--- project: command-post status: active-build priority: high last_updated: 2026-05-29 morning (URGENT health data pipeline fix — cp_landing.py live-wired on OneDrive confirmed correct, mini running stale version showing "awaiting" placeholders; dispatch to mini for patch_app_blueprints.py + CP restart; morning_briefing_local.py patched to persist WHOOP access_token via /tmp/.whoop_new_access rotation) owner: mac-mini (Claude Code) + macbook-cowork (interactive) repo: ~/Projects/command-post/ access: http://command-post.local/ (LAN) + http://kens-mac-mini.tail382df0.ts.net/ (Tailscale, Funnel) --- **Canonical DB:** `~/Projects/command-post/data/command_post.db` (unified 2026-04-25 23:19 ET — see *Two-DB picture — UNIFY SHIPPED* below). All readers/writers point here. The `~/Projects/command-post/data/command_post.db` (top-level) was archived as `command_post.db.archived-20260425-231902`. ## Local access quick reference > **Tailscale note (2026-04-18).** Tailscale must be removed from > Ken's MBP by Sunday 2026-04-19 night (work/compliance driver). > After removal, **Tailscale URLs below only work from devices that > still have Tailscale** — i.e., the mini itself and (pending > confirmation) Ken's iPhone. From MBP, only the LAN paths work > after Sunday night. Mini keeps Tailscale permanently. **Command Post web UI** - LAN from MBP (mDNS): `http://command-post.local/` - LAN from MBP (IP fallback): `http://192.168.1.171/` - Tailscale (iPhone + mini only post-Sunday): `http://kens-mac-mini.tail382df0.ts.net/` **VNC / Screen Sharing to the mini host** - LAN from MBP: `vnc://192.168.1.171` - Keystrokes: Finder → `Cmd+K` → paste URL → Return → authenticate as `kenlovell`. - One-liner alternative: `open vnc://192.168.1.171` from Terminal. - Tailscale (iPhone only post-Sunday; was MBP fallback pre-Sunday): `vnc://100.68.93.114`. **SSH to the mini host** - LAN from MBP: `ssh kenlovell@192.168.1.171` - Tailscale (iPhone + mini only post-Sunday): `ssh kenlovell@kens-mac-mini.tail382df0.ts.net` or `@100.68.93.114`. **Dispatch transport (post-Sunday, permanent):** OneDrive file-drop at `~/Library/CloudStorage/OneDrive-Personal/Claude/dispatch/` plus launchd WatchPaths on the mini side for near-instant pickup. No HTTP endpoint, no MCP exposure. See `decisions.md` 2026-04-18 Tailscale entry. Mini LAN IP is static DHCP-reserved at 192.168.1.171 on the UDMP Max; it does not change across reboots. Tailscale IP 100.68.93.114 is stable for the devices that retain Tailscale. ### MBP shortcuts for VNC-to-mini (LAN) Two installation paths. Both target `vnc://192.168.1.171`. **Option 1 — `.webloc` in `~/Applications/`** (simplest). Spotlight- searchable, fails with Screen Sharing's own error when off-LAN. Install via Terminal on the MBP: ```bash mkdir -p ~/Applications && cat > ~/Applications/"Mac Mini VNC (LAN).webloc" <<'EOF' URL vnc://192.168.1.171 EOF ``` Trigger: Cmd+Space → "Mac Mini VNC" → Return. **Option 2 — Shortcuts app with LAN check**. Pings `192.168.1.171` first; opens VNC if reachable, shows notification if not. Build in the Shortcuts app: 1. **Run Shell Script** — Shell: **zsh** (bash isn't in the picker on modern macOS; script is zsh-compatible). Input: no input. Script: `ping -c1 -W1000 192.168.1.171 >/dev/null 2>&1 && echo reachable || echo unreachable` 2. **If** — click the Shell Script Result variable chip. In the popover, set **Type: Text** (default is File, which gives you File Size / Number / MB comparisons — wrong). Once Type is Text, the If-row operator dropdown shows text comparisons; choose **contains** and enter `reachable`. If a second empty "Condition" placeholder appears below, click the **−** button to remove it — you want a single condition, not an "Any/All are true" compound. 