fix and refactor

tymondesigns · tymondesigns · commit 9713edc6110c · 2026-01-04T15:08:56.000Z
diff --git a/src/Agents/Stages/HandleToolCalls.php b/src/Agents/Stages/HandleToolCalls.php
@@ -39,7 +39,9 @@ public function __construct(
     public function handlePipeable(mixed $payload, RuntimeConfig $config, Closure $next): mixed
     {
         return match (true) {
-            $payload instanceof ChatGenerationChunk && $payload->type === ChunkType::ToolInputEnd => $this->handleStreamingChunk($payload, $config, $next),
+            // We trigger on isFinal (not ToolInputEnd) so that AddMessageToMemoryMiddleware
+            // has already added the assistant message before we process tool calls.
+            $payload instanceof ChatGenerationChunk && $payload->isFinal && $payload->message->hasToolCalls() => $this->handleStreamingChunk($payload, $config, $next),
             $payload instanceof ChatStreamResult => $this->handleStreamingResult($payload),
             default => $this->handleNonStreaming($payload, $config, $next),
         };
@@ -158,7 +160,7 @@ protected function getGeneration(mixed $payload): ChatGeneration|ChatGenerationC
     {
         return match (true) {
             $payload instanceof ChatGeneration => $payload,
-            $payload instanceof ChatGenerationChunk && $payload->type === ChunkType::ToolInputEnd => $payload,
+            $payload instanceof ChatGenerationChunk && $payload->isFinal && $payload->message->hasToolCalls() => $payload,
             $payload instanceof ChatResult => $payload->generation,
             default => null,
         };
diff --git a/src/LLM/Drivers/Anthropic/AnthropicChat.php b/src/LLM/Drivers/Anthropic/AnthropicChat.php
@@ -98,8 +98,6 @@ protected function buildParams(array $additionalParameters): array
 
         if ($this->structuredOutputConfig !== null) {
             $this->structuredOutputMode = StructuredOutputMode::Tool;
-            $params['betas'] ??= [];
-            $params['betas'][] = 'structured-outputs-2025-11-13';
 
             $params['output_format'] = [
                 'type' => 'json_schema',
@@ -148,6 +146,11 @@ protected function buildParams(array $additionalParameters): array
             throw new LLMException('`max_tokens` parameter is required for Anthropic.');
         }
 
+        if ($this->structuredOutputConfig !== null) {
+            $finalParams['betas'] ??= [];
+            $finalParams['betas'][] = 'structured-outputs-2025-11-13';
+        }
+
         return $finalParams;
     }
 
@@ -162,10 +165,15 @@ public static function fake(
     ): self {
         $client = Anthropic::fake($responses, $apiKey);
 
-        return new self(
+        $instance = new self(
             $client,
             $model ?? 'claude-4-5-sonnet-20250926',
             $modelProvider ?? ModelProvider::Anthropic,
         );
+
+        // Set a default max_tokens for testing
+        $instance->withMaxTokens(8096);
+
+        return $instance;
     }
 }
diff --git a/src/LLM/Drivers/Anthropic/Concerns/MapStreamResponse.php b/src/LLM/Drivers/Anthropic/Concerns/MapStreamResponse.php
@@ -316,10 +316,10 @@ private function buildChunk(
                     finishReason: $finishReason,
                     usage: $usage,
                     processingTime: $meta?->processingTime,
-                    providerMetadata: $meta?->raw ?? [],
+                    providerMetadata: $meta->raw ?? [],
                 ),
             ),
-            createdAt: $meta?->createdAt ?? new DateTimeImmutable(),
+            createdAt: $meta->createdAt ?? new DateTimeImmutable(),
             finishReason: $finishReason,
             usage: $usage,
             contentSoFar: $this->buildContentSnapshot(),
diff --git a/src/LLM/Drivers/Anthropic/Concerns/MapsMessages.php b/src/LLM/Drivers/Anthropic/Concerns/MapsMessages.php
@@ -6,8 +6,8 @@
 
 use Cortex\Support\DataUrl;
 use Cortex\LLM\Contracts\Message;
-use Cortex\ModelInfo\Enums\ModelFeature;
 use Cortex\LLM\Enums\MessageRole;
+use Cortex\ModelInfo\Enums\ModelFeature;
 use Cortex\LLM\Data\Messages\ToolMessage;
 use Cortex\LLM\Data\Messages\AssistantMessage;
 use Cortex\LLM\Data\Messages\MessageCollection;
@@ -20,49 +20,80 @@ trait MapsMessages
     /**
      * Map the given messages to the Anthropic API format.
      *
+     * Anthropic requires consecutive tool results to be merged into a single
+     * user message, as it doesn't allow consecutive messages with the same role.
+     *
      * @return array<int, array<string, mixed>>
      */
     protected function mapMessagesForInput(MessageCollection $messages): array
     {
-        return $messages
-            ->map(fn (Message $message): array => $this->mapMessage($message))
-            ->values()
-            ->toArray();
+        $mapped = [];
+        $pendingToolResults = [];
+
+        foreach ($messages as $message) {
+            if ($message instanceof ToolMessage) {
+                $pendingToolResults[] = $this->mapToolResultBlock($message);
+
+                continue;
+            }
+
+            // Flush any pending tool results before adding the next message
+            if ($pendingToolResults !== []) {
+                $mapped[] = $this->createToolResultsMessage($pendingToolResults);
+                $pendingToolResults = [];
+            }
+
+            $mapped[] = $this->mapMessage($message);
+        }
+
+        // Flush any remaining tool results at the end
+        if ($pendingToolResults !== []) {
+            $mapped[] = $this->createToolResultsMessage($pendingToolResults);
+        }
+
+        return $mapped;
     }
 
