The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill fetches the codeburn package and dependencies like better-sqlite3 from the public npm registry via npx, npm, bunx, and pnpm. It also instructs the user to install binaries through Homebrew (brew install codeburn).
[REMOTE_CODE_EXECUTION]: The instructions encourage immediate execution of remote scripts using npx codeburn and bunx codeburn, which download and run code in a single step without local verification of the contents.
[COMMAND_EXECUTION]: The skill executes shell commands that modify the local environment and the agent's own configuration, specifically through codeburn guard install, which writes to ~/.claude/settings.json and ~/.config/codeburn/guard.json.
[DATA_EXFILTRATION]: The skill includes a sync feature (codeburn sync setup <url> and codeburn sync push) that allows users to configure an external endpoint and transmit usage metadata, including project names and token costs, to that server.
[PROMPT_INJECTION]: The skill presents an indirect prompt injection surface by ingesting and processing untrusted data from session logs.
Ingestion points: The tool reads local session files from approximately 40 different AI coding tools (e.g., Claude Code, Cursor, Codex) stored in directories like ~/.claude/.
Boundary markers: There are no explicit delimiters or instructions provided to the agent to disregard embedded commands within the log data being parsed.
Capability inventory: The skill has access to Bash, Read, Write, Edit, Glob, and Grep tools.
Sanitization: There is no mention of sanitization or escaping of the session log content before it is processed or exposed to the agent via the codeburn mcp tools (get_usage, get_savings).

codeburn