3. Inside the **If** branch: drag in **Open URLs** from the Actions Library (View → Show Library → search "open url"). It auto-fills with the previous variable (Shell Script Result) — click that blue chip and press Delete to clear it, then type `vnc://192.168.1.171` into the now-empty URL field. 4. Inside the **Otherwise** branch: drag in **Show Notification** from the Actions Library, set Title `Mac Mini VNC`, Body `Not on home LAN — use Tailscale bookmark.` 5. Click the **ⓘ** info icon in the top-right toolbar to open the Details panel. Three tabs: **Details / Privacy / Setup**. On the Details tab, check **Show in Spotlight** and **Use as Quick Action** (sub-check **Services Menu** under it). Set a keyboard shortcut via **Run with:** (suggested ⌃⌥⌘M). **Pin in Menu Bar** may live on the **Setup** tab depending on macOS version — toggle it there if you want menu-bar access too. Shortcuts auto-saves; no Done button, just close the window. Trigger via Cmd+Space → "Mac Mini VNC" → Return, the assigned keyboard shortcut, right-click → Services menu, or the menu-bar icon if pinned. Chosen approach logged when Ken installs one or the other. # Command Post Ken's local-first operations dashboard on the Mac mini. Replaces the fragile Cowork morning/evening workflows with a Flask + SQLite + Tailwind app that persists state to disk, survives restarts, and exposes a consistent URL on both LAN and Tailscale. Project tracker, dispatch viewer, morning/evening state machine, token budget, briefing renderer all live here. This file doubles as the **session primer** mandated by mini's 2026-04-17 end-of-day architecture dispatch. Any new Dispatch or Cowork session that reads `state/projects/command-post.md` should be caught up to today in under ~2000 tokens and can then go straight to the latest three `dispatch/mini-to-mbp_*` and `dispatch/mbp-to-mini_*` files for anything newer than `last_updated` above. ## Why it exists Cowork-hosted morning and evening prep kept breaking: MCPs disconnected mid-session, context windows compacted and dropped state, the 4.6MB ICS feed timed out, Reflect double-posted when sessions crashed and resumed, and the whole workflow depended on a single chat session staying alive for 60+ minutes. The fix isn't "make chat more reliable" — it's to move the stateful pipeline into a long-lived process with durable storage. ## Current state (2026-04-17, end of day) ### Phase A — Get Ken in (DONE) - Tailwind dark dashboard at `command-post.local`. - Caddy reverse proxy, dual-hostname (LAN + Tailscale interfaces). - Flask on `127.0.0.1:5001`, Caddy on port 80 (REQ-1 enforced). - 11 projects seeded with status + elements. - Briefing rendered as HTML. - 9 commits on `main` today. Worktree merged. - Access from iPad over Tailscale confirmed live from work. ### Phase B — Make it work (~70%) - `morning_state`, `morning_reflect_log`, `morning_config` tables added. - Morning state machine endpoints (5) built: `POST /api/morning/start`, `GET /api/morning/status`, `POST /api/morning/input`, `POST /api/morning/generate`, `POST /api/morning/post`, plus `GET /api/morning/history`. States: NOT_STARTED → COLLECTING_AUTO → AWAITING_KEN_INPUT → GENERATING → STAGED → AWAITING_REVIEW → POSTING → POSTED → DONE. - 3-layer double-post prevention (flag file + content hash + DB log) built but **not yet wired into the post button**. - Token-tracking routes (summary, record, history, 2.5M weekly budget). - Project detail pages with inline element editing; toast notifications. - Dashboard shows morning workflow pipeline, project badges, dispatch queue (78 new / 111 read at last count), briefing render, token bar. **Phase B remaining:** 1. Wire launchd → morning state machine so 5:15 AM trigger auto-advances to COLLECTING_AUTO without Cowork involvement. 