The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted repository files through multiple audit agents, creating a surface for indirect prompt injection where malicious content in a repository could influence agent behavior.
Ingestion points: The entire repository file tree is read during Wave 1 (Reconnaissance), Wave 2 (Code Quality), and Wave 3 (Domain-Specific audits).
Boundary markers: The skill uses structured 'Context blocks' to pass repository metadata to sub-agents, but does not explicitly instruct the agents to use robust delimiters or sanitization when processing the content of audited files.
Capability inventory: The orchestrator has the capability to read files, write audit reports and execution plans to the local file system (docs/local/audits/, docs/local/specs/, docs/local/plans/), and dispatch other agents with full repository access.
Sanitization: The instructions do not specify sanitization or escaping of file content before it is processed by the sub-agents.
[DYNAMIC_EXECUTION]: The skill provides a Bash script within references/detection-patterns.md and instructs the agent to execute it to perform technology stack detection.
Evidence: Step 1 (Reconnaissance) explicitly directs the agent to 'Run the detection script from the repo root' and provides the full script content in the reference file. The script uses grep and git ls-files to match patterns against the file list.
[COMMAND_EXECUTION]: The skill performs several shell commands to establish repository state and manage audit artifacts.
Evidence: Step 0 (Preflight) and Step 1 use git rev-parse, git branch, git ls-files, and mkdir -p to prepare the environment and collect metadata.

deep-audit