The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process content from untrusted external sources, including GitHub repositories, academic papers, and product web pages.
Ingestion points: Discovery and extraction phases utilize WebSearch, WebFetch, and git clone (Phase 2 & 5) to pull external content into the agent's context.
Boundary markers: The skill emphasizes structural delimiters by requiring every extraction to follow defined schemas like MANIFEST.md, KEY-FINDINGS.md, and FEATURE-MATRIX.md.
Capability inventory: Across all scripts and instructions, the skill maintains access to Read, Write, Edit, and Bash tools, providing a surface for action if an injection were to be successful.
Sanitization: While the workflow involves normalization and analysis by the agent, there are no explicit sanitization or escaping protocols defined for handling raw external data before it is interpolated into synthesis prompts.
[EXTERNAL_DOWNLOADS]: The workflow involves identifying and cloning external repositories for analysis.
Evidence: Phase 5 (Gap-Filling Extraction) explicitly instructs the agent to clone promising repositories to a temporary directory for selective pattern extraction.
Context: These downloads target well-known services (GitHub, arXiv) and are central to the skill's primary research purpose.
[COMMAND_EXECUTION]: The skill utilizes shell commands to interact with analysis tools and manage the local environment.
Evidence: SKILL.md describes a "Multi-Model Reconnaissance" approach using bash -lc to invoke tools like gemini and codex for repository and dependency analysis.
Context: Shell access is also used for repository cleanup (deletion of temporary clones) following extraction.

reconnaissance