## Issue summary ### Problem 1: the response is empty for any reasoning model `chat()` reads the answer from content block zero: ```php $message = new ChatMessage($response['output']['message']['role'], $response['output']['message']['content'][0]['text'] ?? ""); ``` Reasoning models put their reasoning in its own content block, and it arrives first. Block zero holds `reasoningContent`, not `text`, so the `?? ""` fires and the provider returns an empty `ChatMessage`. Nothing errors, so the caller sees a successful call that produced no text. That `?? ""` came from #3517795, where block zero held `toolUse` and the undefined key raised a warning. Defaulting to an empty string was right for that case and is the reason this one is silent: the same line now has two ways to miss the answer, and neither reports it. Scanning for the first block that carries text covers both. Observed against `openai.gpt-oss-120b-1:0` in `us-east-1`. Raw Converse response, abbreviated: ```json {"output":{"message":{"role":"assistant","content":[ {"reasoningContent":{"reasoningText":{"text":"..."}}}, {"text":"the actual answer"} ]}}} ``` ### Problem 2: model-specific parameters cannot be sent at all The whole of `$this->configuration` is passed as `inferenceConfig`: ```php 'inferenceConfig' => $this->configuration, ``` Converse validates that key strictly and rejects the request if it carries anything outside `maxTokens`, `temperature`, `topP` and `stopSequences`. Model specific parameters belong in `additionalModelRequestFields` instead. Today there is no way to send one, so `reasoning_effort` cannot be set, and a reasoning model spends its output budget thinking before it answers. ## Steps to reproduce 1. Configure the provider against a Bedrock profile with access to `openai.gpt-oss-120b-1:0`. 2. Call `chat()` with any prompt. 3. The returned `ChatMessage` text is empty. 4. Try to set `reasoning_effort` through the plugin configuration. The request fails validation because it is sent inside `inferenceConfig`. ## Proposed resolution Take the first content block that actually carries text, rather than assuming it is block zero. Split the configuration so the four keys Converse accepts go to `inferenceConfig` and everything else goes to `additionalModelRequestFields`. Verified against live Bedrock: with the patch, the same call returns the answer text, and `additionalModelRequestFields` with `reasoning_effort` is accepted. ## Remaining tasks If MR !12 on #3562254 lands, `hook_ai_provider_aws_bedrock_chat_payload_alter()` would let a site set `additionalModelRequestFields` itself, and the second half of this patch could be dropped in favour of it. Splitting the configuration is still the better default, since a caller should not need a hook to send a parameter the model documents, but that is the maintainers' call. Adjacent problem, deliberately left out of this patch to keep it reviewable. The provider's own `definitions/api_defaults.yml` documents its chat parameters as `max_tokens` and `temperature`, in snake case, but Converse expects `maxTokens`. So a caller who sets a token limit through the AI module has it rejected either way, before this patch inside `inferenceConfig` and after it inside `additionalModelRequestFields`. That looks like it wants its own issue mapping the documented names onto the Converse ones. ## User interface changes None. ## API changes None. Existing callers that set only `temperature` are unaffected, since it is in both lists. ## Patch The API would not let me attach a file, so the diff is inline. It is against `1.x` and applies cleanly to `1.1.0-beta4` as well. Happy to turn it into a merge request if that is preferred. <details> <summary>bedrock-reasoning-model-support.patch</summary> ```diff diff --git a/src/Plugin/AiProvider/BedrockProvider.php b/src/Plugin/AiProvider/BedrockProvider.php index f8c8e1e..afb38fe 100644 --- a/src/Plugin/AiProvider/BedrockProvider.php +++ b/src/Plugin/AiProvider/BedrockProvider.php @@ -351,12 +351,25 @@ class BedrockProvider extends AiProviderClientBase implements // Normalize the configuration. $this->normalizeConfiguration('chat', $model_id); + // Converse validates inferenceConfig strictly and rejects the request if + // it carries anything outside its own four keys. Model specific + // parameters, such as reasoning_effort on the gpt-oss models, belong in + // additionalModelRequestFields instead, so split the configuration rather + // than forcing callers to choose between the two. + $inference_keys = ['maxTokens', 'temperature', 'topP', 'stopSequences']; + $inference_config = array_intersect_key($this->configuration, array_flip($inference_keys)); + $additional_fields = array_diff_key($this->configuration, array_flip($inference_keys)); + $payload = [ 'modelId' => $model_id, 'messages' => $chat_input, - 'inferenceConfig' => $this->configuration, + 'inferenceConfig' => $inference_config, ]; + if ($additional_fields) { + $payload['additionalModelRequestFields'] = $additional_fields; + } + // If we want to add tools to the input. if (method_exists($input, 'getChatTools') && $input->getChatTools()) { $tools = $input->getChatTools()->renderToolsArray(); @@ -391,7 +404,17 @@ class BedrockProvider extends AiProviderClientBase implements else { $response = $this->client->converse($payload); // Text messages is not always, given with tool results. - $message = new ChatMessage($response['output']['message']['role'], $response['output']['message']['content'][0]['text'] ?? ""); + // Reasoning models return their reasoning in its own content block, + // which may come before the answer, so take the first block that + // actually carries text instead of assuming it is block zero. + $text = ""; + foreach ($response['output']['message']['content'] as $block) { + if (isset($block['text'])) { + $text = $block['text']; + break; + } + } + $message = new ChatMessage($response['output']['message']['role'], $text); // Tool usage. if (isset($response['stopReason']) && $response['stopReason'] === 'tool_use') { ``` </details> --- ### AI usage See https://www.drupal.org/about/ai/usage-policy - [x] **AI Assisted Issue** - This issue was generated with AI assistance, but was reviewed and refined by the creator. - [x] **AI Assisted Code** - Code was mainly generated by a human, with AI autocompleting or parts AI generated, but under full human supervision.