Surface chat system_fingerprint and logprobs as response metadata

[PratikDhanave (PratikDhanave)]

The Chat path built its ResponseUpdate with ResponseID/FinishReason/usage but never set AdditionalProperties, so two pieces of response metadata were dropped:

• system_fingerprint (on the completion / chunk), and
• per-choice logprobs (returned when a caller requests them via Logprobs/TopLogprobs).

The Python client carries both as chat response metadata (_get_metadata_from_chat_response → system_fingerprint, _get_metadata_from_chat_choice → logprobs). The Responses path in this provider already populates AdditionalProperties, so this was also a chat-vs-responses asymmetry — the existing chat fixtures even contain system_fingerprint/logprobs values that nothing surfaced.

Change

• Populate AdditionalProperties with SystemFingerprint (when non-empty) and Logprobs (when the choice carries content/refusal logprobs) on both the non-streaming and streaming paths. Keys follow the PascalCase convention already used by the Responses path helper (responsesPopulateAdditionalProperties).

Tests

• TestChatResponseMetadataSurfaced_NonStreaming and _Streaming: assert SystemFingerprint (and, non-streaming, Logprobs) reach the message AdditionalProperties. Both fail before the change (no message carries any), pass after.

[Copilot]

🟡 Changes recommended

Streaming logprobs may be overwritten across chunks, and this path lacks coverage.

Get a fresh assessment by requesting another Copilot review.

Pull request overview

Surfaces Chat response system_fingerprint and logprobs metadata through AdditionalProperties.

Changes:

• Propagates metadata for streaming and non-streaming responses.
• Adds tests for metadata exposure.

File summaries

File	Summary
provider/openaiprovider/chat.go	Populates Chat response metadata.
provider/openaiprovider/chat_test.go	Tests metadata propagation.

Review details

Suppressed comments (2)

provider/openaiprovider/chat.go:227

• The new streaming Logprobs branch is not exercised: the added streaming test only sends/asserts SystemFingerprint, while all existing streaming fixtures use logprobs: null. Add a streamed chunk containing content or refusal logprobs and assert that the yielded update (and collected message) preserves the Logprobs metadata, so this path cannot regress unnoticed.

				if logprobs := chunk.Choices[0].Logprobs; len(logprobs.Content) > 0 || len(logprobs.Refusal) > 0 {
					additionalProperties = map[string]any{"Logprobs": logprobs}
				}

provider/openaiprovider/chat_test.go:457

• This streaming test only sends system_fingerprint; it never supplies a non-empty chunk logprobs value, so it would pass if the new streaming Logprobs handling were removed or broken. Add a chunk with content/refusal logprobs and assert that the streamed update (and, if collection is supported, the collected response) preserves them.

func TestChatResponseMetadataSurfaced_Streaming(t *testing.T) {
	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
		w.Header().Set("Content-Type", "text/event-stream")
		_, _ = io.WriteString(w, "data: {\"id\":\"chatcmpl-md\",\"object\":\"chat.completion.chunk\",\"created\":1727888631,\"model\":\"gpt-4o-mini\",\"system_fingerprint\":\"fp_test\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"ok\"},\"finish_reason\":\"stop\"}]}\n\n")

• Files reviewed: 2/2 changed files
• Comments generated: 1
• Review effort level: Lite

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[github-actions]

Scope: user-visible behavior (bug fix)

Changed Go contract: No exported API added/changed. chatClient.run in provider/openaiprovider/chat.go now populates the existing agent.ResponseUpdate.AdditionalProperties field with SystemFingerprint (from ChatCompletion.SystemFingerprint / ChatCompletionChunk.SystemFingerprint) and Logprobs (from Choice.Logprobs / ChunkChoice.Logprobs) on both the non-streaming and streaming chat paths, matching what the Responses path already does via responsesPopulateAdditionalProperties.

Upstream evidence reviewed:

• Python: python/packages/openai/agent_framework_openai/_chat_completion_client.py, _get_metadata_from_chat_response (system_fingerprint) and _get_metadata_from_chat_choice (logprobs), both merged into response_metadata/chunk_metadata and passed as ChatResponse.additional_properties in _parse_response_from_openai (lines ~999-1015, ~871-895).
• .NET: No same-repo equivalent — Microsoft.Agents.AI.OpenAI builds on the external Microsoft.Extensions.AI OpenAIChatClient, which is not vendored in this repository, so there is no dotnet/src/Microsoft.Agents.AI.OpenAI code path to compare against for this specific mapping.

Result: Aligned. The Go change closes a genuine chat-vs-responses asymmetry (the Responses path already surfaced comparable metadata via responsesPopulateAdditionalProperties) and matches the Python client's behavior of surfacing system_fingerprint and logprobs as response/message metadata. Key naming (SystemFingerprint, Logprobs) follows the existing Go PascalCase convention used by responsesPopulateAdditionalProperties. No exported Go API surface changed (only internal population of an existing field), so public-api-change is not applicable.

Generated by Go API Consistency Review Agent · copilot · auto · 34.8 AIC · ⌖ 8.27 AIC · ⊞ 9.2K · ◷