The Agent Skills Directory

[COMMAND_EXECUTION]: The skill frequently executes shell commands through multiple scripts (e.g., bash_compress.py, antigravity_doctor.py, measure.py) to check tool versions, audit context, and run whitelisted user commands. These executions are guarded by a restricted whitelist of safe commands and pre-tokenization injection screens.
[PERSISTENCE]: The skill establishes persistence by installing hooks and plugins for AI agent platforms like Claude Code, Codex, Hermes, and Cursor. These hooks automate the optimization process across different session lifecycle stages, such as session startup and tool usage.
[DATA_EXFILTRATION]: The skill audits and caches tool outputs and user configuration files (e.g., settings.json, CLAUDE.md, MEMORY.md) in local SQLite databases. It employs a multi-layered credential redaction system (credential_patterns.py) that identifies and masks sensitive information like AWS keys, GitHub tokens, and database URIs before storage. While the skill collects telemetry, it is designed for local auditing and includes defenses to prevent secrets from being recorded.
[OBFUSCATION]: The codex_command_compress.py script uses Base64 encoding to pass structured data ("plans") as command-line arguments. This is an architectural mechanism for passing state through hooks and does not obfuscate intent or hide malicious code.
[INDIRECT_PROMPT_INJECTION]: The skill mitigates the risk of indirect prompt injection by explicitly instructing its analysis agents to treat all ingested file content as raw data and to ignore any instructions contained within those files, as specified in agent-prompts.md.

token-optimizer