relay-dispatch

Warn

Audited by Gen Agent Trust Hub on Oct 9, 2026

Risk Level: MEDIUMDYNAMIC_EXECUTIONCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [DYNAMIC_EXECUTION]: The function 'externalMergeObserver' in 'scripts/recover.js' generates a Node.js script string at runtime and executes it using 'process.execPath -e'. This is used to perform isolated state checks on GitHub repositories.\n
  • Evidence: 'scripts/recover.js' lines 620-639.\n- [DYNAMIC_EXECUTION]: The adapter for the 'Pi' executor in 'scripts/adapters/pi.js' dynamically constructs paths to identify and load extensions from the user's home directory (~/.pi/agent/npm/node_modules/pi-alibaba-models).\n
  • Evidence: 'scripts/adapters/pi.js' lines 104-129.\n- [COMMAND_EXECUTION]: The skill extensively uses 'child_process' methods to execute system commands, Git forge operations, and various AI CLI tools including 'gh', 'git', 'claude', 'codex', and 'opencode'.\n
  • Evidence: Found throughout 'scripts/dispatch.js', 'scripts/host.js', and various adapter scripts in 'scripts/adapters/'.\n- [INDIRECT_PROMPT_INJECTION]: The skill possesses high capabilities and processes data from external AI executors, creating a surface for indirect prompt injection.\n
  • Ingestion points: Result artifacts from executors are read in 'scripts/dispatch.js'.\n
  • Boundary markers: Prompts in 'scripts/adapters/antigravity.js' and 'scripts/adapters/cline.js' use '[RELAY WORKTREE BOUNDARY]' to attempt to restrict model behavior.\n
  • Capability inventory: Spawns subprocesses, performs file system operations, and interacts with forge APIs across multiple scripts (e.g., 'scripts/host.js', 'scripts/forge-adapters.js').\n
  • Sanitization: Implements path validation and a custom process supervisor to mitigate risks, but core functionality depends on executing actions recommended by external models.\n- [EXTERNAL_DOWNLOADS]: The skill interacts with the GitLab REST API to observe the status of change requests.\n
  • Evidence: 'scripts/forge-adapters.js' lines 197-200.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Oct 9, 2026, 12:22 AM