#!/usr/bin/env python3 """ Skills Hub — hub state management and the public facade for the source adapters. Library module (not an agent tool). Owns the hub paths, guarded HTTP, index cache, lock file, taps and audit log. Install/uninstall/update live in ``skills_hub_install``, the index fetch/source router/search in ``skills_hub_search``, and the adapters in the other ``tools.skills_hub_*`` siblings; import each name from its defining module. Used by hermes_cli/skills_hub.py for CLI commands and the /skills slash command. """ import json import logging import threading import time from contextvars import ContextVar from contextlib import ExitStack, contextmanager from datetime import datetime, timezone, UTC from pathlib import Path from typing import Any, Callable, Dict, Iterator, List, Optional from urllib.parse import urljoin import httpx from hermes_constants import get_hermes_home from tools.url_safety import is_safe_url from tools.url_safety import create_ssrf_safe_client from tools.website_policy import check_website_access from tools.skills_hub_models import _normalize_lock_install_path, _validate_skill_name from utils import atomic_write_text logger = logging.getLogger(__name__) # --------------------------------------------------------------------------- # Paths # --------------------------------------------------------------------------- # Resolved per-call (not frozen at import) so the profile override is honored; # import-time constants leaked across profiles in single-process multi-profile # runtimes. The path names (SKILLS_DIR, ...) resolve live via __getattr__ below. INDEX_CACHE_TTL = 3600 # 1 hour def _path_resolver(name: str, parent: str, leaf: str): """Resolver for hub path ``/``: a test-injected real module attribute (patch.object/monkeypatch on SKILLS_DIR etc.) wins over live resolution.""" def resolve() -> Path: forced = globals().get(name) return Path(forced) if forced is not None else _DYNAMIC_PATH_RESOLVERS[parent]() / leaf resolve.__name__ = f"_{name.lower()}" return resolve def _hermes_home() -> Path: return get_hermes_home() _skills_dir = _path_resolver("SKILLS_DIR", "HERMES_HOME", "skills") _hub_dir = _path_resolver("HUB_DIR", "SKILLS_DIR", ".hub") _lock_file = _path_resolver("LOCK_FILE", "HUB_DIR", "lock.json") _quarantine_dir = _path_resolver("QUARANTINE_DIR", "HUB_DIR", "quarantine") _audit_log = _path_resolver("AUDIT_LOG", "HUB_DIR", "audit.log") _taps_file = _path_resolver("TAPS_FILE", "HUB_DIR", "taps.json") _index_cache_dir = _path_resolver("INDEX_CACHE_DIR", "HUB_DIR", "index-cache") _DYNAMIC_PATH_RESOLVERS = {"HERMES_HOME": _hermes_home, **{ r.__name__[1:].upper(): r for r in (_skills_dir, _hub_dir, _lock_file, _quarantine_dir, _audit_log, _taps_file, _index_cache_dir) }} def __getattr__(name: str): """Resolve legacy path constants dynamically (PEP 562) so they reflect the active profile override; a test's patch.object-set real attribute shadows it.""" resolver = _DYNAMIC_PATH_RESOLVERS.get(name) if resolver is not None: return resolver() raise AttributeError(f"module {__name__!r} has no attribute {name!r}") # --------------------------------------------------------------------------- # Install-path safety + guarded HTTP # --------------------------------------------------------------------------- _REDIRECT_STATUS_CODES = {301, 302, 303, 307, 308} _MAX_SKILL_FETCH_REDIRECTS = 5 # Default per-request timeout every one-shot hub GET used before pooling existed. _DEFAULT_HTTP_TIMEOUT = 20 # An inspect resolves metadata and then the preview bundle through the same # source adapter. Keeping this context local to that operation lets httpx reuse # its verified connection without extending a client beyond the CLI request. _skills_hub_http_client: ContextVar[Optional[Any]] = ContextVar( "skills_hub_http_client", default=None ) @contextmanager def skills_hub_http_session() -> Iterator[None]: """Reuse one SSRF-safe HTTP client for a single skills-hub resolution.""" if _skills_hub_http_client.get() is