Renamed Areas to Spaces with a new spaces_projects join table. Projects now live inside a Space with scoped brand and context. Thread orchestration moved off a 5-second poll onto Supabase realtime on postgres replication; Slack mutations land on web surfaces sub-second. Migration ran online behind a feature flag with no downtime.
Inflow is now HumanDeploy. Same codebase, new domain (humandeploy.ai), new name, new story. No data migration. Existing tenants kept their workspace, memory, and billing. Founding-tier pricing holds at $2,500/mo.
", "date_published": "2026-04-10T14:00:00Z", "tags": ["Launch", "v0.7.0"], "authors": [{ "name": "Steffan Howey" }], "_humandeploy": { "version": "0.7.0", "state": "shipped", "signed_by": { "name": "Steffan Howey", "role": "Founder", "kind": "human" } } }, { "id": "humandeploy:changelog:v0.9.0", "url": "https://humandeploy.ai/changelog/#v0-9-0", "title": "v0.9.0 — Credit operations", "summary": "credit-operations with reserve, consume, release. Each call runs inside a SELECT FOR UPDATE transaction with an idempotency key.", "content_html": "Added credit-operations with three primitives: reserve, consume, release. Each call runs inside a SELECT FOR UPDATE transaction with an idempotency key, so a retried network call never double-spends. A scope approval reserves against client balance; delivery consumes; cancellation releases. Failed debits write to a recovery table the admin dashboard reads from. Unlimited-credit allowlist accounts bypass reservation.
Shipped a second Slack app for experts with separate install, separate OAuth, and shared workspace memory. When a handoff moves into assigned, the expert receives a DM with scope, deadline, and a link to the brand profile. Accept flips state to in_progress; decline routes to next-best-fit via the assignment engine. Assignments with no response in 2 hours expire on a cron and re-route to the next expert in rank. Experts see a kanban of the work they own.
integration-oauth now handles the full OAuth lifecycle for any provider listed in providerConfigs.ts. Tokens are AES-encrypted before persistence. Connectors extend BaseConnector and inherit per-provider rate limiting (10 req/s default) and circuit breaking on every outbound call (opens at 5 failures, cools for 60s). A health runner scans active connections every 15 minutes. Notion and Google Drive are live on the new path; adding the next provider is a pull request against registry.ts plus two env vars. Slack and Stripe remain on legacy.
Added generate-handoff-scope. A Slack DM from a client is read alongside the workspace brand guardrails and any prior project context, and returns a HandoffScope validated against a Zod schema: deliverables, exclusions, timeline, credit cost. The scope renders back into the thread via Block Kit with Approve and Adjust actions. Approve writes a row into work_requests on the new slack_conversation intake path and moves the handoff to approved. Adjust opens a modal.
Every inbound Slack message now runs through classifyAndRoute() before reaching a handler. It reads thread history and the workspace brand profile, then returns one of eight intent types with a confidence score. Confidence above 0.7 routes to the owning handler; below that triggers a clarifying question. Runs under 400ms p95. Model: GPT-4.1-mini, temperature 0.2, structured JSON output.
Added the handoffs table with eight states: draft, approved, assigned, in_progress, delivered, revision, complete, cancelled, plus paused. A Postgres trigger validates every transition against an allowed-edges map, so an invalid write fails at the database. Requester-allowed transitions are separated from admin-only ones via RLS. Each state writes its own timestamp column. Downstream surfaces read this row as the source of truth: the client's confirmation, the expert kanban, the admin dashboard.