The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses Bash(python *) to execute local helper scripts (scan_invocations.py and write_session_review.py). Analysis of these scripts shows they are dedicated to reading local log files and writing markdown reports within a specified reports/ directory.
[DATA_EXPOSURE]: The skill reads session logs from specific local agent directories (e.g., ~/.claude/projects, ~/.grok/sessions). This is the intended primary purpose of the skill. The write_session_review.py script includes a proactive REDACT_RE regex that automatically replaces API keys (sk-, ghp_) and Authorization tokens with [REDACTED] before writing them to the report files.
[REMOTE_CODE_EXECUTION]: No remote code execution patterns were detected. The skill operates entirely on local session files and does not perform network requests.
[PRIVILEGE_ESCALATION]: The skill contains specific security checks in its Python scripts, such as refuse_reparse, which prevents directory traversal attacks by refusing to interact with symlinks or reparse points when resolving the repository root or report destination.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data (historical chat logs). It addresses this risk by providing a strict 'finding contract' in references/finding-contract.md, limiting excerpts to 200 characters, and requiring a handoff prompt for human-in-the-loop review before any changes are applied to target skills. It specifically forbids the agent from editing the target skill directly in the current session.

skill-session-review