The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill encourages users to install software by piping a remote script from https://opencode.ai/install directly into the bash shell. This pattern is dangerous as it executes unverified code from an external source with the user's system privileges.
[EXTERNAL_DOWNLOADS]: The documentation directs users to install components from non-standard repositories, specifically a global NPM package (opencode-ai) and a third-party Homebrew tap (anomalyco/tap/opencode).
[COMMAND_EXECUTION]: The skill explicitly details how to use the ! prefix to execute arbitrary shell commands. This capability, when exposed to an AI agent with autonomous file-writing and network tools, significantly expands the system's attack surface.
[PROMPT_INJECTION]: The skill documents an architecture vulnerable to indirect prompt injection by processing untrusted data without sufficient safeguards.
Ingestion points: The tool ingests external data through fuzzy file matching (@filename) and user-provided variables ($ARGUMENTS).
Boundary markers: No boundary markers or delimiters are suggested to isolate untrusted data from instructions.
Capability inventory: The skill environment includes highly sensitive tools such as bash (shell execution), write/edit (filesystem modification), and webfetch (network access).
Sanitization: There is no documentation regarding the sanitization or validation of input received from external files or user arguments.

opencode