The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides configuration examples that allow the agent to execute shell commands via the bash tool. This includes broad patterns like npm * and git *. Additionally, it documents custom formatters that can run arbitrary CLI tools using npx (e.g., ["npx", "prettier", "--write", "$FILE"]).
[CREDENTIALS_UNSAFE]: The configuration reference includes patterns for variable substitution that point to sensitive file paths, specifically providing an example for reading a secret key from {file:~/.secrets/anthropic-key}.
[EXTERNAL_DOWNLOADS]: The documentation describes setting up Model Context Protocol (MCP) servers, which can involve downloading and executing packages from external registries (e.g., npx -y @org/package) or connecting to remote URLs for server functionality.
[REMOTE_CODE_EXECUTION]: By supporting remote MCP servers (type: "remote") and user-defined executable formatters, the skill establishes a configuration path that could allow for remote code execution if configured with untrusted sources.
[INDIRECT_PROMPT_INJECTION]: The skill manages behavioral instructions in AGENTS.md and runtime settings in opencode.json, which are processed by the agent and could serve as an injection surface.
Ingestion points: opencode.json, AGENTS.md, and environment variables referenced in configuration.
Boundary markers: None explicitly defined in the provided examples to delimit instructions from potentially untrusted data.
Capability inventory: Full shell access (bash), file modification (edit, write), and network fetching (webfetch, remote MCP).
Sanitization: The skill advises using restrictive permissions (e.g., denying sudo) and environment variables for secrets, but does not provide specific sanitization logic for data processed within the configuration files.

opencode-configure