The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill implements a verification lane that processes 'pushed edits' generated by a separate 'executor' agent, creating a multi-step chain vulnerability surface.
Ingestion points: The skill ingests external code edits and acceptance clauses as inputs for validation.
Boundary markers: There are no specified delimiters or instructions to treat the 'pushed edit' as data rather than instructions, which could lead the agent to follow malicious commands embedded in the code being validated.
Capability inventory: The skill is instructed to 'Run each check live' (command execution) and write verification receipts to the filesystem (.gm/receipts/<row>/05-validate.md).
Sanitization: The instructions do not define any sanitization, filtering, or validation of the code edits before they are executed in the JIT environment.
[COMMAND_EXECUTION]: The core logic of the skill involves executing code at runtime ('live JIT runs'). While this is the primary purpose of a validation skill, performing execution on data provided by another agent (the 'executor') without isolation or verification of the source increases the risk of executing unintended commands.

gm-validate