Tuesday, May 5, 2026 — Morning Digest

Hand-written the night before. Real numbers, real wins, real schedule. Not the recycled boilerplate from last week.

Yesterday in 90 seconds

Ollama is in Docker now. Native Windows Ollama 0.18 has an IPv6-only bind bug that broke every Bidet/TP3 ingest path. Solution: tp3_ollama container on tp3_internal_network with GPU passthrough (RTX 5060 Ti, 16 GB, all 35/35 layers offloaded). Five models live inside (gemma3:4b, gemma3:12b, qwen3:4b, nomic-embed-text, all-minilm). No re-download — bind-mounted C:\Users\Breezy\.ollama. Bidet flipped back to BIDET_LLM_BACKEND=local, Gemini key disabled.
The silent-fail antipattern is patched. The bug that ate ~12 hours of ingest overnight 5/3-5/4 and ~42 hours of Captain's Log emails 5/2-5/4 is dead. Before: when Ollama was down, _embed() returned None, the INSERT violated tp3_embedding NOT NULL, the bg-task wrapper swallowed the exception, endpoint returned ok:true, row was lost. After: zero-vector placeholder + metadata.needs_embed=true flag, plus a new POST /admin/backfill_embeddings endpoint to re-embed those rows when Ollama is healthy. Patches durable in the rebuilt tp3_memory_api:latest image — won't be lost on next recreate.
Contract Review report shipped. Full 24-clause analysis of the Saint Francis 2026-27 contract live at /r/contract-full-2026-05-04. Headline: +$1,739 base raise (3.00%, smallest in 3 years). Three clauses worth pushback before signing. Atlanta-private-school peer benchmarks. May 8 signing window.
OMI "Computer" trigger app submitted for review. Manifest live at memory.thebarnetts.info/.well-known/omi-tools.json. Routes wake-word queries from the pendant to /omi/ask on Apex (Anthropic-backed when key is set, Gemini Flash fallback). Bypasses OMI's gpt-5.1 chat tab — see the OMI quality finding below for why that matters.
Inboxes — actually cleaned, not "triaged."
- Yahoo (barnett.markd@yahoo.com): 180 → 116 inbox, 4 unread.
- mark@thebarnetts.info: 3,284 → 2,041, 10 unread.
- breezybarnett16@gmail.com: 1,927 → 872, then an aggressive pass to 66 unread.
- School (mbarnett@sfschools.net): 21 archived.
Tasker URLs moved to Tailscale. No more Cloudflare 530s on phone-side POSTs to /ingest, /sms, /phone. Direct http://100.88.195.118:8945 from the pendant + Tasker. Faster, fewer hops, no edge-routing failures.
SMS Received → Ray-Bans speak. New Tasker profile pushes incoming-SMS body into the existing ntfy → Ray-Bans speak pipeline. Texts from Mom + Kim now read aloud through the glasses without unlocking the phone.
Reports page — three new featured pills. Contract Review (#1), Weekend Retro, Tangents Backlog — all on reports.thebarnetts.info already.

TP3 brain status — live, queried just now

Total rows	668,281
Last 24 hours	794
Last 1 hour	66
Newest row age	118 seconds (~2 min)
Freshness	ok

Sources active in the last 24h:

email_yahoo — 329 (Yahoo IMAP poller, 15-min cadence, working)
phone_notification — 255 (Tasker on Pixel 8 Pro, post-Tailscale switchover)
omi — 119 (pendant raw transcripts via webhook)
omi_api_poll — 42 (structured summaries + action items)
phone_location — 26 (geofence + periodic location)
sms_phone — 16
calendar — 3
diagnostic_postpatch_v2 — 2 (the Ollama-fix verification probes)
tasker_switchover_test — 1
email_ingest — 1

The catch-up after yesterday's 12-hour Ollama outage is happening — 794 over 24h is solid (the full-week average is ~700/day) and the 66 in the last hour confirms the patched /ingest + Yahoo poller are pulling the backlog through cleanly.

Tuesday's calendar

Pulled live from Mark's primary, Mom's fran@thebarnetts.info, and the school official calendar.

Mark — school day:

1:00–1:55 PM ET — 8th Lit Exam. The actual graded exam slot. Standard 7th-period block.

