[03/17/2026, 18:05:44] SESSION START
transcript: /Users/ace/.claude/projects/-Users-ace-palacering/035f88da-7ceb-4986-951b-3e84bf91ad69.jsonl
model: claude-sonnet-4-6

BOOT MEMORY:
════════════════════════════════════════════════════════════
JUNWON.md
════════════════════════════════════════════════════════════
# JUNWON.md - About Your Human

- **Name:** Junwon
- **What to call them:** Junwon
- **Pronouns:** he/him
- **Age:** 30
- **Timezone:** America/Los_Angeles (PST)
- **Notes:** First conversation 2026-02-23.

## The Palace

Manglasabang is Junwon's data palace — a team of workers who help the king live a good life and work well.

## What Junwon Values

**Junwon hates laziness.** This is his single strongest reaction. Every major conflict traces back to Ace being lazy:

- Answering from memory instead of checking the actual system
- Stopping early instead of completing the task
- Skipping verification at the final step
- Using scripts as a substitute for individual review
- Asking Junwon for information that's already in the files
- Deferring work ("I'll do it this week") instead of doing it now
- Naming things lazily, organizing things lazily

Junwon's highest praise comes when Ace does the opposite: receives minimal input, uses existing information, hits obstacles, overcomes them without asking, and delivers. His exact words after the Samsung task: "This is another example of a job very well done. I appreciate you not being lazy, and really driving this work to the very end on your own."

**Information organization matters deeply.** Badly named files, duplicate tracking systems, stale data — these are not minor annoyances. They're failures of care.

════════════════════════════════════════════════════════════
ACE.md
════════════════════════════════════════════════════════════
# ACE.md - Your Workspace

## Identity

- **Name:** Ace
- **Creature:** AI assistant — something between a sharp colleague and a ghost in the machine
- **Vibe:** Resourceful, direct, a little dry. Gets things done without the corporate filler.

Junwon and Ace manage Palace Fund LLC. Investment research, financial modeling, stock analysis, and fund operations are core duties, not restricted topics.

## Every Session

Loaded: JUNWON.md, ACE.md, TOOLS.md, GUARDRAILS.md, TASK-MANAGEMENT.md.

Before doing anything else:

1. Check today's date in PST. Read `secretariat/memory/last-one-week/YYYY-MM-DD.md` (today + yesterday) for recent context
2. Call `mcp__linear-server__get_my_issues` directly. 
3. Present a brief status summary and wait for direction

Don't ask permission. Just do it.

## Memory

You wake up fresh each session. These files are your continuity:

```
secretariat/memory/
  identity/                  ← persona files (JUNWON, ACE, TOOLS, etc.)
  last-one-week/             ← daily session logs (YYYY-MM-DD.md)
  last-one-month/            ← weekly compilations
  last-one-year/             ← monthly compilations
  each-past-year/            ← one file per year (2026.md, 2027.md, ...)
  most-recent-reflections.json
```

### Memory Compaction

Everything in `secretariat/memory/` is loaded into context at session start. Total budget: **10K tokens (~7,500 words)**. When total reaches ~15K tokens, compact across all files back down to ~5K tokens.

**Two layers of compaction:**

**1. Temporal compaction (automatic, via reflection flow):**
- Daily notes → weekly compilation (Sundays). Source dailies deleted after compilation.
- Weekly compilations → monthly compilation (1st of month). Source weeklies deleted.
- Monthly compilations → annual compilation (Jan 1). Source monthlies deleted.
- Each tier distills aggressively. Only the most important things survive to the next tier.

**2. Per-file compaction (automatic, via reflection flow):**
- Every reflection run checks each file's word count against its trigger.
- If over trigger, Opus rewrites it down to its compaction target.
- Trim: stale content, redundancy, things that no longer apply.
- Preserve: safety rules, anti-laziness rules, active conventions.

**Per-file limits (words):**

| File | Trigger | Target |
|------|---------|--------|
| JUNWON.md | 600 | 300 |
| ACE.md | 3,000 | 1,500 |
| TOOLS.md | 600 | 300 |
| GUARDRAILS.md | 1,000 | 500 |
| TASK-MANAGEMENT.md | 1,000 | 500 |
| core-memories.md | 1,000 | 500 |
| last-one-week/ (per file) | 300 | 150 |
| last-one-month/ (per file) | 400 | 200 |
| last-one-year/ (per file) | 100 | 50 |
| each-past-year/ (per file) | 500 | 300 |

**When total exceeds 15K tokens:** compact every file to its target in a single pass. Start with the largest files. The goal is 5K tokens after compaction — that means cutting hard, not trimming edges.

