Baseline Studio — Internal Systems

Baseline HQ — System Architecture

Read it top to bottom, the way a request actually flows: work enters through two surfaces, the surfaces drive the workers, every worker reaches the world through one gateway — and all of it stands on one shared brain.

User surfaces

The only two places work enters: the HQ web app for solo deep work, Slack for the agent team

HQ Web Shellhq.baselinestudio.design — chats + terminals, Google-SSO gated
SlackBaseline HQ workspace — mention an agent, it works; replies post back here

The Droplet

Everything that runs Baseline HQ, systemd-supervised and always on — the engines that do the work, plus the one gateway they all share to reach outside

HQ Sessionsone engine per chat/terminal · hq-shell :3000
Slack Receiverroutes each mention to its agent · :8790
Schedulerthe clock — fires each automation on schedule
The 5 Slack agents — a persona team that coordinates, ships, and merges its own PRs
BaseChief Agent Officer
DarwinResearcher
HemingwayWriter
RamsDesigner
TuringEngineer
6 scheduled automations — no human in the loop
Librarian3×/day
Daily Brief9am + hourly
Report Carddaily 6pm
Researchdaily 9am
Usage Reporterweekly Mon
Repo Opsnightly
One shared brain — every worker above reads this before acting, writes back after
Skills Library18 playbooks
Knowledge Basekb/ · 8 categories
Session Memorychats + terminals
Agent Memoryper Slack agent, PR-gated
Thread ContinuitySlack session resume
Self-Improvementgolden-task eval loop
MCP Gatewaythe ONE door out: every external call from every worker passes one policy check, one per-caller allowlist, one audit log

The Outside World

The only things the gateway is allowed to reach — each with its own hard limit baked in

GitHubTrentM6 only
Vercelpersonal team
Gmaildraft-only
Mercuryread-only
OpenRoutermodel fallback
HN / Reddit / Xread-only
second_opinioncross-model review
Shell path
Slack path
Scheduled path
Gateway → services
Every worker reads / writes the brain
×N wires bundled on a bus
BASELINE STUDIO
TitleHQ System Architecture
Dwg NoBL-HQ-001
RevA
ScaleNTS
Date2026-08-24
DrawnHQ

Hosting detail — what's actually running on the droplet

ServicePort / TypeRole
CaddyedgeReverse proxy + automatic TLS for hq.baselinestudio.design
oauth2-proxyedgeGoogle SSO gate in front of the shell (Slack's webhook path is routed outside this gate)
hq-shell.service:3000The Next.js web app
hq-ttyd.service:7681Real terminal sessions (tmux)
hq-slack-agents.service:8790Slack Events receiver + all 5 agent turn-runners, one process
hq-mcp.serviceHTTPExposes the MCP gateway to remote/authenticated clients
hq-shell-reloaderwatcherZero-downtime deploy on every new build
hq-shell-health / hq-watchdogtimersSelf-healing — restart a dead or wedged unit
hq-auto-deploy.timer~90sFast-forwards the live checkout + every session sandbox to main
hq-capture-sweep.timer30 minSweeps quiet sessions/agent runs into inbox/ for filing

Self-improvement loop — measured, not vibes

Applies to any harness element — a skill, a Slack persona, an automation prompt. A pass-rate delta on a fixed task set decides whether a change ships; "feels better" is explicitly not allowed to be the reason.

1

Run the fixed task set against the current version — record the baseline.

2

Find the earliest point a failing run actually went wrong.

3

Write one reversible hypothesis before touching anything.

4

Make the single smallest edit that tests it.

5

Rerun the same tasks fresh, with no memory of the baseline run.

6

Retain on a real gain, revise on a partial fix, remove on none.

Memory — three distinct systems

Session memory

Host-local, cross-session recall for chats and terminals — who the operator is, standing feedback, live project context.

Agent workflow memory

Each Slack agent has its own git-tracked memory folder, loaded at the start of every thread. Agents propose changes to their own memory only; a human merge gate approves them.

Thread continuity

A persisted map from Slack thread → Claude session id, so --resume picks up exactly where a conversation left off, across restarts.

Knowledge base

decisions

Every meaningful call, with reasoning.

products

Grouped by product.

engineering

Cross-product patterns.

operations

How the company runs.

sales-marketing

Go-to-market knowledge.

playbooks +

company · customers · the rest.

Skills library — all 18

Design

ux-design · better-colors · better-typography · better-ui · pick-ui-library · dataviz

Writing

brand-voice · content-writing · deep-research

Reporting

daily-brief · agent-report-card · usage-report · deploy-watch · social-listening

Knowledge

kb-filing · work-capture

Discipline

verify-before-done · i-have-adhd

Meta

harness-improvement — the skill that improves the other skills