fix: preserve the stream property for chat/completions calls (#164)

* fix: preserve the stream property for chat/completions calls

* test: add request body validation to mock server

* document aibridge stream mashalling behaviour for chat completions

---------

Co-authored-by: Susana Cardoso Ferreira <susana@coder.com>

Author: SasSwart

bridge_integration_test.go

@@ -2076,6 +2076,35 @@ func newMockServer(ctx context.Context, t *testing.T, files archiveFileMap, requ
 		defer r.Body.Close()
 		require.NoError(t, err)
 
+		// Validate request body based on endpoint.
+		var validationErr error
+		if strings.Contains(r.URL.Path, "/chat/completions") {
+			validationErr = validateOpenAIChatCompletionRequest(body)
+		} else if strings.Contains(r.URL.Path, "/responses") {
+			validationErr = validateOpenAIResponsesRequest(body)
+		} else if strings.Contains(r.URL.Path, "/messages") {
+			validationErr = validateAnthropicMessagesRequest(body)
+		}
+
+		// If validation failed, return error response
+		if validationErr != nil {
+			// Return HTTP error response
+			w.Header().Set("Content-Type", "application/json")
+			w.WriteHeader(http.StatusBadRequest)
+			errResp := map[string]any{
+				"error": map[string]any{
+					"message": fmt.Sprintf("Request #%d validation failed: %v", ms.callCount.Load(), validationErr),
+					"type":    "invalid_request_error",
+				},
+			}
+			json.NewEncoder(w).Encode(errResp)
+
+			// Mark test as failed with detailed message
+			t.Errorf("Request #%d validation failed: %v\n\nRequest body:\n%s",
+				ms.callCount.Load(), validationErr, string(body))
+			return
+		}
+
 		type msg struct {
 			Stream bool `json:"stream"`
 		}
@@ -2135,6 +2164,72 @@ func newMockServer(ctx context.Context, t *testing.T, files archiveFileMap, requ
 	return ms
 }
 
+// validateOpenAIChatCompletionRequest validates that an OpenAI chat completion request
+// has all required fields. Returns an error if validation fails.
+func validateOpenAIChatCompletionRequest(body []byte) error {
+	var req openai.ChatCompletionNewParams
+	if err := json.Unmarshal(body, &req); err != nil {
+		return fmt.Errorf("request should unmarshal into ChatCompletionNewParams: %w", err)
+	}
+
+	// Collect all validation errors
+	var errs []string
+	if req.Model == "" {
+		errs = append(errs, "model field is required but empty")
+	}
+	if len(req.Messages) == 0 {
+		errs = append(errs, "messages field is required but empty")
+	}
+
+	if len(errs) > 0 {
+		return fmt.Errorf("validation failed: %s", strings.Join(errs, "; "))
+	}
+	return nil
+}
+
+// validateOpenAIResponsesRequest validates that an OpenAI responses request
+// has all required fields. Returns an error if validation fails.
+func validateOpenAIResponsesRequest(body []byte) error {
+	var reqBody map[string]any
+	if err := json.Unmarshal(body, &reqBody); err != nil {
+		return fmt.Errorf("request should be valid JSON: %w", err)
+	}
+
+	// Verify required fields for OpenAI responses
+	// Note: Using map here since there's no specific SDK type for responses endpoint
+	model, ok := reqBody["model"]
+	if !ok || model == "" {
+		return fmt.Errorf("model field is required but missing or empty")
+	}
+	return nil
+}
+
+// validateAnthropicMessagesRequest validates that an Anthropic messages request
+// has all required fields. Returns an error if validation fails.
+func validateAnthropicMessagesRequest(body []byte) error {
+	var req anthropic.MessageNewParams
+	if err := json.Unmarshal(body, &req); err != nil {
+		return fmt.Errorf("request should unmarshal into MessageNewParams: %w", err)
+	}
+
+	// Collect all validation errors
+	var errs []string
+	if req.Model == "" {
+		errs = append(errs, "model field is required but empty")
+	}
+	if len(req.Messages) == 0 {
+		errs = append(errs, "messages field is required but empty")
+	}
+	if req.MaxTokens == 0 {
+		errs = append(errs, "max_tokens field is required but zero")
+	}
+
+	if len(errs) > 0 {
+		return fmt.Errorf("validation failed: %s", strings.Join(errs, "; "))
+	}
+	return nil
+}
+
 const mockToolName = "coder_list_workspaces"
 
