The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill requires the opencode-ai CLI to be installed via npm, which is a third-party dependency used for running worker agents.
[COMMAND_EXECUTION]: The skill executes the opencode CLI to perform tasks. It uses spawnSync and spawn with argument arrays, which prevents shell command injection by avoiding shell interpolation of user-supplied task text.
[DYNAMIC_EXECUTION]: In scripts/oc-worker.mjs, the skill dynamically locates the opencode executable. On Windows systems, it reads and parses npm .cmd shims to identify the correct underlying .exe path for execution.
[INDIRECT_PROMPT_INJECTION]: The skill processes output from external AI workers, which can be influenced by malicious data they ingest. The skill mitigates this through a sanitization layer and human-in-the-loop instructions:
Ingestion points: Worker responses are parsed in scripts/oc-worker.mjs and processed by various command templates in assets/commands/.
Boundary markers: The skill defines a 'Golden Rule' prohibiting raw output streams from entering the orchestrator's context and provides persona-based constraints for workers (e.g., read-only scouts).
Capability inventory: The orchestrator has access to Bash, Edit, and Read tools, while workers can execute shell commands within isolated directories.
Sanitization: Worker output is filtered into a structured JSON format and truncated to 20,000 characters to prevent context flooding and potential injection via raw stream metadata.

hivemind