The Agent Skills Directory

[COMMAND_EXECUTION]: Executes several local development and system tools including git for version control operations, uv and pytest for testing and verification, and the t3 CLI for framework-specific tasks.
[DATA_EXFILTRATION]: Contains logic to commit and push skill improvements to remote git repositories (typically the user's fork or configured upstream). This is governed by a mandatory privacy scan and user confirmation modes.
[PROMPT_INJECTION]: Processes untrusted data from conversation transcripts and PR review comments to synthesize improvements.
Ingestion points: Reads full conversation history and fetches PR review comments via forge clients.
Boundary markers: Relies on internal agent context boundaries; instructions do not specify custom delimiters for injected content.
Capability inventory: Significant capabilities including file system modification, command execution, and network operations (git push).
Sanitization: Implements a dedicated Privacy Scan phase that filters author content for emails, home paths, IP addresses, API keys (regex-based), and banned terms before any external persistence occurs.

retro