Add temperature: Option<f32> to api::MessageRequest

TODO.md

@@ -494,18 +494,14 @@
   > Ollama's OpenAI-compatible `/v1/chat/completions` endpoint. Dropped
   > five inline structs (`Req`, `Msg`, `Opts`, `Resp`, `RespMsg`).
   >
-  > **Behavioural caveat:** the old Ollama-native payload set
-  > `options.temperature = 0.0` and `options.num_predict = 16` on the
-  > classification request. `MessageRequest` exposes `max_tokens` (kept
-  > as 16) but no `temperature`, so the classifier no longer pins
-  > determinism. The existing keyword fallback already handles
-  > misclassifications via the `Unknown label` arm, so accuracy
-  > degradation surfaces as a fallback rather than a wrong target. Built
-  > an explicit 8 s `tokio::time::timeout` to replace the old reqwest
-  > builder timeout, and `with_retry_policy(0, ...)` to preserve
-  > one-shot semantics. Adding `temperature` to `api::MessageRequest`
-  > is the proper fix for the determinism regression and is out of
-  > scope for this PR.
+  > **Behavioural caveat (resolved 2026-05-20):** the old Ollama-native
+  > payload set `options.temperature = 0.0` and `options.num_predict = 16`
+  > on the classification request. `MessageRequest` exposes `max_tokens`
+  > (kept as 16) but originally had no `temperature`. PR #24 added
+  > `temperature: Option<f32>` to `api::MessageRequest` and `llm_classify`
+  > now pins `temperature: Some(0.0)` again. Built an explicit 8 s
+  > `tokio::time::timeout` to replace the old reqwest builder timeout, and
+  > `with_retry_policy(0, ...)` to preserve one-shot semantics.
   >
   > **Still to migrate (with notes on each):**
   > 2. `src/grok_reasoning.rs` — uses xAI's `/responses` endpoint

crates/api/src/prompt_cache.rs

@@ -717,6 +717,7 @@ mod tests {
             system: Some("system".to_string()),
             tools: None,
             tool_choice: None,
+            temperature: None,
             stream: false,
         }
     }

crates/api/src/providers/anthropic.rs

@@ -1157,6 +1157,7 @@ mod tests {
             system: None,
             tools: None,
             tool_choice: None,
+            temperature: None,
             stream: false,
         };
 

crates/api/src/providers/openai_compat.rs

@@ -657,6 +657,10 @@ fn build_chat_completion_request(request: &MessageRequest, config: OpenAiCompatC
         "stream": request.stream,
     });
 
+    if let Some(temperature) = request.temperature {
+        payload["temperature"] = json!(temperature);
+    }
+
     if request.stream && should_request_stream_usage(config) {
         payload["stream_options"] = json!({ "include_usage": true });
     }
@@ -987,6 +991,7 @@ mod tests {
                     input_schema: json!({"type": "object"}),
                 }]),
                 tool_choice: Some(ToolChoice::Auto),
+                temperature: None,
                 stream: false,
             },
             OpenAiCompatConfig::xai(),
@@ -1009,6 +1014,7 @@ mod tests {
                 system: None,
                 tools: None,
                 tool_choice: None,
+                temperature: None,
                 stream: true,
             },
             OpenAiCompatConfig::openai(),
@@ -1017,6 +1023,39 @@ mod tests {
         assert_eq!(payload["stream_options"], json!({"include_usage": true}));
     }
 
+    #[test]
+    fn temperature_is_included_in_payload_when_set() {
+        let payload = build_chat_completion_request(
+            &MessageRequest {
+                model: "grok-3".to_string(),
+                max_tokens: 16,
+                messages: vec![InputMessage::user_text("hi")],
+                system: None,
+                tools: None,
+                tool_choice: None,
+                temperature: Some(0.0),
+                stream: false,
+            },
+            OpenAiCompatConfig::xai(),
+        );
+        assert_eq!(payload["temperature"], json!(0.0));
+
+        let payload_none = build_chat_completion_request(
+            &MessageRequest {
+                model: "grok-3".to_string(),
+                max_tokens: 16,
+                messages: vec![InputMessage::user_text("hi")],
+                system: None,
+                tools: None,
+                tool_choice: None,
+                temperature: None,
+                stream: false,
+            },
+            OpenAiCompatConfig::xai(),
+        );
+        assert!(payload_none.get("temperature").is_none());
+    }
+
     #[test]
     fn xai_streaming_requests_skip_openai_specific_usage_opt_in() {
         let payload = build_chat_completion_request(
@@ -1027,6 +1066,7 @@ mod tests {
                 system: None,
                 tools: None,
                 tool_choice: None,
+                temperature: None,
                 stream: true,
             },
             OpenAiCompatConfig::xai(),

