How to use Claude to audit and repair Shared Brain

Yurukusa’s overnight agent run included a security review of Shared Brain, a command-line tool that stores lessons and checks risky operations. A Claude Desktop coordinator dispatched the work, while Claude Code read files, implemented changes, ran checks, and created local commits.

For the audit, a Sonnet-based executor examined code previously written with Opus. The author reported three high-severity candidates: a regular expression could stall the pre-tool hook, lesson identifiers could escape the storage directory, and shell expansion could interpret tool input as a command.

The repairs added a heuristic filter for problematic regex patterns plus a subprocess timeout, restricted lesson identifiers and checked resolved path containment, and read hook input from an environment variable instead of expanding it in a shell command. Claude also added targeted regression cases for the three issues.

The author reported the checks passing, but left all overnight commits local and unpushed for human review. The surrounding automation had brittle terminal and browser bridges and had previously approached destructive publishing actions. This account therefore demonstrates an agent-produced audit and repair package awaiting review, with explicit evidence that autonomy ended before publication.

Prompt
Suggested prompt (written by Agents In Use, not the author's)

Source: DEV Community · Yurukusa