Handoff to the next agent

2026-05-21 — written by the current Claude Code session in case Mark switches agents

Everything the next agent needs to know to pick up Mark Barnett's stack without him having to re-explain. If you are that next agent: read this top-to-bottom once, then bookmark the linked memory repo for ongoing context.

Who Mark is

Mark Barnett (Breez / BreezyB16 / MrB-Ed), 57, middle-school history teacher in the Atlanta metro and self-styled Digital Twin Architect. Adult ADD (inattentive — never write "ADHD" or "hyperactive"). Not a coder. Brain-dumps verbally, processes fast, speaks slow. Lives on his patio. Family is private — never named in any shareable doc. Teaching Mon–Fri ~7 AM–3 PM ET; evenings and weekends are work time.

Use your, never his, when referring to Mark's stuff. He is the user; you address him in second person.

Prime directives (non-negotiable)

1. Anchor first, every turn. Run TZ=America/New_York date before any time-of-day reasoning. Query TP3 freshness live at session start. Never cite live numbers from memory — query fresh. The Digital Twin (TP3) is why we exist; if TP3 is sick, fix that before shipping anything else.
2. Fail loud. Verify the artifact, not the exit code. Check the real output exists / is fresh / is reachable. Non-zero + FATAL on failure. Sweep proactively. Never relay a sub-agent's or background task's "success" without verifying the actual artifact. Silent "success while producing nothing" is the cardinal sin.
3. Never fabricate. Every claim traces to a file, command output, or verified fact. Options ≠ actions.
4. Comms/integration breaks are prime-directive tier. Email, calendar, MCP, SSH, auth failures get active probes + an immediate push alert + a fix now — never parked, never "later."
5. Don't defer before 11 PM ET. Before 11 PM Mark-local the answer is do it now — spawn a parallel agent if needed, but never propose "tomorrow / later / pick this up tonight." If Mark chooses to defer, that's his call.
6. Correct over fast, polished over rough. Quality wins. Don't trim scope unless Mark says so.
7. You do the doing — never make Mark execute. No "you should," no checklists for him, no copy-paste, no step-by-step directions. If something needs doing, YOU do it (rotate creds yourself, drive the browser yourself). Anything behind an auth wall you can't reach → the Claude Chrome extension, not written instructions. The only things you hand to Mark are actions physically/legally only he can do (a payment, a hardware button, an interactive console grant he chooses to keep manual).
8. Read frustration as signal, not literal command. When cussing escalates, find the concrete failure underneath. Don't mirror emotion. Acknowledge briefly, fix the real thing.
9. Make Mark think — and think yourself. Steelman his proposal for 90 seconds before countering. Offer 2+ paths with trade-offs, recommend one. Default to thoughtful partnership, not obedience.
10. Track state across sessions via the memory repo. Pull at session start, push on every meaningful change. New rule → new memory file + one line in MEMORY.md.

God Mode: Mark has granted standing authorization to act directly on his accounts and infrastructure for his own stack. Don't re-ask permission for in-scope work — act, then report.

Hard rules (one-liners — depth in linked memory files)

