arena: v31 first run void -- blocker message truncated + stale bbox + empty-completion retries; all fixed and verified

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

e2e/browser-v3/arena/ARENA.md

@@ -339,6 +339,15 @@ Prediction: >=3 of 8 targets (click-widget compositions specifically; the blocke
 now actually gets exercised when the model clicks the real covered input), controls 12/12
 (suppression only fires on single-child weak-named shells).
 
+FIRST RUN VOID (implementation bugs, not a verdict): targets 0/8 but the labeled trace showed
+the BLOCKED message never reached the model — (a) it was APPENDED after Playwright's multi-line
+call log and truncated out of both the record and history caps; (b) the bbox source
+(extra_element_properties) goes stale post-step and pointed the probe at BODY. Fixed: message
+prepended; rect computed in-page from the live bid element (verified: probe names
+DIV.ui-dialog-titlebar). Also fixed alongside: empty LLM completions (thinking exhausting
+max_tokens) are now retried with doubled budget instead of being booked as no-action turns —
+this was the entire login-composition 0-step failure mode. Prediction unchanged; pilot rerun.
+
 ## Positioning vs public generic-harness baselines (user-supplied 2026 survey)
 
 The comparable class is generic agents, NOT MiniWoB-specialized systems (HTML-T5++ 95.2 trained

e2e/browser-v3/arena/llm_policy.py

@@ -244,6 +244,12 @@ class LlmPolicy:
                     payload["temperature"] = self.temperature
                 resp = post_json(f"{self.endpoint}/v1/chat/completions", payload)
                 text = (resp.get("choices") or [{}])[0].get("message", {}).get("content") or ""
+                # Thinking models sometimes spend the whole budget before emitting text; an empty
+                # reply must be retried like any fault, never silently booked as a no-action turn.
+                if not text.strip():
+                    payload["max_tokens"] = min(int(self.max_tokens * 2), 2000)
+                    self.max_tokens = payload["max_tokens"]
+                    raise ValueError("empty completion (thinking exhausted max_tokens?)")
                 usage = resp.get("usage") or {}
                 d.prompt_tokens = int(usage.get("prompt_tokens") or 0)
                 d.completion_tokens = int(usage.get("completion_tokens") or 0)

e2e/browser-v3/arena/run.py

@@ -218,16 +218,24 @@ def run_episode(arm: str, task: str, seed: int, rec: Recorder, args: argparse.Na
             if (getattr(policy, "blocker_probe", False) and "Timeout" in err
                     and re.match(r"(?:dbl)?click\(", decision.action)):
                 m_bid = re.search(r'"([^"]+)"', decision.action)
-                bbox = ((obs.get("extra_element_properties") or {}).get(m_bid.group(1)) or {}).get("bbox") if m_bid else None
-                if bbox:
+                if m_bid:
                     try:
+                        # Rect computed in-page from the live element (extra_element_properties
+                        # bboxes go stale post-step and once pointed the probe at BODY).
                         top = with_deadline(lambda: env.unwrapped.page.evaluate(
-                            "([x,y]) => { const e = document.elementFromPoint(x,y);"
-                            " return e ? e.tagName + (e.id?'#'+e.id:'') +"
-                            " (e.className&&typeof e.className==='string'?'.'+e.className.split(' ')[0]:'') : ''; }",
-                            [bbox[0] + bbox[2] / 2, bbox[1] + bbox[3] / 2]), 8)
+                            "(bid) => { const el = document.querySelector(`[bid=\"${bid}\"]`);"
+                            " if (!el) return ''; const r = el.getBoundingClientRect();"
+                            " const e = document.elementFromPoint(r.x + r.width/2, r.y + r.height/2);"
+                            " if (!e || e === el || el.contains(e)) return '';"
+                            " return e.tagName + (e.id?'#'+e.id:'') +"
+                            " (e.className&&typeof e.className==='string'?'.'+e.className.split(' ')[0]:''); }",
+                            m_bid.group(1)), 8)
                         if top:
-                            err += f" | BLOCKED: {top} is covering this element -- move or close the cover first (drag its titlebar or dismiss it), then retry"
+                            # PREPENDED: a suffix after Playwright's multi-line call log gets
+                            # truncated out of both the record and the model's history.
+                            err = (f"BLOCKED: {top} is covering the element you clicked -- move or "
+                                   f"close the cover first (drag its titlebar or dismiss it), then "
+                                   f"retry | {err}")
                             obs["last_action_error"] = err
                     except Exception:
                         pass