Session Handoff — Brain Jarvis V4: unify the brain (voice fix + smart recall wired in)

Where it started

Sondra saw the RoboNuggets "RUBRIC Second Brain" video (youtu.be/VoKiKvgpk78, Jay G) — a layered, voice-driven galaxy of a markdown vault — and wanted her brain to work like that. Key finding: RUBRIC is the same concept as her already-installed Brain Jarvis / Juno (Zubair's Brain Studio). She does NOT want a rebuild — she wants to COMBINE what she owns: the Jarvis galaxy (visual+voice) + this week's scripts/brain/ recall+meaning-search + the Brain Migration corpus (577 old conversations) + a real voice. Constraint: no big Fable spend (she's ~88% Fable usage), and she's non-technical — explain plainly, keep chat readable (she cannot read when AskUserQuestion popups cover the text — do NOT use that tool with her).

Decisions locked + what shipped

• No rebuild — extend Brain Jarvis. These are surgical edits to existing Python/HTML, done by Opus/Claude Code, NOT Fable. Full plan written.
• A — Voice: SHIPPED. She picked Daniel (ElevenLabs British broadcaster, voice_id onwK4e9ZLuTAKqWW03F9). Wired into C:\Users\srver\.claude\skills\brain-jarvis\config.json (elevenlabs.api_key from master .env + voice_id + model_id eleven_turbo_v2_5). Backup: config.json.pre-daniel.bak. Her ElevenLabs account = free tier, 10,000 chars/mo (~10 min speech), 0 used — enough for demos/light use, not heavy daily use. Note: the old memory saying ELEVENLABS_API_KEY is empty is STALE — the key IS set in .env (sk_eb866...).
• B — Smart recall wired into the galaxy: SHIPPED. server.py now calls scripts/brain/brain_recall.py --json as its finder (keyword+meaning-search, cites the exact section) and lights the exact matched galaxy node, with a safe fallback to the old keyword retrieve() on any failure. Backup: server.py.pre-recall.bak. Changes: added sys import (line 23); new helpers _node_for_source() + recall_retrieve() (~lines 129-176, AIOS root config-overridable via "aios_root"); chat() call site (~line 187) tries recall first, skips for chit-chat/synth, falls back on None. Recall --json keys used: source, answer, section, score. Tested via scripts/brain/test_jarvis_recall.py: "what is Stars pricing" → node _idx=59 stars-community (real pricing table); off-wiki query → clean fallback; empty/missing-script → None → fallback.
• Voice preview clips generated at c:\Users\srver\AIOS-Workspace\aios-starter-kit\PROJECTS\Brain Jarvis\voice-previews\ (1-George, 2-Daniel[chosen], 3-Alice, 4-Lily .mp3).

Key files for next session

• Plan file (READ FIRST): c:\Users\srver\AIOS-Workspace\aios-starter-kit\PROJECTS\Brain Jarvis\PLAN-v4-unified-brain-2026-07-08.md — the 4-workstream plan (A voice / B recall / C layered multi-view / D corpus), order, gap check, definition of done.
• C:\Users\srver\.claude\skills\brain-jarvis\server.py — the edited server (recall wiring); backup server.py.pre-recall.bak.
• C:\Users\srver\.claude\skills\brain-jarvis\config.json — voice config; backup config.json.pre-daniel.bak.
• C:\Users\srver\.claude\skills\brain-jarvis\build.py — galaxy generator; has an AIOS "smart" grouping (group_aios/AIOS_GROUPS) that is the seed for the C "Layers" view; README notes --generic is currently used (shows all, but flat).
• c:\Users\srver\AIOS-Workspace\aios-starter-kit\scripts\brain\brain_recall.py + README.md — the deterministic recall+meaning-search now powering Jarvis. Test: scripts\brain\test_jarvis_recall.py.
• c:\Users\srver\AIOS-Workspace\aios-starter-kit\PROJECTS\Brain Migration\INDEX.md — the 577-conversation corpus (workstream D).
• c:\Users\srver\AIOS-Workspace\aios-starter-kit\PROJECTS\Brain Jarvis\README.md — how the skill is installed/run (localhost:4719, Desktop Start Brain Jarvis.bat, --generic build note).
• Memory files touched: none.

Running state

• Background processes: none (the recall-wiring build agent has finished).
• Dev servers / ports: none started this session. Brain Jarvis is NOT running — Sondra must (re)start it (Desktop Start Brain Jarvis.bat → localhost:4719) to load the new voice + recall.
• Open worktrees / branches: none.

Verification — how to confirm things still work

• Restart the brain: close the "Start Brain Jarvis" window, double-click it again; open localhost:4719 in Chrome, hard-refresh Ctrl+Shift+R.
• Voice: ask anything → it should speak in Daniel's British ElevenLabs voice (not the old robotic browser voice).
• Smart recall: ask "what is Stars pricing" → answer = $33/mo · $197/yr AND the stars-community node flies-to/lights up.
• Recall test (no server needed): cd c:\Users\srver\AIOS-Workspace\aios-starter-kit && python -X utf8 scripts\brain\test_jarvis_recall.py.
• Off-wiki query (e.g. "how do I start an agency") → behaves exactly as before (keyword fallback), no crash.

Deferred + open questions

• Open: Rename Juno → Jarvis? She picked a MALE voice but the assistant is still named "Juno" (female). Minor but touches the wake-word regex + HUD labels in viewer/3d.html + PERSONA in server.py (pre-Juno backups exist: server.py.pre-juno.bak, viewer/3d.html.pre-juno.bak). Not done — awaiting her go.
• Deferred: C — layered multi-view (Applications/Routines/Memory/Skills rings, CLAUDE.md center, Rings/Force toggle in 2D+3D, click-node-opens-file). Biggest work; per plan §5, after A+B. Set a /goal for load <10s (~1,550 nodes).
• Deferred: D — fold in the 577-conversation Brain Migration corpus as a Memory/Archive ring (point build.py + brain_index.py at PROJECTS/Brain Migration/_projects-rebuilt/).
• Open: galaxy is ~977 notes as of 2026-07-01, a few days stale (pre-07-05 reorg) — a build.py --generic rebuild refreshes it before demoing (README has the exact command).
• Note: recall covers only Sondra Brain/wiki/ (~90 files); the galaxy covers the full vault — expected; recall returns None (→ keyword fallback) for anything outside the wiki until D brings more content into recall's scope.

Pick up here

Have Sondra restart Brain Jarvis and verify Daniel's voice + the "what is Stars pricing" node-light; then, on her go, do the Juno→Jarvis rename, then build C (layered multi-view) and D (corpus) per the plan.