fix(yutori): treat wait/hold_key duration as seconds, not ms

Yutori's reference impl (frontend-visualqa actions.py:469, 506)
interprets the model-supplied `duration` argument on `wait` and
`hold_key` as seconds — passed straight to asyncio.sleep on the wait
path, and clamped to 100s on the hold_key path. Our handlers were
treating duration as milliseconds, so any model-supplied value was
silently interpreted 1000× too short (`wait { duration: 2 }` slept 2ms
instead of 2s; `hold_key { duration: 0.5 }` held 0.5ms instead of
500ms). Defaults were unaffected because they were pre-computed in ms.

Convert seconds → ms before passing to Kernel's pressKey, and use
seconds directly for asyncio.sleep / setTimeout. Adds a `> 0` guard on
hold_key duration, which also resolves the bugbot nit about negative
values reaching the SDK.

Author: dprevoznik

pkg/templates/python/yutori/tools/computer.py

@@ -251,12 +251,14 @@ async def _handle_hold_key(self, action: N15Action) -> ToolResult:
             raise ToolError("key is required for hold_key action")
 
         mapped_key = self._map_key(key)
-        duration = action.get("duration") or 1000
+        # Yutori emits `duration` in seconds; Kernel SDK's press_key takes ms.
+        duration_s = action.get("duration")
+        duration_ms = int(duration_s * 1000) if duration_s and duration_s > 0 else 1000
 
         self.kernel.browsers.computer.press_key(
             self.session_id,
             keys=[mapped_key],
-            duration=duration,
+            duration=duration_ms,
         )
 
         await asyncio.sleep(SCREENSHOT_DELAY_S)
@@ -276,8 +278,9 @@ async def _handle_drag(self, action: N15Action) -> ToolResult:
         return await self.screenshot()
 
     async def _handle_wait(self, action: N15Action) -> ToolResult:
+        # Yutori emits `duration` in seconds (matches reference impl).
         duration = action.get("duration")
-        seconds = (duration / 1000) if duration and duration > 0 else 2
+        seconds = duration if duration and duration > 0 else 2
         await asyncio.sleep(seconds)
         return await self.screenshot()
 

pkg/templates/typescript/yutori/tools/computer.ts

@@ -285,7 +285,8 @@ export class ComputerTool {
     }
 
     const mappedKey = this.mapKey(key);
-    const durationMs = action.duration && action.duration > 0 ? action.duration : 1000;
+    // Yutori emits `duration` in seconds; Kernel SDK's pressKey takes ms.
+    const durationMs = action.duration && action.duration > 0 ? Math.round(action.duration * 1000) : 1000;
 
     await this.kernel.browsers.computer.pressKey(this.sessionId, {
       keys: [mappedKey],
@@ -310,7 +311,8 @@ export class ComputerTool {
   }
 
   private async handleWait(action: N15Action): Promise<ToolResult> {
-    const durationMs = action.duration && action.duration > 0 ? action.duration : 2000;
+    // Yutori emits `duration` in seconds (matches reference impl).
+    const durationMs = action.duration && action.duration > 0 ? Math.round(action.duration * 1000) : 2000;
     await this.sleep(durationMs);
     return this.screenshot();
   }