Robin Researcher
A daily research agent that turns forwarded links into structured, fully-sourced vault notes — phase one of the Robin project, and adopted rather than built from scratch.
This prototype is private and not meant to be shared. It's listed here only as a placeholder for structure.
Context
Robin is a long-horizon agent project: a premortem first, then a roadmap where every node has to be earned before the next gets built, then a constitution with kill criteria. The plan's first node was a researcher that reads a research inbox and files what matters — then an existing pipeline turned out to already do most of it, so it was adopted instead of building a new one. It's been running daily at 5pm since July; the build work went into the scaffolding around it — the constitution, the run ledger, and a frozen test set for scoring.
What it tests
Whether one strict-sourcing agent holds up under real daily traffic before any second node exists: can it handle whatever actually lands in the inbox, including bare share-sheet links with no context; does headless scheduling stay reliable when the platform's own trigger list lies about it; and can a fixed test set turn "it seems fine" into an actual score.
Notes
Real traffic runs 10–14 items a day, and bare share-sheet links are now the norm, not the exception. The scheduler platform has a bug — triggers are invisible to its own API and 404 in the desktop UI — so the lesson so far is that the inbox and Gmail labels are ground truth, not the trigger list; no third blind trigger gets added. The frozen test set was backfilled from real runs on 21 August but hasn't been formally scored yet — that scoring is the gate for building node two.