vision: real-time perception — auto frame attach every N turns (--look-every) and on collision, frame context cap (6), digest fallback for text models

testbed/chat.py

@@ -122,6 +122,7 @@ async def chat_loop(client: AICCClient, manifest, args: argparse.Namespace) -> i
     # The REPL stays responsive: /status to check progress, /stop to cancel.
     mission: asyncio.Task | None = None
     mission_goal = "reach the beacon and activate it"
+    user_turns = 0
 
     def mission_log(role: str, msg: str) -> None:
         print(f"  [{role}] {msg}")
@@ -159,6 +160,7 @@ async def chat_loop(client: AICCClient, manifest, args: argparse.Namespace) -> i
                         args.mission_steps,
                         log=mission_log,
                         nudge_limit=args.mission_nudges,
+                        look_every=args.look_every,
                     )
                 except RuntimeError as exc:
                     print(f"  [mission] LLM error: {exc}")
@@ -330,6 +332,13 @@ async def chat_loop(client: AICCClient, manifest, args: argparse.Namespace) -> i
             if not auto_steps:
                 break
             # The model acted without commenting; let it keep going for 'go to X'
+        # Real-time perception in manual chat: refresh the model's view of the
+        # world on a cadence, so it reacts to what it sees without being asked.
+        user_turns += 1
+        if args.look_every and user_turns % args.look_every == 0:
+            await controller.auto_frame(
+                "Fresh camera frame for your reference — react if the world changed."
+            )
         if controller.pos is not None:
             save_map()
 
@@ -431,6 +440,12 @@ def main() -> int:
         default=None,
         help="always include the color-grid digest alongside images (off by default for multimodal models)",
     )
+    parser.add_argument(
+        "--look-every",
+        type=int,
+        default=3,
+        help="attach a fresh camera frame every N turns (0 disables; default 3)",
+    )
     args = parser.parse_args()
     try:
         return asyncio.run(run(args))