     /**
+     * Create a single user message containing all tool results.
+     *
+     * @param array<int, array<string, mixed>> $toolResults
+     *
      * @return array<string, mixed>
      */
-    private function mapMessage(Message $message): array
+    private function createToolResultsMessage(array $toolResults): array
     {
-        return match (true) {
-            $message instanceof ToolMessage => $this->mapToolMessage($message),
-            $message instanceof AssistantMessage => $this->mapAssistantMessage($message),
-            default => $this->mapGenericMessage($message),
-        };
+        return [
+            'role' => MessageRole::User->value,
+            'content' => $toolResults,
+        ];
     }
 
     /**
-     * Map a tool message (tool result) to Anthropic format.
-     *
-     * Tool results must be sent as a user message with tool_result content blocks.
+     * Map a tool message to a tool_result content block.
      *
      * @return array<string, mixed>
      */
-    private function mapToolMessage(ToolMessage $message): array
+    private function mapToolResultBlock(ToolMessage $message): array
     {
         return [
-            'role' => MessageRole::User->value,
-            'content' => [
-                [
-                    'type' => 'tool_result',
-                    'tool_use_id' => $message->id,
-                    'content' => $message->text() ?? '',
-                ],
-            ],
+            'type' => 'tool_result',
+            'tool_use_id' => $message->id,
+            'content' => $message->text() ?? '',
         ];
     }
 
+    /**
+     * @return array<string, mixed>
+     */
+    private function mapMessage(Message $message): array
+    {
+        return match (true) {
+            $message instanceof AssistantMessage => $this->mapAssistantMessage($message),
+            default => $this->mapGenericMessage($message),
+        };
+    }
+
     /**
      * Map an assistant message to Anthropic format.
      *
@@ -130,7 +161,7 @@ private function mapMessageContent(string|array|null $content): array
         }
 
         return array_values(array_filter(
-            array_map(fn (mixed $item): ?array => $this->mapContentBlock($item), $content),
+            array_map(fn(mixed $item): ?array => $this->mapContentBlock($item), $content),
         ));
     }
 
@@ -196,4 +227,3 @@ private function isDataUrl(string $url): bool
         return str_starts_with($url, 'data:');
     }
 }
-
diff --git a/tests/Unit/LLM/Drivers/Anthropic/AnthropicChatTest.php b/tests/Unit/LLM/Drivers/Anthropic/AnthropicChatTest.php
@@ -498,7 +498,9 @@ public function __construct(
         new UserMessage('What is the weather?'),
         new AssistantMessage(
             toolCalls: new ToolCallCollection([
-                new ToolCall('tool_123', new FunctionCall('get_weather', ['location' => 'London'])),
+                new ToolCall('tool_123', new FunctionCall('get_weather', [
+                    'location' => 'London',
+                ])),
             ]),
         ),
         new ToolMessage('The weather in London is sunny and 22°C', 'tool_123'),
@@ -527,7 +529,9 @@ public function __construct(
         new AssistantMessage(
             content: 'Let me check the weather for you.',
             toolCalls: new ToolCallCollection([
-                new ToolCall('tool_abc', new FunctionCall('get_weather', ['location' => 'Paris'])),
+                new ToolCall('tool_abc', new FunctionCall('get_weather', [
+                    'location' => 'Paris',
+                ])),
             ]),
         ),
         new ToolMessage('Rainy, 15°C', 'tool_abc'),
@@ -551,7 +555,9 @@ public function __construct(
             if ($block['type'] === 'tool_use'
                 && $block['id'] === 'tool_abc'
                 && $block['name'] === 'get_weather'
-                && $block['input'] === ['location' => 'Paris']) {
+                && $block['input'] === [
+                    'location' => 'Paris',
+                ]) {
                 $hasToolUse = true;
             }
         }
@@ -569,8 +575,12 @@ public function __construct(
         new UserMessage('Compare weather in London and Paris'),
         new AssistantMessage(
             toolCalls: new ToolCallCollection([
-                new ToolCall('tool_1', new FunctionCall('get_weather', ['location' => 'London'])),
-                new ToolCall('tool_2', new FunctionCall('get_weather', ['location' => 'Paris'])),
+                new ToolCall('tool_1', new FunctionCall('get_weather', [
+                    'location' => 'London',
+                ])),
+                new ToolCall('tool_2', new FunctionCall('get_weather', [
+                    'location' => 'Paris',
+                ])),
             ]),
         ),
         new ToolMessage('Sunny, 20°C', 'tool_1'),
@@ -585,7 +595,7 @@ public function __construct(
         // Should have two tool_use blocks
         $toolUseBlocks = array_filter(
             $assistantMessage['content'],
-            fn ($block) => $block['type'] === 'tool_use',
+            fn(array $block): bool => $block['type'] === 'tool_use',
         );
 