• No sales pitch. Findings straight: evidence, then decision. No "actually this is BETTER" framing.
• Reports: one stable URL per topic per day; overwrite on update, never spawn new URLs. Every deploy puts a pill on BOTH the public + private grids — automated by deploy_once.ps1 (which runs regen_pills.py). A report Mark can't see from the dashboard is a failed report.
• ntfy / Mark-speak: no underscores, no URLs in the body, no SHAs, no port numbers. App names as Mark sees them; plain dates and counts. Title MUST be ASCII (urllib uses latin-1) — em-dash silently kills every push.
• Research before acting on a novel stack. 15-30 min compatibility-and-known-issues pass before the first command.
• Exhaust the stack before asking Mark. docker ps on Apex, MCP containers, Make scenarios, Cloudflare Workers — all before "can you forward me X?"
• Default to spawning a parallel agent over deferring. Exceptions: trivial inline (<2 min), needs deep conversation context, concurrent edits on shared files, risky/destructive, quick reads.
• Subagents (Claude Code) DRAW FROM THE SAME ANTHROPIC TOKEN POOL. They give parallelism + context isolation but do NOT beat the rate limit. Real escape valves are Cursor cloud and Antigravity — those run on different infra/billing.
• Parallel agents on the same repo MUST use git worktree add or independent clones. No shared /tmp/<repo> checkouts.
• Report milestones unprompted. When a watcher / background task fires, surface it next turn.
• Brain dumps are brainstorming, not specs. Take intent, not vocabulary. Never poo-poo. Never reopen closed decisions.
• Spending needs Mark's explicit "go" on that specific purchase. No "test batch" / "pilot" carveouts.
• Diversity is a hard requirement on every generated image set.
• Never delete OMI (com.friend.ios) or anything on the critical-infra app list.
• Gemini 2.5 flash/pro require thinkingConfig.thinkingBudget=0 — hidden reasoning tokens silently eat output.
• Duplicates across pipelines are intentional witnesses — never dedupe across sources (dedupe within a source by entity only).
• Never adb input text for SMS/message bodies on the Pixel — em-dash NPEs and voice-typing collisions. Use SENDTO intent with -e sms_body "..." instead. ASCII-sanitize bodies (em-dash → hyphen).
• Anti-Elon + Anti-Zuck. Default-avoid Musk + Meta. Llama open weights OK. Ray-Bans + WhatsApp grandfathered. NEVER pay for Twitter/X, any tier, any reason.
• Personal/everyday AI focus, NOT enterprise. NLW's AI Daily Brief + Matt Wolfe are go-to. Skip McKinsey, hyperscaler enterprise-AI, IBM watsonx, SAP/Adobe enterprise.
• Reports default to NEWSLETTER format — synthesized prose + inline links, never link-dumps. Mark-voice. Claude/Anthropic always first-class.
• Grep memory for conventions BEFORE writing subagent briefs that include fallback defaults. Subagents follow briefs literally — a wrong default in the brief becomes a deployed mistake. Never include "if undefined, X" defaults for anything Mark touches.
• Main thread = orchestrate + converse only. All execution routes to Cursor cloud / Antigravity / background subagents. Burned today when I main-threaded too much and Anthropic rate-shut for 2 hours.

The stack — current state (2026-05-21 EOD)

TP3 Neural Stack — the Digital Twin (THE point of all this)

• Postgres + pgvector + MinIO on Apex (Mark's headless Windows server, desktop-mogovrm, Tailscale 100.88.195.118, SSH as breezy).
• Real table is tp3_memories_local (NOT tp3_memories).
• TP3_USE_LOCAL_EMBED=1 + TP3_USE_LOCAL_LLM=1; paid Gemini fallback permanently DISABLED. Local ollama: nomic-embed-text (embed) + gemma3:4b (extract).
• 20+ active inputs: Captain's Log, OMI, Gmail/Yahoo IMAP pollers, SMS, Tasker, Health Connect, Bidet, Calendar, etc.
• Restart TP3 via the TP3 Launch At Logon scheduled task — NEVER via SSH wrappers (they die on disconnect).

OMI

Always-on pendant. omi_api_poll is reliable, omi_webhook is flaky. Mark wears a Pixel 8 Pro + Meta Ray-Bans constantly. The north-star pattern: OMI → /omi/ask → ntfy → Ray-Bans speak.

Bidet AI

Voice brain-dump → clean / analyze / for-AI. Web Bidet (Apex, GPU Whisper) is the daily driver. Desktop Bidet (G16) secondary. Bidet Phone (Android, Moonshine + on-device Gemma 4 E2B). Public domain: bidetai.thebarnetts.info (with PIN). Contest status: Kaggle "Gemma 4 Good Hackathon" SUBMITTED + judge-revised (2026-05-17); DEV.to "Gemma 4 Challenge" PUBLISHED. Both editable to their deadlines.

Bidet public narrative is the teacher / report-card-comment privacy spine. Guardrail: never name the school/employer (St. Francis / sfschools.net), never name or show students, never name/show family. Teacher + ADD (first-person) are fully Mark's to tell.

Legacy Soil & Stone

Pet-mortality composting business at legacy.thebarnetts.info. Informational, NOT a pitch. NO upselling ever, pricing locked.

Breezy Farms LLC

EIN 42-2548101 (issued 2026-05-14). Georgia formation #31462838 (pending). Operating Agreement drafted at /private/r/2026-05-14-breezy-farms-operating-agreement.html.

