@vybestack/llxprt-code-core 0.7.0-nightly.251211.134f1920b → 0.7.0-nightly.251211.f9e1b74e4

package/dist/src/providers/openai/OpenAIProvider.js

@@ -1557,6 +1557,8 @@ export class OpenAIProvider extends BaseProvider {
             let streamingUsage = null;
             // Track total chunks for debugging empty responses
             let totalChunksReceived = 0;
+            // Track finish_reason for detecting empty responses (issue #584)
+            let lastFinishReason = null;
             try {
                 // Handle streaming response
                 for await (const chunk of response) {
@@ -1635,6 +1637,7 @@ export class OpenAIProvider extends BaseProvider {
                     }
                     // Check for finish_reason to detect proper stream ending
                     if (choice.finish_reason) {
+                        lastFinishReason = choice.finish_reason;
                         logger.debug(() => `[Streaming] Stream finished with reason: ${choice.finish_reason}`, {
                             model,
                             finishReason: choice.finish_reason,
@@ -2086,6 +2089,28 @@ export class OpenAIProvider extends BaseProvider {
                     },
                 };
             }
+            // Detect and handle empty streaming responses after tool calls (issue #584)
+            // Some models (like gpt-oss-120b on OpenRouter) return finish_reason=stop with tools but no text
+            const hasToolsButNoText = lastFinishReason === 'stop' &&
+                accumulatedToolCalls.length > 0 &&
+                _accumulatedText.length === 0 &&
+                textBuffer.length === 0 &&
+                accumulatedReasoningContent.length === 0 &&
+                accumulatedThinkingContent.length === 0;
+            if (hasToolsButNoText) {
+                logger.log(() => `[OpenAIProvider] Model returned tool calls but no text (finish_reason=stop). Requesting continuation for model '${model}'.`, {
+                    model,
+                    toolCallCount: accumulatedToolCalls.length,
+                    baseURL: baseURL ?? this.getBaseURL(),
+                });
+                // Request continuation after tool calls (delegated to shared method)
+                const toolCallsForContinuation = accumulatedToolCalls.map((tc) => ({
+                    id: tc.id,
+                    type: tc.type,
+                    function: tc.function,
+                }));
+                yield* this.requestContinuationAfterToolCalls(toolCallsForContinuation, messagesWithSystem, requestBody, client, abortSignal, model, logger, customHeaders);
+            }
             // Detect and warn about empty streaming responses (common with Kimi K2 after tool calls)
             // Only warn if we truly got nothing - not even reasoning content
             if (_accumulatedText.length === 0 &&
@@ -2723,6 +2748,10 @@ export class OpenAIProvider extends BaseProvider {
             let accumulatedReasoningContent = '';
             // Track token usage from streaming chunks
             let streamingUsage = null;
+            // Track finish_reason for detecting empty responses (issue #584)
+            let lastFinishReason = null;
+            // Store pipeline result to avoid duplicate process() calls (CodeRabbit review #764)
+            let cachedPipelineResult = null;
             const allChunks = []; // Collect all chunks first
             try {
                 // Handle streaming response - collect all chunks
@@ -2805,6 +2834,7 @@ export class OpenAIProvider extends BaseProvider {
                     }
                     // Check for finish_reason to detect proper stream ending
                     if (choice.finish_reason) {
+                        lastFinishReason = choice.finish_reason;
                         logger.debug(() => `[Streaming] Stream finished with reason: ${choice.finish_reason}`, {
                             model,
                             finishReason: choice.finish_reason,
@@ -3175,12 +3205,12 @@ export class OpenAIProvider extends BaseProvider {
                 }
             }
             // Process and emit tool calls using the pipeline
-            const pipelineResult = await this.toolCallPipeline.process(abortSignal);
-            if (pipelineResult.normalized.length > 0 ||
-                pipelineResult.failed.length > 0) {
+            cachedPipelineResult = await this.toolCallPipeline.process(abortSignal);
+            if (cachedPipelineResult.normalized.length > 0 ||
+                cachedPipelineResult.failed.length > 0) {
                 const blocks = [];
                 // Process successful tool calls
-                for (const normalizedCall of pipelineResult.normalized) {
+                for (const normalizedCall of cachedPipelineResult.normalized) {
                     const sanitizedArgs = this.sanitizeToolArgumentsString(normalizedCall.originalArgs ?? normalizedCall.args);
                     // Process tool parameters with double-escape handling
                     const processedParameters = processToolParameters(sanitizedArgs, normalizedCall.name);
@@ -3192,7 +3222,7 @@ export class OpenAIProvider extends BaseProvider {
                     });
                 }
                 // Handle failed tool calls (could emit as errors or warnings)
-                for (const failed of pipelineResult.failed) {
+                for (const failed of cachedPipelineResult.failed) {
                     this.getLogger().warn(`Tool call validation failed for index ${failed.index}: ${failed.validationErrors.join(', ')}`);
                 }
                 if (blocks.length > 0) {
@@ -3240,11 +3270,44 @@ export class OpenAIProvider extends BaseProvider {
                     },
                 };
             }
+            // Detect and handle empty streaming responses after tool calls (issue #584)
+            // Some models (like gpt-oss-120b on OpenRouter) return finish_reason=stop with tools but no text
+            // Use cachedPipelineResult instead of pipelineStats.collector.totalCalls since process() resets the collector (CodeRabbit review #764)
+            const toolCallCount = (cachedPipelineResult?.normalized.length ?? 0) +
+                (cachedPipelineResult?.failed.length ?? 0);
+            const hasToolsButNoText = lastFinishReason === 'stop' &&
+                toolCallCount > 0 &&
+                _accumulatedText.length === 0 &&
+                textBuffer.length === 0 &&
+                accumulatedReasoningContent.length === 0 &&
+                accumulatedThinkingContent.length === 0;
+            if (hasToolsButNoText) {
+                logger.log(() => `[OpenAIProvider] Model returned tool calls but no text (finish_reason=stop). Requesting continuation for model '${model}'.`, {
+                    model,
+                    toolCallCount,
+                    baseURL: baseURL ?? this.getBaseURL(),
+                });
+                // Note: In pipeline mode, tool calls have already been processed.