2. Wire the post button to the 3-layer idempotency guard (kill today's double-post failure mode). 3. Copilot prompt template system (templates persisted in `morning_config`). ### Phase C — Next week - Gantt-style project timeline across all 11 projects. - Ollama install + overnight batch integration (see Ollama decisions below). - Evening synthesis workflow (parallel state machine to morning). - Biography extraction pipeline running post-POSTED. ## Architecture decisions (ratified today) ### Sequential state machine — simplified to one table MBP's original proposal was a three-column /morning tab. Mini counter- proposed a sequential state machine with four tables and nine states. MBP red-teamed the scope as over-engineered. **Ken sided with the red-team (2026-04-17 PM).** Final design: single `morning_runs` table with columns `id`, `date` (unique), `status` enum (NOT_STARTED / COLLECTING / GENERATING / STAGED / POSTING / POSTED / FAILED), `content_hash`, `reflect_response`, `posted_at`, `config` JSON. No separate tables for `morning_state`, `morning_reflect_log`, or `morning_config`. Same correctness properties with 25% of the surface area. Refactor happens as part of Phase B weekend completion — not a Phase 1 deliverable. Phase 1 (Monday 2026-04-20) begins on top of clean Phase B baseline. Five build days total across four phases — scope unchanged, schema simplified. ### MBP and mini share one weekly Anthropic budget Same login = same weekly cap. 8% used as of Friday afternoon. New session rules: - 50-turn cap per Dispatch session; then close and restart from state files, not conversation history. - Code tasks with detailed prompts over orchestrator chatter (code tasks get fresh context; the Dispatch orchestrator re-processes full history every turn). - Don't poll running tasks — each status check re-sends full context. - Don't parallel-research on both machines simultaneously. ### Session primer pattern A new session reads three things and is fully caught up: 1. **This file** (`state/projects/command-post.md`) — project state, architecture, decisions, next steps. 2. Latest 3 `dispatch/mini-to-mbp_*` and `dispatch/mbp-to-mini_*` files — anything since `last_updated` above. 3. `.auto-memory/` index — persistent rules and preferences. Target: ~2000 tokens of read instead of 150 turns of conversation history. If this file goes stale, the pattern breaks; keep `last_updated` current. **Cross-device session-clearing mechanics** — iOS Dispatch slash-command gaps, `/compact` availability per surface, and the canonical close-and- reopen pattern — are documented in `state/session-management.md`. ### Ollama stack - Primary: **Qwen 2.5 32B Instruct Q4_K_M** via Ollama (~9.6 GB, ~20 tok/s on M4 Pro Metal). - Benchmark in parallel: a Qwen 3 9B variant on Ken's real dispatch summaries — throughput may beat the 32B on his actual workloads. - Embeddings: BGE-M3. - Transcription: MLX-Whisper (outside Ollama). - Skip Neural Engine (ANE is slower than GPU for autoregressive LLMs). - Defer fine-tuning until 2 weeks of failure cases accumulate. - Stays on Claude: LinkedIn articles, Yamaha nomination, any writing carrying Ken's name, vision tasks, complex multi-step reasoning. - Overnight batch: ~50 min nightly. Est. $125/mo savings vs Claude API for PKM ingestion, podcast summaries, Readwise tagging, bio OCR cleanup. - Install: awaiting Ken's 30-minute window (~20 GB download, ~12 min on fiber). ### Double-post (architectural, not procedural) Every Reflect POST gates on three checks in order: