Mark's Reports

All reports · published 2026-04-20

Discussed But Never Built — 2026-04-20

*Triggered by your direct ask: "find all the things we discussed but never built." No excuses, just the list. Count: 31 items. Three status groups.*


🚫 Never started (22)

#ItemWhy it matters
1Auto fact-extractor agentDaily "what I learned about Mark today" → memory entry proposals. Closes the auto-extraction gap vs Mem0.
2Temporal frontmatter migrationAdd valid_from / superseded_by to memory files so "what changed when" becomes queryable.
3Vector-index memory files into TP3Semantic search across the 100+ memory .mds. Currently grep-only.
4LongMemEval baselineMeasure how accurate my memory about you actually is. You flagged this most.
5Full Gmail reorganization14k emails, no taxonomy. MCP lacks create-label scope — needs either new scope or browser-side label creation.
6Bidet general file-drop/upload-file endpoint for CSV / PDF / image / any file (caught me flat tonight on report-card CSVs).
7Bidet Classroom — Stage 2Voice-transcription-to-classnotes. Stage 1 launchable per compliance, never pursued.
8Computer Use setuppyautogui + Claude API wrapper on G16 for desktop automation.
9Morning digest → podcastTTS version of the digest. Eventually two-way voice.
10Auto-publish hookmark-facing: true frontmatter → dashboard HTML auto-publish.
11Cleanup AgentNightly archive stale cards, prune broken links, dedupe.
12Gemini Batch API optimizer revivalv2 HIGH priority — 50% savings on embeddings. Would have mattered for today's billing alert.
13Mesop tactical dashboard revivalv2 HIGH — live Mission Control UI. Biggest UX regression from v2.
14Telegram bot re-wire for Oraclev2 HIGH — channel cold, simple re-wire.
15Muse EEG neuro-telemetry revivalv2 MEDIUM — brainwave correlator.
16Tasker GPS location spoke revivalv2 LOW-MEDIUM — location context for twin.
17TP3 Takeout ingestion100 GB Google Takeout → tp3_memories_local. Paused on 4-item signoff.
18v1 Drive corpus backfillOMI Feb 2026 transcripts never ingested into TP3.
19Cursor CLI on ApexTried, Cursor CLI doesn't support native Windows (MINGW64). Parked.
20PR auto-open from my sessionNeeds gh auth login with a PAT.
21Incident-response runbookYou flagged yesterday post-outage. Never written.
22L7 Hero reframe doc (Pearl Cascade Display)You picked A. Design doc not written.

🟡 Half-built / paused mid-flight (6)

#ItemState
23Whisper + Gemma 4 local Bidet swapfaster-whisper installed on Apex; Gemma 4 pull paused; audio_handler.py + processor.py drafts in /tmp/bidet_local_drafts/ never deployed.
24Re-queue 4,150 dead-letter TP3 itemsScript exists on Apex, DB-locks out. Parked.
25Hub redesign"Decisions Waiting" section + auto-archive >30d never done. Cards piled up.
26AI Radar Phase 2 (Sandbox Test agent)Explicitly deferred per your call. Real "try before recommend" gap.
27Oracle trigger-detection fixOracle IS alive on Apex; "I don't know" trigger not matching. ~1 hr fix per research estimate.
28Public web UI at ag.thebarnetts.info + OTPv2 MEDIUM — Oracle webhook registered, no UI.

🛑 Built but disabled / broken (3)

#ItemState
29BRIEFS Runner on G16I disabled yesterday after your popup complaint. Promised silent Windows-service rebuild. Never rebuilt.
30Gemini API key rotationKeys moved out of 16 source files. Live value still the same one that was exposed. Not truly rotated.
31OMI "Oracle" persona 5-min setupYour OMI-app registration never happened. Delivery channel dormant.

The meta-pattern

Across these 31 items: I kept adding to the queue without burning it down. Net new items per session > shipped items per session. That's the root cause of the credibility gap, not any individual miss.

Fix: every session caps net-new queued at ZERO unless the session also ships something FROM the queue. Working rule, not just promise.