The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill facilitates an orchestration loop where the orchestrator agent ingests and acts on data produced by an external implementer (OpenCode CLI), creating a potential for indirect prompt injection attacks.
Ingestion points: The orchestrator reads implementation results from result.json (specifically the finalMessage and touchedFiles fields) and file diffs in the working tree.
Boundary markers: The instructions recommend structured briefs using XML tags (e.g., <task>, <verification_loop>) to provide task boundaries.
Capability inventory: The orchestrator can execute shell commands via a relay script, read/write repository files, and commit changes using git.
Sanitization: The workflow includes mandatory human review and the re-execution of project gates (tests, linting, builds) to verify implementer output before committing changes.
[EXTERNAL_DOWNLOADS]: The skill documentation guides the user to install the opencode-ai package via NPM and references the amElnagdy/delegate-skills repository. These community resources are outside the trusted organization scope.
[COMMAND_EXECUTION]: The skill's operation is based on the execution of local shell commands and a Node.js relay script that orchestrates the OpenCode CLI implementer.

opencode-delegate