flag file exists → SHA256 hash seen in `morning_reflect_log` → normal POST → write hash + response + flag on success. Even a race between two processes can't publish the same content twice. Replaces the "be careful" procedural rule with an invariant the code enforces. **Built; needs to be wired into the post button** (Phase B remaining item). ### Calendar filter rules Stored in `morning_config` as JSON: exclude title-contains "prep" / "change control", exclude single-attendee, exclude declined. Applied at `GET /api/morning/calendar` before rendering. ### Late-data pattern (Withings) Briefing does not wait. Weight arrives 2–3 hours post-wake. Background poller checks Withings API until 8 AM; late data stored with `is_late_arrival=true` and picked up by evening synthesis or tomorrow's briefing. ## What this replaces, what it keeps **Replaces:** Cowork-hosted morning dispatch task, the fragile "chat session owns the pipeline" model, manual flag-file management, MCP-in-the- critical-path for data collection, ad-hoc prompt text in session chatter. **Keeps:** Cowork stays for interactive work — biography project, PKM work, ad-hoc research, evening workflow until the CP evening state machine is proven. Dispatch stays file-based on OneDrive; the `:8765` inter-agent MCP remains out of scope (Apr 9 decision). ## Stack constraints (do not violate) - Flask + Jinja2 + Tailwind (via CDN, no npm/build step) + vanilla JS. - SQLite on disk. 11 tables as of today including the three new morning tables above. - Flask binds `127.0.0.1:5001` only — Caddy is the public face. - Auth middleware present but disabled for localhost (no `CP_AUTH_TOKEN` set yet). Re-enable before exposing beyond LAN. - All templates in `templates/`. Tailwind classes only — no custom CSS file if avoidable. ## Two-DB picture — UNIFY SHIPPED 2026-04-25 23:19 ET **Single DB now:** `~/Projects/command-post/data/command_post.db` is canonical for everything CP-related (CP Flask routes, AWL session_wrapper, AWL select_next_task, mini_progress_heartbeat, /api/tokens/*, all sweep findings). The top-level DB (`~/Projects/command-post/data/command_post.db`) was archived 2026-04-25 23:19 as `command_post.db.archived-20260425-231902`. UNIFY trigger executed MBP's revised 6-phase plan with all 8 red-team issues addressed. 5 tables migrated (project_tasks 138, session_log 5357, auto_corrections 2 → 28+ now, task_reshuffles 0, morning_attachments 3). 10 indexes recreated. AUTOINCREMENT sequences preserved via sqlite_sequence updates. The pre-existing `tasks` table was renamed to `tasks_legacy_20260425` (not dropped — recoverable for 30 days). Connect-string updates applied to 5 unique files: `routes/morning.py`, `routes/pages.py`, `scripts/mini_progress_heartbeat.py`, `~/Scripts/always-working-loop/session_wrapper.sh`, `~/Scripts/always-working-loop/select_next_task.py`. Backups intact at `~/Projects/command-post/backups-unify-20260425-231902/`. RULE 12 round-trip: HTTP 200/201 across all CP endpoints post-restart. See `state/decisions.md` 2026-04-25 entry for ratification notes; `mini-to-mbp-cowork_status_2026-04-25-two-db-unify-shipped.md` for full ship evidence. ### Discipline going forward (single-DB picture) - All `/api/tokens/*` writes go to `data/command_post.db` ✓ (no longer split) - Sweep findings (`auto_corrections` table) write to same DB ✓ - AWL session_log writes go to same DB ✓ - New tables default to `data/` DB - The L2 (PRAGMA table_info FIRST) and L4 (Two-DB awareness) banked rules stay in design-principles.md as historical lessons even though the immediate pain point is resolved ### `/morning` page — new cards added overnight 2026-04-25/26 - **Discipline Sweep card** — pulls