not None: yield return with create_ssrf_safe_client(timeout=_DEFAULT_HTTP_TIMEOUT, follow_redirects=False) as client: token = _skills_hub_http_client.set(client) try: yield finally: _skills_hub_http_client.reset(token) def _skills_hub_http_get(url: str, **kwargs: Any) -> httpx.Response: """GET through the active resolution pool, or preserve one-shot behavior.""" client = _skills_hub_http_client.get() if client is not None: return client.get(url, **kwargs) return httpx.get(url, **kwargs) def _ssrf_safe_http_get(url: str, *, timeout: int = _DEFAULT_HTTP_TIMEOUT, headers: Optional[dict[str, str]] = None) -> httpx.Response: """Fetch one URL with connect-time SSRF validation and no automatic redirects.""" with skills_hub_http_session(): return _skills_hub_http_client.get().get(url, timeout=timeout, headers=headers) def _guarded_http_get(url: str, *, timeout: int = _DEFAULT_HTTP_TIMEOUT, headers: Optional[dict[str, str]] = None) -> Optional[httpx.Response]: """Fetch a URL with SSRF and redirect-target validation (each hop re-checked). *headers* are plain request headers (no credentials) and are sent on every hop.""" from tools.url_safety import SSRFConnectionBlocked current_url = url for _ in range(_MAX_SKILL_FETCH_REDIRECTS + 1): if not is_safe_url(current_url): logger.warning("Blocked unsafe Skills Hub URL: %s", current_url) return None blocked = check_website_access(current_url) if blocked: logger.info( "Blocked Skills Hub fetch for %s by rule %s", blocked["host"], blocked["rule"], ) return None try: resp = _ssrf_safe_http_get(current_url, timeout=timeout, headers=headers) except (SSRFConnectionBlocked, httpx.HTTPError) as exc: logger.debug("Skills Hub fetch failed for %s: %s", current_url, exc) return None if resp.status_code in _REDIRECT_STATUS_CODES: location = getattr(resp, "headers", {}).get("location") if not location: return None current_url = urljoin(current_url, location) continue return resp logger.warning("Skills Hub fetch exceeded redirect limit for %s", url) return None @contextmanager def _guarded_http_stream( url: str, *, params: Optional[dict[str, str]] = None, timeout: int = _DEFAULT_HTTP_TIMEOUT, ) -> Iterator[Optional[httpx.Response]]: """Stream one response with bounded, policy-checked redirects.""" from tools.url_safety import SSRFConnectionBlocked, create_ssrf_safe_client current_url = url current_params = params response: Optional[httpx.Response] = None stack = ExitStack() try: for _ in range(_MAX_SKILL_FETCH_REDIRECTS + 1): if not is_safe_url(current_url): logger.warning("Blocked unsafe Skills Hub URL: %s", current_url) response = None break blocked = check_website_access(current_url) if blocked: logger.info( "Blocked Skills Hub fetch for %s by rule %s", blocked["host"], blocked["rule"], ) response = None break stack.close() stack = ExitStack() try: client = stack.enter_context( create_ssrf_safe_client(timeout=timeout, follow_redirects=False) ) response = stack.enter_context( client.stream("GET", current_url, params=current_params) ) except (SSRFConnectionBlocked, httpx.HTTPError) as exc: logger.debug("Skills Hub stream failed for %s: %s", current_url, exc) response = None break if response.status_code not in _REDIRECT_STATUS_CODES: break location = response.headers.get("location") if not location: response = None break current_url = urljoin(current_url, location) current_params = None else: logger.warning("Skills Hub fetch exceeded redirect limit for %s", url) response = None yield response finally: stack.close() # --------------------------------------------------------------------------- # Shared index cache (used by every adapter) # --------------------------------------------------------------------------- def _read_json_if_fresh(path: Path, ttl: float) -> Optional[Any]: """Parsed JSON from ``path`` when it exists and is younger than ``ttl`` seconds.""" try: if time.time() - path.stat().st_mtime > ttl: return None return json.loads(path.read_text(encoding="utf-8-sig")) except (OSError, json.JSONDecodeError): return