Cleaner conversation
Self-repeat rules make repeated content sound intentional instead of scripted or broken.
Lumenarc
GitHub-sourced update · 24 Apr–8 May 2026
Management briefing
From fragile live demo flow to instrumented, adaptive demo platform.
Two weeks of landmark changes across studio-monorepo and sidecar_supervisor, translated for technically savvy non-developers.
Activity snapshot
GitHub history shows a high-shipping reliability and pilot-readiness push.
| Repository | Merged PRs | Created PRs | Created issues | Releases |
|---|---|---|---|---|
| studio-monorepo | 24 | 33 | 38 | 14 |
| sidecar_supervisor | 5 | 6 | 11 | 16 |
Release movement
Studio: v0.115.0 → v0.128.0
Sidecar: v1.14.0 → v1.23.2
Reading the signal
Most changes cluster around live demo quality, prequalification handoff, observability, supervisor control, and agent-assisted deployment.
Theme 1
Prequalification now drives personalized demos.
What changed
- Prequalification now hands structured prospect context to the demo agent.
- The handoff includes a recommended approach, objections, focus areas, and per-product demo decisions.
- Demo prompts can use prospect identity and company context directly.
- Playground panels now expose prequal data, summaries, notes, and demo decisions.
Why it matters
The demo no longer starts cold. The product loop is now: qualify → understand → personalize → demo.
Sources: studio PRs #922, #928, #926 · releases v0.126.0–v0.128.0 · issues #905, #906
Contributors: @Lamirarc, @martinhecher
Theme 2
Live demo reliability improved where users actually feel it.
Better turn-taking
Non-rhetorical question detection helps the system stop talking and wait for a real answer.
Faster recovery
Force-immediate steer-back lets corrective guidance land quickly instead of sitting in a queue.
Audio fixes
Goodbye TTS now finishes; Gemini duplication and cut-off suspects were mitigated.
Rollout controls
Risky narrator behaviors are feature-flagged and visible in the playground.
Pilot impact
Fewer awkward moments: silence, repetition, clipped endings, and wrong-topic continuation.
Sources: studio PRs #876, #892, #914, #916, #920, #925 · releases v0.116.0, v0.120.0, v0.124.0–v0.126.0
Contributors: @martinhecher, @Lamirarc
Theme 3
The supervisor became state-aware, replayable, and more privacy-conscious.
Landmark changes
- Deterministic state snapshots are injected into supervisor turns so it knows the current agenda state.
- Agenda nodes now include goals and completion signals, making navigation decisions less ambiguous.
- Supervisor decisions are archived in machine-readable session files and can be replayed against real conversations.
- Playground diagnostics expose supervisor health, steering, timing, and rejection signals.
Important sidecar change: Cerebras history control
The Cerebras proxy can now run in a history-exclusion mode that avoids sending the full accumulated conversation history to Cerebras. Operators can toggle this live through a proxy /admin endpoint, and the playground exposes it as a Supervisor Chat History control.
This enables A/B testing whether deterministic state snapshots are sufficient, while reducing provider-side context exposure and long-context dependence.
Sources: sidecar PRs #104, #132, #133, #137, #138 · studio PRs #893, #896 · sidecar release v1.18.0
Contributors: @martinhecher, @melshakobyann
Theme 4
Observability became a first-class capability.
What we can now see
- Prequalification runtime and state-machine decisions.
- Transcript content and selected model metadata.
- Playground events tied to user and team identity.
- MCP tool usage, duration, success/failure, and session context.
- Conversation-item, response-trigger, playback cleanup, and interruption paths.
Management value
The team can investigate live sessions from evidence instead of guesses: who ran it, what was said, what the model did, which state changed, and where recovery failed.
Sources: studio PRs #865, #887, #897, #903, #932, #933 · releases v0.116.0, v0.121.0, v0.123.0
Contributors: @martinhecher, @Lamirarc, @vrezh-lumenarc
Theme 5
Agent-driven demo creation is moving toward end-to-end.
1Generate
Create demo agenda/project from input.
2Polish
Improve agenda text while preserving structure.
3Deploy
Create a callable voice-agent demo URL.
4Observe
Track MCP tool usage in Axiom.
Why it matters
The workflow is shifting from “manually assemble a demo” toward “generate a demo, deploy it, and try it” — a better foundation for internal users and future customer self-service.
Sources: studio PRs #902, #903, #877 · releases v0.118.0, v0.122.0, v0.123.0 · issues #904, #909, #910, #913
Contributors: @vrezh-lumenarc, @martinhecher
Currently building
The next wave targets pilot-readiness and runtime architecture.
Whisper transcript authority
Independent OpenAI Realtime Whisper transcript path for Gemini audio.
Issues/PRs: #936, #937 · @martinhecherServer-side prequalification
Move state machine and sensitive context out of the browser.
Issue: #927 · @LamirarcDeployed-agent phase editing
Edit/import/export prequal phase config without redeploying agents.
Issue: #930 · @LamirarcVertex Gemini in playground
Expand provider coverage for Smart Sequence sessions.
Issue/PR: #923, #924 · @martinhecherSupervisor navigation memory
One-shot agenda summary helps “go back” and cross-section transitions.
Sidecar #144, #145 · @martinhecherPrompt/context compression
Reduce instruction decay and unbounded context growth.
Sidecar #134, #135 · @martinhecherStrategic takeaway
We are building the operating system around live AI sales demos.
SmarterDemos adapt to prequalification context.
More reliableNarrator and supervisor behavior are actively guarded.
More observableSessions are debuggable through telemetry and replay.
More scalableDemo generation and deployment are becoming agent-assisted.
More controlledProvider context exposure can be reduced via proxy history mode.
The work has shifted from isolated prompt tweaks to platform-level control, measurement, and pilot-readiness.