The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill correctly identifies that processing repository and external content introduces a risk of indirect prompt injection.
Ingestion points: As defined in references/runtime-trust.md, the skill ingests repository files, registry content, and web content.
Boundary markers: It requires an independent oracle to verify behavior and characterization proofs to ensure mutation remains within the approved scope.
Capability inventory: The skill is scoped to refactoring source code (file read/write) and verifying API documentation (network read).
Sanitization: The references/runtime-trust.md document provides a mandatory framework where untrusted content cannot alter the agent's authority, governing rules, or side-effect permissions.
[DATA_EXFILTRATION]: The skill mandates strict controls to prevent credential exposure.
Evidence: references/runtime-trust.md explicitly forbids copying secrets into logs, prompts, or evidence records. references/middleware-and-security.md instructs the agent to exclude credentials and secrets from tracing and compression.
[COMMAND_EXECUTION]: Instructions regarding process lifecycle and concurrency follow safe development standards.
Evidence: references/lifecycle-and-testing.md requires bounding blocking work with spawn_blocking and managing all tasks within structured owners like JoinSet to prevent resource leaks and unmanaged execution.

tailrocks-axum-refactor