         return count($toolUseBlocks) === 2;
@@ -658,7 +668,7 @@ public function __construct(
     $llm->addFeature(ModelFeature::Vision);
 
     $base64Data = base64_encode('fake-image-data');
-    $dataUrl = "data:image/png;base64,{$base64Data}";
+    $dataUrl = 'data:image/png;base64,' . $base64Data;
 
     $llm->invoke([
         new UserMessage([
@@ -725,3 +735,105 @@ public function __construct(
             && $assistantMessage['content'][0]['text'] === 'Hi there!';
     });
 });
+
+test('it merges consecutive tool messages into a single user message', function (): void {
+    $llm = AnthropicChat::fake([
+        CreateMessage::class => MockResponse::fixture('anthropic/messages/simple'),
+    ]);
+
+    $llm->invoke([
+        new UserMessage('Compare weather in London and Paris'),
+        new AssistantMessage(
+            toolCalls: new ToolCallCollection([
+                new ToolCall('tool_1', new FunctionCall('get_weather', [
+                    'location' => 'London',
+                ])),
+                new ToolCall('tool_2', new FunctionCall('get_weather', [
+                    'location' => 'Paris',
+                ])),
+            ]),
+        ),
+        new ToolMessage('Sunny, 20°C', 'tool_1'),
+        new ToolMessage('Rainy, 15°C', 'tool_2'),
+    ]);
+
+    MockClient::getGlobal()->assertSent(function (CreateMessage $request): bool {
+        $messages = $request->body()->get('messages');
+
+        // Should have exactly 3 messages: user, assistant, user (merged tool results)
+        if (count($messages) !== 3) {
+            return false;
+        }
+
+        // Third message should be a single user message with both tool results
+        $toolResultsMessage = $messages[2];
+
+        if ($toolResultsMessage['role'] !== 'user') {
+            return false;
+        }
+
+        // Should have 2 tool_result content blocks
+        if (count($toolResultsMessage['content']) !== 2) {
+            return false;
+        }
+
+        $firstResult = $toolResultsMessage['content'][0];
+        $secondResult = $toolResultsMessage['content'][1];
+
+        return $firstResult['type'] === 'tool_result'
+            && $firstResult['tool_use_id'] === 'tool_1'
+            && $firstResult['content'] === 'Sunny, 20°C'
+            && $secondResult['type'] === 'tool_result'
+            && $secondResult['tool_use_id'] === 'tool_2'
+            && $secondResult['content'] === 'Rainy, 15°C';
+    });
+});
+
+test('it does not merge non-consecutive tool messages', function (): void {
+    $llm = AnthropicChat::fake([
+        CreateMessage::class => MockResponse::fixture('anthropic/messages/simple'),
+    ]);
+
+    $llm->invoke([
+        new UserMessage('What is the weather?'),
+        new AssistantMessage(
+            toolCalls: new ToolCallCollection([
+                new ToolCall('tool_1', new FunctionCall('get_weather', [
+                    'location' => 'London',
+                ])),
+            ]),
+        ),
+        new ToolMessage('Sunny', 'tool_1'),
+        new AssistantMessage('The weather in London is sunny. Want to check another city?'),
+        new UserMessage('Yes, check Paris'),
+        new AssistantMessage(
+            toolCalls: new ToolCallCollection([
+                new ToolCall('tool_2', new FunctionCall('get_weather', [
+                    'location' => 'Paris',
+                ])),
+            ]),
+        ),
+        new ToolMessage('Rainy', 'tool_2'),
+    ]);
+
+    MockClient::getGlobal()->assertSent(function (CreateMessage $request): bool {
+        $messages = $request->body()->get('messages');
+
+        // Should have 7 messages (tool messages not merged because they're not consecutive)
+        if (count($messages) !== 7) {
+            return false;
+        }
+
+        // First tool result at index 2
+        $firstToolResult = $messages[2];
+        // Second tool result at index 6
+        $secondToolResult = $messages[6];
+
+        return $firstToolResult['role'] === 'user'
+            && count($firstToolResult['content']) === 1
+            && $firstToolResult['content'][0]['tool_use_id'] === 'tool_1'
+            && $secondToolResult['role'] === 'user'
+            && count($secondToolResult['content']) === 1
+            && $secondToolResult['content'][0]['tool_use_id'] === 'tool_2';
+    });
+});
