evalops-mirror · Sep 23, 2026 · Sep 23, 2026
@@ -3,11 +3,11 @@
   "projection": "deixic-code",
   "projectionSchemaVersion": 1,
   "sourceRepository": "dx-corp/mono",
-  "sourceSha": "06b74765cc5e7353134d868de4140e69fa634f38",
+  "sourceSha": "5f0f02dc74958fce77b77d166a14dc09ec490804",
   "destinationRepository": "dx-corp/code",
-  "priorProjectedBase": "85879e61077a7fd73f0dedcfd21e7c06e7f2f280",
+  "priorProjectedBase": "acfb663d7933c3028edc4e5315ce353159715c68",
   "definitionDigest": "82936441c776e3e8edb5d215a75007ec9714a233f489d460075d79d5ef5ba32f",
   "toolDigest": "45502ff0478e541d02f34cc39ac332935f8d3c0030fa221afe5bcc3b5e51f88e",
-  "contentDigest": "65e63c3da4396f8ae48cd156df59c171580f0159ba3d1d943b9fe667879f9856",
+  "contentDigest": "e71c46bc28e9552284f4b3d89e26852771d0d14cee7df7337f4876d4712d8f72",
   "publicationEligible": true
 }
@@ -822,6 +822,18 @@ fn distinct_output_tokens(context: Option<u32>, output: Option<u32>) -> Option<u
     }
 }
 
+fn omit_unverified_sonar_output(id: &str, output: Option<u32>) -> Option<u32> {
+    // Perplexity publishes a 128K context for Sonar Reasoning Pro but no
+    // 4,096-token output ceiling. The models.dev Perplexity row supplies 4,096
+    // for OpenRouter's fallback after its derived 90% value is rejected.
+    // Keep this exact disputed value unknown until the vendor
+    // publishes a cap: https://docs.perplexity.ai/docs/sonar/models/sonar-reasoning-pro
+    if id == "perplexity/sonar-reasoning-pro" && output == Some(4_096) {
+        return None;
+    }
+    output
+}
+
 fn parse_limit_tokens(model: &serde_json::Value, field: &str) -> Option<u32> {
     model
         .get("limit")
@@ -1058,9 +1070,12 @@ fn map_openrouter_model(
         .and_then(serde_json::Value::as_u64)
         .and_then(|output| u32::try_from(output).ok())
         .filter(|output| *output > 0);
-    let output_tokens = limits.map_or_else(
-        || distinct_output_tokens(Some(context_tokens), advertised_output),
-        |limits| limits.distinct_output(id, context_tokens, advertised_output),
+    let output_tokens = omit_unverified_sonar_output(
+        id,
+        limits.map_or_else(
+            || distinct_output_tokens(Some(context_tokens), advertised_output),
+            |limits| limits.distinct_output(id, context_tokens, advertised_output),
+        ),
     );
     let name = model
         .get("name")
@@ -1652,6 +1667,35 @@ mod tests {
         }
     }
 
+    #[test]
+    fn unverified_sonar_output_limit_stays_unknown_after_refresh() {
+        assert_eq!(
+            catalog_model("perplexity/sonar-reasoning-pro")
+                .capabilities
+                .output_tokens,
+            None
+        );
+        let models_dev = serde_json::json!({
+            "perplexity": {"models": {
+                "sonar-reasoning-pro": {"limit": {"context": 128_000, "output": 4_096}}
+            }}
+        });
+        let openrouter = serde_json::json!({"data": [
+            {"id": "perplexity/sonar-reasoning-pro", "context_length": 128_000,
+             "top_provider": {"max_completion_tokens": 115_200}},
+            {"id": "perplexity/another-model", "context_length": 128_000,
+             "top_provider": {"max_completion_tokens": 4_096}}
+        ]});
+        let limits = DevTokenLimits::from_models_dev(&models_dev);
+        let refreshed = map_openrouter_catalog_with_limits(&openrouter, Some(&limits)).unwrap();
+        assert_eq!(refreshed[0].capabilities.output_tokens, None);
+        assert_eq!(refreshed[1].capabilities.output_tokens, Some(4_096));
+        assert_eq!(
+            omit_unverified_sonar_output("perplexity/sonar-reasoning-pro", Some(8_192)),
+            Some(8_192)
+        );
+    }
+
     #[test]
     fn managed_model_routes_survive_cache_and_preserve_provider_and_limits() {
         let mut models = Vec::new();

