The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/agent_usage.py executes the local codex CLI tool to retrieve account quota information.
Evidence: The capture_codex function uses subprocess.Popen to launch the codex app-server and communicate with it via a JSON-RPC interface on stdin/stdout. The executable path is resolved using shutil.which.
[INDIRECT_PROMPT_INJECTION]: The skill processes agent conversation transcripts (e.g., from Claude or Codex rollouts) which could contain malicious instructions designed to influence the agent during the analysis process.
Ingestion points: scripts/usage_collectors.py reads local log files and databases from multiple agent types (Codex, Claude, Copilot, Grok, and Antigravity).
Boundary markers: The skill performs deterministic analysis (token counting, character measurement, and regex-based classification) rather than passing the raw content back to an LLM for interpretation.
Capability inventory: The skill writes local reports (JSON, Markdown, HTML), maintains a local SQLite ledger, and interacts with the codex CLI tool.
Sanitization: Input data is sanitized using the safe_name utility, Markdown output is escaped in scripts/usage_render.py, and the HTML dashboard implements a strict Content Security Policy.
[SAFE]: The skill implements a guarded_path function in scripts/agent_usage.py that explicitly prevents the reading of common credential and configuration files (e.g., .env, auth.json, secrets.env, credentials.json, hosts.yml).
[SAFE]: The generated HTML dashboard (assets/dashboard.html) includes a highly restrictive Content Security Policy (CSP) that prevents the execution of external scripts, image loading from remote sources, and form actions, significantly mitigating XSS and data exfiltration risks.

agent-usage-analytics