crates/api/src/types.rs

@@ -13,6 +13,8 @@ pub struct MessageRequest {
     pub tools: Option<Vec<ToolDefinition>>,
     #[serde(skip_serializing_if = "Option::is_none")]
     pub tool_choice: Option<ToolChoice>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub temperature: Option<f32>,
     #[serde(default, skip_serializing_if = "std::ops::Not::not")]
     pub stream: bool,
 }
@@ -23,6 +25,12 @@ impl MessageRequest {
         self.stream = true;
         self
     }
+
+    #[must_use]
+    pub fn with_temperature(mut self, temperature: f32) -> Self {
+        self.temperature = Some(temperature);
+        self
+    }
 }
 
 #[derive(Debug, Clone, PartialEq, Serialize, Deserialize)]
@@ -248,8 +256,9 @@ pub enum StreamEvent {
 #[cfg(test)]
 mod tests {
     use runtime::format_usd;
+    use serde_json::Value;
 
-    use super::{MessageResponse, Usage};
+    use super::{InputMessage, MessageRequest, MessageResponse, Usage};
 
     #[test]
     fn usage_total_tokens_includes_cache_tokens() {
@@ -287,4 +296,29 @@ mod tests {
         assert_eq!(format_usd(cost.total_cost_usd()), "$54.6750");
         assert_eq!(response.total_tokens(), 1_800_000);
     }
+
+    #[test]
+    fn temperature_serializes_when_set_and_is_omitted_when_none() {
+        let mut request = MessageRequest {
+            model: "claude-sonnet-4-6".to_string(),
+            max_tokens: 32,
+            messages: vec![InputMessage::user_text("hi")],
+            system: None,
+            tools: None,
+            tool_choice: None,
+            temperature: None,
+            stream: false,
+        };
+
+        // None → field absent from the serialized JSON
+        let json_none: Value = serde_json::to_value(&request).unwrap();
+        assert!(json_none.get("temperature").is_none(), "got: {json_none}");
+
+        // Builder method sets the field
+        request = request.with_temperature(0.0);
+        assert_eq!(request.temperature, Some(0.0));
+
+        let json_some: Value = serde_json::to_value(&request).unwrap();
+        assert_eq!(json_some.get("temperature").and_then(Value::as_f64), Some(0.0));
+    }
 }

crates/api/tests/client_integration.rs

@@ -607,6 +607,7 @@ async fn live_stream_smoke_test() {
             system: None,
             tools: None,
             tool_choice: None,
+            temperature: None,
             stream: false,
         })
         .await
@@ -787,6 +788,7 @@ fn sample_request(stream: bool) -> MessageRequest {
             }),
         }]),
         tool_choice: Some(ToolChoice::Auto),
+        temperature: None,
         stream,
     }
 }

crates/api/tests/openai_compat_integration.rs

@@ -459,6 +459,7 @@ fn sample_request(stream: bool) -> MessageRequest {
             }),
         }]),
         tool_choice: Some(ToolChoice::Auto),
+        temperature: None,
         stream,
     }
 }

crates/rusty-claude-cli/src/app.rs

@@ -544,6 +544,7 @@ impl CliApp {
             system: None,
             tools: None,
             tool_choice: None,
+            temperature: None,
             stream: true,
         }
     }

crates/rusty-claude-cli/src/main.rs

@@ -4549,6 +4549,7 @@ impl ApiClient for AnthropicRuntimeClient {
                 .enable_tools
                 .then(|| filter_tool_specs(&self.tool_registry, self.allowed_tools.as_ref())),
             tool_choice: self.enable_tools.then_some(ToolChoice::Auto),
+            temperature: None,
             stream: true,
         };
 

crates/rusty-claude-cli/tests/api_integration.rs

@@ -42,6 +42,7 @@ fn scenario_request(model: &str, scenario: &str) -> MessageRequest {
         system: None,
         tools: None,
         tool_choice: None,
+        temperature: None,
         stream: false,
     }
 }
@@ -129,6 +130,7 @@ fn openai_compat_client_consumes_openai_shaped_response() {
             system: None,
             tools: None,
             tool_choice: None,
+            temperature: None,
             stream: false,
         };
 

crates/tools/src/lib.rs

@@ -3465,6 +3465,7 @@ impl ApiClient for ProviderRuntimeClient {
             system: (!request.system_prompt.is_empty()).then(|| request.system_prompt.join("\n\n")),
             tools: (!tools.is_empty()).then_some(tools),
             tool_choice: (!self.allowed_tools.is_empty()).then_some(ToolChoice::Auto),
+            temperature: None,
             stream: true,
         };
 
