Improve inference script robustness and update defaults

- Use Groq API with Llama 3.3 70B as default (faster, best scores)
- Add retry with backoff for rate limits (429) and connection errors
- Improve JSON parsing: handle markdown fences, try full-text parse first
- Set ENV_BASE_URL default to live HF Space URL

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

inference.py

@@ -15,14 +15,10 @@ import os
 import re
 import sys
 import textwrap
-from typing import List, Optional
+import time
 
-from openai import OpenAI
-
-# ---------------------------------------------------------------------------
-# Inline client (HTTP) so inference.py is self-contained
-# ---------------------------------------------------------------------------
 import requests
+from openai import OpenAI
 
 
 class _StepResult:
@@ -39,16 +35,25 @@ class _SimpleClient:
         self.base_url = base_url.rstrip("/")
         self.s = requests.Session()
 
+    def _post(self, path: str, payload: dict) -> dict:
+        """POST with retry on transient errors."""
+        for attempt in range(3):
+            try:
+                r = self.s.post(f"{self.base_url}{path}", json=payload, timeout=60)
+                r.raise_for_status()
+                return r.json()
+            except (requests.ConnectionError, requests.Timeout) as exc:
+                if attempt < 2:
+                    time.sleep(2 ** attempt)
+                    continue
+                raise
+
     def reset(self, task_name: str = "easy") -> _StepResult:
-        r = self.s.post(f"{self.base_url}/reset", json={"task_name": task_name}, timeout=30)
-        r.raise_for_status()
-        d = r.json()
+        d = self._post("/reset", {"task_name": task_name})
         return _StepResult(d.get("observation", {}), float(d.get("reward", 0)), bool(d.get("done", False)))
 
     def step(self, action: dict) -> _StepResult:
-        r = self.s.post(f"{self.base_url}/step", json=action, timeout=30)
-        r.raise_for_status()
-        d = r.json()
+        d = self._post("/step", action)
         return _StepResult(d.get("observation", {}), float(d.get("reward", 0)), bool(d.get("done", False)))
 
     def close(self):
@@ -58,12 +63,11 @@ class _SimpleClient:
 # ---------------------------------------------------------------------------
 # Configuration
 # ---------------------------------------------------------------------------
-API_BASE_URL = os.getenv("API_BASE_URL", "https://router.huggingface.co/v1")
+API_BASE_URL = os.getenv("API_BASE_URL", "https://api.groq.com/openai/v1")
 API_KEY = os.getenv("HF_TOKEN") or os.getenv("API_KEY")
-MODEL_NAME = os.getenv("MODEL_NAME")
+MODEL_NAME = os.getenv("MODEL_NAME", "llama-3.3-70b-versatile")
 
-# Where the DataClean env server is running
-ENV_BASE_URL = os.getenv("ENV_BASE_URL", "http://localhost:7860")
+ENV_BASE_URL = os.getenv("ENV_BASE_URL", "https://glitchghost-dataclean-openenv.hf.space")
 
 MAX_STEPS_PER_TASK = {"easy": 12, "medium": 20, "hard": 30}
 TEMPERATURE = 0.1
@@ -109,21 +113,34 @@ RULES:
 # ---------------------------------------------------------------------------
 # Helpers
 # ---------------------------------------------------------------------------
-ACTION_JSON_RE = re.compile(r"\{[^}]+\}", re.DOTALL)
+ACTION_JSON_RE = re.compile(r"\{[^{}]*\}", re.DOTALL)
+# Also match JSON that may span multiple lines or have nested content
+ACTION_JSON_GREEDY_RE = re.compile(r"\{.*?\}", re.DOTALL)
 
 
 def parse_action(text: str) -> dict:
     """Extract the first JSON object from the model response."""
     if not text:
         return {"action_type": "noop"}
-    m = ACTION_JSON_RE.search(text)
-    if m:
-        try:
-            obj = json.loads(m.group(0))
-            if "action_type" in obj:
-                return obj
-        except json.JSONDecodeError:
-            pass
+    # Strip markdown code fences if present
+    cleaned = re.sub(r"```(?:json)?\s*", "", text)
+    cleaned = re.sub(r"```", "", cleaned).strip()
+    # Try parsing the whole cleaned text as JSON first
+    try:
+        obj = json.loads(cleaned)
+        if isinstance(obj, dict) and "action_type" in obj:
+            return obj
+    except (json.JSONDecodeError, ValueError):
+        pass
+    # Try regex extraction
+    for pattern in [ACTION_JSON_RE, ACTION_JSON_GREEDY_RE]:
+        for m in pattern.finditer(cleaned):
+            try:
+                obj = json.loads(m.group(0))
+                if isinstance(obj, dict) and "action_type" in obj:
+                    return obj
+            except (json.JSONDecodeError, ValueError):
+                continue
     return {"action_type": "noop"}
 
 
@@ -182,18 +199,26 @@ def run_task(
             {"role": "user", "content": user_prompt},
         ]
 
-        try:
-            completion = llm_client.chat.completions.create(
-                model=MODEL_NAME,
-                messages=messages,
-                temperature=TEMPERATURE,
-                max_tokens=MAX_TOKENS,
-                stream=False,
-            )
-            response_text = completion.choices[0].message.content or ""
-        except Exception as exc:
-            print(f"  Step {step}: LLM error ({exc}), using noop")
-            response_text = '{"action_type": "noop"}'
+        for _attempt in range(3):
+            try:
+                completion = llm_client.chat.completions.create(
+                    model=MODEL_NAME,
+                    messages=messages,
+                    temperature=TEMPERATURE,
+                    max_tokens=MAX_TOKENS,
+                    stream=False,
+                )
+                response_text = completion.choices[0].message.content or ""
+                break
+            except Exception as exc:
+                if "429" in str(exc) and _attempt < 2:
+                    wait = 5 * (2 ** _attempt)
+                    print(f"  Step {step}: Rate limited, waiting {wait}s...")
+                    time.sleep(wait)
+                    continue
+                print(f"  Step {step}: LLM error ({exc}), using noop")
+                response_text = '{"action_type": "noop"}'
+                break
 
         action = parse_action(response_text)
         print(f"  Step {step}: {action.get('action_type', '?')}", end="")