The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill facilitates the creation of plugins that process external data and execute shell commands. This creates a surface where untrusted data could influence agent actions if not properly sanitized.
Ingestion points: Plugin hooks and custom tool arguments defined in references/hooks.md and references/tool-helper.md.
Boundary markers: While Zod schemas are recommended for validation, the skill does not enforce explicit boundary markers for free-text inputs.
Capability inventory: The SDK provides access to the BunShell utility for command execution and the full OpenCode client API across all plugin scripts.
Sanitization: The documentation explicitly requires Zod schema validation for all external data and API boundaries.
[COMMAND_EXECUTION]: Documentation and code examples in references/examples.md and references/tool-helper.md demonstrate the use of the $ (Bun shell) utility to execute local commands such as osascript, wc, and eslint for development tasks.
[EXTERNAL_DOWNLOADS]: The references/update-notifications.md file contains logic to fetch package metadata from the official npm registry (registry.npmjs.org) to check for available updates, which is a standard check for pinned dependencies.

create-opencode-plugin