprovider/​chat.go	Emit accumulated tool calls and refusals after streaming.
provider/​openaiprovider/​chat_test.go	Cover packed calls, interleaved arguments, refusals, and stream failures.
docs/​agent-harness.md	Explain partial outcomes, cancellation, and external history.
agent/​messageinjection.go	Document persistence requirements for injected content.
agent/​messageinjection_test.go	Test serialized image history across tool rounds.
agent/​harness/​toolautocall/​autocall.go	Preserve tool outcomes and buffered updates before terminal errors.
agent/​harness/​toolautocall/​autocall_test.go	Cover cancellation, approvals, concurrent outcomes, and provider failures.

_{Review effort: Balanced}