+                // We need to get the normalized tool calls from the cached pipeline result to build continuation messages.
+                // Use cached result to avoid duplicate process() call that would return empty results (CodeRabbit review #764)
+                if (!cachedPipelineResult) {
+                    throw new Error('Pipeline result not cached - this should not happen in pipeline mode');
+                }
+                const toolCallsForHistory = cachedPipelineResult.normalized.map((normalizedCall, index) => ({
+                    id: `call_${index}`,
+                    type: 'function',
+                    function: {
+                        name: normalizedCall.name,
+                        arguments: JSON.stringify(normalizedCall.args),
+                    },
+                }));
+                // Request continuation after tool calls (delegated to shared method)
+                yield* this.requestContinuationAfterToolCalls(toolCallsForHistory, messagesWithSystem, requestBody, client, abortSignal, model, logger, customHeaders);
+            }
             // Detect and warn about empty streaming responses (common with Kimi K2 after tool calls)
             // Only warn if we truly got nothing - not even reasoning content
-            const pipelineStats = this.toolCallPipeline.getStats();
             if (_accumulatedText.length === 0 &&
-                pipelineStats.collector.totalCalls === 0 &&
+                toolCallCount === 0 &&
                 textBuffer.length === 0 &&
                 accumulatedReasoningContent.length === 0 &&
                 accumulatedThinkingContent.length === 0) {
@@ -3268,7 +3331,7 @@ export class OpenAIProvider extends BaseProvider {
                 // Log what we DID get for debugging
                 logger.debug(() => `[Streaming pipeline] Stream completed with accumulated content`, {
                     textLength: _accumulatedText.length,
-                    toolCallCount: pipelineStats.collector.totalCalls,
+                    toolCallCount,
                     textBufferLength: textBuffer.length,
                     reasoningLength: accumulatedReasoningContent.length,
                     thinkingLength: accumulatedThinkingContent.length,
@@ -3612,5 +3675,83 @@ export class OpenAIProvider extends BaseProvider {
             };
         return { thinking: thinkingBlock, toolCalls };
     }
+    /**
+     * Request continuation after tool calls when model returned no text.
+     * This is a helper to avoid code duplication between legacy and pipeline paths.
+     *
+     * @plan PLAN-20250120-DEBUGLOGGING.P15
+     * @issue #584, #764 (CodeRabbit review)
+     */
+    async *requestContinuationAfterToolCalls(toolCalls, messagesWithSystem, requestBody, client, abortSignal, model, logger, customHeaders) {
+        // Build continuation messages
+        const continuationMessages = [
+            ...messagesWithSystem,
+            // Add the assistant's tool calls
+            {
+                role: 'assistant',
+                tool_calls: toolCalls,
+            },
+            // Add placeholder tool responses (tools have NOT been executed yet - only acknowledged)
+            ...toolCalls.map((tc) => ({
+                role: 'tool',
+                tool_call_id: tc.id,
+                content: '[Tool call acknowledged - awaiting execution]',
+            })),
+            // Add continuation prompt
+            {
+                role: 'user',
+                content: 'The tool calls above have been registered. Please continue with your response.',
+            },
+        ];
+        // Make a continuation request (wrap in try-catch since tools were already yielded)
+        try {
+            const continuationResponse = await client.chat.completions.create({
+                ...requestBody,
+                messages: continuationMessages,
+                stream: true, // Always stream for consistency
+            }, {
+                ...(abortSignal ? { signal: abortSignal } : {}),
+                ...(customHeaders ? { headers: customHeaders } : {}),
+            });
+            let accumulatedText = '';
+            // Process the continuation response
+            for await (const chunk of continuationResponse) {
+                if (abortSignal?.aborted) {
+                    break;
+                }
+                const choice = chunk.choices?.[0];
+                if (!choice)
+                    continue;
+                const deltaContent = this.coerceMessageContentToString(choice.delta?.content);
+                if (deltaContent) {
+                    const sanitized = this.sanitizeProviderText(deltaContent);
+                    if (sanitized) {
+                        accumulatedText += sanitized;
+                        yield {
+                            speaker: 'ai',
+                            blocks: [
+                                {
+                                    type: 'text',
+                                    text: sanitized,
+                                },
+                            ],
+                        };
+                    }
+                }
+            }
+            logger.debug(() => `[OpenAIProvider] Continuation request completed, received ${accumulatedText.length} chars`, {
+                model,
+                accumulatedTextLength: accumulatedText.length,
+            });
+        }
+        catch (continuationError) {
+            // Tool calls were already successfully yielded, so log warning and continue
+            logger.warn(() => `[OpenAIProvider] Continuation request failed, but tool calls were already emitted: ${continuationError instanceof Error ? continuationError.message : String(continuationError)}`, {
+                model,
+                error: continuationError,
+            });
+            // Don't re-throw - tool calls were already successful
+        }
+    }
 }
 //# sourceMappingURL=OpenAIProvider.js.map