Allow stopping of inference

sia/web/api.py

@@ -30,6 +30,7 @@ class Api:
     def _init_routes(self):
         """Initialize REST API and WebSocket routes."""
         self._app.router.add_post("/api/inference/{llm}", self._run_inference)
+        self._app.router.add_post("/api/inference/{llm}/stop", self._stop_inference)
         self._app.router.add_post("/api/approve/{llm}", self._approve_response)
         self._app.router.add_post("/api/context", self._modify_context)
         self._app.router.add_post("/api/input", self._send_input)
@@ -60,6 +61,15 @@ class Api:
         except (ValueError, RuntimeError) as e:
             return web.Response(status=400, text=str(e))
 
+    async def _stop_inference(self, request: web.Request) -> web.Response:
+        """Stop inference on specified LLM."""
+        llm_name = request.match_info["llm"]
+        try:
+            self._agent.stop_inference(llm_name)
+            return web.Response(status=200)
+        except ValueError as e:
+            return web.Response(status=400, text=str(e))
+
     async def _approve_response(self, request: web.Request) -> web.Response:
         """Approve response from specified LLM."""
         llm_name = request.match_info["llm"]