← חזרה לכל השלבים

Orion Studio — Research (P5-1)

Stage: RESEARCH. Goal (Ben): find open-source projects + companies to learn from for a
local-first control room / IDE for Orion + PRISM, with VPS terminal access, secured,
reading the existing stores (Orion Postgres + Directus) — not a parallel data store.
Reference shape: nesquena/hermes-webui. Feeds VISION.

A. Observability / trace layer — "see all the runs" (manage Orion)

The market has mature open-source trace UIs; we should study, not reinvent:

end-to-end tracing, prompt mgmt, eval, datasets. The reference for the runs/traces panel.

For us: Orion already has the data (shadow_logs + Sherlock). The Observe panel should

present Orion's own runs in a Langfuse-style waterfall (cost/tokens/steps), via the Observe

API (which wraps Sherlock's functions). Possibly emit OpenTelemetry so Phoenix/Langfuse could

plug in later — but the native panel reads Orion directly.

B. Agent IDE / control-room shape — the layout

+ workspace (VS Code integration) + Jupyter + app viewer + browser + **terminal for command

execution + web GUI + Terminal UI + CLI. Self-hostable, free local tier. Study its panel

layout and terminal integration closely.**

workspace), control-center modal, approval gates, passkey auth, mobile. The

self-hosted-shell + security ethos.

For us: hermes-weight shell (no heavy framework) + OpenHands-style panels (terminal,

workspace) + Langfuse-style trace views. OpenHands = the ceiling, hermes = the floor.

C. VPS terminal in the browser (Ben's core requirement) — embed, don't build

embeddable in an iframe/panel. Top pick for the terminal panel.

"manage the VPS" surface.

container ops (restart Orion, view logs).

For us: embed ttyd for the VPS terminal; borrow Cockpit/Portainer patterns for

service/container control. xterm.js is the standard terminal frontend.

D. Security / remote access (Ben: "everything secured")

zero open ports, NAT traversal) to reach the VPS; open alternatives: NetBird (BSD-3,

self-host), Headscale (self-host Tailscale control server), Pangolin (identity-aware).

For us — defense in depth: Tailscale to the VPS (no public ports) + Caddy TLS + passkey

on the Studio + ttyd's own auth on the terminal. Local-first by default; remote via Tailscale.

E. Companies to learn from

Cognition (Devin) · Factory.ai · All-Hands (OpenHands commercial) · Langfuse ·

Arize — their "agent mission-control / control-room" UX is the bar.

F. Integration constraints (from Ben)

+ the existing Directus CRM where things are already documented. The Studio is a *lens +

controls*, never a second source of truth ("UI never computes status" — it reads Orion's).

it also needs the Observe API (wrap Sherlock functions) before the UI can read live.

G. Synthesis → recommended v1 (input to VISION)

Shape: hermes-weight self-hosted shell, dark "tool" aesthetic (own identity, divergent

from portfolio/foodflow). Two halves in one UI (manage Orion / manage PRISM).

v1 = Observe + Terminal (A), then grow to act (B):

container/service status (Portainer/Cockpit pattern) · HITL approvals (tier 2) next.

Don't reinvent: embed ttyd; study Langfuse (traces) / OpenHands (IDE) / hermes (shell+auth);

build only the thin Orion+PRISM-specific glue.

Sources