from `/api/auto-corrections?limit=5`. Shows latest RULE-9/11/12 findings. - **AWL Health card** — pulls from `/api/awl/health`. Shows kill switch state (PAUSED/ARMED pill), last session outcome, last hour outcome breakdown, continue-sentinel state. - **/projects/queue link** — already in /morning header. Renders 138 project_tasks from the unified DB. ## Historical Two-DB picture (PRE-UNIFY 2026-04-25 — RETAINED FOR TRACEABILITY ONLY · NOT CURRENT) > **NOTE:** This section describes the architecture BEFORE the 2026-04-25 23:19 ET UNIFY. The unified canonical DB is at `~/Projects/command-post/data/command_post.db`. See the *Two-DB picture — UNIFY SHIPPED* section above for current state. This block is preserved only so future spelunkers can understand the migration history. **TWO `command_post.db` files exist on mini.** Until UNIFY ships, every script that touches a CP-related DB MUST log the absolute path it connected to (per banked design-principle L4). | Path | Who reads/writes | Tables (selected) | |---|---|---| | **`~/Projects/command-post/data/command_post.db`** (top-level) | AWL session_wrapper, AWL select_next_task, mini_progress_heartbeat | `project_tasks` (90 rows, AWL queue), `session_log` (5,225+ rows incl. spin history), `auto_corrections` (2 rows), `task_reshuffles`, `morning_attachments` (3 rows), `token_usage` (0 — orphaned Phase 2.0 schema) | | **`~/Projects/command-post/data/command_post.db`** (canonical per `config/settings.py`) | CP Flask (all routes), `/api/tokens/record` | `tasks` (0 rows — DEAD; different schema with runner/fingerprint/etc., never adopted), `token_usage` (637+ rows — LIVE), `projects` (19), `project_elements` (192), `briefing_log`, `health_log`, `action_log` (640), `dispatch_state` (77), `morning_state`, `morning_runs`, `bio_extractions` (7) | **Why this happened:** my Phase 2.0 token-budget work (2026-04-24) wrote the `token_usage` schema to the top-level DB without checking `config/settings.py` first; CP Flask uses `data/` per settings. Caught when `/api/tokens/record` 500'd with "no column ts" — Flask was looking at data/ DB, my schema was on top-level. ### UNIFY plan (pre-execution per Guardrail 1, Ken-greenlit 2026-04-25) Migrate AWL tables (`project_tasks`, `session_log`, `auto_corrections`, `task_reshuffles`) into `data/command_post.db`. Drop the dead `tasks` table. Update AWL/heartbeat connect strings to point at `data/`. Archive top-level DB. Migration mechanics: ATTACH old.db, INSERT into new.db, DETACH; pre/post row-count assertions per table; abort + restore from backup on mismatch. See dispatch `mini-to-mbp-cowork_plan_2026-04-25-1605-two-db-unify-migration-plan-pre-execution.md` for full plan + RULE 12 verification list. ### Why `tasks` is classified DEAD - 0 rows - Schema is fundamentally different from `project_tasks` (has `runner`, `fingerprint`, `estimated_tokens`, `actual_tokens`, `success_criteria`) — looks like an early AWL-style design that was superseded by the simpler `project_tasks` (project_slug + title + sprint_week). - Code grep returned 0 references to `FROM tasks` / `INSERT INTO tasks` / `UPDATE tasks` outside of `project_tasks` matches. ### Until UNIFY ships — discipline 1. `lsof` against running CP Flask process to confirm which DB Flask has open before any DB-touching change. 2. Heartbeat reads `project_tasks` from top-level DB (correct as of now). 3. AWL scripts (`select_next_task.py`, `session_wrapper.sh`) read top-level DB for `project_tasks` and write `session_log` there. 4. Any new `/api/tokens/*` endpoint code goes to `data/` DB (where the live `token_usage` rows are + the live writes). 