relay-fleet

Pass

Audited by Gen Agent Trust Hub on Oct 9, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill ingests external task definitions from a JSON file (specified via --leaves-file) and from child run records (run.json) in ~/.relay/runs/. This content is used to drive the orchestration loop and is passed as arguments to subprocesses.
  • Ingestion points: The --leaves-file JSON input in relay-fleet.js and run.json files in the runs directory serve as ingestion points for data that influences agent behavior.
  • Boundary markers: The skill does not implement explicit boundary markers or instructions to ignore embedded content when passing data to sub-scripts.
  • Capability inventory: The skill spawns multiple Node.js subprocesses (dispatch.js, review-runner.js, finalize-run.js), performs file system operations (read/write in ~/.relay and the repository), and executes git commands.
  • Sanitization: The skill performs validation for slugs, single-line strings, and integers. It also employs a trustedDirectory check and O_NOFOLLOW flag to prevent symlink-based path traversal and hijacking.
  • [COMMAND_EXECUTION]: The script relay-fleet.js dynamically constructs and executes command lines for sibling scripts using child_process.spawn. Although it uses an argument array to mitigate shell injection, the parameters are derived from the input JSON, which could potentially lead to argument injection if the receiving scripts do not perform strict validation of their CLI flags.
Audit Metadata
Risk Level
SAFE
Analyzed
Oct 9, 2026, 12:31 AM