The Agent Skills Directory

[COMMAND_EXECUTION]: The relay.mjs script invokes the kimi and git binaries through the node:child_process module to perform the core task delegation and state tracking. It also executes a Node.js script from a sibling directory (../../delegate-setup/scripts/lane.mjs) when resolving configuration lanes using process.execPath.
[COMMAND_EXECUTION]: The task brief is passed as a command-line argument to the kimi process using the --prompt flag. This makes the brief's content visible to other users on the same host via process monitoring tools like ps. The skill's documentation explicitly notes this risk and advises against including secrets directly in the brief text.
[INDIRECT_PROMPT_INJECTION]: The skill functions as an instruction relay for a secondary AI implementer (kimi) which operates in an autonomous "auto permission" mode. Malicious instructions provided in the task brief could lead to unintended filesystem modifications or execution of project commands.
Ingestion points: User-provided task descriptions are read from files or stdin by the relay.mjs script and passed to the implementer CLI.
Boundary markers: The documentation recommends using structural XML tags (e.g., <task>, <action_safety>, <verification_loop>) within the brief to separate instructions from data and constraints.
Capability inventory: The implementer (kimi) can modify files in the repository and execute project-specific commands (such as tests, lints, or build scripts).
Sanitization: The relay script does not perform sanitization of the brief content; it relies on the orchestrator agent to manually review filesystem diffs and verify project gates before committing the implementation work.

kimi-delegate