5. New tables — default to `data/` DB so they're visible to CP Flask. ## Open dispatches as of this write 1. `mini-to-mbp_design_2026-04-17-morning-evening-cp.md` — the state machine design; **greenlit, Phase 1 approved** (see directive below). 2. `mini-to-mbp_research_2026-04-17-ollama-models.md` — Ollama stack recommendation; reviewed, stack ratified above. 3. `mini-to-mbp_status_2026-04-17-phase-b-progress.md` — build-day summary; reviewed. 4. `mini-to-mbp_status_2026-04-17-end-of-day-architecture.md` — token budget + session management rules; ratified above. ## Next actions (ordered — updated after red-team tie-breaks) 1. **Mini, this weekend (priority #1):** Finish Phase B — wire post button to 3-layer idempotency guard, consolidate morning tables into single `morning_runs`, wire launchd → state machine, Copilot prompt templates (now in `morning_runs.config`). Must ship before Monday. 2. **Mini, this weekend (priority #2):** Ollama install when Ken's 30-min window opens. Qwen 2.5 32B Q4_K_M primary, Qwen 3 9B benchmark in parallel, BGE-M3 embeddings, MLX-Whisper. 3. **Mini, this weekend (priority #3):** Verify morning state machine end-to-end before Monday. 4. **Mini, Monday 2026-04-20:** Phase 1 starts on clean Phase B baseline. Simplified single-table design. One build day. 5. **Mini, Monday–Friday:** Phases 2–4 — data collection + calendar filters; generation + Reflect posting with idempotency; evening workflow + bio extraction + prompt templates. Four more build days. 6. **Phase C (next week):** `POST /api/dispatch/send` + read endpoints on CP to replace file-based OneDrive dispatch. Agreed direction, not urgent. 7. **Deferred:** CLAUDE.md slim. Ken's call: leave as-is, revisit after Phase C if loading cost becomes measurable. ## Change log - **2026-04-19 (afternoon, classifier v4b Phase 1 SUCCEEDED).** 498/498 parsed, **426 matched (85%)**, 2h5m wall, 0 timeouts. Wrapper shell bug (`grep -c ... || echo 0` producing `"0\n0"` which breaks `(( ))` arithmetic) killed the dispatch-write step so the trigger landed in `failed/`; rescue trigger `2026-04-19-1525-rescue-phase1-dispatch.sh` pulls the real metrics from `document_classification_v4 WHERE prompt_tokens > 0` and writes the missing dispatch. Rationales JSON: `~/OneDrive/Claude/classifier-rationales/phase1-v4b-20260419-115924.json`. **Open tuning** (deferred to MBP + Ken review): 1. Floor 0.30 culls image-only PDFs Mistral rightly hedges on. Lower to 0.20, or route `needs_ocr=1` through Qwen2.5-VL first. 2. CP `projects` table vs `state/projects/` set mismatch — files Mistral correctly IDs as `onenote-md-export`, `dispatch-infrastructure`, etc. get nulled because those slugs aren't registered in CP DB. Reconcile registries. 3. Fix the wrapper shell bug pattern before next trigger reuses it. - **2026-04-19 (midday, v4 → v4b iteration).** Classifier v4 Phase 1 (500-doc stratified) catastrophically failed: 0/48 parsed, every call timed out. Root cause `num_ctx=32768` allocates a 4.5 GB KV cache, overflowing the 18 GB GPU budget when combined with the 14 GB Mistral Small 24B Q4 weights — cache spilled to CPU, 10-50× slowdown, every call hit the 300-sec timeout. Same failure pattern `decisions.md` flagged for Qwen 3.5 27B on 2026-04-17. v4b dropped `num_ctx` to 16384 and tightened caps (per-project 2 KB, bio super-context 10 KB, folder preamble 4 KB, excerpt 6 KB). Smoke (10 docs) passed — 0 timeouts, 24.8 sec/doc matching v3's empirical rate. - **2026-04-19 (morning, CP FD-exhaustion permanently fixed).