fix(openrouter): stop silently ignoring thinking effort off (#10991)

2026-08-05 19:11:14 -07:00
parent 1c1bd5299a
commit 86eec2a8ed
3 changed files with 175 additions and 56 deletions
@@ -88,6 +88,7 @@ impl RetryConfig {
 const PERMANENT_REQUEST_FAILURE_MARKERS: &[&str] = &[
    "blocks in the latest assistant message cannot be modified",
    "must remain as they were in the original response",
+    "Reasoning is mandatory for this endpoint",
 ];

 fn is_permanent_request_failure(message: &str) -> bool {
@@ -99,25 +99,40 @@ fn reasoning_effort_for_openrouter(effort: ThinkingEffort) -> Option<&'static st
    }
 }

-pub fn apply_reasoning_config(payload: &mut Value, model_config: &ModelConfig) {
+/// Returns true when a reasoning disable request was inserted, which
+/// mandatory-reasoning endpoints reject; the provider downgrades those to
+/// the lowest effort on OpenRouter's mandatory-reasoning error.
+pub fn apply_reasoning_config(payload: &mut Value, model_config: &ModelConfig) -> bool {
    let Some(effort) = model_config.thinking_effort() else {
-        return;
+        return false;
    };

    if let Some(obj) = payload.as_object_mut() {
        if obj.contains_key("reasoning") {
            obj.remove("reasoning_effort");
-            return;
+            return false;
        }

        let clamped_effort = obj
            .remove("reasoning_effort")
            .and_then(|value| value.as_str().map(str::to_owned));
        if effort == ThinkingEffort::Off {
-            return;
+            if !model_config.is_reasoning_model() {
+                return false;
+            }
+            return match clamped_effort {
+                Some(clamped) => {
+                    obj.insert("reasoning".to_string(), json!({ "effort": clamped }));
+                    false
+                }
+                None => {
+                    obj.insert("reasoning".to_string(), json!({ "enabled": false }));
+                    true
+                }
+            };
        }
        if clamped_effort.is_none() && !model_config.is_reasoning_model() {
-            return;
+            return false;
        }

        let effort = clamped_effort
@@ -127,6 +142,7 @@ pub fn apply_reasoning_config(payload: &mut Value, model_config: &ModelConfig) {
            obj.insert("reasoning".to_string(), json!({ "effort": effort }));
        }
    }
+    false
 }

 #[cfg(test)]
@@ -230,23 +246,6 @@ mod tests {
        assert!(payload.get("reasoning_effort").is_none());
    }

-    #[test]
-    fn test_apply_reasoning_config_omits_off_reasoning_capable_model() {
-        let mut payload = json!({
-            "model": "google/gemini-2.5-flash",
-            "messages": []
-        });
-        let mut model_config = ModelConfig::new("google/gemini-2.5-flash");
-        model_config.reasoning = Some(true);
-        let mut params = HashMap::new();
-        params.insert("thinking_effort".to_string(), json!("off"));
-        model_config.request_params = Some(params);
-
-        apply_reasoning_config(&mut payload, &model_config);
-
-        assert!(payload.get("reasoning").is_none());
-    }
-
    #[test]
    fn test_apply_reasoning_config_uses_reasoning_metadata() {
        let mut payload = json!({
@@ -298,24 +297,7 @@ mod tests {
    }

    #[test]
-    fn test_apply_reasoning_config_off_omits_reasoning() {
-        let mut payload = json!({
-            "model": "x-ai/grok-4",
-            "messages": []
-        });
-        let mut model_config = ModelConfig::new("x-ai/grok-4");
-        let mut params = HashMap::new();
-        params.insert("thinking_effort".to_string(), json!("off"));
-        model_config.request_params = Some(params);
-        model_config.reasoning = Some(true);
-
-        apply_reasoning_config(&mut payload, &model_config);
-
-        assert!(payload.get("reasoning").is_none());
-    }
-
-    #[test]
-    fn test_apply_reasoning_config_off_ignores_clamped_effort() {
+    fn test_apply_reasoning_config_off_keeps_clamped_effort() {
        let mut payload = json!({
            "model": "openai/gpt-5",
            "messages": [],
@@ -327,9 +309,67 @@ mod tests {
        model_config.request_params = Some(params);
        model_config.reasoning = Some(true);

+        let sent_disable = apply_reasoning_config(&mut payload, &model_config);
+
+        assert!(!sent_disable);
+        assert_eq!(payload["reasoning"], json!({ "effort": "low" }));
+        assert!(payload.get("reasoning_effort").is_none());
+    }
+
+    #[test]
+    fn test_apply_reasoning_config_kimi_k3_off_disables_reasoning() {
+        let mut payload = json!({
+            "model": "moonshotai/kimi-k3",
+            "messages": []
+        });
+        let mut model_config =
+            ModelConfig::new("moonshotai/kimi-k3").with_canonical_limits("openrouter");
+        let mut params = HashMap::new();
+        params.insert("thinking_effort".to_string(), json!("off"));
+        model_config.request_params = Some(params);
+
+        let sent_disable = apply_reasoning_config(&mut payload, &model_config);
+
+        assert!(sent_disable);
+        assert_eq!(payload["reasoning"], json!({ "enabled": false }));
+    }
+
+    #[test]
+    fn test_apply_reasoning_config_off_skips_non_reasoning_model() {
+        // OpenAI-shaped but canonically non-reasoning; off must drop the clamp.
+        let mut payload = json!({
+            "model": "openai/gpt-5.1-chat",
+            "messages": [],
+            "reasoning_effort": "low"
+        });
+        let mut model_config = ModelConfig::new("openai/gpt-5.1-chat");
+        let mut params = HashMap::new();
+        params.insert("thinking_effort".to_string(), json!("off"));
+        model_config.request_params = Some(params);
+        model_config.reasoning = Some(false);
+
        apply_reasoning_config(&mut payload, &model_config);

