The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes the opencode and claude CLI tools via node:child_process. These calls are used to spawn background workers, check service status, and perform benchmarks. Arguments are passed as arrays, which avoids shell-based command injection vulnerabilities.
[INDIRECT_PROMPT_INJECTION]: The skill orchestrates external AI workers and brings their output back into the primary agent's context. This constitutes an indirect prompt injection surface where a sub-agent's output could potentially influence the primary agent's instructions.
Ingestion points: Worker results are parsed from NDJSON streams in scripts/oc-worker.mjs and returned as a single JSON line to the primary agent.
Boundary markers: The skill enforces a 'Golden Rule' where worker output only enters the context via the script's structured JSON response, providing a clear boundary and avoiding context flooding.
Capability inventory: The orchestration commands have access to file system tools (Read, Edit, Grep, Glob) and execution tools (Bash).
Sanitization: The oc-worker.mjs script truncates worker output to a maximum of 20,000 characters.
[EXTERNAL_DOWNLOADS]: The skill requires the opencode-ai global NPM package to be installed by the user. At runtime, the CLI communicates with the opencode.ai service to facilitate task delegation to remote models.
[DATA_EXFILTRATION]: As an orchestration layer for an external service, the skill sends task-related data to opencode.ai endpoints via the CLI. The SKILL.md documentation includes a clear warning to users not to send secrets or private code to the workers.

hivemind