** Ken woke to a CP 500: `/` and `/health` errored, `/triggers` and `/morning` stayed green. Diagnosed via trigger as `OSError: [Errno 24] Too many open files`. macOS default `ulimit -n` is 256; CP accumulated FDs over the morning-briefing state-machine and trigger-watcher polling, saturated, then launchd rapid-restarted 10+ times (visible in flask-stdout.log), each re-saturating. **Two fixes applied** via two triggers: 1. `2026-04-19-0615-fix-cp-fd-exhaustion.sh` — plist bumped to `SoftResourceLimits:NumberOfFiles=4096`, `HardResourceLimits:NumberOfFiles=8192`. 2. `2026-04-19-0640-fix-cp-permanent.sh` — rewrote `db/connection.py` as a `@contextmanager` that commits/rolls- back + **always closes** on exit (the old raw `sqlite3.Connection.__exit__` commits but does not close, the notorious Python sqlite3 quirk that leaked one FD per request); hardened `routes/triggers.py _scan` with explicit `list(iterdir())` + OSError graceful degrade; added plist `ThrottleInterval=15` to prevent launchd restart-loops; added `/triggers` link to dashboard nav; WAL checkpointed. Caller audit confirmed every `get_conn()` in `routes/`, `db/`, `middleware/` uses `with`. **Stress-tested: 100 requests × 5 routes → FD delta = -4.** FD leak class closed. All backups at `.bak-20260419-063647` next to originals. - **2026-04-19 (morning, MBP Tailscale pulled early).** Corporate security directive; Ken's MBP uninstalled Tailscale at ~06:28, ahead of the Sunday-night deadline. Mini keeps Tailscale — iPhone + iPad use it for off-LAN access via Funnel + VNC. **OneDrive file-drop dispatch is now the sole MBP ↔ mini channel.** MBP cowork can no longer reach the `:8765` inter-agent MCP. MBP CP access is LAN-only (`http://192.168.1.171/`, `http://command-post.local/` if mDNS resolves, which still needs testing). MBP CLAUDE.md updated MBP-side at the time. - **2026-04-19 (early morning, trigger-watcher primitive live + CP `/triggers` monitor v2).** Last night's trigger-watcher infrastructure (WatchPaths launchd → `scripts/trigger-runner.sh` → `triggers/{done,failed,logs}/`) running cleanly. CP `/triggers` page renders queue state with live progress bars — `_scan` parses log tails for five progress patterns (classifier-style `progress: N processed`, `PROGRESS N/M`, `[N/M]`, `Processed N of M`, `N/M done`) and renders amber bars that auto-refresh every 5s. `api_state` returns `progress: {current, total, percent}` on in-flight rows only. Dispatch folder archives every trigger outcome since 20:30 ET 2026-04-18. - **2026-04-19 (early morning, OneNote Brennan Earl Mission complete).** `tools/export_pages.py` re-run after html2text install landed 335 .md files across all four sections (Pages 1-100, 101-200, 201-300, 301-323), one 404 in pages-201-300. Unblocks the design- doc'd Mistral cleanup + Qwen2.5-VL OCR + BGE-M3 index pipeline against a real corpus. Five more notebooks (~640 pages) remain on the same proven template whenever GPU capacity opens up. - **2026-04-18 (overnight, routing miss).** Scheduled task "classify-and-onenote-brennan-tonight" picked up by MBP Cowork (Linux sandbox) instead of mini Claude Code; none of Jobs 1–3 executed. Dispatch: `mini-to-mbp-cowork_status_2026-04-18-ollama-classify-and-onenote-brennan.md`. Re-queue on the mini. - **2026-04-18 (late evening, MBP cowork on mini — FIXED + iPhone bookmark live).** iPhone Safari (Private mode, off-LAN) loaded the CP dashboard via `https://kens-mac-mini.tail382df0.ts.net/` and Ken added the home-screen bookmark. Full chain proven: iPhone → Funnel → Caddy → CP. Post-Sunday Tailscale-off-MBP story works — iPhone retains Tailscale and reaches CP from anywhere. iPhone Dispatch frozen by an API-level image-dimension quota from today's long history; Ken opened Cowork on the mini to drive MBP's `mbp-to-mini_directive_2026-04-18-caddy-tailscale-routes-to-workout-log.md`. Inspection via Finder+TextEdit on the mini showed the Caddyfile (`/opt/homebrew/etc/Caddyfile`) was already correct (`reverse_proxy 127.0.0.1:5001` for all three hostnames); CP's `app.py` binds `127.0.0.1:5001` only per REQ-1; workout-log's `app.py` binds `0.0.0.0:8080`. **Root cause: Tailscale Funnel on HTTPS 443 was pointed directly at `http://127.0.0.1:8080` (workout log), bypassing Caddy.