Skip to main content

​
Brain Changelog

Every shipped workflow, feature, or capability lands here. The bar is:
  • Tracing wired up
  • Correction logging in place (where it applies)
  • A workflow or architecture page updated
  • This changelog updated
If those four aren’t true, it isn’t shipped.
Status legend:
  • 🟢 Live — running with real users on real data
  • 🟡 In Development — being built now
  • ⚪ Planned — next up
  • 🔵 Future — backlog

​
v0.1.1 — Aggregator pattern + name corrections

Date: 2026-05-19 Status: ⚪ Documentation update

​
Changes

  • New architecture page: Aggregator Agent + Workflow Registry. Documents the supervisor pattern that routes chat messages to registered workflows, with slash-command and approval-click bypass paths.
  • Updated architecture index: at-a-glance diagram now shows the aggregator layer between UI surfaces and workflows, plus the external-services row for the BOM agent, KG service, and CutMake API.
  • Updated Principles #2 (Scoped chat): clarified that “scoped” doesn’t mean “slash commands only” — natural-language chat is allowed; the aggregator enforces scope by only routing to registered workflows.
  • Updated UI Structure (Ask surface): now describes how chat goes through the aggregator and how slash commands bypass it.
  • Name corrections: merchandiser is Mai (was incorrectly listed as “Mile” in v0.1.0 from the briefing doc). Manager is Zean / Zean (was “Zin / Sae Zin”). Updated across all 25 pages.

​
Engineering decisions captured this update

From the /plan-eng-review session on the v0.1.0 docs:
  • CutMake integration boundary: reads go direct from CutMake’s Supabase; writes go through CutMake’s HTTP API or MCP tools (no direct cross-project writes). Custom AI endpoints added when existing endpoints don’t fit.
  • Observability for Phase 1: LangSmith stays during early development (already wired in deer-flow); migration to Langfuse self-hosted before merchandiser #3 onboards or any data-residency clause lands.
  • Deployment: managed PaaS (Fly.io / Render / Railway) for the three services; not Kubernetes.
  • Repo structure: four repos — agent-research (Brain framework, deer-flow fork), bom-agent (Ku Kue), brain-kg (Andy), brain-contracts (shared Pydantic + Markdown).
  • Contracts: only at service boundaries (Brain ↔ BOM agent, Brain ↔ KG service). Internal Brain features don’t need contracts.

​
v0.1.0 — Project Bootstrap

Date: May 2026 Status: ⚪ Initial documentation snapshot

​
Overview

Brain enters formal development. This release is a documentation and planning snapshot — no production workflows are live yet. It captures:
  • The vision and bet behind Brain (Vision)
  • The non-negotiable principles every workflow follows (Principles)
  • The phase-by-phase roadmap (Roadmap)
  • The full set of planned workflows (Workflows)
  • The architecture choices (deer-flow runtime, dual Supabase, scoped UI surfaces, Langfuse observability) (Architecture)
  • The role and permission model (Roles)

​
What’s in development

  • Tech Pack → BOM — 🟡 In Development. The anchor workflow. PDF parsing, Excel parsing, image analysis, garment template matching, BOM draft generation with confidence scoring. Built by Ku Kue on a dedicated agent track.

​
What’s planned next (Phase 2)

​
Infrastructure decisions made

  • Runtime: deer-flow (forked, private at vinmaketeam/agent-research)
  • Primary observability tool: Langfuse, self-hosted (privacy for client data)
  • Dev observability: LangGraph Studio
  • Eval platform: Braintrust — not yet, added at Phase 2-3 when correction data exists
  • Brain database: dedicated Supabase project, separate from CutMake’s
  • CutMake integration: read for context, promote drafts on approval
See Observability and Data Integration for the reasoning.

​
How to read this changelog

Each release entry tells you:
  • What shipped — concrete workflows, capabilities, or infrastructure
  • What’s in development — what’s actively being built next
  • Decisions made — architecture or process choices worth recording
  • Links to the relevant workflow or architecture page
The changelog is the commitment record. If something is described as live here, Mai and Zean should be able to use it. If it’s not here, it doesn’t exist yet.

​
Upcoming releases

The format below will be filled in as Brain ships.

​
v0.2.0 — Tech Pack → BOM (target)

The first live workflow.

​
v0.3.0 — Style Creation + Email Intake (target)

The framework layer around the anchor feature.

​
v0.4.0 — True Cost + Quotation (target)

Closing the intake-to-quotation loop.

​
v0.5.0 — Oversight + Manager Queries (target)

Pipeline visibility.

​
v1.0.0 — Pricing Intelligence (target)

Once the data substrate exists.