coordinator
Pass
Audited by Gen Agent Trust Hub on Nov 21, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONPERSISTENCE
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The coordinator is designed to read and process
standing-rules.mdand reports generated by 'worker' agents, which could contain instructions to influence the coordinator's behavior. - Ingestion points: Reads
project.md,standing-rules.md, and markdown reports from thereports/directory. - Boundary markers: No specific escaping or boundary markers are used when interpolating report content into the coordinator's context.
- Capability inventory: The coordinator can execute shell commands (
t3,gh,git), delegate new tasks, and schedule timers. - Sanitization: While the
project.pyscript escapes markdown pipes for table formatting, it does not sanitize natural language instructions, potentially allowing a malicious worker to influence the coordinator's behavior through its report content. - [COMMAND_EXECUTION]: The skill uses a Python script (
scripts/project.py) to execute shell commands viasubprocess.run. These commands interact withgit,gh(GitHub CLI), andt3(a thread management tool). - Evidence: The script extracts thread links and PR references using regular expressions (
LINKandPR_REFinproject.py) to validate parameters before passing them to CLI tools. - [PERSISTENCE]: The skill includes a background 'tick' mechanism for project monitoring that persists across sessions.
- Evidence:
references/tick.mdandscripts/tick.shprovide instructions for installing OS-level services (macOS LaunchAgents or Linux systemd timers). This behavior is documented and requires manual user intervention to set up.
Audit Metadata