Panopticon

An operating system for AI-agent work. Talk in a room; the system passively listens, floats idea bubbles proposing things to build, and lets you spawn and steer many concurrent agent processes at once — by click+type or by voice.

This is the V0 of the design captured in .context/attachments/.../panopticon-session-1-guide.md — implemented as a runnable system today, with clean seams where Smithers, cue and local-inference plug in (see ARCHITECTURE.md). Smithers is the described backend now; Eliza support is planned later.

                          META-SESSION  (always-on outer loop · Smithers-backed)
                                │
        ┌───────────────────────┼───────────────────────────┐
   SUGGESTION ENGINE        INPUT ROUTER                PROCESS MANAGER
   (cue-style wake policy)  (steer vs. suggest, C2/C3)  (create/fork/kill/…)
   always-on transcript →   targeted → process queue          │
   idea bubbles             ambient  → suggestion engine   PROCESS #n
                                                            (session loop:
                                                             input→pre→action→post→output)

Run it

Requires Bun (≥ 1.2). No build step.

bun install            # types only; the app needs no runtime deps
bun start              # → http://localhost:7777

Optional: set ANTHROPIC_API_KEY to use real models instead of the deterministic mock brain. With no key it runs fully offline with a built-in mock that still exercises the entire loop.

ANTHROPIC_API_KEY=sk-... bun start
# model tiers (spec §5.9 / P-Cost-fit):
PANOPTICON_IO_MODEL=claude-haiku-4-5-20251001 \   # the cheap always-on suggest loop
PANOPTICON_PROCESS_MODEL=claude-fable-5 \         # per-process orchestration
ANTHROPIC_API_KEY=sk-... bun start

Try the flow

1. Open http://localhost:7777.
2. In the room bar at the bottom, type something buildable — e.g. "we should build a tool to track all our running agents" — or hit 🎤 to talk (Web Speech API). This is the ambient channel.
3. An idea bubble appears on the right with a live demo + multiple-choice questions. Answer some, click Accept → spawn. A process is born and auto-selected.
4. Steer it: with the process selected, type into the steer panel and hit Enter (the spec's highest-value flow: click → type → Enter → effect).
5. Fork / pause / kill from the card. Click ▦ to get a QR code that pairs a phone (/m/:token) as a mic+chat steering device.

Or drive it hands-free against a running server:

bun run seed           # plays a scripted room conversation

What's implemented (spec §6.1 V0)

• Always-on suggestion engine → idea bubbles with demo + clarifying questions, TTL, merge-in-place, model-initiated cadence (§5.5)
• Process Manager: create · modify · kill · fork/spawn · merge · import/export · pause · resume · switch_mode (§5.1)
• Per-process session loop with pre/post hooks incl. pre-spawn resource check & pre-kill context archive (§5.3, C6)
• Input router with the two-channel rule: ambient→suggest, selected→steer (§5.4, C2/C3)
• Pro click→type→Enter UI on a spatial board with live per-process visualizers that auto-pick by artifact kind (§5.6, G5)
• Mobile QR pairing → mic+chat device feeding a process input queue (§5.7)
• Tunable knobs: bubbles/min, suggestion TTL, safe/dangerous, optimistic/explicit (§4)
• Swappable brain (mock ↔ Anthropic) and agent/model/container metadata fields (C7)

See ARCHITECTURE.md for the file-by-file map to the spec and the integration seams for Smithers / cue / plugin-local-inference. Eliza support is planned later.

Layout

src/core/      domain: meta-session, process-manager, process, suggestion-engine,
               input-router, hooks, brain (mock | anthropic)
src/server/    Bun.serve — REST commands + /ws event stream + mobile pairing
src/web/       Pro spatial UI (index/app/style) + mobile device (mobile.*)
src/scripts/   seed-demo — scripted room conversation