fix(provider): retain full request on continuation mismatch

commit: 59535ff0d74e48c04c85803ea8657a74727c07f3
author: brice · 2mo ago · 2026-08-18 09:25 UTC
parent: 54f793e
signature: GPG VerifiedCD88F9E564E0102646FB6077A1A07C7C7C3E349D · arch-gpg Verified for the committer email

2 files · +32 -1

M packages/provider/openai_responses_websocket.go +2 -1


@@ -467,10 +467,11 @@ func (s *responsesWebSocketSession) incrementalRequest(req Request, cacheSession
467	467		// The cached response already contains the exact assistant output from the
468	468		// last generation. It must be replayed verbatim in the caller transcript
469	469		// before we can safely send only the strict new tool-result/user suffix.
	470	+	fullRequest := req
470	471		req.Messages = append([]Message(nil), req.Messages[len(previous.Messages):]...)
471	472		if len(req.Messages) > 0 && req.Messages[0].Role == RoleAssistant {
472	473		if !sameResponsesOutput(req.Messages[0], s.lastOutput) {
473		-	return req, ""
	474	+	return fullRequest, ""
474	475		}
475	476		req.Messages = req.Messages[1:]
476	477		}

 	// The cached response already contains the exact assistant output from the
 	// last generation. It must be replayed verbatim in the caller transcript
 	// before we can safely send only the strict new tool-result/user suffix.
+	fullRequest := req
 	req.Messages = append([]Message(nil), req.Messages[len(previous.Messages):]...)
 	if len(req.Messages) > 0 && req.Messages[0].Role == RoleAssistant {
 		if !sameResponsesOutput(req.Messages[0], s.lastOutput) {
-			return req, ""
+			return fullRequest, ""
 		}
 		req.Messages = req.Messages[1:]
 	}

M packages/provider/openai_responses_websocket_test.go +30 -0


@@ -261,6 +261,36 @@ func TestResponsesWebSocketWireBaselineChangeForcesFullRequest(t *testing.T) {
261	261		}
262	262		}
263	263
	264	+	func TestResponsesWebSocketAssistantMismatchForcesCompleteTranscript(t *testing.T) {
	265	+	client := newResponsesWebSocketClient(&codexClient{
	266	+	providerName: "openai", capabilities: responsesCapabilities{StablePromptCacheKey: true},
	267	+	}).(*responsesWebSocketClient)
	268	+	first := Request{
	269	+	Model: "gpt-5.6-sol", Context: RequestContext{CacheSessionID: "cache-root-1", ThreadID: "thread-1", TurnID: "turn-1"},
	270	+	Messages: []Message{{Role: RoleUser, Content: []Content{TextBlock{Text: "first"}}}},
	271	+	}
	272	+	baseline, err := client.canonicalResponsesBaseline(first, first.Context.CacheSessionID)
	273	+	if err != nil {
	274	+	t.Fatal(err)
	275	+	}
	276	+	session := &responsesWebSocketSession{}
	277	+	session.recordCompleted(first, baseline, Message{Role: RoleAssistant, Content: []Content{TextBlock{Text: "actual"}}}, "resp_1")
	278	+	second := Request{
	279	+	Model: first.Model, Context: RequestContext{CacheSessionID: "cache-root-1", ThreadID: "thread-1", TurnID: "turn-2"},
	280	+	Messages: append(append([]Message(nil), first.Messages...),
	281	+	Message{Role: RoleAssistant, Content: []Content{TextBlock{Text: "modified"}}},
	282	+	Message{Role: RoleUser, Content: []Content{TextBlock{Text: "second"}}},
	283	+	),
	284	+	}
	285	+	streamReq, previousResponseID := session.incrementalRequest(second, second.Context.CacheSessionID, client.canonicalResponsesBaseline)
	286	+	if previousResponseID != "" {
	287	+	t.Fatalf("previous response ID = %q, want complete request", previousResponseID)
	288	+	}
	289	+	if !reflect.DeepEqual(streamReq.Messages, second.Messages) {
	290	+	t.Fatalf("stream messages = %#v, want complete transcript %#v", streamReq.Messages, second.Messages)
	291	+	}
	292	+	}
	293	+
264	294		func TestResponsesWebSocketCancellationClosesOnlyActiveConnection(t *testing.T) {
265	295		closed := make(chan struct{})
266	296		var closeOnce sync.Once


@@ -261,6 +261,36 @@ func TestResponsesWebSocketWireBaselineChangeForcesFullRequest(t *testing.T) {
261	}	261		}
262	}	262		}
263		263
		264	+	func TestResponsesWebSocketAssistantMismatchForcesCompleteTranscript(t *testing.T) {
		265	+	client := newResponsesWebSocketClient(&codexClient{
		266	+	providerName: "openai", capabilities: responsesCapabilities{StablePromptCacheKey: true},
		267	+	}).(*responsesWebSocketClient)
		268	+	first := Request{
		269	+	Model: "gpt-5.6-sol", Context: RequestContext{CacheSessionID: "cache-root-1", ThreadID: "thread-1", TurnID: "turn-1"},
		270	+	Messages: []Message{{Role: RoleUser, Content: []Content{TextBlock{Text: "first"}}}},
		271	+	}
		272	+	baseline, err := client.canonicalResponsesBaseline(first, first.Context.CacheSessionID)
		273	+	if err != nil {
		274	+	t.Fatal(err)
		275	+	}
		276	+	session := &responsesWebSocketSession{}
		277	+	session.recordCompleted(first, baseline, Message{Role: RoleAssistant, Content: []Content{TextBlock{Text: "actual"}}}, "resp_1")
		278	+	second := Request{
		279	+	Model: first.Model, Context: RequestContext{CacheSessionID: "cache-root-1", ThreadID: "thread-1", TurnID: "turn-2"},
		280	+	Messages: append(append([]Message(nil), first.Messages...),
		281	+	Message{Role: RoleAssistant, Content: []Content{TextBlock{Text: "modified"}}},
		282	+	Message{Role: RoleUser, Content: []Content{TextBlock{Text: "second"}}},
		283	+	),
		284	+	}
		285	+	streamReq, previousResponseID := session.incrementalRequest(second, second.Context.CacheSessionID, client.canonicalResponsesBaseline)
		286	+	if previousResponseID != "" {
		287	+	t.Fatalf("previous response ID = %q, want complete request", previousResponseID)
		288	+	}
		289	+	if !reflect.DeepEqual(streamReq.Messages, second.Messages) {
		290	+	t.Fatalf("stream messages = %#v, want complete transcript %#v", streamReq.Messages, second.Messages)
		291	+	}
		292	+	}
		293	+
264	func TestResponsesWebSocketCancellationClosesOnlyActiveConnection(t *testing.T) {	294		func TestResponsesWebSocketCancellationClosesOnlyActiveConnection(t *testing.T) {
265	closed := make(chan struct{})	295		closed := make(chan struct{})
266	var closeOnce sync.Once	296		var closeOnce sync.Once