Skip to content

FE-1789: Experiment with end-to-end Brunch Voice steering - #9826

Closed
kostandinang wants to merge 3 commits into
mainfrom
kostandin/fe-1789-brunch-voice-steering
Closed

kostandinang wants to merge 3 commits into
mainfrom
kostandin/fe-1789-brunch-voice-steering

Conversation

@kostandinang

@kostandinang kostandinang commented Sep 24, 2026 •

Copy link
Copy Markdown
Contributor

🌟 What is the purpose of this PR?

Brunch Voice previously received Brunch replies but not enough confirmed context to guide a complete modeling session safely.

This experiment gives Voice the context needed to describe model progress, control simulation playback, and distinguish proposed experiments from completed results without inventing actions or outcomes.

🔗 Related links

🚫 Blocked by

  • Nothing.

🔍 What does this change?

  • Draft: send quiet Ledger coverage after each applied revision so Voice can describe what is settled, open, or missing.
  • Check: route explicit Play and Pause commands directly to the canvas and report playback state and token counts to Voice.
  • Experiment: tell Voice when an experiment is drafted but unrun, and speak a summary only after a person runs it successfully.
  • Extend the Live prompt so Voice does not claim unconfirmed model changes, playback actions, or experiment results.
  • Keep failures, cancellations, invalid drafts, and dismissals quiet.

Pre-Merge Checklist 🚀

🚢 Has this modified a publishable library?

  • No publishable library changes.

📜 Does this require a change to the docs?

  • Internal experimental behavior; no documentation change required.

🕸️ Does this require a change to the Turbo Graph?

  • No execution-graph changes.

⚠️ Known issues

  • The flow has unit coverage but has not been exercised against a real GPT-Live session.
  • Playback notes use frame indices because model time is unavailable.
  • Run remains a button action; Voice never starts an experiment.
  • Changing the model during playback is outside this experiment.

🐾 Next steps

Record the support-desk scenario with Live, tune the context notes, and decide whether spoken Run should be explored separately.

🛡 What tests cover this?

  • Focused tests cover Ledger notes, playback command routing and failures, experiment draft/result relay, and editor integration.
  • 574 focused tests pass; TypeScript and ESLint checks pass with two pre-existing warnings.

❓ How to test this?

  1. Start Brunch Voice with the support-desk scenario and confirm each applied model revision sends one quiet coverage note.
  2. Say “Play” and “Pause”. Confirm the canvas changes without a Brunch turn and Voice reports the action.
  3. Ask Brunch to draft an experiment. Confirm Voice describes it without claiming it ran.
  4. Press Run and confirm Voice summarizes only the completed result.

📹 Demo

The RFC contains an illustrative walkthrough. No real-session recording exists yet.

The Live bridge now derives a short note from the latest applied
mutate_workpiece revision in the conversation snapshot (which workpiece
sections hold claims, which are annotated as open, which are empty, how
many cross-cutting issues remain) and sends it once per revision through
session.thinking.append with no delegation. Quiet context lands before any
spoken answer from the same update and is re-offered once the Live channel
connects. The session prompt tells GPT-Live to use it only to describe
progress; Brunch stays the author of substantive questions.

Remove interview-coverage.ts, which read the retired sweep tool.

Co-authored-by: Amp <amp@ampcode.com>
@vercel

vercel Bot commented Sep 24, 2026 •

Copy link
Copy Markdown

The latest updates on your projects. Learn more about Vercel for GitHub.

Project Deployment Actions Updated
hash Ready Ready Preview Oct 5, 2026 8:43am UTC
hashdotdesign-tokens Ready Ready Preview Oct 5, 2026 8:43am UTC
petrinaut Ready Ready Preview Oct 5, 2026 8:43am UTC
petrinaut-docs Ready Ready Preview Oct 5, 2026 8:43am UTC

Request Review

…ted experiments to GPT-Live

Co-authored-by: Amp <amp@ampcode.com>
@kostandinang kostandinang self-assigned this Sep 24, 2026
@kostandinang kostandinang changed the title FE-1789: Steer Brunch Voice through Draft, Check and Experiment FE-1789: Guide Voice through Draft, Check, and Experiment Sep 25, 2026
@kostandinang kostandinang changed the title FE-1789: Guide Voice through Draft, Check, and Experiment FE-1789: Steer Brunch Voice through Draft, Check, and Experiment Sep 25, 2026
@kostandinang kostandinang changed the title FE-1789: Steer Brunch Voice through Draft, Check, and Experiment FE-1789: Experiment with end-to-end Brunch Voice steering Sep 28, 2026
@kostandinang
kostandinang added this pull request to stack #9861 September 28, 2026 18:43
@kostandinang
kostandinang removed this pull request from stack #9861 September 28, 2026 19:19
Co-authored-by: Cursor <cursoragent@cursor.com>
@kostandinang

Copy link
Copy Markdown
Contributor Author

Closing as an experiment that didn't progress.

Not validated: Ledger coverage notes, Play/Pause routed to the canvas, and the drafted-vs-completed experiment relay were never tried in a real GPT-Live session.

FE-1795 already stops Voice from claiming Brunch work that hasn't finished. The rest may be useful inspiration for future work.

This branch was successfully deployed

4 active deployments
Preview – petrinaut-docs — 88cb2980 Deployed Oct 5, 2026 by vercel[bot]
Preview – hash — 88cb2980 Deployed Oct 5, 2026 by vercel[bot]
Preview – hashdotdesign-tokens — 88cb2980 Deployed Oct 5, 2026 by vercel[bot]
Preview – petrinaut — 88cb2980 Deployed Oct 5, 2026 by vercel[bot]
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Development

Successfully merging this pull request may close these issues.

2 participants