diff --git a/.claude/skills/cogames-watch-replay/SKILL.md b/.claude/skills/cogames-watch-replay/SKILL.md
new file mode 100644
index 000000000..876e9e25c
--- /dev/null
+++ b/.claude/skills/cogames-watch-replay/SKILL.md
@@ -0,0 +1,36 @@
+---
+name: cogames-watch-replay
+description: Run a cogames episode and capture emoji map frames so you can observe agent behavior without a GUI. Use this to diagnose navigation, gear acquisition, or routing issues.
+argument-hint: "[--steps 500] [--every 50] [--agents 4] [--policy class=cogames.policy.starter_agent.StarterPolicy]"
+---
+
+Run the frame capture script with the provided arguments (or defaults), then read and analyze the output.
+
+```bash
+python scripts/capture_frames.py $ARGUMENTS --out docs/replay_frames.txt 2>&1
+```
+
+After the script completes, read the frames file and analyze it:
+
+## How to read the output
+
+**Parse programmatically, not visually** — the emoji grid is 98×98, too large to eyeball. Extract agent positions by searching for each symbol (`🟦`, `🟧`, `🟩`, `🟨`) and record their (row, col) coordinates across frames. Then compute movement deltas to detect stuck agents.
+
+**Questions to answer:**
+- Are agents moving between frames, or returning to the same position?
+- Are they spreading across the map or clustering near the hub?
+- Do they reach gear stations (🔗⛏️🔭🌀) and change gear?
+- When do alignments happen (watch reward jumps in the step header)?
+
+**Zoom into stuck areas** — once you identify where an agent is frozen, extract just the 15×15 subgrid around that position to see what's blocking it (walls ⬛, resource extractors 📦, stations).
+
+## Recommended configs
+
+- Short episode, fine-grained: `--steps 200 --every 10`
+- Standard diagnostic: `--steps 500 --every 50`
+- Full episode: `--steps 1000 --every 100`
+- Isolate single agent: `--agents 1 --steps 500 --every 50`
+- Multi-agent contention: run 1-agent vs 3-agent vs 8-agent and compare reward/alignment counts
+
+**Default policy**: `StarterPolicy` (scripted heuristics, no LLM required).
+To watch an LLM policy, pass `--policy class=your.policy.ClassName` and ensure any required API keys are set.
diff --git a/pyproject.toml b/pyproject.toml
index d4fa6aa48..e2f18fc30 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -9,7 +9,7 @@ readme = "README.md"
 requires-python = ">=3.12,<3.13"
 classifiers = ["Programming Language :: Python :: 3", "Programming Language :: Python :: 3.12"]
 dependencies = [
-  "mettagrid==0.23.1",
+  "mettagrid==0.23.3",
   "packaging>=24.0.0",
   "pufferlib-core",
   "pydantic>=2.11.5",
@@ -68,7 +68,7 @@ testpaths = ["tests"]
 source = ["cogames"]
 
 [tool.uv.sources]
-mettagrid = { workspace = true }
+mettagrid = {git = "https://github.com/Metta-AI/mettagrid.git"}
 
 [tool.ruff]
 extend = "../../.ruff.toml"
diff --git a/scripts/capture_frames.py b/scripts/capture_frames.py
new file mode 100644
index 000000000..5bd979dc5
--- /dev/null
+++ b/scripts/capture_frames.py
@@ -0,0 +1,159 @@
+"""
+Frame capture script — watch the policy play without a GUI.
+
+Runs an episode and saves emoji map snapshots at regular intervals to a text file,
+so you (or a Claude agent) can read what agents are doing across the episode.
+
+No TTY, no interaction required.
+
+Usage:
+    python scripts/capture_frames.py [options]
+
+Examples:
+    # Watch the starter policy on the default mission
+    python scripts/capture_frames.py --steps 500 --every 50
+
+    # Use your own policy class
+    python scripts/capture_frames.py --policy class=cogames.policy.starter_agent.StarterPolicy
+
+    # More agents, more frequent snapshots
+    python scripts/capture_frames.py --agents 8 --steps 1000 --every 100
+"""
+
+import argparse
+import sys
+from datetime import datetime
+from pathlib import Path
+from typing import Optional
+
+sys.path.insert(0, str(Path(__file__).parent.parent / "src"))
+
+from mettagrid.renderer.miniscope.buffer import MapBuffer
+from mettagrid.renderer.miniscope.symbol import DEFAULT_SYMBOL_MAP
+from mettagrid.renderer.renderer import Renderer
+from mettagrid.simulator.interface import SimulatorEventHandler
+from typing_extensions import override
+
+
+class FrameCaptureRenderer(Renderer):
+    """Non-interactive renderer that captures emoji map frames to a file."""
