The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The skill includes a native helper (one_conv/native/CredentialHelper.swift and one_conv/native_auth.py) that accesses the macOS Keychain to retrieve encrypted browser cookies (Chrome, Arc, Brave, etc.). This is used to authenticate and sync web-based ChatGPT and Claude conversations without requiring user login prompts.
[EXTERNAL_DOWNLOADS]: The installer script (one_conv/control/installer.py) downloads a binary executable from boat.dev, which is an unverified source. This binary is used as part of the 'cloud spawner' feature to manage sandbox environments.
[PERSISTENCE]: The skill can install itself as a persistent background service using launchd on macOS or systemd on Linux (one_conv/live/service.py). This allows the 'bridge' component to run automatically at login and restart after crashes.
[PRIVILEGE_ESCALATION]: The cloud controller installer (one_conv/control/installer.py) uses sudo -n to perform administrative actions such as enabling lingering for user services and installing systemd units.
[COMMAND_EXECUTION]: The skill makes extensive use of subprocess and os.execvp to interact with local agent CLIs (claude, codex, agent), system tools (lsof, ps, git, tmux, codesign), and its own helper scripts.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data by reading local and cloud conversation histories. While the agent is instructed to treat this history as data and not instructions, the large ingestion surface combined with command execution capabilities creates a potential vector for indirect prompt injection if a previous conversation contains malicious instructions.
[DYNAMIC_EXECUTION]: The scripts/archive-sync.py script uses SourceFileLoader to dynamically load and execute code from one_conv_cli.py at runtime. The skill also uses ctypes.CDLL to interface with macOS system frameworks.

one-conv-cli