Session audit

Answer questions about the user's own work from their local Claude Code transcripts (~/.claude/projects/<project>/<session>.jsonl, subagents under <session>/subagents/). Raw transcripts are gigabytes of JSON and full of secrets, so the bundled script reduces them to masked digests and counts first; readers and the synthesis work from those, and every count in the report is re-checked against them.

The evidence is local. When a research or deep-research skill was invoked on such a question, say that its web fan-out would answer a question about the world, not about the user's history, and run this skill instead. Web search adds nothing here.

$SKILL_DIR below is the directory holding this file.

Privacy

Transcripts hold pasted tokens, env values and tool output. These rules bind every step, every subagent brief and the final reply:

• All output goes to one mktemp -d directory under /var/tmp, outside any checkout. The script refuses a directory inside a git work tree, and one an earlier run already filled.
• Quote only user turns, as masked in the digests. Describe a tool error by tool name and error class. Tool output, command lines, tool inputs and env values stay out of the report and the chat.
• Read transcripts only through the script. To find a file, use grep -l or ls, which print paths, never matching lines.

1. Scope

Settle three things with the user, recommending a default for each: