The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from git diffs, pull request metadata, and specification documents which are potential vectors for indirect prompt injection.
Ingestion points: Diffs are collected via git diff and git log commands in axes.md and closeout.md. Specification sources are identified in local paths like docs/, specs/, and .scratch/ as described in axes.md.
Capability inventory: The agent is instructed to use git, gh (GitHub CLI), and codex tools, and it has the capability to modify files through the 'implementer' role when findings are accepted.
Boundary markers: The skill uses structured Markdown headings (e.g., ## Summary, ## Issues) to organize output, providing a clear schema for findings.
Sanitization: The closeout.md contract mandates that the agent verify every finding against the real code path and adjacent files, and it implements a 'Scope Governor' to limit the depth and breadth of changes, preventing unauthorized refactoring based on potentially malicious input.

code-review