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/demo.py

@@ -141,6 +141,7 @@ async def run_llm_agent(
     recorder: FrameRecorder | None = None,
     vision: bool | None = None,
     digest: bool | None = None,
+    look_every: int = 0,
 ) -> dict[str, Any]:
     """Autonomous LLM run: controller + nudge/correct loop (see llm_agent)."""
     from testbed.llm_agent import LLMController, run_llm_agent_loop
@@ -313,6 +314,7 @@ async def run_demo(args: argparse.Namespace) -> dict[str, Any]:
                 recorder=recorder,
                 vision=args.vision,
                 digest=args.digest,
+                look_every=args.look_every,
             )
         elif args.agent == "scripted":
             summary = await run_scripted_agent(
@@ -335,6 +337,7 @@ async def run_demo(args: argparse.Namespace) -> dict[str, Any]:
                     recorder=recorder,
                     vision=args.vision,
                     digest=args.digest,
+                    look_every=args.look_every,
                 )
                 if summary.get("interacted"):
                     return summary
@@ -408,6 +411,12 @@ 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 agent steps (0 disables; default 3)",
+    )
     parser.add_argument(
         "--frame",
         default="demo_final_frame.png",