Herald · Stages H0–H5

AI Marketing Manager

Herald routes work to specialist employee agents, assembles a walled per-task briefing packet (one company per context, always), brand-checks output, runs the authorization chain to Nick, and learns from real outcomes.

H0 · Foundation (the walls)

✓ Schema + RLS migrations — supabase/migrations
✓ Briefing Packet Assembler (the runtime wall) — lib/packet.ts
✓ Isolation test — npm run test:isolation
✓ Seed loader (agents · knowledge · OKRs · authority) — npm run seed
✓ Ingestion worker (content + SOPs, walled) — npm run ingest

H1 · Content agent loop (GMRV)

drafted → brand_checked → pending_authorization → authorized → published

✓ Employee runtime runEmployee — packet → draft, never publishes
✓ Herald brand-check — reviews vs overlay, returns notes, never rewrites
✓ Authority routing — L0 → Nick's queue; L1/L2 graduate
✓ Nick authorizes → publish with a tracking_id
✓ Lifecycle test (nothing auto-publishes) — npm test

H2 · Attribution + KPIs + brief

✓ Attribution — ARC revenue_events → asset_metrics via tracking_id, idempotent
✓ Scorecard rollup — department → company → agent (outcomes vs activity)
✓ Performance brief — Sonnet composes from real numbers
✓ Nightly rollup cron — /api/cron/rollup (CRON_SECRET-guarded)
✓ CLIs — npm run attribute · npm run brief

H3 · Authority graduation + performance (IDPs)

✓ Graduation — N proven outcome deltas → Herald proposes L0→L1→L2; Nick approves
✓ Pinned L0 — ad_spend · positioning never graduate
✓ Reversible — a bad streak drops a level
✓ IDPs — missed KPI → diagnosis + upgrade action + leadership style (we upgrade, don't fire)
✓ Weekly cron + CLIs — npm run review · npm run graduate

H4 · Gated learning loop

✓ Surface — a tracked win becomes a candidate_lesson (Frontier→Playbook)
✓ Gate — promotes into the knowledge brain only on Nick's approval
✓ Commons bar — 2+ companies OR structural, always a metric delta
✓ Closes the loop — the promoted lesson rides in the next packet
✓ Weekly cron + CLIs — npm run lessons · npm run promote

H5 · Hire #2 + multi-agent routing

✓ SEO Strategist hired — added via the template, no code change
✓ Routing — one owner per category; Herald picks the specialist
✓ No freelancing — cross-domain/unknown work escalates to Herald
✓ Isolated per company and per agent (playbooks don't leak)
✓ Generic CLI — npm run task -- --category seo_brief …

Hard rules: exactly one companyId enters any packet — two companies never coexist in a context window. Nothing publishes without an authorized proposal. Herald brand-checks and returns; it never rewrites a specialist's output. One owner per domain — no cross-domain freelancing. Outcomes, not activity. Nothing self-promotes.