 // callAccumulator tracks all tool invocations by name and each instance's arguments.

go.mod

@@ -94,4 +94,4 @@ require (
 replace github.com/anthropics/anthropic-sdk-go v1.13.0 => github.com/dannykopping/anthropic-sdk-go v0.0.0-20251230111224-88a4315810bd
 
 // https://github.com/openai/openai-go/pull/602
-replace github.com/openai/openai-go/v3 => github.com/SasSwart/openai-go/v3 v3.0.0-20260202093810-72af3b857f95
+replace github.com/openai/openai-go/v3 => github.com/SasSwart/openai-go/v3 v3.0.0-20260204134041-fb987b42a728

go.sum

@@ -7,8 +7,8 @@ cloud.google.com/go/logging v1.8.1 h1:26skQWPeYhvIasWKm48+Eq7oUqdcdbwsCVwz5Ys0Fv
 cloud.google.com/go/logging v1.8.1/go.mod h1:TJjR+SimHwuC8MZ9cjByQulAMgni+RkXeI3wwctHJEI=
 cloud.google.com/go/longrunning v0.5.1 h1:Fr7TXftcqTudoyRJa113hyaqlGdiBQkp0Gq7tErFDWI=
 cloud.google.com/go/longrunning v0.5.1/go.mod h1:spvimkwdz6SPWKEt/XBij79E9fiTkHSQl/fRUUQJYJc=
-github.com/SasSwart/openai-go/v3 v3.0.0-20260202093810-72af3b857f95 h1:HVJp3FanNaeFAlwg0/lkdkSnwFemHnwwjXBM8KRj540=
-github.com/SasSwart/openai-go/v3 v3.0.0-20260202093810-72af3b857f95/go.mod h1:cdufnVK14cWcT9qA1rRtrXx4FTRsgbDPW7Ia7SS5cZo=
+github.com/SasSwart/openai-go/v3 v3.0.0-20260204134041-fb987b42a728 h1:FOjd3xOH+arcrtz1e5P6WZ/VtRD5KQHHRg4kc4BZers=
+github.com/SasSwart/openai-go/v3 v3.0.0-20260204134041-fb987b42a728/go.mod h1:cdufnVK14cWcT9qA1rRtrXx4FTRsgbDPW7Ia7SS5cZo=
 github.com/aws/aws-sdk-go-v2 v1.30.3 h1:jUeBtG0Ih+ZIFH0F4UkmL9w3cSpaMv9tYYDbzILP8dY=
 github.com/aws/aws-sdk-go-v2 v1.30.3/go.mod h1:nIQjQVp5sfpQcTc9mPSr1B0PaWK5ByX9MOoDadSN4lc=
 github.com/aws/aws-sdk-go-v2/aws/protocol/eventstream v1.6.3 h1:tW1/Rkad38LA15X4UQtjXZXNKsCgkshC3EbmcUmghTg=

intercept/chatcompletions/streaming.go

@@ -125,6 +125,16 @@ func (i *StreamingInterception) ProcessRequest(w http.ResponseWriter, r *http.Re
 			opts = append(opts, intercept.ActorHeadersAsOpenAIOpts(actor)...)
 		}
 
+		// We take control of request body here and pass it to the SDK as a raw byte slice.
+		// This is because the SDK's serialization applies hidden request options that result in
+		// unexpected, breaking behaviour. See https://github.com/coder/aibridge/pull/164
+		body, err := json.Marshal(i.req.ChatCompletionNewParams)
+		if err != nil {
+			return fmt.Errorf("marshal request body: %w", err)
+		}
+		opts = append(opts, option.WithRequestBody("application/json", body))
+		opts = append(opts, option.WithJSONSet("stream", true))
+
 		stream = i.newStream(streamCtx, svc, opts)
 		processor := newStreamProcessor(streamCtx, i.logger.Named("stream-processor"), i.getInjectedToolByName)
 
@@ -380,7 +390,7 @@ func (i *StreamingInterception) newStream(ctx context.Context, svc openai.ChatCo
 	_, span := i.tracer.Start(ctx, "Intercept.ProcessRequest.Upstream", trace.WithAttributes(tracing.InterceptionAttributesFromContext(ctx)...))
 	defer span.End()
 
-	return svc.NewStreaming(ctx, i.req.ChatCompletionNewParams, opts...)
+	return svc.NewStreaming(ctx, openai.ChatCompletionNewParams{}, opts...)
 }
 
 type streamProcessor struct {