+
+    def __init__(self, output_path: Path, capture_every: int = 100):
+        super().__init__()
+        self._output_path = output_path
+        self._capture_every = capture_every
+        self._map_buffer: Optional[MapBuffer] = None
+        self._frames: list[tuple[int, str, str]] = []  # (step, reward_str, grid)
+
+    @override
+    def on_episode_start(self) -> None:
+        symbol_map = DEFAULT_SYMBOL_MAP.copy()
+        # Extend with game-specific render symbols if the config exposes them
+        for obj in self._sim.config.game.objects.values():
+            render_symbol = getattr(obj, "render_symbol", None)
+            if render_symbol:
+                key = getattr(obj, "render_name", None) or obj.name
+                symbol_map[key] = render_symbol
+                if getattr(obj, "render_name", None) and obj.render_name != obj.name:
+                    symbol_map[obj.name] = render_symbol
+        self._map_buffer = MapBuffer(
+            symbol_map=symbol_map,
+            initial_height=self._sim.map_height,
+            initial_width=self._sim.map_width,
+        )
+        self._frames = []
+        self._capture(0)
+
+    @override
+    def on_step(self) -> None:
+        step = self._sim.current_step
+        if step % self._capture_every == 0:
+            self._capture(step)
+
+    @override
+    def on_episode_end(self) -> None:
+        self._capture(self._sim.current_step)
+        self._write()
+
+    def _capture(self, step: int) -> None:
+        if self._map_buffer is None:
+            return
+        grid_objects = self._sim.grid_objects()
+        grid = self._map_buffer.render_full_map(grid_objects)
+        rewards = self._sim.episode_rewards
+        reward_str = " ".join(f"{r:.4f}" for r in rewards) if rewards is not None else "n/a"
+        self._frames.append((step, reward_str, grid))
+
+    def _write(self) -> None:
+        self._output_path.parent.mkdir(parents=True, exist_ok=True)
+        with open(self._output_path, "w") as f:
+            f.write(f"# Frame Capture — {datetime.now().isoformat()}\n")
+            f.write(f"# map={self._sim.map_height}x{self._sim.map_width}  "
+                    f"agents={self._sim.num_agents}  frames={len(self._frames)}\n")
+            f.write("# Symbols: 🟦=agent0 🟧=agent1 🟩=agent2 🟨=agent3  ⬛=wall  · =empty\n\n")
+            for step, reward_str, grid in self._frames:
+                f.write(f"{'='*60}\n")
+                f.write(f"Step {step:4d}  rewards=[{reward_str}]\n")
+                f.write(f"{'='*60}\n")
+                f.write(grid)
+                f.write("\n\n")
+        print(f"Saved {len(self._frames)} frames → {self._output_path}")
+
+
+def main():
+    parser = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
+    parser.add_argument("--mission", default="machina_1",
+                        help="Mission name (default: machina_1)")
+    parser.add_argument("--policy", default="class=cogames.policy.starter_agent.StarterPolicy",
+                        help="Policy spec (default: starter policy, no LLM required)")
+    parser.add_argument("--agents", type=int, default=4,
+                        help="Number of agents (default: 4)")
+    parser.add_argument("--steps", type=int, default=500,
+                        help="Episode length in steps (default: 500)")
+    parser.add_argument("--every", type=int, default=50,
+                        help="Capture a frame every N steps (default: 50)")
+    parser.add_argument("--seed", type=int, default=42)
+    parser.add_argument("--out", default="docs/replay_frames.txt",
+                        help="Output file path (default: docs/replay_frames.txt)")
+    args = parser.parse_args()
+
+    output_path = Path(args.out)
+    capturer = FrameCaptureRenderer(output_path=output_path, capture_every=args.every)
+
+    from cogames.cli.mission import get_mission
+    from cogames.cli.policy import parse_policy_spec
+    from mettagrid.policy.loader import initialize_or_load_policy
+    from mettagrid.policy.policy_env_interface import PolicyEnvInterface
+    from mettagrid.runner.rollout import resolve_env_for_seed, Rollout
+    from mettagrid.simulator.time_averaged_stats import TimeAveragedStatsHandler
+
+    _name, env_cfg, _mission = get_mission(args.mission, cogs=args.agents)
+    env_cfg.game.max_steps = args.steps
+
+    policy_spec = parse_policy_spec(args.policy).to_policy_spec()
+
+    env_for_rollout = resolve_env_for_seed(env_cfg, args.seed)
+    env_interface = PolicyEnvInterface.from_mg_cfg(env_for_rollout)
+    multi_policy = initialize_or_load_policy(env_interface, policy_spec)
+
+    num_agents = env_for_rollout.game.num_agents
+    agent_policies = [multi_policy.agent_policy(i) for i in range(num_agents)]
+
+    stats_handler = TimeAveragedStatsHandler()
+    rollout = Rollout(
+        env_for_rollout,
+        agent_policies,
+        render_mode="none",
+        seed=args.seed,
+        event_handlers=[stats_handler, capturer],
+        autostart=True,
+    )
+    rollout.run_until_done()
+
+    rewards = list(rollout._sim.episode_rewards)
+    print(f"\nEpisode done. Steps: {rollout._sim.current_step}  "
+          f"Avg reward/agent: {sum(rewards)/len(rewards):.4f}")
+
+
+if __name__ == "__main__":
+    main()