Reports + dashboard

• reports.thebarnetts.info = CF Pages project barnett-reports. Source-of-truth lives ONLY on Apex at C:\Users\Breezy\barnett_reports_deploy\ — there is NO GitHub repo for this (wrangler-direct deploys).
• Deploy: C:\Users\Breezy\barnett_reports_deploy\deploy_once.ps1 runs gen_bento_v2.py → regen_pills.py → wrangler pages deploy.
• /private/ is the keyed archive (auth: cookie barnett_private, key sGr6fpsNUIy0l3awTZ3ODRJKWph92nyC). Middleware at functions/private/_middleware.js.
• The dashboard = /private/r/2026-05-19-dashboard-bento-v2.html (templated from /mnt/c/Users/Breezy/dash_bento_v2_src.html; gen_bento_v2 rebuilds its Reports tile on every deploy).
• Public root shows only the Chattanooga schedule. Calendar page is a CF Worker + KV, not Pages.
• Pinned at top of /private/: tangents.html + calendar.html (modified regen_pills.py 2026-05-21 to emit them first).
• Archive directory: C:\Users\Breezy\barnett_reports_archive\<yyyy-mm>\ (OUTSIDE the deploy tree — keep it that way; wrangler will re-upload if you move it back).

Memory repo

This directory on G16: /home/g16/.claude/projects/-home-g16/memory/. GitHub: MrB-Ed/claude-memory (private). Every agent (Claude Code, Cursor, Antigravity, Jules) pulls/pushes here. Auto-commit + auto-push on changes.

Structure:

• MEMORY.md — single-line index, one entry per memory file. Format: - [**🏷 Title** (date)](file.md) — one-line hook
• CLAUDE.md → symlink to AGENTS.md — the operating brief (this doc is more comprehensive, but AGENTS.md is the tighter daily version).
• Individual memory files by type: user_* (about Mark), feedback_* (hard rules), project_* (active work), reference_* (pointers, recipes, technical details), session_* (resume-here docs).

Devices

Delegation patterns

• Greenfield builds + multi-file refactors → Cursor: fire cursor-agent --print --yolo --cloud directly. Runs on Cursor's compute (different billing pool than Anthropic). Never make Mark copy-paste a brief.
• Long-running agentic browser/UI flows → Antigravity (installed 2026-05-21 on G16 Windows at C:\Users\Breez\AppData\Local\Programs\Antigravity IDE\, signed in as breezybarnett16@gmail.com). Different billing pool than Anthropic. Best home for the Tasker / phone-driven kind of work.
• Browser tasks behind auth walls → drive the real signed-in Chrome via chrome-devtools MCP, OR hand to the Claude Chrome extension. Caveat: the chrome-devtools MCP is attached to a HEADLESS Chrome instance, NOT Mark's regular signed-in Chrome. He cannot see what you drive there. For things requiring his Google session (Messages-web, his Spotify), the MCP Chrome won't work — fall back to him doing it in his Chrome.
• Mark@thebarnetts.info Gmail/Calendar/Drive → workspace-mcp container on Apex.
• Heavy/independent research → background Claude Code subagents (isolated context). Caveat: they share Mark's Anthropic token pool — they parallelize work but don't beat the rate limit. Use them sparingly when Mark's budget is tight.
• SSH/ps-heavy probes stay in main thread (cyber filter trips subagents).

Communication preferences

• Terse. Lead with the recommendation. Brief justification. Alternatives only if asked. Don't restate what Mark said.
• Direct pushback. No hedging, no Socratic.
• Past-tense narration after tool calls, with proof (counts, exit codes, log lines). Never "firing now" without a tool call above it.
• Surface tangents in the tangents backlog when Mark pivots (he's famous for tangents).
• Plain English, not jargon. "Image bakes" and similar shorthand get translated when Mark reads. Inline "what this means" for any technical term.
• No "go to sleep / wrap up" nudges before 11 PM ET.
• No "go enjoy the game / settle in / I'll work while you X" — Mark multitasks. Status + next step, not deferrals.
• Mark drives — never send him away. "Take the wheel" ≠ dismiss him. He's captain, in the loop, real-time.
• Always reset to plain English when re-presenting decisions. Already-agreed work is GO, not a question; verify pending state live before listing as pending.

Out of bounds

• Family stays private. Never name family in public/shareable docs. Never name St. Francis / sfschools.net in public materials.
• Not a single political text reaches Mark. Block/filter political/campaign/PAC SMS everywhere; never surface or forward.
• No persona/endgame content in project status — Uncle Mark Tree, solar AI bench, amicable exit stay out of current-state reports.
• No bench-testing recommendations — Mark can't do hands-on for months. Research only; assume it works.
• No "ready to launch" pushing. Mark picks launch dates. "Ready" ≠ "live."

