The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses the Windows Package Manager (winget) to install software and utilizes various opencode CLI commands to perform tasks such as running agents, managing sessions, and starting a local server.
[EXTERNAL_DOWNLOADS]: Fetches the OpenCode package from the Windows Package Manager repository using the command winget install --id SST.opencode.
[REMOTE_CODE_EXECUTION]: Provides functionality to install arbitrary NPM modules as plugins (opencode plugin <npm-module>) and add Model Context Protocol (MCP) servers from remote URLs (opencode mcp add --url), which allows for the execution of external, unverified code within the agent's environment.
[PRIVILEGE_ESCALATION]: Documents the use of the --auto flag to "Auto-approve permissions," which bypasses user confirmation prompts for potentially dangerous agent actions, effectively increasing the autonomy of the skill without oversight.
[DYNAMIC_EXECUTION]: Launches a local HTTP server (opencode serve) and dynamic coding agents (opencode run), creating a surface for runtime code execution and network interaction.
[INDIRECT_PROMPT_INJECTION]:
Ingestion points: Accepts user-provided strings as input for the opencode run command in SKILL.md.
Boundary markers: Lacks explicit delimiters or instructions to ignore embedded commands within the input strings passed to the CLI.
Capability inventory: The OpenCode CLI has extensive capabilities including file system modification, network access, and plugin execution as described across the skill instructions.
Sanitization: While the skill mentions a --sanitize flag for redacting sensitive paths during session exports, it does not provide mechanisms to sanitize the primary input strings provided to the agents.

opencode-harness