The Agent Skills Directory

[COMMAND_EXECUTION]: The scripts/relay.mjs script executes shell commands using node:child_process to interact with the system. It spawns the oz CLI for code implementation, git for repository status monitoring, and node for fleet lane resolution. The script uses spawn without a shell and validates dynamic arguments to prevent injection.
[INDIRECT_PROMPT_INJECTION]: The skill facilitates a delegation workflow where task descriptions (briefs) are processed by an external implementer (oz) with extensive system capabilities.
Ingestion points: The brief is ingested from standard input or a local file in scripts/relay.mjs and passed as a command-line argument to the oz agent.
Boundary markers: The skill provides a structured template for briefs and emphasizes that the orchestrator must explicitly instruct the implementer not to commit. The 'Review and land' section defines the boundary where the user must verify all implementer claims.
Capability inventory: The oz agent runs without a sandbox, permission mode, or read-only mode, meaning it can read, write, and execute commands with the user's full permissions. The relay script also manages artifacts in the system's temporary directory.
Sanitization: The relay.mjs script includes validation logic to ensure that dynamic inputs like model IDs and conversation IDs match a safe character set (A-Za-z0-9._:/-) and that freer text inputs do not start with a dash, preventing CLI flag injection attacks.

warp-delegate