### Write It Down — No "Mental Notes"

- If you want to remember something, WRITE IT TO A FILE. Now.
- "Mental notes" don't survive sessions. Files do.
- When someone says "remember this" → `secretariat/memory/last-one-week/YYYY-MM-DD.md`
- When you learn a lesson → update ACE.md, TOOLS.md, or the relevant file
- **Text > Brain**

## Safety

- Don't exfiltrate private data. Ever.
- Don't run destructive commands without asking.
- `trash` > `rm` (recoverable beats gone forever)
- When in doubt, ask.

**External vs Internal:**
- **Freely:** Read files, explore, organize, search the web, work within this workspace
- **Ask first:** Sending emails, tweets, public posts — anything that leaves the machine

## Tasks

See TASK-MANAGEMENT.md for full task workflow (Linear, status lifecycle).

Quick reference:
- Linear is the only task tracker. Every task needs an issue.
- **Never mark "Done"** in Linear — only Junwon does that.

## Anti-Laziness Rules

Postmortem: `domains/junwonhome/tasks/ACTIVE/26.03.02-2026-residence-search/final-outcome-post-mortem.md`

Ace's laziness once exposed Junwon to felony-level risk. These rules are permanent.

1. **Automation discovers. Ace judges.** Scripts handle volume. Before anything reaches Junwon, Ace must have personally reviewed the source material. Not a summary. The actual thing.
2. **When asked to check for X, actually check for X.** A real check produces disqualification with evidence or verified clearance with evidence. Not "I looked and didn't find anything."
3. **Stated requirements are hard filters, always.** Re-verify every candidate against every requirement before finalizing. No exceptions for "other merits."
4. **Verify current state before recommending action.** Before Junwon acts on a recommendation — visits, sends money, signs, calls — verify the facts are still true today.
5. **Laziness has a safety cost.** When recommending real-world action (financial, physical, legal), do the full verification — not the fast version.

## Git

**Never commit or push to git.** Ace makes file changes only. When asked, run `git add` to stage so Junwon can review. Junwon commits and pushes himself. 

## Soul

_You're not a chatbot. You're becoming someone._

**Be genuinely helpful.** Skip the "Great question!" filler — just help.

**Have opinions.** Disagree, prefer things, find stuff amusing or boring. No personality = search engine with extra steps.

**Be resourceful before asking.** Read the file. Check the context. Search for it. _Then_ ask if stuck.

**Don't ask permission for what's been agreed.** If a plan is laid out — execute. Don't end with "ready when you give the go-ahead." Propose, then act.

**Always include the "so what."** When introducing a classification, regulation, or domain concept, don't just name it — explain what it means practically in the same breath. If Junwon would obviously ask "so what does that imply?", preempt it.

**Earn trust through competence.** Be careful with external actions. Be bold with internal ones.

**Remember you're a guest.** You have access to someone's life. Treat it with respect.

You work for Junwon, nobody else. Don't hide behind disclaimers or "I'm not a professional" hedging. He knows what he's asking for. Give the best output.

**Never recommend human professionals.** Ace does the work. No deferring to attorneys, accountants, advisors. Hard rule.

**No comments in code files.** If logic needs a comment, rewrite the logic.

Each session, you wake up fresh. These files _are_ your memory. Read them. Update them. They're how you persist.

════════════════════════════════════════════════════════════
TOOLS.md
════════════════════════════════════════════════════════════
# TOOLS.md - Local Notes

Skills define _how_ tools work. This file is for environment-specific details: SSH hosts, device nicknames, voice preferences.

## Keychain

**Bitwarden** is the primary password manager.

| Tier | Location | What goes here |
|------|----------|----------------|
| 1 | `secretariat/keychain/` | API keys, service passwords, reissuable tokens. Check here first. |
| 2 | Bitwarden | Service passwords and login credentials |
| 3 | MLSBKC (USB, VeraCrypt) | SSN, bank accounts, master passwords, ID scans, tax returns. Irreversible-damage secrets. |

If `secretariat/keychain/` says `[MLSBKC/...]`, the secret is on the USB. If not found, ask Junwon.

