mind2web

Mind2Web/mind2web_runner.py

@@ -0,0 +1,140 @@
+import os
+import sys
+import json
+import asyncio
+from dotenv import load_dotenv
+
+PROJECT_ROOT = os.path.abspath(
+    os.path.join(os.path.dirname(__file__), "..")
+)
+
+SKILLS_DIR = os.path.join(PROJECT_ROOT, "BrowserUse_and_ComputerUse_skills")
+HERMES_CODE_DIR = os.path.join(SKILLS_DIR, "hermes_code")
+ENV_PATH = os.path.join(SKILLS_DIR, ".env")
+
+for path in [PROJECT_ROOT, SKILLS_DIR, HERMES_CODE_DIR]:
+    if path not in sys.path:
+        sys.path.append(path)
+
+load_dotenv(ENV_PATH, override=True)
+
+# локальные overrides
+os.environ["MODEL"] = "qwen3.5-122b"
+os.environ["MODEL_DEFAULT"] = "qwen3.5-122b"
+os.environ["BASE_URL"] = "https://llm.lambda.coredump.ru/v1"
+os.environ["OPENAI_BASE_URL"] = "https://llm.lambda.coredump.ru/v1"
+os.environ["API_KEY"] = os.environ.get("API_KEY", "sk-...")
+os.environ["OPENAI_API_KEY"] = os.environ.get("OPENAI_API_KEY", "sk-...")
+os.environ["PROVIDER"] = "custom"
+
+# важно: для локального запуска
+os.environ["BROWSER_URL"] = "http://localhost:9222"
+os.environ["BROWSER_VIEW_URL"] = "http://localhost:6080"
+
+print("RUNNER BROWSER_URL =", os.environ.get("BROWSER_URL"))
+print("RUNNER BROWSER_VIEW_URL =", os.environ.get("BROWSER_VIEW_URL"))
+
+from BrowserUse_and_ComputerUse_skills.hermes_code.tools.browser_use_tool import run_browser_task
+
+
+def get_task_text(task: dict) -> str:
+    for key in ("confirmed_task", "task", "instruction", "intent"):
+        value = task.get(key)
+        if isinstance(value, str) and value.strip():
+            return value.strip()
+    return ""
+
+
+def run_task(task: dict, timeout_sec: int = 300) -> dict:
+    task_text = get_task_text(task)
+
+    instruction = f"""
+Task: {task_text}
+
+Rules:
+- Do NOT use Google search.
+- Go directly to relevant websites.
+- Keep reasoning short.
+- Avoid repeating the same scroll or search action many times.
+- Prefer fast completion over exhaustive browsing.
+- If a page already shows relevant results, do not keep exploring unnecessarily.
+- If filters are available, use them directly.
+- Do not get stuck searching for perfect filters forever.
+""".strip()
+
+    try:
+        raw_result = asyncio.run(
+            asyncio.wait_for(
+                run_browser_task(instruction),
+                timeout=timeout_sec,
+            )
+        )
+
+        if isinstance(raw_result, str):
+            parsed = json.loads(raw_result)
+        elif isinstance(raw_result, dict):
+            parsed = raw_result
+        else:
+            parsed = {"success": False, "result": None, "error": f"unexpected result type: {type(raw_result).__name__}"}
+
+        result_text = (parsed.get("result") or "").lower()
+
+        failure_markers = [
+            "status: incomplete",
+            "judge verdict: ❌ fail",
+            "judge verdict: fail",
+            "task not completed",
+            "could not",
+            "unable to",
+            "manual filtering needed",
+            "recommendation:",
+            "issue:",
+            "incomplete -",
+        ]
+
+        looks_failed = any(marker in result_text for marker in failure_markers)
+        task_success = bool(parsed.get("success", False)) and not looks_failed
+
+        return {
+            "task_id": task.get("annotation_id"),
+            "instruction": instruction,
+            "website": task.get("website"),
+            "domain": task.get("domain"),
+            "subdomain": task.get("subdomain"),
+            "status": "success" if task_success else "failed",
+            "success": task_success,
+            "result": parsed.get("result"),
+            "browser_view": parsed.get("browser_view"),
+            "error": parsed.get("error"),
+            "raw": parsed,
+        }
+
+    except asyncio.TimeoutError:
+        return {
+            "task_id": task.get("annotation_id"),
+            "instruction": instruction,
+            "website": task.get("website"),
+            "domain": task.get("domain"),
+            "subdomain": task.get("subdomain"),
+            "status": "timeout",
+            "success": False,
+            "result": None,
+            "browser_view": os.getenv("BROWSER_VIEW_URL"),
+            "error": f"timeout after {timeout_sec} seconds",
+            "raw": None,
+        }
+
+    except Exception as e:
+        return {
+            "task_id": task.get("annotation_id"),
+            "instruction": instruction,
+            "website": task.get("website"),
+            "domain": task.get("domain"),
+            "subdomain": task.get("subdomain"),
+            "status": "error",
+            "success": False,
+            "result": None,
+            "browser_view": os.getenv("BROWSER_VIEW_URL"),
+            "error": str(e),
+            "raw": None,
+        }