@@ -9189,7 +9189,6 @@
         "reasoning": true,
         "streaming": true,
         "context_tokens": 128000,
-        "output_tokens": 4096,
         "temperature": true
       },
       "cost": {

@@ -344,7 +344,7 @@ pub static MODEL_FACTS: &[(&str, ModelFacts)] = &[
     ("perplexity/sonar-deep-research", ModelFacts { context_window: 128_000, max_output_tokens: Some(32_768), input_per_million: Some(2.0), output_per_million: Some(8.0), cache_read_per_million: None, cache_write_per_million: None }),
     ("perplexity/sonar-pro", ModelFacts { context_window: 200_000, max_output_tokens: Some(8_000), input_per_million: Some(3.0), output_per_million: Some(15.0), cache_read_per_million: None, cache_write_per_million: None }),
     ("perplexity/sonar-pro-search", ModelFacts { context_window: 200_000, max_output_tokens: Some(8_000), input_per_million: Some(3.0), output_per_million: Some(15.0), cache_read_per_million: None, cache_write_per_million: None }),
-    ("perplexity/sonar-reasoning-pro", ModelFacts { context_window: 128_000, max_output_tokens: Some(4_096), input_per_million: Some(2.0), output_per_million: Some(8.0), cache_read_per_million: None, cache_write_per_million: None }),
+    ("perplexity/sonar-reasoning-pro", ModelFacts { context_window: 128_000, max_output_tokens: None, input_per_million: Some(2.0), output_per_million: Some(8.0), cache_read_per_million: None, cache_write_per_million: None }),
     ("poolside/laguna-s-2.1", ModelFacts { context_window: 1_048_576, max_output_tokens: Some(131_072), input_per_million: Some(0.09), output_per_million: Some(0.18), cache_read_per_million: Some(0.009), cache_write_per_million: None }),
     ("poolside/laguna-s-2.1:free", ModelFacts { context_window: 262_144, max_output_tokens: Some(32_768), input_per_million: Some(0.0), output_per_million: Some(0.0), cache_read_per_million: None, cache_write_per_million: None }),
     ("poolside/laguna-xs-2.1", ModelFacts { context_window: 262_144, max_output_tokens: Some(32_768), input_per_million: Some(0.06), output_per_million: Some(0.12), cache_read_per_million: Some(0.03), cache_write_per_million: None }),

@@ -127,7 +127,11 @@ fn read_http_request(stream: &mut TcpStream) -> String {
     let mut request = Vec::new();
     loop {
         let mut chunk = [0_u8; 4096];
-        let bytes = stream.read(&mut chunk).expect("read test request");
+        let bytes = match stream.read(&mut chunk) {
+            Ok(bytes) => bytes,
+            Err(error) if error.kind() == std::io::ErrorKind::Interrupted => continue,
+            Err(error) => panic!("read test request: {error}"),
+        };
         if bytes == 0 {
             break;
         }

@@ -1218,6 +1218,11 @@ pub(crate) async fn gh_repo(
 mod tests {
     use super::*;
 
+    // Process scheduling on a loaded CI runner can outlast the command's own
+    // shutdown bound. Wait for the fake command to start before testing it.
+    #[cfg(unix)]
+    const GH_TEST_PROCESS_START_TIMEOUT: Duration = Duration::from_secs(10);
+
     #[cfg(unix)]
     struct TestGhOverride;
 
@@ -1260,7 +1265,7 @@ mod tests {
         let execution =
             tokio::spawn(async move { run_command_output(command, Some(&cancel_for_task)).await });
 
