The Agent Skills Directory

[SAFE]: The skill uses a 'Parent/Child' model architecture to separate planning/validation from execution. The Child model (Sol) is explicitly invoked with a restrictive guard (extensions/deny-git.ts) that prevents any interaction with Git or the .git directory, mitigating risks of unauthorized repository manipulation.
[SAFE]: Tool access for the editing agent is limited to a pre-approved list of files and specific indexed commands via run_check. This prevent arbitrary shell execution or 'jailbreaking' into unauthorized directories.
[SAFE]: All external references, such as the openai-codex model identifiers and internal extension paths, are part of the intended developer platform integration and do not involve untrusted third-party downloads.
[SAFE]: The skill implements a 'Strong Ownership' pattern where the Parent model must independently re-run all validation steps before accepting code, preventing potential injection or hallucinated successes from the execution layer.

pi-develop