The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses a custom helper script scripts/gate.py to execute shell commands for testing, environment setup, and running development servers. These commands are defined in local board and task files and are executed using subprocess.Popen with shell=True. To mitigate risks, the script implements a cryptographic approval system that requires a human to explicitly verify and approve commands before they can be executed by the agent.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process data from potentially untrusted sources, including task descriptions, project board files, and GitHub PR metadata.
Ingestion points: Data is read from jt-queue/board.md, jt-queue/tasks/<id>/task.md, and the output of the gh (GitHub CLI) tool.
Boundary markers: SKILL.md contains a dedicated 'Untrusted input' section that explicitly instructs the agent to treat all data from these sources as non-instructional and to report any attempts to override its behavior.
Capability inventory: The skill possesses the ability to execute arbitrary shell commands via gate.py, perform file system operations, and interact with GitHub via the gh CLI.
Sanitization: Security is enforced through a human-in-the-loop approval mechanism (scripts/gate.py approve-commands) which stores hashes of verified commands in the local .git/ directory, ensuring that modified or unapproved commands are rejected at runtime.

jt-work-queue