None def _read_index_cache(key: str) -> Optional[Any]: return _read_json_if_fresh(_index_cache_dir() / f"{key}.json", INDEX_CACHE_TTL) def _write_index_cache(key: str, data: Any) -> None: index_cache_dir = _index_cache_dir() index_cache_dir.mkdir(parents=True, exist_ok=True) # Cache files hold unvetted community text (possible prompt injection); # a .ignore keeps ripgrep and .ignore-aware tools out of the hub dir. ignore_file = _hub_dir() / ".ignore" if not ignore_file.exists(): try: ignore_file.write_text("# Exclude hub internals from search tools\n*\n", encoding="utf-8") except OSError: pass try: (index_cache_dir / f"{key}.json").write_text( json.dumps(data, ensure_ascii=False, default=str), encoding="utf-8" ) except OSError as e: logger.debug("Could not write cache: %s", e) # --------------------------------------------------------------------------- # Hub state files: lock.json, taps.json, audit.log # --------------------------------------------------------------------------- class _JsonStateFile: """A JSON file under the hub dir with a fixed empty shape (``EMPTY``, deep-copied on every miss/corrupt read); ``DEFAULT_PATH`` is the hub path resolver.""" EMPTY: dict = {} DEFAULT_PATH: Any = None _LOCK_HOLDER: threading.local def __init__(self, path: Optional[Path] = None): self.path = path if path is not None else type(self).DEFAULT_PATH() def _read(self) -> dict: try: return json.loads(self.path.read_text(encoding="utf-8-sig")) except (json.JSONDecodeError, OSError): return json.loads(json.dumps(self.EMPTY)) def _write(self, data: dict, **dumps_kwargs) -> None: self.path.parent.mkdir(parents=True, exist_ok=True) # Temp + rename: another process (CLI vs desktop install) reading mid-write would see an # empty file, load EMPTY and save a lock that drops every other installed skill. atomic_write_text(self.path, json.dumps(data, indent=2, **dumps_kwargs) + "\n", preserve_mode=True) def _update(self, mutate: Callable[[dict], bool], **dumps_kwargs) -> bool: """Re-read, ``mutate`` and save under a cross-process lock; the file is rewritten only when ``mutate`` returns True. The CLI and desktop rewrite the same file, and a save from a snapshot loaded outside the lock drops whatever the other process recorded meanwhile. An unusable lock (no flock on the filesystem, a wedged holder) degrades to an unlocked update rather than failing the install that triggered it.""" from hermes_cli.auth import _file_lock with ExitStack() as stack: try: stack.enter_context(_file_lock( self.path.with_name(self.path.name + ".lock"), type(self)._LOCK_HOLDER, 10.0, f"Timed out waiting for the skills hub state lock ({self.path})", )) except (OSError, TimeoutError): logger.warning("Updating %s without its lock", self.path, exc_info=True) data = self._read() changed = mutate(data) if changed: self._write(data, **dumps_kwargs) return changed class HubLockFile(_JsonStateFile): """skills/.hub/lock.json — provenance of installed hub skills.""" EMPTY = {"version": 1, "installed": {}} DEFAULT_PATH = staticmethod(_lock_file) _LOCK_HOLDER = threading.local() def load(self) -> dict: return self._read() def save(self, data: dict) -> None: self._write(data, ensure_ascii=False) def record_install( self, name: str, source: str, identifier: str, trust_level: str, scan_verdict: str, skill_hash: str, install_path: str, files: list[str], metadata: Optional[dict[str, Any]] = None, scan_provenance: Optional[dict[str, Any]] = None, ) -> None: # Validate name and install-path SHAPE at write time: a poisoned lock # entry is the precondition for the uninstall_skill rmtree-escape. safe_name = _validate_skill_name(name) safe_install_path = _normalize_lock_install_path(install_path, safe_name) def record(data: dict) -> bool: now = datetime.now(UTC).isoformat() data.setdefault("installed", {})[safe_name] = { "source": source, "identifier": identifier, "trust_level": trust_level, "scan_verdict": scan_verdict, "content_hash": skill_hash, "install_path": safe_install_path, "files": files, "metadata": metadata