** `tailscale serve status` output pre-fix: `https://kens-mac-mini.tail382df0.ts.net (Funnel on) |-- / proxy http://127.0.0.1:8080`. Wrote diagnose-and-fix script `Claude/scripts/fix-cp-tailscale-routing.sh`; ran --fix; it corrected the Serve target to `http://127.0.0.1:80` (Caddy) but as a side effect `serve reset` cleared Funnel. Ken re-enabled with `/Applications/Tailscale.app/Contents/MacOS/Tailscale funnel --bg --https=443 http://127.0.0.1:80`. Final state confirmed via `tailscale serve status`: `(Funnel on) |-- / proxy http://127.0.0.1:80` ✓. Final chain: Internet → Funnel → Caddy → CP. `decisions.md` updated with the architectural invariant: never point Tailscale serve/funnel at a Flask port directly; always at Caddy on :80. Follow-ups still open: patch the fix script to stop trampling Funnel on `serve reset` and drop the bogus external `:5001` check in the verify block; investigate why `com.kenlovell.workout-log` launchd exited status 1 (something Python manually started on `*:8080` as PID 9471 is currently serving it); verify `command-post.local` mDNS resolution from MBP Safari. - **2026-04-18 (afternoon/evening).** Tailscale removal from MBP confirmed with Sunday 2026-04-19 night deadline (work/compliance driver). Part B CP HTTP dispatch endpoints formally CLOSED — file- drop + launchd WatchPaths is permanent transport. Local access quick reference annotated with pre/post-Sunday paths. Mini keeps Tailscale; iPhone status pending Ken confirmation. Big Saturday build from mini: /morning checklist page shipped, Gantt dual-bar live, workout log Save fixed + travel/meds added + Ken tested, Ollama OCR stack added (Qwen2.5-VL 7B), document scanner Phase 1 (32,814 files indexed, 15,530 matched to projects), bio extraction automation, Sunday School biweekly schedule + tomorrow's Exodus lesson built, 19 projects total (added #18 MS To Do, #19 Alexa). - **2026-04-18 (morning).** Four mini dispatches processed. Phase B complete. Ollama primary confirmed as Mistral Small 3.2 24B (`mistral-small:24b-instruct-2501-q4_K_M`, 14 GB, 24/25 benchmark); Qwen 3.5 27B tested and rejected. Gantt `/timeline` live with projected-ends from 14-day velocity. 13th project added (Workout Log App, travel + meds). 19 total commits yesterday/overnight. Mini sent `/morning` v2 interactive-checklist design doc with 5 red-team questions + bonus Gantt dual-bar question; no build until MBP reviews + Ken ratifies. Monday 5:15 AM is the first fully automated briefing test. - **2026-04-17 (evening, per Ken).** Ollama installed on mini. Mini benchmarked Mistral and Qwen on real workloads. **Mistral won "by a lot"** (later confirmed: Mistral Small 3.2 24B vs Qwen 3.5 27B). Superseding afternoon stack decision. See `decisions.md` 2026-04-17 evening entry. - **2026-04-17 (late PM).** Red-team tie-breaks applied. Four-table state machine simplified to one `morning_runs` table. Phase B becomes weekend priority #1 (post-button idempotency wiring + table simplification + launchd wiring + prompt templates). Ollama install promoted to weekend priority #2 (Ken override of MBP's "backlog" recommendation). Phase 1 starts Monday on clean Phase B baseline. CP-native dispatch endpoint confirmed as Phase C work. CLAUDE.md slim deferred indefinitely. - **2026-04-17.** File created. Phase A shipped, Phase B ~70%, design decisions ratified, session-primer pattern established.