-        tokio::time::timeout(std::time::Duration::from_secs(2), async {
+        tokio::time::timeout(GH_TEST_PROCESS_START_TIMEOUT, async {
             while !pid_path.exists() {
                 tokio::time::sleep(std::time::Duration::from_millis(10)).await;
             }
@@ -1325,7 +1330,7 @@ mod tests {
             )
             .await
         });
-        tokio::time::timeout(Duration::from_secs(2), async {
+        tokio::time::timeout(GH_TEST_PROCESS_START_TIMEOUT, async {
             while !pid_path.exists() {
                 tokio::time::sleep(Duration::from_millis(10)).await;
             }
@@ -1388,7 +1393,7 @@ mod tests {
             )
             .await
         });
-        tokio::time::timeout(Duration::from_secs(2), async {
+        tokio::time::timeout(GH_TEST_PROCESS_START_TIMEOUT, async {
             while !pid_path.exists() {
                 tokio::time::sleep(Duration::from_millis(10)).await;
             }

@@ -29,11 +29,21 @@ fn argv(words: &[&str]) -> Vec<OsString> {
 fn every_native_utility_command_routes_to_the_utility_handler() {
     assert_eq!(
         NATIVE_UTILITY_COMMANDS.len(),
-        45,
-        "this test's evidence baseline is 45 utility commands; update it deliberately \
+        46,
+        "this test's evidence baseline is 46 utility commands; update it deliberately \
          if the canonical table changes size"
     );
 
+    let thread_args = argv(&["thread", "attach", "thread-id"]);
+    assert_eq!(
+        native_utility_tokens(&thread_args[1..]),
+        Some(vec![
+            "thread".to_owned(),
+            "attach".to_owned(),
+            "thread-id".to_owned()
+        ])
+    );
+
     assert!(NATIVE_UTILITY_COMMANDS.contains(&"experiments"));
     let experiment_args = argv(&["experiments", "off"]);
     assert_eq!(

@@ -6666,6 +6666,17 @@ pub(crate) const MESSAGES: &[(&str, [&str; 6])] = &[
             "完成",
         ],
     ),
+    (
+        "Complete Deixic sign-in in your browser.",
+        [
+            "Completa el inicio de sesión de Deixic en el navegador.",
+            "Terminez la connexion à Deixic dans votre navigateur.",
+            "Schließen Sie die Deixic-Anmeldung im Browser ab.",
+            "ブラウザーで Deixic へのサインインを完了してください。",
+            "브라우저에서 Deixic 로그인을 완료하세요.",
+            "请在浏览器中完成 Deixic 登录。",
+        ],
+    ),
     (
         "Complete setup to start.",
         [
@@ -15692,6 +15703,17 @@ pub(crate) const MESSAGES: &[(&str, [&str; 6])] = &[
             "偏好设置",
         ],
     ),
+    (
+        "Preparing browser sign-in…",
+        [
+            "Preparando el inicio de sesión en el navegador…",
+            "Préparation de la connexion dans le navigateur…",
+            "Browser-Anmeldung wird vorbereitet…",
+            "ブラウザーでのサインインを準備しています…",
+            "브라우저 로그인을 준비하는 중…",
+            "正在准备浏览器登录…",
+        ],
+    ),
     (
         "Preparing workspace search",
         [
@@ -15769,6 +15791,17 @@ pub(crate) const MESSAGES: &[(&str, [&str; 6])] = &[
             "按 a 以批准控股。",
         ],
     ),
+    (
+        "Press c to copy the link or o to open the browser again.",
+        [
+            "Pulsa c para copiar el enlace u o para abrir el navegador de nuevo.",
+            "Appuyez sur c pour copier le lien ou sur o pour rouvrir le navigateur.",
+            "Drücken Sie c, um den Link zu kopieren, oder o, um den Browser erneut zu öffnen.",
+            "c でリンクをコピーするか、o でブラウザーを再度開いてください。",
+            "c를 눌러 링크를 복사하거나 o를 눌러 브라우저를 다시 여세요.",
+            "按 c 复制链接，或按 o 再次打开浏览器。",
+        ],
+    ),
     (
         "Preview and select a color theme",
         [
@@ -20738,6 +20771,17 @@ pub(crate) const MESSAGES: &[(&str, [&str; 6])] = &[
             "选中的账户已更改 。 完成前再检查一遍",
         ],
     ),
+    (
+        "The session will continue when sign-in finishes.",
+        [
+            "La sesión continuará cuando finalice el inicio de sesión.",
+            "La session reprendra une fois la connexion terminée.",
+            "Die Sitzung wird fortgesetzt, sobald die Anmeldung abgeschlossen ist.",
+            "サインインが完了するとセッションが続行されます。",
+            "로그인이 완료되면 세션이 계속됩니다.",
+            "登录完成后，会话将继续。",
+        ],
+    ),
     (
         "The test sends a small fixed request to your selected model and may incur usage charges. It does not send repository contents.",
         [
@@ -23842,6 +23886,17 @@ pub(crate) const MESSAGES: &[(&str, [&str; 6])] = &[
             "等待ChatGPT登录完成...",
         ],
     ),
+    (
+        "Waiting for browser sign-in…",
+        [
+            "Esperando el inicio de sesión en el navegador…",
+            "En attente de la connexion dans le navigateur…",
+            "Warten auf die Browser-Anmeldung…",
+            "ブラウザーでのサインインを待っています…",
+            "브라우저 로그인을 기다리는 중…",
+            "正在等待浏览器登录…",
+        ],
+    ),
     (
         "Waiting for the browser callback…",
         [
@@ -24744,6 +24799,17 @@ pub(crate) const MESSAGES: &[(&str, [&str; 6])] = &[
             "内置专家简介必须来自汇编的指令",
         ],
     ),
+    (
+        "c copy   o open   u show link   enter retry   esc close",
+        [
+            "c copiar   o abrir   u mostrar enlace   enter reintentar   esc cerrar",
+            "c copier   o ouvrir   u afficher le lien   enter réessayer   esc fermer",
+            "c kopieren   o öffnen   u Link anzeigen   enter erneut versuchen   esc schließen",
+            "c コピー   o 開く   u リンク表示   enter 再試行   esc 閉じる",
+            "c 복사   o 열기   u 링크 표시   enter 재시도   esc 닫기",
+            "c 复制   o 打开   u 显示链接   enter 重试   esc 关闭",
+        ],
+    ),
     (
         "cancel",
         [

