fix: harden inference against validator runtime failures

inference.py

@@ -19,6 +19,8 @@ MODEL_NAME = os.getenv("MODEL_NAME", "Qwen/Qwen2.5-72B-Instruct")
 API_KEY = os.getenv("HF_TOKEN") or os.getenv("API_KEY")
 LOCAL_IMAGE_NAME = os.getenv("LOCAL_IMAGE_NAME")
 ENV_BASE_URL = os.getenv("ENV_BASE_URL")
+DEFAULT_ENV_BASE_URL = os.getenv("DEFAULT_ENV_BASE_URL", "http://127.0.0.1:8000")
+DEFAULT_LOCAL_IMAGE_NAME = os.getenv("DEFAULT_LOCAL_IMAGE_NAME", "code-security-auditor-env:latest")
 TASK_IDS = [t.strip() for t in os.getenv("TASK_IDS", "easy,medium,hard").split(",") if t.strip()]
 MAX_STEPS = int(os.getenv("MAX_STEPS", "12"))
 TEMPERATURE = 0.0
@@ -65,6 +67,13 @@ def _default_action() -> Dict[str, Any]:
     }
 
 
+def _safe_error(exc: Exception) -> str:
+    msg = str(exc).strip()
+    if not msg:
+        msg = exc.__class__.__name__
+    return msg.replace("\n", " ")[:240]
+
+
 def _parse_action(raw: str, available_files: List[str]) -> Dict[str, Any]:
     try:
         parsed = json.loads(raw)
@@ -145,13 +154,17 @@ def _query_model(client: OpenAI, obs: Any, step: int) -> Dict[str, Any]:
 
 
 async def _create_env() -> CodeSecurityAuditorEnv:
-    if LOCAL_IMAGE_NAME:
-        return await CodeSecurityAuditorEnv.from_docker_image(LOCAL_IMAGE_NAME)
+    # Prefer explicit configuration, then fall back to common local defaults.
     if ENV_BASE_URL:
         return CodeSecurityAuditorEnv(base_url=ENV_BASE_URL)
-    raise RuntimeError(
-        "Set LOCAL_IMAGE_NAME (docker mode) or ENV_BASE_URL (remote mode) to run inference."
-    )
+
+    if LOCAL_IMAGE_NAME:
+        return await CodeSecurityAuditorEnv.from_docker_image(LOCAL_IMAGE_NAME)
+
+    try:
+        return CodeSecurityAuditorEnv(base_url=DEFAULT_ENV_BASE_URL)
+    except Exception:
+        return await CodeSecurityAuditorEnv.from_docker_image(DEFAULT_LOCAL_IMAGE_NAME)
 
 
 async def run_task(env: CodeSecurityAuditorEnv, client: OpenAI, task_id: str) -> float:
@@ -191,6 +204,14 @@ async def run_task(env: CodeSecurityAuditorEnv, client: OpenAI, task_id: str) ->
         score = float(obs.reward or 0.0)
         score = min(max(score, 0.0), 1.0)
         success = score >= 0.6
+    except Exception as exc:
+        # Keep evaluator contract: do not crash inference.py on transient/runtime errors.
+        log_step(step=max(1, steps_taken), action="{}", reward=0.0, done=True, error=_safe_error(exc))
+        if not rewards:
+            rewards.append(0.0)
+        steps_taken = max(1, steps_taken)
+        score = 0.0
+        success = False
     finally:
         log_end(success=success, steps=steps_taken, score=score, rewards=rewards)
 
@@ -198,11 +219,21 @@ async def run_task(env: CodeSecurityAuditorEnv, client: OpenAI, task_id: str) ->
 
 
 async def main() -> None:
-    if not API_KEY:
-        raise RuntimeError("HF_TOKEN (or API_KEY) is required for inference.")
+    # Keep script resilient in validators even if a key is temporarily unavailable.
+    api_key = API_KEY or "missing"
 
-    client = OpenAI(base_url=API_BASE_URL, api_key=API_KEY)
-    env = await _create_env()
+    client = OpenAI(base_url=API_BASE_URL, api_key=api_key)
+
+    try:
+        env = await _create_env()
+    except Exception as exc:
+        # Emit structured logs for each task and exit cleanly.
+        err = _safe_error(exc)
+        for task_id in TASK_IDS:
+            log_start(task=task_id, env=BENCHMARK, model=MODEL_NAME)
+            log_step(step=1, action="{}", reward=0.0, done=True, error=err)
+            log_end(success=False, steps=1, score=0.0, rewards=[0.0])
+        return
 
     try:
         scores: List[float] = []