FE-1828: Add interview length controls to Brunch

[kostandinang]

🌟 What is the purpose of this PR?

Add experimental Interview length controls to Brunch Chat and Voice, so people can choose how much questioning to do before wrapping up. Minutes are an estimate derived from replies remaining, not a timer.

🔗 Related links

• Tracking: FE-1828, following FE-1787.
• Shares the version-two contextual envelope with FE-1834: Teach Brunch voice agent custom words #9899 (FE-1834, Custom words); see Known issues.

🚫 Blocked by

• Run the support-desk-staffing persona at Quick, Standard and Thorough and verify observed question counts. Real-model adherence is not yet verified.

🔍 What does this change?

• Add a separate, default-disabled Interview length experiment under User settings → Labs → Use Brunch. Remember the chosen level separately; Standard is the initial selection when enabled.
• Put an icon-only picker at the leading edge of the composer and Voice dock, with five stops: Off, Quick (~5 min), Standard (~10 min), Thorough (~20 min), and Deep (no limit). Preserve the collapsed Voice dock height.
• Count completed canonical Brunch replies against text caps of 3 / 6 / 10, or Voice caps of 2 / 4 / 7. Confirmations and grouped questions each count once; tool-only messages, hidden messages and Live captions do not count, and the wrap-up summary does not spend a question. Earlier replies carry over when the level changes, including after reload.
• Show a question-derived estimate, final-question state and wrap-up state; add compact, accessible transcript notes when the level changes.
• Send the current allowance on every submission as opaque, host-owned submission context. The AI SDK transport gains a generic submissionContext option and a single version-two envelope that it bounds (plain JSON object, finite numbers, depth and size limits) but never interprets. Diagnostics-only bodies stay byte-identical on version one, and the transport's dependency rules are unchanged.
• The SDCPN plugin owns the allowance: interviewBudgetContextKey, interviewBudgetSchema, parseInterviewBudget and the instruction it mounts. The Brunch app reads the key from the current delivery; an absent or malformed allowance means Off. The allowance is no longer copied into creation-only initialData. The model is told the level, cap and questions remaining, never minutes; minutes are website UI only.
• At the cap, settle the latest answer and list stated facts, labelled assumptions and open items. Standard does not fill missing essentials with unconfirmed assumptions. Never invent operational facts, ranges or units. Deep offers pauses between topics instead of count-based closure.
• Pass the Voice level through a validated startup header and quiet session.thinking.append updates, with a non-interrupting warning if an update fails. Keep instructions.append reserved for existing failure redirects.
• Add generic Petrinaut slots: renderComposerStatus, renderSystemMessage, and inputMode in the composer control context. Petrinaut describes no interview-length behaviour.
• Preserve the legacy request and instruction path when Off or the experiment is disabled; remove obsolete sweep-stream formatting and its schema.

Pre-Merge Checklist 🚀

🚢 Has this modified a publishable library?

• Modifies an npm-publishable library and includes a patch changeset for @hashintel/petrinaut describing only the new generic slots. The Brunch packages are private.

📜 Does this require a change to the docs?

• The user guide is in the website's docs, apps/petrinaut-website/docs/interview-length.md, linked from the website README. Petrinaut's model-visible guide (libs/@hashintel/petrinaut/docs/) is unchanged, so the stock assistant never reads about this website Labs feature.

🕸️ Does this require a change to the Turbo Graph?

• Does not affect the execution graph.

⚠️ Known issues

• Caps are model instructions, not a hard runtime cutoff. Persona counts and paid live-Voice behaviour remain unverified; browser evidence uses fixture history.
• Level-change notes are local to the mounted session. Per-question records, durable note storage and disclosure changes are outside this phase.
• FE-1834: Teach Brunch voice agent custom words #9899 currently defines its own petrinaut-contextual-user-message:v2 with a words payload. Both cannot merge as written; FE-1834: Teach Brunch voice agent custom words #9899 should carry its words as submissionContext.words in this envelope instead of adding a transport option and validator.
• apps/brunch-agent's provider-accounting.test.ts ("two real processes cannot interleave ledger transactions") fails locally on this branch with or without these changes.

🐾 Next steps

• Complete persona validation before marking this draft ready for review.
• Align FE-1834: Teach Brunch voice agent custom words #9899 on the shared submission-context envelope.

🛡 What tests cover this?

Fresh local checks on the PR head:

Package	Checks
AI SDK transport	54 unit tests passed; build, lint and type check pass
SDCPN plugin	29 unit tests passed; build, lint and type check pass
Brunch agent core	69 unit tests passed; build, lint and type check pass
Brunch app	349 of 350 passed (one pre-existing failure, above); lint and type check pass
Petrinaut Core	ai.test.ts passed, stock prompt contract unchanged
Petrinaut	304 AI assistant panel tests passed, including docs content
Petrinaut website	867 budget, transport, Labs, Live and server Voice tests passed; type check passes

Tests cover submission-context bounds and version-one byte equality, Off request serialization and instruction equality, malformed allowances falling back to Off, mode-specific caps, canonical counting across wrap-up and reload, level changes, transcript projection, Labs preferences and quiet Voice updates. Browser checks cover picker interaction, keyboard focus, progress text, compact notes and the collapsed dock height.

❓ How to test this?

1. Start the existing local Brunch development setup with yarn dev:brunch.
2. Open User settings → Labs, enable Use Brunch, then Interview length. Open Chat and use the icon at the left of the composer.
3. Select each stop by name, rail or keyboard. Verify that the description follows the hovered stop and Escape returns focus to the trigger.
4. With fixture history or an approved model session, verify the estimate changes with replies, not elapsed time. Switch to a lower cap and confirm earlier replies still count.
5. Check the same control in Voice and confirm collapsing the conversation does not increase the dock height. With an approved Live session, change the level and verify it does not interrupt speech.
6. Select Off, then disable the Labs experiment. Verify Off leaves no estimate, disabling hides the control, and re-enabling restores the saved level.

📹 Demo

interview-length-demo.mp4

[vercel]

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

4 Skipped Deployments

Project	Deployment	Actions	Updated
hash	Ignored	Preview	Oct 7, 2026 5:37pm UTC
hashdotdesign-tokens	Ignored	Preview	Oct 7, 2026 5:37pm UTC
petrinaut	Skipped		Oct 7, 2026 5:37pm UTC
petrinaut-docs	Skipped		Oct 7, 2026 5:37pm UTC

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit b25edb6. Configure here.}