@@ -0,0 +1,12 @@
+/** Keep disputed aggregator output caps unknown until a vendor publishes them. */
+export function omitUnverifiedSonarOutput(id, output) {
+	// Perplexity documents a 128K context for Sonar Reasoning Pro, but no
+	// 4,096-token output ceiling. OpenRouter's derived 90%-of-context value
+	// falls back to the models.dev Perplexity row, which supplies 4,096.
+	// Scope the omission to this exact model and disputed value.
+	// https://docs.perplexity.ai/docs/sonar/models/sonar-reasoning-pro
+	if (id === "perplexity/sonar-reasoning-pro" && output === 4096) {
+		return undefined;
+	}
+	return output;
+}
@@ -0,0 +1,10 @@
+import assert from "node:assert/strict";
+import test from "node:test";
+
+import { omitUnverifiedSonarOutput } from "./catalog-output-limits.mjs";
+
+test("only the unverified Sonar Reasoning Pro output cap is omitted", () => {
+	assert.equal(omitUnverifiedSonarOutput("perplexity/sonar-reasoning-pro", 4096), undefined);
+	assert.equal(omitUnverifiedSonarOutput("perplexity/sonar-reasoning-pro", 8192), 8192);
+	assert.equal(omitUnverifiedSonarOutput("perplexity/another-model", 4096), 4096);
+});
@@ -1,5 +1,7 @@
 #!/usr/bin/env node
 
+import { omitUnverifiedSonarOutput } from "./catalog-output-limits.mjs";
+
 /**
  * Regenerate the bundled model catalog snapshot consumed by
  * `packages/local-host-rs/src/model_catalog.rs` via `include_str!`.
@@ -244,7 +246,7 @@ function mapOpenRouterModel(model, tokenLimits, vendorFacts) {
 				(model.reasoning != null && typeof model.reasoning === "object"),
 			streaming: true,
 			context_tokens: context,
-			output_tokens: resolveOpenRouterOutput(tokenLimits, id, context, advertised),
+			output_tokens: omitUnverifiedSonarOutput(id, resolveOpenRouterOutput(tokenLimits, id, context, advertised)),
 			// OpenRouter lists `temperature` in `supported_parameters` for
 			// routes that accept it. Omitted when the route lists no
 			// parameters at all.