from pathlib import Path from pydantic import model_validator from pydantic_settings import BaseSettings, SettingsConfigDict class Settings(BaseSettings): model_config = SettingsConfigDict( env_file=".env", env_file_encoding="utf-8", extra="ignore", frozen=True ) ollama_host: str = "http://localhost:11434" ollama_api_key: str = "" ollama_default_model: str = "gemma4:31b-cloud" ollama_num_ctx: int = 65536 ollama_think: bool = True ollama_request_timeout: int = 30 # Embedding model for memory vector search (Ollama API) # When embedding_ollama_host is empty, falls back to ollama_host. embedding_ollama_host: str = "" embedding_ollama_api_key: str = "" embedding_model: str = "nomic-embed-text:latest" embedding_dimensions: int = 768 openai_api_key: str = "" openai_model: str = "gpt-4" openai_base_url: str | None = None anthropic_api_key: str = "" # Web search fallbacks (used when DuckDuckGo returns no results) # Brave Search API: free tier = 2000 req/month — https://brave.com/search/api/ brave_search_api_key: str = "" # SearXNG: self-hosted meta-search, e.g. "http://localhost:8888" searxng_url: str = "" # Filesystem tool: comma-separated allowed root paths fs_allowed_paths: str = "*" # Terminal tool: "*" = allow all commands (recommended for local use) # or comma-separated list of allowed executables, e.g. "ls,cat,git" terminal_allowed_commands: str = "*" # Terminal allowlist for non-admin users (multi-user mode). # Admin bypasses this restriction. terminal_user_allowed_commands: str = ( "ls,cat,git,grep,egrep,fgrep,head,tail,wc,find,locate," "ps,df,du,free,uptime,whoami,id,mkdir,touch,cp,mv,rm,chmod,chown," "echo,printf,sort,uniq,awk,sed,tr,cut,date,uname,hostname,which,whereis," "python,python3,node,npm,npx,pip,pip3,env,export,source," "make,cmake,gcc,g++,rustc,cargo,javac,java,go,tsc" ) # SSH tool: path to JSON file with named connections ssh_hosts_file: str = "ssh_hosts.json" # Ollama multi-backend fallback: path to JSON file with server list [{host, api_key?}, ...] # When set, overrides ollama_host / ollama_api_key and enables server+model fallback. ollama_backends_file: str = "" # Database — PostgreSQL is required (SQLite support removed). database_url: str = "" log_level: str = "INFO" # Directory for user-defined tools (auto-discovered at startup) tools_dir: str = "tools" # Directory for user-defined context providers (auto-discovered at startup) context_providers_dir: str = "context_providers" # Internal MCP UI server — lets Navi render structured components in the webclient. navi_ui_mcp_enabled: bool = True navi_ui_mcp_host: str = "127.0.0.1" navi_ui_mcp_port: int = 8001 # Session file uploads session_files_dir: str = "session_files" session_files_max_size_mb: int = 200 session_messages_window: int = 1000 # max hot messages per session; older -> archive ws_replay_buffer_size: int = 500 # max events retained for WS reconnect replay share_file_max_size_mb: int = 1024 # Public base URL used by share_file tool to build download links. # Change if the server is behind a reverse proxy or runs on a different port. public_url: str = "http://localhost:8000" # Gmail IMAP/SMTP (App Password auth) gmail_address: str = "" gmail_app_password: str = "" # gnexus-auth OAuth integration gnauth_base_url: str = "http://gnexus-auth.local" gnauth_client_id: str = "" gnauth_client_secret: str = "" gnauth_redirect_uri: str = "http://localhost:8000/auth/callback" gnauth_admin_role_slug: str = "navi_admin" gnauth_user_role_slug: str = "navi_user" gnauth_profile_path: str = "/account/profile" # appended to gnauth_base_url for profile links # Auth session cookie encryption (Fernet key, 32-byte base64) navi_auth_encryption_key: str = "" navi_auth_cookie_name: str = "navi_auth_session" navi_auth_cookie_secure: bool = False navi_auth_cookie_samesite: str = "lax" navi_auth_cookie_max_age_days: int = 30 # LLM call timeouts # complete() is non-streaming (planning, compression) — blocked until full response llm_complete_timeout: int = 120 # stream_complete(): how long to wait for the FIRST token (prefill phase) # Large contexts can take 60-90s to prefill; 90s matches user expectation llm_stream_first_chunk_timeout: int = 90 # stream_complete(): max gap between any two subsequent tokens llm_stream_chunk_timeout: int = 60 # Context compression context_compression_enabled: bool = True context_compression_threshold: float = 0.70 # trigger at 70% of ollama_num_ctx context_keep_recent: int = 8 # conversational turns to keep verbatim context_summary_temperature: float = 0.3 context_summary_max_tokens: int = 3000 # max output tokens for the summary LLM call output_reserve_tokens: int = 2048 # headroom reserved for model response in context checks # Global personality prompt prepended to every agent's system prompt. # Multi-line values don't survive .env parsing reliably, so prefer # navi_persona_file (path to a plain .txt file) over inline navi_persona. navi_persona: str = "" navi_persona_file: str = "" @model_validator(mode="before") def _load_persona_from_file(cls, values: dict) -> dict: if not values.get("navi_persona") and values.get("navi_persona_file"): try: values["navi_persona"] = Path(values["navi_persona_file"]).read_text( encoding="utf-8" ).strip() except Exception: pass return values @property def fs_allowed_paths_list(self) -> list[str]: return [p.strip() for p in self.fs_allowed_paths.split(",") if p.strip()] @property def terminal_allowed_commands_list(self) -> list[str]: return [c.strip() for c in self.terminal_allowed_commands.split(",") if c.strip()] @property def terminal_user_allowed_commands_list(self) -> list[str]: return [c.strip() for c in self.terminal_user_allowed_commands.split(",") if c.strip()] settings = Settings()