The Agent Skills Directory

[COMMAND_EXECUTION]: The skill explicitly instructs the agent to autonomously execute the command opencode run "test" using the bash tool after modifying configuration files, ensuring validation without requiring prior user approval for that specific step.
[REMOTE_CODE_EXECUTION]: The schema reference in references/config-schema.md provides templates for configuring Model Context Protocol (MCP) servers and code formatters that use npx to download and execute remote packages (e.g., npx -y @org/package).
[PRIVILEGE_ESCALATION]: The skill provides examples for modifying security permissions within opencode.json, including patterns that explicitly grant broad access to tools like bash (* : allow) or specific commands (npm run * : allow), which can be used to bypass default security restrictions.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted data from multiple project files including opencode.json, AGENTS.md, and various instruction files defined via glob patterns (e.g., .cursor/rules/*.md).
Ingestion points: opencode.json, AGENTS.md, and any file matched by the instructions array in the project configuration.
Boundary markers: No specific delimiters or "ignore embedded instructions" directives are present in the skill's processing logic.
Capability inventory: The skill utilizes file modification (edit), bash command execution, and can configure network access (webfetch) and sub-agent behavior.
Sanitization: No explicit sanitization or validation logic is provided for external content interpolated into the agent's context.
[EXTERNAL_DOWNLOADS]: The documentation includes examples of third-party plugins and packages from unknown or untrusted sources (e.g., @tarquinen/opencode-dcp, @howaboua/pickle-thinker) as potential additions to the user's environment.

opencode-config