arena: v8 ties v7's 56.8 with a different shape -- drag and spatial now BEAT the real browser-use

Scripted exact-match fastpath + exact-name guard moved the wins where near-miss clicks
were terminal: drag 8/13 vs their 6/13, spatial 7/13 vs their 4/13, at 4.8s and zero
false claims. The remaining systematic gap is runway, not ability: 21 of 54 losses die
exactly at the 12-step cap, all in the email/forms/long-flow cluster their 44s median
happily outlasts. v9 sweeps with a 24-step cap (wins end at median 3 steps, so the win
path pays nothing) plus progressive vision -- any episode that burns 6 actions gets a
screenshot every turn thereafter.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WsbS5x2rYsMDxP2kW3qqmQ

e2e/browser-v3/arena/llm_policy.py

@@ -270,8 +270,12 @@ class OpenSwarmLlmPolicy(LlmPolicy):
                 return d
         last_err = bool(str(obs.get("last_action_error") or "").strip())
         image = ""
-        if self.vision == "always" or (self.vision == "adaptive"
-                                       and self.stuck_or_spatial(page, goal, last_err)):
+        # progressive: adaptive triggers PLUS any task that has already burned 6 actions gets eyes
+        # every turn -- a dragging episode is by definition one the text view is not solving.
+        struggling = self.vision == "progressive" and len(self.history) >= 6
+        if (self.vision == "always"
+                or (self.vision in ("adaptive", "progressive")
+                    and (self.stuck_or_spatial(page, goal, last_err) or struggling))):
             image = encode_screenshot(obs)
         raw, d = self.call(goal, page, image_b64=image)
         d.vision = 1 if image else 0
@@ -398,4 +402,7 @@ def build(name: str, model: str = "", endpoint: str = "", **_: Any) -> Any:
     if name == "osw-llm-v8":
         v8 = dict(v7, system=OSW_SYSTEM_V8)
         return OpenSwarmLlmPolicy(name=name, multi=True, vision="adaptive", fastpath=True, **v8)
+    if name == "osw-llm-v9":  # v8 + progressive vision; run with --max-steps 24 (21 losses were step-capped)
+        v9 = dict(v7, system=OSW_SYSTEM_V8)
+        return OpenSwarmLlmPolicy(name=name, multi=True, vision="progressive", fastpath=True, **v9)
     raise SystemExit(f"unknown arm: {name}")

e2e/browser-v3/arena/supervisor.py

@@ -54,7 +54,8 @@ def spawn(arm: str, tasks: list[str], args: argparse.Namespace) -> subprocess.Po
                "--episode-timeout", str(args.episode_timeout)]
     else:
         cmd = [PY, str(ARENA / "run.py"), "--arm", arm, "--tasks", ",".join(tasks), "--seeds", "1",
-               "--seed-base", str(args.seed), "--model", args.model, "--shots", args.shots]
+               "--seed-base", str(args.seed), "--model", args.model, "--shots", args.shots,
+               "--max-steps", str(args.max_steps)]
     env = dict(os.environ)
     env.setdefault("MINIWOB_URL", "http://localhost:8099/miniwob/")
     env.setdefault("BROWSER_USE_LOGGING_LEVEL", "warning")
@@ -81,6 +82,7 @@ def main() -> None:
     ap.add_argument("--stall", type=float, default=240.0, help="kill if no new episode lands this long")
     ap.add_argument("--rounds", type=int, default=12)
     ap.add_argument("--episode-timeout", type=float, default=100.0)
+    ap.add_argument("--max-steps", type=int, default=12)
     ap.add_argument("--shots", default="first-last")
     ap.add_argument("--log", default="")
     # Resume window: episodes recorded after this epoch count as done, so a restarted supervisor