Face & Brain UI

frontend/ is a Vite + React + TypeScript app that exposes the runtime as named rooms instead of hiding behavior in a chat transcript. It reads the telemetry stream from backend/ (see Backend & Telemetry); in the packaged desktop app it runs in a native pywebview window over an in-process bridge with no open port.

The rooms (frontend/src/pages/)

Room	What it shows
Face	The person-facing surface: conversation and expressions composed through the expression membrane
Brain	The live internals: workspace/thought stream, control-signal rings, demands, attention, goals, internal state
Cognition	Function-selection stats, per-function EMAs, thinking cost (tokens/cache), deliberation activity
Memory	The memory inspector: working memory, long-term retrievals, consolidation activity
Learning	Behavior changes and belief revisions as before→after→because diffs
Life	The existence view: runtime-lifetime phase, restoration, and run history
Timeline	The event timeline across the run
Watch	A passive observation screen built around the live thought line
Voice	The transcript: only Orrin's own utterances, verbatim, newest first (see below)
Settings	Provider/API-key management (keys go to the OS keychain), RAM budget, runtime options

Two different lines: the status label and the transcript

These are easy to confuse and are not the same thing.

The Watch/Face thought line is a UI-authored status label. frontend/src/lib/thoughts.ts (sibling to lexicon.ts) maps the reported active_fn to a phrase a developer wrote — generate_intrinsic_goals → "surfacing self-set goals from drives", and a companion-register rewrite of the same act that only /orrin renders. Nothing in the runtime ever said it. It follows the same hard rule as lexicon.ts: translate the chrome, never the mind.

The Voice room (/voice) is the other half of that rule — the mind, untranslated. It renders the voice telemetry field, fed by brain/cognition/voice.py, which only accepts utterances from call sites that already hold first-person content:

Kind	Source
felt	the global-workspace winner when it is a felt state ("a strong sense of being stuck")
prediction	an introspection miss — felt yes, behaved no
intent	the goal he just committed to, in its own words
closeout	a question closed by epistemic close-out (brain/cognition/epistemic_closeout.py), answered or not
speech	anything composed and delivered through the expression membrane
final	last words, written once at the end of a life

voice.py never composes — it veils (the felt lexicon membrane, so a raw signal identifier never reaches the transcript), drops anything still carrying backend markers rather than cleaning it up, suppresses consecutive repeats, and records. The stitched narrative composer is deliberately not a source: its prose is working-memory strings read back as content (the self-echo family), so it would broadcast that bug as speech.

The transcript keeps its own append-only file (voice_transcript.jsonl), so private_thoughts.txt — which interleaves his lines with the runtime's third-person narration at roughly one to fifty — is no longer the only place his own words exist.

Panels

frontend/src/components/brain/ holds the panel library: control-signal rings, demands, attention, goals + goal health, internal state, memory inspector, learning, language, predictions, relationships, self-model, symbolic-model, tensions, resource signs, live console, and the cognitive-sphere visualization.

Type safety across the wire

The telemetry schema is defined once in backend/server/schema.py; generate_telemetry_ts.py generates the frontend's TypeScript types from it (make telemetry-types), so the two sides cannot silently drift.

Modes

• Native window (default): pywebview over the built frontend/dist, no browser, no port.
• Developer (ORRIN_UI_DEV=1): Vite dev server on :5173 with hot reload, backend on :8800.
• Headless (ORRIN_UI=0): no UI at all.