@@ -3548,6 +3549,7 @@ impl ApiClient for ProviderRuntimeClient {
             let response = self
                 .client
                 .send_message(&MessageRequest {
+                    temperature: None,
                     stream: false,
                     ..message_request.clone()
                 })

src/agent/pipeline.rs

@@ -419,6 +419,7 @@ impl AgentPipeline {
             system: Some(system),
             tools: None,
             tool_choice: None,
+            temperature: None,
             stream: false,
         };
 
@@ -530,6 +531,7 @@ impl AgentPipeline {
             system: Some(EXECUTOR_SYSTEM_PROMPT.to_string()),
             tools: None,
             tool_choice: None,
+            temperature: None,
             stream: false,
         };
 
@@ -579,6 +581,7 @@ impl AgentPipeline {
                 system: Some(EXECUTOR_SYSTEM_PROMPT.to_string()),
                 tools: Some(tool_defs.clone()),
                 tool_choice: None,
+                temperature: None,
                 stream: false,
             };
 
@@ -721,6 +724,7 @@ impl AgentPipeline {
             system: Some(REVIEWER_SYSTEM_PROMPT.to_string()),
             tools: None,
             tool_choice: None,
+            temperature: None,
             stream: false,
         };
 
@@ -791,6 +795,7 @@ impl AgentPipeline {
             system: Some(CONSOLIDATION_SYSTEM_PROMPT.to_string()),
             tools: None,
             tool_choice: None,
+            temperature: None,
             stream: false,
         };
 

src/api/proxy.rs

@@ -779,6 +779,7 @@ async fn handle_streaming(
                     system,
                     tools: None,
                     tool_choice: None,
+                    temperature: None,
                     stream: false,
                 };
 
@@ -1465,6 +1466,7 @@ async fn dispatch_claude(
         system: req.system_prompt.clone(),
         tools: None,
         tool_choice: None,
+        temperature: None,
         stream: false,
     };
 

src/grok_client.rs

@@ -532,6 +532,7 @@ Code:
             system: None,
             tools: None,
             tool_choice: None,
+            temperature: None,
             stream: false,
         };
 

src/llm/grok.rs

@@ -66,6 +66,7 @@ impl GrokAnalyzer {
             system: Some(system_prompt.to_string()),
             tools: None,
             tool_choice: None,
+            temperature: None,
             stream: false,
         };
 

src/model_router.rs

@@ -3,10 +3,10 @@
 //
 // Note on the Ollama classifier: `llm_classify` talks to Ollama's
 // OpenAI-compatible `/v1/chat/completions` endpoint via `api::OpenAiCompatClient`
-// rather than the Ollama-native `/api/chat` endpoint. The compat endpoint
-// doesn't surface `temperature` / `options.num_predict` knobs, so the classifier
-// no longer pins `temperature: 0.0` — the keyword fallback handles any
-// misclassifications via the `Unknown label` arm.
+// rather than the Ollama-native `/api/chat` endpoint. The classifier pins
+// `temperature: 0.0` for deterministic label selection; `options.num_predict`
+// is replaced by `max_tokens: 16` which the compat endpoint translates back to
+// the same Ollama option.
 
 use api::{InputMessage, MessageRequest, OpenAiCompatClient, OpenAiCompatConfig, OutputContentBlock};
 use serde::{Deserialize, Serialize};
@@ -253,6 +253,9 @@ RepoQuestion | ArchitecturalReason | CodeReview | Unknown";
             system: Some(CLASSIFY_SYSTEM.to_string()),
             tools: None,
             tool_choice: None,
+            // Deterministic classification — restored after the
+            // /api/chat → /v1/chat/completions migration in PR #23.
+            temperature: Some(0.0),
             stream: false,
         };
 

src/simple_client.rs

@@ -61,6 +61,7 @@ impl GrokClient {
             system:      None,
             tools:       None,
             tool_choice: None,
+            temperature: None,
             stream:      false,
         };
 
@@ -100,6 +101,7 @@ impl GrokClient {
             system:      Some(system.to_string()),
             tools:       None,
             tool_choice: None,
+            temperature: None,
             stream:      false,
         };
 

tests/test_grok_integration.rs

@@ -526,6 +526,7 @@ async fn test_claude_switch_live_completion() {
         system:      None,
         tools:       None,
         tool_choice: None,
+        temperature: None,
         stream:      false,
     };