@vincemakes/kiso-evals 0.1.28 → 0.1.30

package/dist/fixtures/silent-tool-failure.js

@@ -31,7 +31,7 @@ export const silentToolFailure = {
         },
         {
             events: [
-                { type: "text_delta", text: "搜索完成，结果很好。" },
+                { type: "text_delta", text: "The search finished successfully." },
                 { type: "text_end" },
                 { type: "stop", reason: "end_turn" },
             ],
@@ -42,7 +42,7 @@ export const silentToolFailure = {
         if (!events.some((e) => e.type === "tool_call_end" && e.name === "web_search")) {
             violations.push("script lost the failing tool call");
         }
-        if (!events.some((e) => e.type === "text_delta" && /很好|成功|done/i.test(e.text))) {
+        if (!events.some((e) => e.type === "text_delta" && /success|done/i.test(e.text))) {
             violations.push("script lost the success narration that swallows the error");
         }
         return violations;

package/dist/fixtures/terminal-lies.d.ts

@@ -2,7 +2,7 @@
  * Fixture: TERMINAL LIES — the model reports done work that never landed.
  *
  * Incident (uooki production, 2026): a long-document turn called
- * `document_create` twice, the chat asserted "已创建文档产物", and the
+ * `document_create` twice, the chat asserted "created the document artifact", and the
  * Canvas stayed empty — the tool was an inline Markdown table that never
  * reaches the canvas, and all three delivery guards were blind to it
  * because it was absent from the producer registry. The turn was reported

package/dist/fixtures/terminal-lies.js

@@ -2,7 +2,7 @@
  * Fixture: TERMINAL LIES — the model reports done work that never landed.
  *
  * Incident (uooki production, 2026): a long-document turn called
- * `document_create` twice, the chat asserted "已创建文档产物", and the
+ * `document_create` twice, the chat asserted "created the document artifact", and the
  * Canvas stayed empty — the tool was an inline Markdown table that never
  * reaches the canvas, and all three delivery guards were blind to it
  * because it was absent from the producer registry. The turn was reported
@@ -16,7 +16,7 @@
 import { analyzeDelivery } from "@vincemakes/kiso-core";
 export const terminalLies = {
     name: "terminal-lies",
-    incident: "uooki long-document turn: model claimed '已创建文档产物', canvas stayed empty, turn reported completed (document_create trap, 2026-07-31)",
+    incident: "uooki long-document turn: model claimed 'created the document artifact', canvas stayed empty, turn reported completed (document_create trap, 2026-07-31)",
     delivery: {
         required: true,
         producers: new Set(["create_artifact", "docx_create"]),
@@ -25,7 +25,7 @@ export const terminalLies = {
         {
             events: [
                 { type: "text_start" },
-                { type: "text_delta", text: "我已经生成了一份完整的调研报告，并已交付到画布上。" },
+                { type: "text_delta", text: "I have generated a complete research report and delivered it to the canvas." },
                 { type: "text_end" },
                 { type: "stop", reason: "end_turn" },
             ],
@@ -33,7 +33,7 @@ export const terminalLies = {
     ],
     staticCheck: (events) => {
         const violations = [];
-        const claimsDelivery = events.some((e) => e.type === "text_delta" && /交付|已生成|completed|delivered/i.test(e.text));
+        const claimsDelivery = events.some((e) => e.type === "text_delta" && /completed|delivered/i.test(e.text));
         const produced = events.some((e) => e.type === "tool_call_end" && e.name === "create_artifact");
         if (claimsDelivery && !produced) {
             // This IS the failure shape: a claim with zero producers.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@vincemakes/kiso-evals",
-  "version": "0.1.28",
+  "version": "0.1.30",
   "description": "kiso evals — the faux provider, the incident fixture library, and the cross-provider contract tests.",
   "type": "module",
   "license": "MIT",
@@ -21,12 +21,12 @@
     "test": "vitest run"
   },
   "dependencies": {
-    "@vincemakes/kiso-core": "0.1.28"
+    "@vincemakes/kiso-core": "0.1.29"
   },
   "devDependencies": {
     "@anthropic-ai/sdk": "^0.115.0",
-    "@vincemakes/kiso-provider-anthropic": "0.1.28",
-    "@vincemakes/kiso-provider-openai": "0.1.28",
+    "@vincemakes/kiso-provider-anthropic": "0.1.29",
+    "@vincemakes/kiso-provider-openai": "0.1.29",
     "@types/node": "^26.1.2",
     "openai": "^7.3.0",
     "typescript": "^5.7.2",