Fix/responses streaming

juce_llm/llm/LLMTypes.h

@@ -17,7 +17,7 @@ struct ProviderConfig {
     juce::String apiKey;
     juce::String model;
 
-    int maxTokens = 4096;  // max output tokens (0 = provider default)
+    int maxTokens = 0;  // max output tokens (0 = provider default — don't send a cap)
 
     // Provider-specific options
     bool noTemperature = false;    // GPT-5 doesn't support temperature

juce_llm/llm/providers/AnthropicClient.cpp

@@ -10,7 +10,8 @@ juce::String AnthropicClient::buildRequestBody(const Request& request) const {
 
     auto* payload = new juce::DynamicObject();
     payload->setProperty("model", config_.model);
-    int maxTok = request.maxTokens > 0 ? request.maxTokens : (config_.maxTokens > 0 ? config_.maxTokens : 4096);
+    int maxTok = request.maxTokens > 0 ? request.maxTokens
+                                       : (config_.maxTokens > 0 ? config_.maxTokens : 4096);
     payload->setProperty("max_tokens", maxTok);
     payload->setProperty("temperature", (double)request.temperature);
     payload->setProperty("messages", messagesArray);
@@ -31,12 +32,9 @@ juce::String AnthropicClient::buildRequestBody(const Request& request) const {
         payload->setProperty("system", systemArray);
     }
 
-    // Output effort — low/medium/high/max (similar to OpenAI reasoning_effort)
-    if (config_.reasoningEffort.isNotEmpty()) {
-        auto* outputConfig = new juce::DynamicObject();
-        outputConfig->setProperty("effort", config_.reasoningEffort);
-        payload->setProperty("output_config", juce::var(outputConfig));
-    }
+    // NOTE: Anthropic's Messages API does not accept an `effort` / `output_config`
+    // field — that is an OpenAI-ism. `reasoningEffort` is deliberately ignored here.
+    // (Extended thinking uses a separate `thinking` block on models that support it.)
 
     // App identification for abuse tracking
     if (config_.userAgent.isNotEmpty()) {

juce_llm/llm/providers/OpenAIResponsesClient.cpp

@@ -114,4 +114,17 @@ Response OpenAIResponsesClient::parseResponseBody(const juce::String& jsonString
     return response;
 }
 
+/** Responses API SSE chunks look like:
+      data: {"type":"response.output_text.delta","delta":"token","item_id":"..."}
+      data: {"type":"response.completed", ...}
+    The base LLMClient class only reads "data: " lines, so we just need to pull
+    the "delta" field out of the JSON when the event type matches. */
+juce::String OpenAIResponsesClient::parseStreamChunk(const juce::String& dataLine) const {
+    auto json = juce::JSON::parse(dataLine);
+    auto type = json["type"].toString();
+    if (type == "response.output_text.delta")
+        return json["delta"].toString();
+    return {};
+}
+
 }  // namespace llm

juce_llm/llm/providers/OpenAIResponsesClient.h

@@ -15,6 +15,7 @@ class OpenAIResponsesClient : public LLMClient {
     juce::String getEndpointUrl() const override;
     juce::StringPairArray getHeaders() const override;
     Response parseResponseBody(const juce::String& jsonString) const override;
+    juce::String parseStreamChunk(const juce::String& dataLine) const override;
 };
 
 }  // namespace llm