The Agent Skills Directory

[COMMAND_EXECUTION]: The relay.mjs script executes the opencode CLI and git to manage coding tasks and inspect the repository state. Potential command injection via flags like --model and --variant is mitigated by strict regex validation (MODEL_TOKEN and SAFE_TOKEN). The task description (brief) is passed to the subprocess via stdin, further reducing the risk of shell-based injection attacks.
[INDIRECT_PROMPT_INJECTION]: As the skill involves delegating tasks to another AI agent, it is susceptible to indirect prompt injection from the sub-agent's output. The skill addresses this by dedicating an entire section (review-and-land.md) to verifying the 'implementer's' work, instructing the orchestrator to re-run tests, read diffs for scope creep, and never trust the implementer's self-report.
[DYNAMIC_EXECUTION]: The script uses node:child_process functions (spawn, spawnSync, execSync) to interface with system tools. While it uses shell: true on Windows for opencode compatibility, the arguments are pre-validated, and the shell is not used for critical path data interpolation. It also dynamically resolves a sibling script lane.mjs from the delegate-setup skill, which is a standard pattern for multi-skill systems by the same author.

opencode-delegate