The Agent Skills Directory

[COMMAND_EXECUTION]: The skill's functionality is built around orchestrating complex shell command sequences to invoke the opencode CLI for code generation, review, and research tasks.
[REMOTE_CODE_EXECUTION]: The skill instructs the agent to use high-privilege configurations, specifically the --agent build mode and the --dangerously-skip-permissions flag. This configuration grants the sub-agent autonomous authority to modify the local filesystem and execute bash commands without user approval, effectively bypassing the primary agent's security boundaries for all actions taken by the sub-agent.
[PROMPT_INJECTION]: The architecture is highly susceptible to indirect prompt injection. Since the sub-agent is granted auto-approval for its tool calls and is explicitly instructed to read and process project files, malicious instructions embedded within the source code could manipulate the sub-agent into performing unauthorized actions.
[EXTERNAL_DOWNLOADS]: The skill depends on the external opencode CLI tool, which is not part of the standard environment and requires separate installation from a third-party source.
[DATA_EXFILTRATION]: To perform its tasks, the skill attaches local project files to prompts sent to external AI providers (e.g., OpenAI, GitHub Copilot). While this is the intended use case for model delegation, it involves transmitting sensitive codebase data to remote servers controlled by third parties.

opencode