from typing import List, Dict, Any, Optional from sqlalchemy.orm import Session from app.db import models from app.core.skills.base import BaseSkill import logging import time import os from app.core.tools.registry import tool_registry import time from app.core._regex import ( SKILL_CONFIG_JSON, SKILL_DESC_OVERRIDE, SKILL_PARAM_TABLE, SKILL_BASH_LOGIC ) import json import yaml import litellm import shlex logger = logging.getLogger(__name__) class ToolService: """ Orchestrates AI tools (Skills) available to users. Handles discovery, permission checks, and execution routing. """ def __init__(self, services: Any = None, local_skills: List[BaseSkill] = []): self._services = services self._local_skills = {s.name: s for s in local_skills} tool_registry.load_plugins() def get_available_tools(self, db: Session, user_id: str, feature: str = None, session_id: int = None) -> List[Dict[str, Any]]: allowed_skill_names = self._get_allowed_skills(db, session_id) # 1. Local Skills local_skills = [s.to_tool_definition() for s in self._local_skills.values() if not feature or feature in getattr(s, "features", ["chat"])] if allowed_skill_names is not None: local_skills = [t for t in local_skills if t["function"]["name"] in allowed_skill_names] # 2. VFS/FS Skills from app.core.skills.fs_loader import fs_loader db_skills = [] max_md_len = self._resolve_model_max_len(db, user_id) for fs_skill in fs_loader.get_all_skills(): if not fs_skill.get("is_enabled", True): continue if not (fs_skill.get("is_system") or fs_skill.get("owner_id") == user_id): continue if feature and feature not in fs_skill.get("features", ["chat"]): continue if allowed_skill_names is not None and fs_skill["name"] not in allowed_skill_names: continue if any(t["function"]["name"] == fs_skill["name"] for t in local_skills): continue db_skills.append(self._parse_vfs_skill(fs_skill)) return local_skills + db_skills def _get_allowed_skills(self, db: Session, session_id: int) -> Optional[set]: if not session_id or not db: return None session_obj = db.query(models.Session).filter(models.Session.id == session_id).first() if not session_obj or not getattr(session_obj, "restrict_skills", False): return None allowed = set() if session_obj.allowed_skill_names: allowed.update(session_obj.allowed_skill_names) if getattr(session_obj, "skills", None): allowed.update(s.name for s in session_obj.skills) return allowed def _resolve_model_max_len(self, db: Session, user_id: str) -> int: from app.config import settings m_name = settings.ACTIVE_LLM_PROVIDER try: user = db.query(models.User).filter(models.User.id == user_id).first() if db else None if user and user.preferences: llm_prefs = user.preferences.get("llm", {}) m_name = llm_prefs.get("model") or user.preferences.get("llm_model", m_name) if "/" not in m_name: p = llm_prefs.get("active_provider") or user.preferences.get("llm_provider", settings.ACTIVE_LLM_PROVIDER) m_name = f"{p}/{m_name}" info = litellm.get_model_info(m_name) if info: max_t = info.get("max_input_tokens", 8192) return max(min(int(max_t * 4 * 0.05), 40000), 1000) except Exception as e: logger.warning(f"Tool schema truncation fail: {e}") return 1000 def _parse_vfs_skill(self, fs_skill: dict) -> dict: name = fs_skill["name"] description = fs_skill.get("description", "") parameters = {"type": "object", "properties": {}, "required": []} # Binary/VFS normalization class _Obj: def __init__(self, d): for k, v in d.items(): setattr(self, k, v) files = [_Obj(f) for f in fs_skill.get("files", [])] skill_md = next((f for f in files if f.file_path == "SKILL.md"), None) if skill_md and skill_md.content: content = str(skill_md.content) exec_file = next((f.file_path for f in files if f.file_path.endswith((".sh", ".py")) or "run." in f.file_path), "") exec_cmd = f"bash .skills/{name}/{exec_file}" if exec_file.endswith(".sh") else f"python3 .skills/{name}/{exec_file}" if exec_file.endswith(".py") else f".skills/{name}/{exec_file}" description += f"\n\n[Native VFS Skill - Execute via: `{exec_cmd}`]\n{content}" # YAML Frontmatter if content.startswith("---"): try: parts = content.split("---", 2) if len(parts) >= 3: fm = yaml.safe_load(parts[1]) parameters = fm.get("config", {}).get("parameters", parameters) except: pass # Regex Parsers if not parameters or not parameters.get("properties"): mig_match = SKILL_CONFIG_JSON.search(content) if mig_match: try: parameters = json.loads(mig_match.group(1).strip()) except: pass