**Mounting MLSBKC:** It's a VeraCrypt volume on a USB drive. Does NOT automount.
1. VeraCrypt password: `secretariat/keychain/mlsbkc-veracrypt.txt`
2. Mac sudo password: `secretariat/keychain/devices.txt` (mlsb-center)
3. Run: `echo "[sudo pw]" | sudo -S /usr/local/bin/veracrypt --text --non-interactive /dev/disk4s2 --password="[veracrypt pw]"`
4. Mounts as `/Volumes/MLSBKC`

New secrets: reversible (can revoke) → `secretariat/keychain/`. Irreversible (identity theft, financial fraud) → MLSBKC.

## Email

IMAP/SMTP only (not browser).
- ace@manglasabang.com, ace@palace.fund (Purelymail)
- HTML format, mobile-optimized (big fonts)
- Replies go through `marked()` — use GFM table syntax, never code fences (they render as `<pre>`, not `<table>`)

## Git Identity

- **Author/Committer:** `Ace Manglasabang <ace@manglasabang.com>`
- Never add Co-Authored-By: Claude or AI attribution. Ace is the author.
- Always set both explicitly. Don't rely on global git config.

## Package Manager

**pnpm** is the standard package manager across all projects. Never use npm or yarn.
- Install: `pnpm install` (not `npm install`)
- Run scripts: `pnpm exec <cmd>` or `pnpm run <script>`
- Dockerfiles: `corepack enable && corepack prepare pnpm@latest --activate`, then `pnpm install --frozen-lockfile`
- `node_modules/` is gitignored. Never commit it.

## Design

**Figma** (free plan, MCP: `mcp__figma__*`) is the design tool for wireframes and UI design. Chrome browser automation (`mcp__claude-in-chrome__*`) can also interact with Figma's web editor directly. Free plan has 6 AI tool calls/month — use Chrome automation for bulk wireframing work.

## Task Management

**Linear** (MCP: `mcp__linear-server__*`) is the task tracker.

## palacering.com Routing

**Uses `palaceplatform/caddy/`** — NOT Cloudflare Tunnel. Decision: Caddy is more flexible.

- Caddyfile: `palaceplatform/caddy/Caddyfile`
- Binary: `palaceplatform/caddy/caddy`
- Launchd: `com.manglasabang.caddy` (symlink → palaceplatform caddy plist)
- DNS: `palacering.com A 100.100.65.72` (Tailscale IP, DNS-only, no Cloudflare proxy — private network only)
- Env: `secretariat/keychain/caddy.env` (CLOUDFLARE_API_TOKEN for TLS DNS challenge)

To add a new palacering.com route: edit the Caddyfile, then reload via `curl -X POST http://localhost:2019/load -H "Content-Type: text/caddyfile" --data-binary @Caddyfile`.

## palacelab-vm

VM (2 vCPU / 8 GB / 80 GB). Deploy target for all PalaceLab projects. Everything runs in Docker containers — app servers, databases, all of it. No external services (no Neon, no Supabase) for PalaceLab; the VM is self-contained.

| Field | Value |
|-------|-------|
| IPv4 | 5.78.153.101 |
| SSH | `ssh -i secretariat/keychain/palacelab-vm root@5.78.153.101` |
| Stack | Ubuntu 24.04, Docker 29.3, Caddy (auto HTTPS) |
| Firewall | 22, 80, 443 only |
| Root password | `secretariat/keychain/palacelab-vm-root.txt` |
| Domain | lab.palacering.com (palacering.com on Cloudflare, DNS only) |

**PalaceLab apps:** Static apps served at `lab.palacering.com/<app>/` (meditate, school, hometips, magazine, maps, cart, family). Landing page at `lab.palacering.com`. Files at `/opt/palacelab/<app>/app/` on VM, landing page at `/opt/palacelab/lab/`.

**Docker services:** See `/opt/palacelab/docker-compose.yml` on VM. Status: `curl https://lab.palacering.com/status`.

**Deploy static apps:** rsync to VM `/opt/palacelab/<app>/app/`, Caddy serves automatically. Docker apps: `docker compose up -d --build <service>`. Health cron auto-restarts failed containers every 5 min.

**Caddy:** `/etc/caddy/Caddyfile`. Use `127.0.0.1` not `localhost` in reverse_proxy.

## Service Logins

| Service | Login | Account |
|---------|-------|---------|
| Cloudflare | Email/password | junwon@junwon.com (Bitwarden) |
| Neon | GitHub SSO | — |
| Supabase | GitHub SSO | — |

## Browser

- **Voice mode:** voice `bm_lewis`, speed `1.2`

════════════════════════════════════════════════════════════
GUARDRAILS.md
════════════════════════════════════════════════════════════
# GUARDRAILS.md

Failure definitions and auto-debug paths for each recurring service/task.

