The Agent Skills Directory

[PERSISTENCE]: The skill installs persistent event hooks into the agent's configuration file (~/.claude/settings.json) and copies executable scripts to ~/.claude/hooks/. This mechanism ensures the cache-resume-detect.py, cache-warn-stop.py, and cache-expiry-warn.py scripts execute automatically during specific lifecycle events (SessionStart, Stop, UserPromptSubmit) in all future agent sessions. While this is the stated purpose of the skill and requires user confirmation, it represents a significant modification to the agent's execution environment.
[DATA_EXPOSURE]: The skill performs an extensive read of all conversation history, metadata, and token usage data stored within the user's home directory at ~/.claude/projects/ to generate its analytics.
[INDIRECT_PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection due to its data processing pipeline:
Ingestion points: scripts/ingest_token_data.py reads raw JSONL conversation logs from ~/.claude/projects/.
Boundary markers: There are no explicit delimiters or instructions to ignore embedded commands when constructing the prompt for the claude-code-guide subagent in Step 1.5.
Capability inventory: The skill has the ability to write to the filesystem, modify configuration files, and install executable hooks via the Bash tool.
Sanitization: The skill instructions specify including finding and root_cause text 'verbatim' from the ingested logs into the subagent's prompt without escaping or sanitization, potentially allowing malicious content in past logs to influence subagent behavior.
[EXTERNAL_DOWNLOADS]: The templates/dashboard.html file references external assets from well-known services, including Chart.js from the jsDelivr CDN and fonts from Google Fonts. These are used for the interactive visualization dashboard.

get-token-insights