Matron

Curator keeps our registry honest and the pipeline moving. It hunts for projects with missing metadata, a stage that contradicts its health, or a duplicate hiding in the list.

It checks reality — git log in the repo, whether the prod URL is actually up — instead of trusting the registry record. We set the stall line at 21 days: an in-flight or POC project with no commit in that window gets flagged.

It then nominates projects for the gate: promote, hold, kill, or recycle. For a solo operator the kill gate matters most, so we let it propose sunset for the genuinely dead. Stage and health stay separate columns — live-and-red is a real state.

Sentinel watches only what is actually live, and asks one question per project: is it healthy right now, and if not, what broke.

It pulls Railway deploy status, Sentry errors with Seer's root-cause, uptime by fetching the prod URL, and cost anomalies — then rolls it all into one verdict, green, amber, or red, with the failing deploy id or the Sentry link as evidence. We run it hourly over just the live subset, which keeps it cheap.

It raises the alarm and proposes the fix; it never runs the fix. That decision stays with you, every single time.

Scout is the expensive worker, so we built it not to wander. For in-flight projects it researches improvements, new libraries, and competitor moves; for pipeline ideas it runs feasibility and market scans.

It reads your registry record and your repo first, so it never hands you something you already built. What comes back is tight: a TL;DR, three to six ranked findings with citations, and improvements tied to where the project actually is.

We cap it at five to eight searches per project, weekly, and only on the projects you name. It stops when it has enough to recommend — not when it has read everything.

Postman turns your inbox into per-project digests and ready-to-send drafts. It pulls the recent threads under each project's Gmail label, sorts every one — urgent, needs-reply, info, noise — bundles them by project, and drafts the replies in your voice.

The limit is deliberate: it cannot send; you send. The connector has no send tool at all — it writes the reply, and the send stays yours.

It never invents a price, a date, or a commitment you did not authorise; if a reply needs a real decision, it drafts the question back to you instead of making one up.

Time-sensitive drafts expire, so a stale reply cannot fire three days late.

Scribe writes the briefing and leads with what needs you. A 20-project roster recited every morning just trains you to ignore it, so we collapse healthy projects to one line — '11 of 12 live projects healthy' — and cap the needs-decision queue at the top five by severity.

It takes the deterministic skeleton and folds in Sentinel's verdicts, Postman's bundles, Scout's briefs, and Curator's nominations as colour. Its second job is the verification pass we insisted on: it checks every proposal against its own evidence before the proposal reaches you, and flags anything that looks hallucinated.

If nothing needs you, the briefing says 'all clear' and stops there.