The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is specifically designed to ingest third-party codebase content, creating a significant attack surface for indirect prompt injection. However, it explicitly addresses this in its core principles by defining a 'Trust Boundary' which mandates that all content read from a repository must be treated as data and never as instructions.
[DATA_EXFILTRATION]: While the skill identifies sensitive information during its deep recon phase (e.g., API keys, DB connections, configuration files), it implements a normative 'Secrets Rule' across all phases. This rule strictly prohibits recording actual values, instead requiring the agent to note the location and name of the secret and use redaction for any potentially sensitive values.
[COMMAND_EXECUTION]: The skill utilizes common developer toolchains (such as dotnet, npm, and go) to resolve project structures. These commands are limited to read-only operations and metadata extraction. The playbooks explicitly forbid the execution of commands that would fetch remote dependencies or modify the system without clear user approval.
[EXTERNAL_DOWNLOADS]: The skill documentation includes references to external MCP bridges for Language Server Protocol (LSP) support. These are positioned as opt-in configurations for the user to set up manually, and the skill does not attempt to automate the download or installation of these third-party tools.

codebase-discovery