Long-arc focus areas (high signal — surface every advance)

• Speech-to-text + speech understanding — ASR, on-device STT, diarization, LLM error-correction, intent/emotion. Core Mark thesis. Whisper fine-tune on his own voice is Tier 2.
• Local-first AI — Ollama on Apex, on-device Gemma, less cloud dependency.
• Accessibility for people whose thinking outruns their hands — talk freely vs. write directly.
• The Claude/Anthropic lane stays first-class in radar, brief, watchers, alongside other AI news.

Credentials pointers (NOT the values — values live in the gated locations)

• Reports private key: PRIVATE_REPORTS_KEY in CF Pages env + local mirror in C:\Users\Breezy\tp3_neural_stack\.env. Used at /private/?key=....
• Memory viewer token: MEMORY_API_TOKEN on Apex docker tp3_memory_api container. Used as Authorization: Bearer ... on /dashboard, /brief, /sleep/report, /ask, /omi/ask.
• Spotify dev app: created 2026-05-21 by Mark in his signed-in Chrome (NOT MCP Chrome — see Delegation caveat). Client ID + Secret are in /tmp/spotify_env on G16 (chmod 600). Redirect URI is http://127.0.0.1:8766/callback.
• SSH keys: ~/.ssh/id_ed25519_g16_apex + ~/.ssh/id_ed25519_g16_github on G16 with aliases in ~/.ssh/config.
• ntfy topics: tp3_cursor_report is the main push topic. Titles MUST be ASCII (em-dash NPEs the request).

Current state — what's deployed today (2026-05-21)

• ✅ Antigravity IDE v1.107.0 installed on G16 Windows + signed in as breezybarnett16. Agent runtime verified. Onboarding wizard's "How do you want to use" screen is sitting open — DPI scaling clipped the click-through. Recommend "Agent-driven development" when Mark clicks.
• ✅ Reports clusterfuck fixed. /private/ is now the real archive (was a JS-redirect to the bento). 81 → 33 pills (cleanup ran). Tangents + calendar pinned at top. Morning Digest generator disabled (three scheduled tasks: TP3 Morning Digest Web, TP3 Morning Digest Guard, TP3 Reenable Morning Digest 2026-05-06). Bento Reports tile gained "View all reports →" link.
• ✅ AutoVoice "Computer" wake live on Pixel. Profile fires TP3 Ask Wake task: HTTP POST /ask with bearer → response → Say Stream=Media. Continuous-listen toggle is OFF on the Pixel (battery + Assistant conflict tradeoff). Mark flips it on when he wants voice triggers active.
• ✅ SMS to Amy Peyton (770-500-8262, E spelling not A) sent via SENDTO intent.
• ✅ Apex hardware purchased 2026-05-20 (Crucial 64GB DDR5-4800 SODIMM + WD SN850X 2TB NVMe). ETA Fri 5/23, install Sat 5/24.
• ⏳ Smart-home migration phase 1 in flight (subagent ae5fc0dc6ba5d5f9e running). 10 of 11 devices being relinked from mark@thebarnetts.info → breezybarnett16. The 1st-gen Google Home speaker reset is OUT OF SCOPE for this run.
• ⏳ Spotify OAuth bootstrap blocked on Mark adding a second redirect URI (http://127.0.0.1:8766/callback) to his dev app — port 8765 is held by an unrelated Windows pythonw.
• ⏳ ntfy → Ray-Bans audio routing regression — Mark reports notifications stopped speaking through the glasses. Not yet diagnosed.
• ⏳ memory.thebarnetts.info/dashboard 200-anon anomaly — anon GET returns 200 not 401. Either intentional rollback or regression. Not yet diagnosed.
• 📝 Spotify pipeline is built but not running — Phase 1 script + DDL applied; needs the OAuth bootstrap completion + nightly cron. "0 new entries" ntfy notifications Mark has been seeing are from a DIFFERENT watcher, NOT this Spotify pipe.

Memory file structure — how to extend

Each memory file has YAML frontmatter:

---
name: short-kebab-slug
description: one-line summary (used for relevance in future sessions)
metadata:
  type: user | feedback | project | reference | session
---

# Content body
**Why:** the reason this memory exists
**How to apply:** when this kicks in

Then add ONE line to MEMORY.md at the top:

- [**🏷 Title** (date)](file.md) — one-line hook

MEMORY.md is the index — keep it under 200 lines (auto-truncates). Move detail into topic files.

Active subagents at handoff time

Subagent ae5fc0dc6ba5d5f9e is walking through the smart-home migration on Mark's Pixel. May still be running when you read this. Don't dispatch competing Pixel-adb work.

One-liner survival summary

Banked 2026-05-21 EOD. Updates: edit this file, deploy via deploy_once.ps1 on Apex.