        assert!(payload.get("reasoning").is_none());
        assert!(payload.get("reasoning_effort").is_none());
    }
+
+    #[test]
+    fn test_apply_reasoning_config_off_preserves_user_reasoning() {
+        let mut payload = json!({
+            "model": "x-ai/grok-4",
+            "messages": [],
+            "reasoning": { "max_tokens": 2000 }
+        });
+        let mut model_config = ModelConfig::new("x-ai/grok-4");
+        let mut params = HashMap::new();
+        params.insert("thinking_effort".to_string(), json!("off"));
+        model_config.request_params = Some(params);
+        model_config.reasoning = Some(true);
+
+        let sent_disable = apply_reasoning_config(&mut payload, &model_config);
+
+        assert!(!sent_disable);
+        assert_eq!(payload["reasoning"], json!({ "max_tokens": 2000 }));
+    }
 }
@@ -74,6 +74,28 @@ impl OpenRouterProvider {
            configured_parameters,
        })
    }
+
+    async fn post_chat_completions(
+        &self,
+        model_config: &ModelConfig,
+        payload: &Value,
+    ) -> Result<reqwest::Response, ProviderError> {
+        self.with_retry(|| async {
+            let resp = self
+                .api_client
+                .request("api/v1/chat/completions")
+                .model_headers(model_config)?
+                .streaming(true)
+                .response_post(payload)
+                .await?;
+            handle_status(resp).await
+        })
+        .await
+    }
+}
+
+fn is_mandatory_reasoning_error(error: &ProviderError) -> bool {
+    matches!(error, ProviderError::RequestFailed(message) if message.contains("Reasoning is mandatory"))
 }

 /// Update the request when using anthropic model.
@@ -335,7 +357,8 @@ impl Provider for OpenRouterProvider {
        if is_gemini_model(&model_config.model_name) {
            openrouter_format::add_reasoning_details_to_request(&mut payload, messages);
        }
-        openrouter_format::apply_reasoning_config(&mut payload, model_config);
+        let sent_reasoning_disable =
+            openrouter_format::apply_reasoning_config(&mut payload, model_config);

        if let Some(obj) = payload.as_object_mut() {
            obj.insert("transforms".to_string(), json!(["middle-out"]));
@@ -344,21 +367,20 @@ impl Provider for OpenRouterProvider {

        let mut log = start_log(model_config, &payload)?;

-        let response = self
-            .with_retry(|| async {
-                let resp = self
-                    .api_client
-                    .request("api/v1/chat/completions")
-                    .model_headers(model_config)?
-                    .streaming(true)
-                    .response_post(&payload)
-                    .await?;
-                handle_status(resp).await
-            })
-            .await
-            .inspect_err(|e| {
-                let _ = log.error(e);
-            })?;
+        let response = match self.post_chat_completions(model_config, &payload).await {
+            // Mandatory-reasoning endpoints reject the disable request, so
+            // downgrade to the lowest effort they all accept and retry once.
+            Err(error) if sent_reasoning_disable && is_mandatory_reasoning_error(&error) => {
+                let _ = log.error(&error);
+                payload["reasoning"] = json!({ "effort": "low" });
+                log = start_log(model_config, &payload)?;
+                self.post_chat_completions(model_config, &payload).await
+            }
+            result => result,
+        }
+        .inspect_err(|e| {
+            let _ = log.error(e);
+        })?;

        stream_openai_compat(response, log)
    }
@@ -448,4 +470,60 @@ mod tests {
        assert_eq!(request_params["plugins"], json!([{ "id": "web" }]));
        assert_eq!(request_params["verbosity"], json!("xhigh"));
    }
+
+    #[tokio::test]
+    async fn stream_downgrades_reasoning_disable_on_mandatory_endpoint() {
+        use wiremock::matchers::{body_partial_json, method, path};
+        use wiremock::{Mock, MockServer, ResponseTemplate};
+
+        let server = MockServer::start().await;
+        Mock::given(method("POST"))
+            .and(path("/api/v1/chat/completions"))
+            .and(body_partial_json(
+                json!({ "reasoning": { "enabled": false } }),
+            ))
+            .respond_with(ResponseTemplate::new(400).set_body_json(json!({
+                "error": { "message": "Reasoning is mandatory for this endpoint and cannot be disabled." }
+            })))
+            .expect(1)
+            .mount(&server)
+            .await;
+        Mock::given(method("POST"))
+            .and(path("/api/v1/chat/completions"))
+            .and(body_partial_json(
+                json!({ "reasoning": { "effort": "low" } }),
+            ))
+            .respond_with(
+                ResponseTemplate::new(200)
+                    .insert_header("content-type", "text/event-stream")
+                    .set_body_string("data: [DONE]\n\n"),
+            )
+            .expect(1)
+            .mount(&server)
+            .await;
+
+        let provider = OpenRouterProvider {
+            api_client: ApiClient::new_with_tls(
+                server.uri(),
+                AuthMethod::BearerToken("test-key".to_string()),
+                None,
+            )
+            .unwrap(),
+            supports_streaming: true,
+            name: OPENROUTER_PROVIDER_NAME.to_string(),
+            configured_parameters: None,
+        };
+
+        let mut config = model_config("google/gemini-3.5-flash");
+        config.reasoning = Some(true);
+        config.request_params = Some(HashMap::from([(
+            "thinking_effort".to_string(),
+            json!("off"),
+        )]));
+
+        let _stream = provider
+            .stream(&config, "system", &[Message::user().with_text("hi")], &[])
+            .await
+            .unwrap();
+    }
 }