agent-session-auditorlisted

# Agent Session Auditor Turn repeated user requests into evidence-backed skill candidates. The audit is read-only, local-first, privacy-preserving, and honest about coverage. ## Safety - Get scope from the request: machine, user account, time range, agent families, and whether archived sessions count. Do not broaden to other users, cloud accounts, or network stores. - Never modify, migrate, compact, or delete a session store. Open databases read-only and copy a database to temporary storage before queries when its client may create journal files. - Extract user-authored prompts only. Exclude system/developer messages, assistant replies, tool calls/results, injected repo instructions, retries, and generated continuation noise. - Redact secrets, tokens, credentials, personal/customer data, private URLs, and sensitive absolute paths before examples or intermediate artifacts leave their source. - Keep clustering local by default. Obtain explicit approval before sending transcript-derived text to a hosted embedding or model service. ## Workflow 1. Discover stores from installed commands, documented paths, app data, config, and recent modification evidence. Do not guess unsupported directories. 2. Create a coverage manifest per agent: detected version, store paths, format, sessions found, sessions parsed, prompts extracted, skipped records, errors, and unsupported surfaces. 3. Sample each format before bulk parsing. Identify role fields, content blocks, timestamps, session IDs