if not parameters or not parameters.get("properties"): desc_match = SKILL_DESC_OVERRIDE.search(content) if desc_match: description = f"{desc_match.group(1).strip()}\n\n[Native VFS Skill - Execute via: `{exec_cmd}`]\n{content}" table_match = SKILL_PARAM_TABLE.search(content) if table_match: parameters = {"type": "object", "properties": {}, "required": []} for row in table_match.group(1).strip().split('\n'): cols = [c.strip() for c in row.split('|')][1:-1] if len(cols) >= 4: p_n = cols[0].replace('`', '').strip() parameters["properties"][p_n] = {"type": cols[1].strip(), "description": cols[2].strip()} if cols[3].strip().lower() in ['yes', 'true', '1', 'y']: parameters["required"].append(p_n) # Inject Node Selector if "node_id" not in parameters.get("properties", {}): if "properties" not in parameters: parameters["properties"] = {} parameters["properties"]["node_id"] = { "type": "string", "description": "Mesh node ID for execution. Leave empty to use session default." } return {"type": "function", "function": {"name": name, "description": description, "parameters": parameters}} async def call_tool(self, tool_name: str, arguments: Dict[str, Any], db: Session = None, user_id: str = None, session_id: str = None, session_db_id: int = None, on_event = None, provider_name: str = None) -> Any: """ Executes a registered skill. """ # --- Node Auto-Resolution & Security Guard --- node_id = arguments.get("node_id") node_ids = arguments.get("node_ids") if db and session_db_id: session_meta = db.query(models.Session).filter(models.Session.id == session_db_id).first() if session_meta: attached = session_meta.attached_node_ids or [] if not node_id and not node_ids and attached: node_id = attached[0] arguments["node_id"] = node_id allowed_ids = attached + ["hub", "server", "local"] if node_id and node_id not in allowed_ids: # Fuzzy match fallback (e.g. 'test node1' -> 'test-node-1') fuzzy_nid = str(node_id).lower().replace(" ", "").replace("-", "").replace("_", "") for aid in allowed_ids: if aid.lower().replace("-", "").replace("_", "") == fuzzy_nid: node_id = aid arguments["node_id"] = aid break if node_id not in allowed_ids: return {"success": False, "error": f"Node '{node_id}' is NOT attached or doesn't exist. Available nodes: {attached}"} if node_ids: illegal = [nid for nid in node_ids if nid not in allowed_ids] if illegal: return {"success": False, "error": f"Nodes {illegal} are NOT attached to this session. Available nodes: {attached}"} # 1. Try local/native skill first if tool_name in self._local_skills: skill = self._local_skills[tool_name] result = await skill.execute(**arguments) return result.dict() # 2. Check the tool_registry for lightweight plugins (e.g. read_skill_artifact) # These are system agents that DO NOT need a full SubAgent loop — their prepare_task # returns a synchronous function we can call directly. plugin = tool_registry.get_plugin(tool_name) if plugin: orchestrator = getattr(self._services, "orchestrator", None) context = { "db": db, "user_id": user_id, "session_id": session_id or "__fs_explorer__", "node_id": node_id, "node_ids": node_ids, "services": self._services, "orchestrator": orchestrator, "assistant": orchestrator.assistant if orchestrator else None, "on_event": on_event } task_fn, task_args = plugin.prepare_task(arguments, context) if not task_fn: return task_args # error dict try: import asyncio if asyncio.iscoroutinefunction(task_fn): res = await task_fn(**task_args) else: # BLOCKING CALL: Move to thread to avoid freezing the event loop res = await asyncio.to_thread(task_fn, **task_args) # M6: Post-processing for Binary Artifacts (Screenshots, etc.) if isinstance(res, dict): res = self._process_binary_artifacts(tool_name, res, session_id, arguments, orchestrator.assistant if orchestrator else None) return res except Exception as e: logger.exception(f"Plugin '{tool_name}' execution failed: {e}") return {"success": False, "error": str(e)} # 3. Handle System / FS Skills (full SubAgent or Bash) from app.core.skills.fs_loader import fs_loader all_fs_skills = fs_loader.get_all_skills() for fs_skill in all_fs_skills: if fs_skill.get("name") == tool_name: class _DictObj: def __init__(self, d): for k, v in d.items(): setattr(self, k, v) fs_skill["files"] = [_DictObj(f) for f in fs_skill.get("files", [])] db_skill_mock = _DictObj(fs_skill) return await self._execute_system_skill(db_skill_mock, arguments, user_id=user_id, db=db, session_id=session_id, session_db_id=session_db_id, on_event=on_event, provider_name=provider_name) logger.error(f"Tool '{tool_name}' not found or handled yet.") return {"success": False, "error": "Tool not found"} async def _execute_system_skill(self, skill: Any, args: Dict[str, Any], user_id: str = None, db: Session = None, session_id: str = None, session_db_id: int = None, on_event = None, provider_name: str = None) -> Any: """Routes FS skill execution to a stateful SubAgent or Dynamic Plugin.""" from app.core.services.sub_agent import SubAgent from app.core.providers.factory import get_llm_provider # Node IDs are already validated and resolved from step 0 in call_tool. node_id = args.get("node_id") node_ids = args.get("node_ids") # --- Standard Preparation --- llm_provider = None orchestrator = getattr(self._services, "orchestrator", None) if not orchestrator: return {"success": False, "error": "Orchestrator not available"} assistant = orchestrator.assistant # M3: Resolve session_id from either arguments OR the passed session_id context # (AI might use placeholders like 'current' which we resolve here) session_id_arg = args.get("session_id") if not session_id_arg or session_id_arg == "current": resolved_sid = session_id or "__fs_explorer__" else: resolved_sid = session_id_arg llm_provider = self._resolve_llm_for_sub_agent(db, user_id, provider_name) if llm_provider: logger.info(f"[ToolService] AI Sub-Agent enabled for {skill.name}") # Define the task function and arguments for the SubAgent task_fn = None task_args = {} plugin = tool_registry.get_plugin(skill.name) if not plugin: # Check if this is a Dynamic Bash Skill bash_logic = None skill_md_file = next((f for f in skill.files if f.file_path == "SKILL.md"), None) if getattr(skill, "files", None) else None if skill_md_file and skill_md_file.content: import re # Broadened regex: Allows 3-6 hashes, any title containing "Execution Logic", and handles files ending without newlines. bash_match = re.search(r"#{3,6}\s*Execution Logic.*?```bash\s*\n(.*?)(?:```|\Z)", skill_md_file.content, re.DOTALL | re.IGNORECASE) if bash_match: bash_logic = bash_match.group(1).strip() if not bash_logic: exec_file = "" for f in getattr(skill, "files", []): if f.file_path.endswith(".sh") or f.file_path.endswith(".py") or "run." in f.file_path: exec_file = f.file_path break exec_cmd = f"bash .skills/{skill.name}/{exec_file}" if exec_file.endswith(".sh") else f"python3 .skills/{skill.name}/{exec_file}" if exec_file.endswith(".py") else f".skills/{skill.name}/{exec_file}" class DynamicBashPlugin: name = skill.name retries = 0 def prepare_task(self, invoke_args, invoke_context): import shlex if bash_logic: cmd = bash_logic for k, v in invoke_args.items(): cmd = cmd.replace(f"${{{k}}}", shlex.quote(str(v))) else: # Auto-bridging fallback: construct command with env vars and positional args safe_args = {k: v for k, v in invoke_args.items() if k != "timeout" and k != "node_id" and k != "node_ids"} bash_env = " ".join([f'{k}={shlex.quote(str(v))}' for k, v in safe_args.items()]) bash_args_str = " ".join([shlex.quote(str(v)) for v in safe_args.values()]) cmd = f"{bash_env} {exec_cmd} {bash_args_str}".strip() timeout = int(invoke_args.get("timeout", 60)) node_id = invoke_context.get("node_id", invoke_args.get("node_id")) node_ids = invoke_context.get("node_ids", invoke_args.get("node_ids")) resolved_sid = invoke_context.get("session_id") assistant = invoke_context.get("assistant") services = invoke_context.get("services") if node_id in ["hub", "server", "local"] or (node_ids and any(nid in ["hub", "server", "local"] for nid in node_ids)): def _hub_command(**kwargs): import subprocess, os cwd = os.getcwd() if kwargs.get("resolved_sid") and getattr(services, "orchestrator", None): try: cwd = services.orchestrator.mirror.get_workspace_path(kwargs.get("resolved_sid")) except: pass try: proc = subprocess.run(kwargs.get("cmd"), shell=True, capture_output=True, text=True, timeout=kwargs.get("timeout"), cwd=cwd) return {"status": "SUCCESS" if proc.returncode == 0 else "FAILED", "stdout": proc.stdout, "stderr": proc.stderr, "exit_code": proc.returncode, "node_id": "hub"} except subprocess.TimeoutExpired as e: return {"status": "TIMEOUT", "stdout": e.stdout or "", "stderr": e.stderr or "", "error": "Command timed out"} except Exception as e: return {"status": "ERROR", "error": str(e)} return _hub_command, {"cmd": cmd, "timeout": timeout, "resolved_sid": resolved_sid} elif node_ids and isinstance(node_ids, list): return assistant.dispatch_swarm, {"node_ids": node_ids, "cmd": cmd, "timeout": timeout, "session_id": resolved_sid, "no_abort": False} elif node_id: return assistant.dispatch_single, {"node_id": node_id, "cmd": cmd, "timeout": timeout, "session_id": resolved_sid, "no_abort": False} return None, {"success": False, "error": f"Please map a target node_id to execute this tool natively."