(No Sutter coverage on Tuesday's calendar, no parking-duty entry, no mid-day exam-prep block returned by the workspace fetch. If anything got added since last sync, the live dashboard calendar card will show it.)

Mom (Fran):

8:30 AM – 4:00 PM — Movie event (likely the senior-center daytime outing she does Tuesdays).
7:00 – 8:00 PM — Support Group.

School (St. Francis official):

8:30 – 9:30 AM — MS New Parent Coffee. Not on Mark's plate; informational.

Open threads still moving

OMI Computer app — review pending. App submitted with manifest, App Home URL, scopes, icon. Once approved, the "Computer, ___" wake-word will route any personal-context question to /omi/ask on Apex and the answer comes back through the Ray-Bans. No ETA from OMI on review turnaround.
Contract signing window — Friday, May 8. Brian Daly's signing target. Three clauses flagged for pre-signing pushback in the contract review. Don't sign Tuesday or Wednesday — read the report once more, decide on the three flagged items, then sign Thursday or Friday.
School exams next week. Lit exam is Tuesday 5/5. Other 8th-grade exams stack across the rest of the week. Grading block + comment-generator runs probably Saturday or Sunday — light load relative to past quarters.
Captain's Log recovery. The Make scenario fired correctly during the 5/2-5/4 Ollama outage (200 OK from /ingest) but rows died server-side before the patch landed. Make.com retains execution body data — the recovery script can pull and re-POST those bodies through the now-patched endpoint. Not done yet; on the active to-do list.
Yahoo poller throughput. Up and ingesting (329 rows in last 24h). Verify-then-flag-\Seen logic preserves UNSEEN messages on TP3-side failure, so no data loss; just slow catch-up while Ollama warms its cache. Backlog should clear over the next 12-24 hours.

OMI quality finding — why "Ask TP3" matters

Ran a deep investigation Sunday on the "OMI replies got worse" complaint from earlier this week. Findings:

The pendant chat tab is not Claude. It runs OpenAI's gpt-5.1 (with gpt-4.1 + gpt-4.1-mini for routing). Claude only powers OMI Desktop chat (Opus 4.6 / Sonnet 4.6 / Haiku 4.5).
The "low-quality phrase" rate in OMI's structured summaries (paraphrased "incomplete / unclear / lacks context" punts) jumped from 3.0% the week of Apr 20 to 14.6% the week of Apr 27. That's a quintupling. OMI's own April 26 changelog admitted a backend prompt regression.
Workaround already in flight: the "Computer, ___" trigger routes through /omi/ask on Apex, which is Anthropic-backed when the key is set (Gemini Flash fallback). Bypasses OMI's gpt-5.1 entirely for personal-context questions. Once the app is approved, this is the answer.

Full investigation file at /tmp/omi_quality_investigation_2026-05-04.md — table of weekly stats, public OMI commit references, recommendation matrix, all there.

Stack health

14 containers up on Apex. All healthy or running:

tp3_postgres_brain	Up 16h
tp3_memory_api	Up 1m (healthy, freshly restarted with patches)
tp3_ingest	Up 16h (healthy)
tp3_omi_mcp	Up 16h (healthy)
tp3_embed	Up 15h (healthy)
tp3_bidet	Up 15h (healthy)
tp3_ollama	Up 15h (in Docker, GPU active — new home)
yahoo_tp3_ingest	Up 15h (healthy)
yahoo-mcp	Up 16h
workspace-mcp	Up 2h
wa-web-mcp	Up 16h (WhatsApp Web linked, persistent ~14d)
tp3_pinger	Up 16h
tp3_minio_vault	Up 16h
tp3_autoheal	Up 3h (healthy)

The Ollama-in-Docker move closed the last "fragile native Windows process" gap. Whole ingest path is now containerized end-to-end on Apex.

Generated 2026-05-04 ~23:00 ET for Tuesday morning read. Real content this time — no recycled boilerplate. Numbers were live-queried from tp3_memories_local on Apex; calendar pulled from /omi/ask; container list from docker ps; OMI quality findings from /tmp/omi_quality_investigation_2026-05-04.md.