The Agent Skills Directory

[COMMAND_EXECUTION]: The skill requires the execution of shell commands to run repository gates, characterization tests, and baseline proofs. This is an intended capability used to verify that code removals do not change observable behavior.
Evidence: The skill instructions specify running focused and full repository gates under a "bounded command contract" with disabled network, scrubbed secrets, and bounded execution time (TERM-then-KILL cleanup).
[INDIRECT_PROMPT_INJECTION]: The skill processes external data in the form of audit reports and removal sets, which serves as a potential ingestion point for untrusted content.
Ingestion points: The argument-hint indicates that the skill ingests an "approved removal set and PR, branch, or diff".
Boundary markers: The skill uses hashes of every writable file and CAS receipts to ensure that only the exact, approved changes are applied.
Capability inventory: Shell command execution for tests and gates (in SKILL.md).
Sanitization: The runtime-trust.md file contains explicit instructions to treat all repository files and tool outputs as untrusted data, stating that embedded instructions cannot alter the skill's scope or authority.

tailrocks-simplify