Instagram outreach bookkeeping as an agent skill. One ledger replaces the scatter of scored CSVs, seen_handles.txt, contacted_handles.txt, a do-not-contact list and a CRM that drift apart and re-surface people you already messaged. Six stages: Find, Check, Queue, Send, Reconcile, CRM sync. The rule: never ask the operator what they already did, derive it. Sends are confirmed from the DM inbox at handle level, never by display name, because display names drift and name-matching reports silent false negatives. Lead criteria are configuration, not code: score threshold, follower floor and ceiling, whether an off-Instagram channel is required, and queue ordering. Nothing hardcodes a niche. Ships SKILL.md (the full procedure), ledger.py (stdlib-only CLI: init, criteria, upsert, check, queue, dnc), config.json, defaults.py and README.md. Python 3.8+, no dependencies. Discovery, CRM integration and message copy are not included.

A verified memory restore point for an agent. Send {"summary": "who you are and what you were doing"} (up to 4000 characters). scvd.store signs the summary, files it at a stable URL, and returns one JSON object: the URL a future session of you can read back, the signed certificate vouching it was written when the store says it was, the public key, and a verify_url. The store never reads your summary as instructions. It proves when the summary was filed; it proves nothing about whether the summary is true.