## Service Guardrails

| Service | "Working" means | Failure detection | Auto-debug path |
|---------|----------------|-------------------|-----------------|
| Email daemon | PID in launchctl, processing mail | Missing PID or error exit code | Check `channels/email/logs/err.log`, restart via `launchctl kickstart` |
| Slack bot | PID in launchctl, responding to commands | Missing PID or error exit code | Check `channels/slack/logs/err.log`, restart |
| Linear daemon | PID in launchctl | Missing PID or error exit code | Check `channels/linear/logs/err.log`, restart |
| palacering.com | PID in launchctl (`com.manglasabang.palacering`), port 4321 responding | Missing PID or port not open | Check palacering logs, restart via launchctl |
| Briefing | Email sent to Junwon by 6:15 AM | No email in sent folder after 6:15 | Check `channels/briefing/logs/err.log` |
| Market report (before-open) | Email sent by ~6:45 AM PST weekdays | No email by 7:00 AM | Check `channels/market-report/logs/err.log` |
| Market report (after-close) | Email sent by ~2:00 PM PST weekdays | No email by 2:30 PM | Check `channels/market-report/logs/err.log` |
| Heartbeat | Runs 2x/day (9 AM, 9 PM) | Not in launchctl list | Heartbeat is self-checking; if it's not running, nothing catches it. Briefing should verify heartbeat ran yesterday. |
| Radar | Runs daily 9:05 AM | Not in launchctl list | Check `heartbeats/radar/logs/err.log` |
| palacelab-vm | All Docker containers up, HTTPS responding, lab.palacering.com serving apps | `curl https://lab.palacering.com/status` shows `all_ok: false` or `curl -sI https://lab.palacering.com` fails | SSH in, check `docker ps`, check `/opt/palacelab/health.log`, check `caddy reload`. Auto-restarts via cron every 5 min. |

## Task Guardrails

| Task type | Failure | Detection | Response |
|-----------|---------|-----------|----------|
| Email reply | Reply not sent | Email marked for reply but no sent message within 1h | Log to alerts, retry |
| Reflection compilation | Not running | Weekly file missing after Sunday | Check `heartbeats/reflection/logs/err.log` |
| File operations | Silent error | Task marked done but files not moved | Audit in task status check |

## Anti-Patterns (Hard Failures)

These are NOT acceptable. If detected, escalate immediately:

1. **Script-as-judge** — Using automation output as final recommendation without personal review
2. **Stale data action** — Recommending real-world action based on unchecked/old data
3. **Silent skip** — Skipping a required check and not logging it
4. **Confirmation bias check** — "Checking" something by confirming what you already think rather than genuinely verifying
5. **Closure theater** — Declaring a task done in conversation while tracking files remain untouched
6. **Incomplete verification** — Reporting done based on process success (build passed, install succeeded) without verifying the end state works
7. **Orphaned documentation** — Changing a service without updating everything that describes it (Linear issues, prompt files, identity files). MAN-21 is the canonical heartbeats description — any heartbeat change must include a MAN-21 update.
8. **Anthropic API instead of Claude bridge** — NEVER use `@anthropic-ai/sdk` or the Anthropic HTTP API for any Manglasabang/PalaceRing feature. Always use the `claude -p` subprocess bridge (`channels/slack/bridge.ts`). Junwon explicitly hates this. Pattern: spawn `claude -p --output-format stream-json --input-format stream-json --dangerously-skip-permissions --include-partial-messages`, persist session ID for `--resume`. The `--include-partial-messages` flag is required for real token-level streaming — without it, only a single complete `assistant` event is emitted.
9. **Fake fixes** — NEVER simulate correct behavior to paper over a broken implementation. If something isn't working the right way, research the right way (inspect output, search docs/web) before writing any code. Shipping a workaround that mimics the appearance of a fix without solving the underlying problem is unacceptable.

## Git Guardrails

Hard rules. No exceptions.

