The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes complex command strings on remote hosts via ssh and docker exec to interact with the agent-vault service. It mitigates command injection risks by applying strict regular expression validation to all user-controllable inputs such as hostnames, usernames, container names, and vault names (e.g., requireSafeToken in agent-vault-add-secret-core.mjs).
[EXTERNAL_DOWNLOADS]: The skill requires several external dependencies for its operation, including glimpseui (referenced in MAINTENANCE.md), wl-paste for clipboard access, hyprctl for window management in Hyprland environments, and the agent-vault CLI on the remote host.
[DYNAMIC_EXECUTION]: The JavaScript intake script dynamically loads the glimpseui module from a list of candidate file system paths at runtime (resolveGlimpse in agent-vault-add-secret.mjs). It also generates and injects HTML into the UI window to facilitate secret entry, utilizing an escapeHtml helper to prevent cross-site scripting (XSS) or HTML injection.
[INDIRECT_PROMPT_INJECTION]: The skill has a potential attack surface for indirect injection via configuration files and environment variables, handled as follows:
Ingestion points: CLI flags, environment variables, and the ~/.config/agent-vault-secret/config file (processed by resolveSettings in agent-vault-add-secret-core.mjs and resolve_setting in agent-vault-set-secret.sh).
Boundary markers: The skill does not use LLM-specific boundary markers, but treats inputs as literal tokens for shell execution.
Capability inventory: Remote SSH execution, Docker container interaction, and local file access (reading configuration and writing temporary CA certificates).
Sanitization: Robust sanitization is implemented through regex-based validation (e.g., requireSafeToken and requireCredentialName) which ensures that ingested data conforms to safe identifier patterns before being interpolated into shell commands.

agent-vault-secret