The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses a Node.js helper script to execute the opencode and git command-line tools. This allows the AI agent to initiate and monitor external processes to perform coding tasks and inspect the repository state.
Evidence: scripts/relay.mjs uses node:child_process methods such as spawn and execSync to call opencode run and git status --porcelain.
[DYNAMIC_EXECUTION]: The relay.mjs script dynamically constructs and executes shell commands with arguments derived from user input and configuration flags.
Evidence: The buildArgv function in scripts/relay.mjs assembles the command-line arguments for the implementer process, including model names, agent types, and session identifiers.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted task descriptions (briefs) that are passed directly to an LLM-backed implementer with file-writing capabilities. This creates a surface for instructions embedded in the codebase or the brief to influence the implementer's actions.
Ingestion points: scripts/relay.mjs reads the content of a task brief from a file (brief.txt) or via stdin.
Boundary markers: Absent. The script transmits the brief content directly to the implementer's standard input without additional delimiters or protective system instructions.
Capability inventory: The implementer (opencode) is typically run with the build agent, which has capabilities to modify files in the working directory and potentially execute shell commands.
Sanitization: Absent. The script performs no filtering, escaping, or sanitization of the brief content before processing.
[EXTERNAL_DOWNLOADS]: The skill instructions direct users to download and install an external CLI tool from the public npm registry as a prerequisite for operation.
Evidence: SKILL.md and references/dispatch-and-poll.md explicitly instruct the user to install the opencode-ai package using npm i -g opencode-ai.

opencode-delegate