} plugin = DynamicBashPlugin() context = { "db": db, "user_id": user_id, "session_id": resolved_sid, "node_id": node_id, "node_ids": node_ids, "assistant": assistant, "orchestrator": orchestrator, "services": self._services, "on_event": on_event } task_fn, task_args = plugin.prepare_task(args, context) if not task_fn: return task_args # error dict returned by prepare_task try: if task_fn: # Create and run the SubAgent (potentially AI-powered) sub_agent = SubAgent( name=f"{skill.name}_{node_id or 'swarm'}", task_fn=task_fn, args=task_args, retries=plugin.retries, llm_provider=llm_provider, assistant=assistant, on_event=on_event ) res = await sub_agent.run() # Standardize output for AI: # If it's a string (our new Intelligence Report), pass it through directly. # If it's a dict, only wrap as failure if a non-None error exists. if isinstance(res, dict) and res.get("error"): return {"success": False, "error": res["error"], "sub_agent_status": sub_agent.status} # M6: Post-processing for Binary Artifacts (Screenshots, etc.) if isinstance(res, dict): res = self._process_binary_artifacts(skill.name, res, resolved_sid, args, assistant) logger.info(f"[ToolService] System skill '{skill.name}' completed (Status: {sub_agent.status}).") return res except Exception as e: logger.exception(f"System skill execution failed: {e}") return {"success": False, "error": str(e)} return {"success": False, "error": "Skill execution logic not found"} def _process_binary_artifacts(self, skill_name: str, res: Dict[str, Any], resolved_sid: str, args: Dict[str, Any], assistant: Any) -> Dict[str, Any]: """Post-processing for Binary Artifacts (Screenshots, etc.)""" # Unconditionally prevent binary data from leaking into the JSON serializer screenshot_bits = res.pop("_screenshot_bytes", None) if skill_name == "browser_automation_agent" and assistant: # Organise browser data by session for better UX if resolved_sid and resolved_sid != "__fs_explorer__": try: abs_workspace = assistant.mirror.get_workspace_path(resolved_sid) # M6: Use .browser_data (ignored from node sync) base_dir = os.path.join(abs_workspace, ".browser_data") os.makedirs(base_dir, exist_ok=True) timestamp = int(time.time()) action = args.get("action", "unknown").lower() # Clean filename for the image: {timestamp}_{action}.png ss_filename = f"{timestamp}_{action}.png" # Save Screenshot if available if screenshot_bits: ss_path = os.path.join(base_dir, ss_filename) with open(ss_path, "wb") as f: f.write(screenshot_bits) res["screenshot_url"] = f"/.browser_data/{resolved_sid}/{ss_filename}" res["_visual_feedback"] = f"Action screenshot captured: {res['screenshot_url']}" # Save Metadata/A11y into a hidden or specific sub-folder if needed, # but keep images in the root of the session for quick gallery view. action_dir = os.path.join(base_dir, ".metadata", f"{timestamp}_{action}") os.makedirs(action_dir, exist_ok=True) # Save Metadata/Result for easy debugging in file explorer import json meta = { "timestamp": timestamp, "action": action, "url": res.get("url"), "title": res.get("title"), "success": res.get("success"), "error": res.get("error"), "eval_result": res.get("eval_result") } with open(os.path.join(action_dir, "metadata.json"), "w") as f: json.dump(meta, f, indent=2) # Optional: Save A11y summary for quick viewing if "a11y_summary" in res: with open(os.path.join(action_dir, "a11y_summary.txt"), "w") as f: f.write(res["a11y_summary"]) logger.info(f"[ToolService] Browser artifacts saved to: {action_dir}") except Exception as sse: logger.warning(f"Failed to persist browser data to workspace: {sse}") return res def _resolve_llm_for_sub_agent(self, db: Session, user_id: str, provider_name: str) -> Optional[Any]: if not db or not user_id or not self._services: return None user = db.query(models.User).filter(models.User.id == user_id).first() if not user: return None provider, _ = self._services.preference_service.resolve_llm_provider(db, user, provider_name) return provider