FEATURE: Add support for 'minimal' reasoning effort in OpenAI models (#34179)

xfalcox · pento · commit a6af10f679e4 · 2025-08-11T11:28:06.000+10:00
- Add 'minimal' as a supported reasoning effort value in LlmModel
configuration
- Update OpenAI endpoint validation to accept 'minimal' effort level
- Implement dual payload format support:
  * Responses API: `{ reasoning: { effort: "value" } }`
  * Standard API: `{ reasoning_effort: "value" }`
- Add comprehensive tests for reasoning effort payload formats
- Update GPT-5 model configurations with max_output_tokens and pricing
fixes
diff --git a/plugins/discourse-ai/app/models/llm_model.rb b/plugins/discourse-ai/app/models/llm_model.rb
@@ -55,7 +55,7 @@ def self.provider_params
         enable_responses_api: :checkbox,
         reasoning_effort: {
           type: :enum,
-          values: %w[default low medium high],
+          values: %w[default minimal low medium high],
           default: "default",
         },
       },
diff --git a/plugins/discourse-ai/lib/completions/endpoints/open_ai.rb b/plugins/discourse-ai/lib/completions/endpoints/open_ai.rb
@@ -62,7 +62,7 @@ def disable_streaming?
         def reasoning_effort
           return @reasoning_effort if defined?(@reasoning_effort)
           @reasoning_effort = llm_model.lookup_custom_param("reasoning_effort")
-          @reasoning_effort = nil if !%w[low medium high].include?(@reasoning_effort)
+          @reasoning_effort = nil if !%w[minimal low medium high].include?(@reasoning_effort)
           @reasoning_effort
         end
 
@@ -80,7 +80,13 @@ def model_uri
         def prepare_payload(prompt, model_params, dialect)
           payload = default_options.merge(model_params).merge(messages: prompt)
 
-          payload[:reasoning_effort] = reasoning_effort if reasoning_effort
+          if reasoning_effort
+            if responses_api?
+              payload.merge!({ reasoning: { effort: reasoning_effort } })
+            else
+              payload.merge!({ reasoning_effort: reasoning_effort })
+            end
+          end
 
           if @streaming_mode
             payload[:stream] = true
diff --git a/plugins/discourse-ai/lib/completions/llm.rb b/plugins/discourse-ai/lib/completions/llm.rb
@@ -130,8 +130,9 @@ def presets
                       tokens: 400_000,
                       display_name: "GPT-5",
                       input_cost: 1.25,
-                      cached_input_cost: 0.124,
+                      cached_input_cost: 0.125,
                       output_cost: 10,
+                      max_output_tokens: 128_000,
                     },
                     {
                       name: "gpt-5-mini",
@@ -140,14 +141,16 @@ def presets
                       input_cost: 0.25,
                       cached_input_cost: 0.025,
                       output_cost: 2.0,
+                      max_output_tokens: 128_000,
                     },
                     {
                       name: "gpt-5-nano",
                       tokens: 400_000,
-                      display_name: "GPT-4.1 Nano",
+                      display_name: "GPT-5 Nano",
                       input_cost: 0.05,
                       cached_input_cost: 0.005,
                       output_cost: 0.40,
+                      max_output_tokens: 128_000,
                     },
                   ],
                   tokenizer: DiscourseAi::Tokenizer::OpenAiTokenizer,
diff --git a/plugins/discourse-ai/spec/lib/completions/endpoints/open_ai_spec.rb b/plugins/discourse-ai/spec/lib/completions/endpoints/open_ai_spec.rb
@@ -995,4 +995,61 @@ def request_body(prompt, stream: false, tool_call: false)
       end
     end
   end
+
+  describe "reasoning effort payload format" do
+    let(:prompt) { compliance.generic_prompt }
+    let(:dialect) { compliance.dialect(prompt: prompt) }
+
+    it "uses reasoning object format for responses API" do
+      model.update!(provider_params: { enable_responses_api: true, reasoning_effort: "minimal" })
+
+      parsed_body = nil
+      stub_request(:post, "https://api.openai.com/v1/chat/completions").with(
+        body:
+          proc do |req_body|
+            parsed_body = JSON.parse(req_body, symbolize_names: true)
+            true
+          end,
+      ).to_return(status: 200, body: { choices: [{ message: { content: "test" } }] }.to_json)
+
+      endpoint.perform_completion!(dialect, user)
+
+      expect(parsed_body[:reasoning]).to eq({ effort: "minimal" })
+      expect(parsed_body).not_to have_key(:reasoning_effort)
+    end
+
+    it "uses reasoning_effort field for standard API" do
+      model.update!(provider_params: { reasoning_effort: "low" })
+
+      parsed_body = nil
+      stub_request(:post, "https://api.openai.com/v1/chat/completions").with(
+        body:
+          proc do |req_body|
+            parsed_body = JSON.parse(req_body, symbolize_names: true)
+            true
+          end,
+      ).to_return(status: 200, body: { choices: [{ message: { content: "test" } }] }.to_json)
+
+      endpoint.perform_completion!(dialect, user)
+
+      expect(parsed_body[:reasoning_effort]).to eq("low")
+      expect(parsed_body).not_to have_key(:reasoning)
+    end
+
+    it "omits reasoning parameters when not configured" do
+      parsed_body = nil
+      stub_request(:post, "https://api.openai.com/v1/chat/completions").with(
+        body:
+          proc do |req_body|
+            parsed_body = JSON.parse(req_body, symbolize_names: true)
+            true
+          end,
+      ).to_return(status: 200, body: { choices: [{ message: { content: "test" } }] }.to_json)
+
+      endpoint.perform_completion!(dialect, user)
+
+      expect(parsed_body).not_to have_key(:reasoning)
+      expect(parsed_body).not_to have_key(:reasoning_effort)
+    end
+  end
 end
