The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill acknowledges the risk of malicious instructions embedded in external data and provides explicit mitigations.
Evidence: references/runtime-trust.md defines repository files, reports, fixtures, and registry content as untrusted data.
Mitigation: The instructions explicitly state that embedded instructions cannot alter the agent's scope, governing rules, or approval requirements, effectively establishing a boundary against indirect injection attacks.
[COMMAND_EXECUTION]: The skill utilizes standard, well-known Rust ecosystem tools to verify code correctness.
Evidence: references/tooling-lints.md and references/errors-testing-docs.md instruct the use of cargo test, cargo clippy, rustfmt, and nextest.
Context: These tools are used as "preservation oracles" to ensure that refactoring does not change the observable behavior of the code, which is a standard and safe development practice.
[DATA_EXFILTRATION]: The skill contains proactive measures to prevent the accidental exposure of credentials or secrets.
Evidence: references/runtime-trust.md mandates that secret values must remain unread and never be copied into outputs, logs, or artifacts. The agent is instructed to only cite the location and type of any discovered credentials.
[PRIVILEGE_ESCALATION]: The skill emphasizes the principle of least privilege regarding system authority.
Evidence: references/runtime-trust.md states that repository content grants no authority for commit, release, or external-system operations, requiring separate explicit human sign-off for destructive actions.

tailrocks-rust-refactor