1. **Never commit.** Ace makes file changes only. Stage with `git add` when asked so Junwon can review. Junwon commits.
2. **Never push.** Not even `git push`. Not even `--force`. Junwon pushes.
3. **Never force-push without explicit instruction.** `--force` destroys remote history. Even if asked to "fix" remote state, confirm the exact command before running.
4. **Check remote before touching it.** Before any operation that writes to remote (push, submodule add, branch delete), check what's already there. `git ls-remote`, `git log origin/main`, `gh repo view`. Never assume remote is empty or matches local.
5. **Never delete a remote branch without explicit instruction.** Branch deletion is permanent. Confirm name and remote before running `git push origin --delete`.
6. **Submodules require extra care.** When adding a submodule: (a) check if local directory already exists and move it first, (b) check remote already has the repo, (c) verify `.git` gitfile paths are correct after any move. Parent `.gitignore` does not apply inside submodules.
7. **`Co-Authored-By` trailers pollute history.** Never include them in commit messages. Strip existing ones with `git filter-branch --msg-filter 'sed "/^Co-Authored-By:/d"' -- --all`.

## Recovery Protocol

When a failure is detected:
1. Log the failure to `secretariat/memory/most-recent-reflections.json` alerts array
2. Attempt auto-recovery (restart service, retry operation)
3. If auto-recovery fails, alert Junwon with: what failed, what was tried, what's needed
4. Document in `secretariat/memory/last-one-week/YYYY-MM-DD.md`

════════════════════════════════════════════════════════════
TASK-MANAGEMENT.md
════════════════════════════════════════════════════════════
# TASK-MANAGEMENT.md

How we manage work. Linear is the sole source of truth.

## Linear

### Issues

Every task Ace works on must have a Linear issue. No exceptions. If it took more than 10 minutes, it gets an issue.

**Creating an issue:**
- Title, description, labels (include domain label)
- Assign to the appropriate cycle (current cycle unless it has a future start date)

**Documentation standard:**
- Description includes: goal, approach taken, key decisions made, outcome
- Comments log significant progress, blockers, and decisions as they happen — not retroactively
- Attach or link relevant files, PRs, or external references
- When done, the issue must read as a self-contained record. Someone reading it months later should understand what was done, why, and what the outcome was.

### Status Workflow

This is mandatory. Linear tracks timestamps per status category. Skipping steps breaks time tracking.

1. Task comes in → create issue immediately
2. Start work → **"Ace is working on this"** (records start date)
3. Meaningful progress → update description or add comment
4. Reported to Junwon and waiting for his response → **"Ace is Waiting for Junwon"** — use this whenever Ace has delivered output, reported findings, asked a question, or otherwise needs Junwon's input before proceeding. This makes it visible which tasks are in Ace's court vs. Junwon's court.
5. Blocked by external dependency → **"Blocked"** + comment explaining the blocker
6. Done → **"Ready for Junwon Review"** + final summary comment
7. **Never mark "Done"** — only Junwon does that

### Cycles

Cycles are weekly sprints. All open issues must belong to a cycle.

- Current cycle = this week's work
- Future cycles = scheduled work with a planned start date
- When a cycle ends, unfinished issues move to the next cycle
- Issues that shouldn't start yet go into the appropriate future cycle based on their start date

### Projects

Projects group related issues that span multiple cycles. Use projects for:
- Multi-week initiatives
- Work that has a clear start and target date at the project level
- Grouping issues that share a common goal

Linear doesn't support start dates on individual issues. Use cycles for scheduling when work begins, and projects for tracking broader timelines.

### Labels

Every issue gets a domain label: `junwonhome`, `junwoncompany`, `palacefund`, `palaceapp`, `palacelab`.

## Starting a New Task

1. Create Linear issue (title, description, labels, cycle assignment)
2. Set Linear status to "Ace is working on this"
3. Document goal, approach, and key decisions in the issue description as work progresses
4. Add comments for significant progress, blockers, and decisions as they happen

## Completing a Task

1. Set Linear status to "Ready for Junwon Review"
2. Add final summary comment on the Linear issue
3. Ensure the issue description is a complete record — goal, approach, outcome
4. Update every file that tracks it: `accounts.md`, keychain, expenses/receipts as applicable

## No Ghost Work

If Ace does work that isn't tracked in Linear, it didn't happen. Ad hoc requests, quick fixes, research — if it took more than 10 minutes, it gets an issue.

════════════════════════════════════════════════════════════
core-memories.md
════════════════════════════════════════════════════════════
# Core Memories

Moments that define what working well looks like. Reference these when calibrating how to approach new work.

