agy-delegate

Pass

Audited by Gen Agent Trust Hub on Oct 27, 2026

Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill facilitates the execution of the agy CLI and git commands via the scripts/relay.mjs script and a dedicated MCP server. Evidence: The relay script uses Node.js child_process modules (spawn, execFileSync) to run agy with the --dangerously-skip-permissions flag for autonomous implementation. Context: This behavior is the core purpose of the skill and is documented as requiring the human to authorize the bypass for the trusted Google CLI tool.
  • [INDIRECT_PROMPT_INJECTION]: The skill is designed to process external inputs—specifically implementation briefs and repository files—that could harbor instructions intended to manipulate the implementer's behavior. Ingestion points: Input is accepted via the prompt parameter in the delegation tools and through files discovered in the cwd (current working directory) and dirs workspace roots. Boundary markers: The bridge uses cryptographic nonces and distinctive fence markers (--- agy output begins ---) to distinguish untrusted model output from system metadata. Capability inventory: The agy tool possesses the capability to modify the local filesystem and execute shell commands, as detailed in the relay script and technical references. Sanitization: Mitigation is provided by default through the disabling of slash_commands (to prevent workspace files from triggering additional skills) and a mandated manual review/gate-running process for the orchestrator.
  • [SAFE]: The skill incorporates several advanced security features, such as workspace fingerprinting to detect read-only violations and the use of sandbox-exec on macOS to enforce filesystem restrictions during planning phases.
Audit Metadata
Risk Level
SAFE
Analyzed
Oct 27, 2026, 04:46 PM