The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted repository data (Markdown, package manifests, and Makefiles) to generate context proposals and verify factual claims about the codebase.
Ingestion points: Repository files are read into the tool's context via runtime/scripts/lib/repo.mjs for analysis.
Boundary markers: The skill employs Markdown parsing and structured identification of code spans and links to separate instructions from data.
Capability inventory: The skill executes git commands through execFileSync and spawns Node.js audit scripts. It can create review branches and commit file changes to the local repository.
Sanitization: The skill implements strong architectural safeguards. runtime/scripts/lib/claims.mjs includes logic (verifySpans) to ensure the AI agent only uses verbatim substrings from the source file, preventing the model from inventing or paraphrasing claims. All findings are validated against strict JSON schemas before being presented.
[SAFE]: No confirmed security vulnerabilities were detected.
The skill does not perform network operations or access sensitive credentials outside the analyzed repository.
Subprocess execution of git and node is handled safely without shell interpolation.
All writes to the repository are gated by a content-addressed plan ID (plan_id) and require a review of the exact bytes being changed, as implemented in runtime/scripts/lib/context-init.mjs.
Core libraries (Ajv, js-yaml) are vendored as static files, mitigating external dependency risks.

context-init