The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill references and encourages the download of a repository from an untrusted external source (github.com/vibeforge1111/keep-codex-fast.git). The account 'vibeforge1111' is not verified or associated with the stated skill author ('reason-machines' or 'ara.so').
[REMOTE_CODE_EXECUTION]: The skill requires running python scripts/keep_codex_fast.py, which is part of the externally downloaded package. This constitutes the execution of untrusted remote code on the local machine with user-level permissions.
[COMMAND_EXECUTION]: The maintenance functionality relies on executing shell commands and Python's subprocess module to manipulate the local file system, interact with system processes, and modify database states.
[DYNAMIC_EXECUTION]: Code snippets within the skill demonstrate the use of sys.path.insert to dynamically load modules from computed paths and subprocess.run to execute external scripts, which is a pattern often used to hide or inject malicious logic at runtime.
[INDIRECT_PROMPT_INJECTION]: The skill has a surface for indirect prompt injection because it processes local SQLite databases (state_5.sqlite, logs_2.sqlite) which contain untrusted content from past chat interactions.
Ingestion points: The skill reads and parses chat session metadata and log files through the scripts/keep_codex_fast.py script.
Boundary markers: No delimiters or safety instructions are provided to the processing script to distinguish between data and potential commands within the logs.
Capability inventory: The skill possesses the ability to move worktrees, archive folders, modify database fields (title/preview), and execute subprocesses.
Sanitization: There is no evidence of validation or sanitization of the database content before it is processed or modified by the script.

keep-codex-fast-maintenance