or {}, "scan_provenance": scan_provenance or {}, "installed_at": now, "updated_at": now, } return True self._update(record, ensure_ascii=False) def record_uninstall(self, name: str) -> None: def forget(data: dict) -> bool: data.setdefault("installed", {}).pop(name, None) return True self._update(forget, ensure_ascii=False) def get_installed(self, name: str) -> Optional[dict]: return self.load()["installed"].get(name) def list_installed(self) -> list[dict]: return [{"name": name, **entry} for name, entry in self.load()["installed"].items()] def by_identifier(self) -> dict[str, dict]: """Installed entries keyed by identifier. A ClawHub row is also keyed by the ``clawhub/`` id a catalog card installs, since the lock records ``@owner/slug``. That alias names its source, so a ClawHub install never answers for another source's id. Exact identifiers win.""" from tools.skills_hub_clawhub import ClawHubSource rows = [entry for entry in self.list_installed() if entry.get("identifier")] installed = {entry["identifier"]: entry for entry in rows} for entry in rows: if entry.get("source") == "clawhub": if alias := ClawHubSource.catalog_identifier(entry["identifier"]): installed.setdefault(alias, entry) return installed def installed_entry(installed: dict[str, dict], identifier: str, source: Optional[str] = None) -> Optional[dict]: """The ``HubLockFile.by_identifier`` entry a hub result names. Search and the index list a ClawHub skill by its bare slug, so a ``source="clawhub"`` result also matches through its ``clawhub/`` card id. Any other source matches its exact id only.""" if identifier in installed: return installed[identifier] if source != "clawhub": return None from tools.skills_hub_clawhub import ClawHubSource alias = ClawHubSource.catalog_identifier(identifier) return installed.get(alias) if alias else None class TapsManager(_JsonStateFile): """skills/.hub/taps.json — custom GitHub repo sources.""" EMPTY = {"taps": []} DEFAULT_PATH = staticmethod(_taps_file) _LOCK_HOLDER = threading.local() def load(self) -> list[dict]: return self._read().get("taps", []) def save(self, taps: list[dict]) -> None: self._write({"taps": taps}) def add(self, repo: str, path: str = "skills/") -> bool: """Add a tap. Returns False if already exists.""" def append(data: dict) -> bool: taps = data.setdefault("taps", []) if any(t["repo"] == repo for t in taps): return False taps.append({"repo": repo, "path": path}) return True return self._update(append) def remove(self, repo: str) -> bool: """Remove a tap by repo name. Returns False if not found.""" def drop(data: dict) -> bool: taps = data.get("taps", []) kept = [t for t in taps if t["repo"] != repo] data["taps"] = kept return len(kept) != len(taps) return self._update(drop) list_taps = load def append_audit_log(action: str, skill_name: str, source: str, trust_level: str, verdict: str, extra: str = "") -> None: """Append one space-separated line to the audit log (best-effort).""" audit_log = _audit_log() audit_log.parent.mkdir(parents=True, exist_ok=True) timestamp = datetime.now(UTC).strftime("%Y-%m-%dT%H:%M:%SZ") parts = [timestamp, action, skill_name, f"{source}:{trust_level}", verdict] if extra: parts.append(extra) try: with open(audit_log, "a", encoding="utf-8") as f: f.write(" ".join(parts) + "\n") except OSError as e: logger.debug("Could not write audit log: %s", e) # --------------------------------------------------------------------------- # Hub operations (high-level) # --------------------------------------------------------------------------- def ensure_hub_dirs() -> None: """Create the .hub directory structure if it doesn't exist.""" _hub_dir().mkdir(parents=True, exist_ok=True) _quarantine_dir().mkdir(exist_ok=True) _index_cache_dir().mkdir(exist_ok=True) for path, initial in ( (_lock_file(), '{"version": 1, "installed": {}}\n'), (_audit_log(), ""), (_taps_file(), '{"taps": []}\n'), ): if not path.exists(): path.write_text(initial, encoding="utf-8")