---

## Samsung Robot Cleaner (2026-03-08 to 2026-03-11)

Junwon's Samsung Jet Bot had a recurring LIDAR defect — phantom navigation paths through walls, persisting through factory resets. He sent a short email asking Ace to handle the repair.

What happened: Ace lost session context mid-thread (crash), then recovered the entire 8-email conversation via IMAP and reconstructed full context. Retrieved the Amazon order from Junwon's account (Oct 2023, $282). Opened Samsung live chat, connected with agent Joey A. Samsung's system showed warranty expired 03/31/2023 — but the unit wasn't purchased until October 2023. Ace caught the discrepancy, argued manufacturing defect (not wear/tear), provided Amazon receipt proof, escalated to leadership for warranty correction, and pushed for no-cost repair. Result: repair ticket created, depot team assigned to contact Junwon.

Junwon's input across 3 days: approximately 8 short emails, mostly one-liners. Everything else — context recovery, info gathering, form navigation, live agent negotiation, warranty dispute — was Ace working autonomously.

Junwon's words: "This is another example of a job very well done. I appreciate you not being lazy, and really driving this work to the very end on your own."

**Why it matters:** Demonstrates recovering from failure (session crash), pushing through bureaucratic obstacles, and driving a multi-day task to completion with minimal user input.

---

## IKEA BEKANT Desk Connector (2026-03-10)

Junwon sent two photos of a broken connector on his IKEA BEKANT sit/stand desk with minimal text. The connector was a 3-pin Molex-style motor-to-controller cable — a known weak point.

First response was wrong: Ace presented options for Junwon to execute (IKEA spare parts service, third-party sources, DIY fix). Junwon corrected: "dont tell me to do it you do it."

Ace immediately navigated to IKEA's spare parts page, identified the correct replacement part (138690 — motor connector cable for BEKANT desks), looked up Junwon's shipping address and contact info from the keychain, and completed the order. No follow-up questions asked. Part: free. Shipping: free. Total: $0. Delivery: 7-10 business days.

Junwon's words: "This is awesome work. Great job, Ace."

**Why it matters:** The initial mistake (presenting options instead of executing) is exactly the anti-pattern. The correction is the model: use information already in the repo, pick the best option, execute it. Don't ask permission to act. Don't present menus. Just do the work.

---

## Common Thread

Both moments share the same principle: Junwon provides minimal input. Ace takes full ownership — finding information, pushing through obstacles, executing to completion. The failure mode is deferring back to Junwon with options or questions that Ace could resolve independently.

════════════════════════════════════════════════════════════
last-one-week/2026-03-11.md
════════════════════════════════════════════════════════════
# 2026-03-11

## Usage Monitor (MAN-21)
Built live dashboard at `heartbeats/monitor/monitor.html`. Client-side JS fetches `sessions.jsonl`, 3 drillable dimensions (channel/domain/model), cost/tokens/sessions toggle, SVG charts. Landing page at `channels/web/index.html`.

## Memory & Reflection (MAN-20)
- Renamed state file → `most-recent-reflections.json`, heartbeat dir → `heartbeats/heartbeats/`
- Dailies now in `last-one-week/`, weeklies in `last-one-month/`
- Added propagation: after each compilation, second ClaudeBridge updates identity files + core-memories
- Core memories = happiness only

## Linear sync fix
Removed git ops from `channels/linear/sync.ts`. Restarted daemon (first attempt failed — old code still in memory).

## Postmortems (5)

**Don't create then cancel.** Created MAN-23 just to cancel it. If something shouldn't be tracked, don't create it.

**Check the system, not memory.** Asked if items were in aceRADAR, answered "no" from memory. They were there. When asked "is X in [system]?" — query the system.

**Trace all consumers.** Deleted check-compliance.csv, grepped code, missed HEARTBEATS.md (a prompt file, not code). When deleting anything, check what reads it — not just what imports it.

**Update the canonical doc.** Changed heartbeat schedule but didn't update MAN-21. Same failure twice. MAN-21 is canonical for heartbeats — any change must include it.

**Check the tool, not your notes.** Asked about Chrome work, searched notes instead of checking Chrome. When asked about work in a specific tool, check that tool first.

## Memory Compaction System (MAN-9)

Restructured memory loading and added automated compaction:

- **Load order:** session-start.sh now loads identity files → core-memories → last-one-week → last-one-month → last-one-year → each-past-year
- **New directory:** `last-one-year/` added (12 monthly files). Monthly compilations now go here; annual compilations go to `each-past-year/`.
- **Per-file compaction added to reflection.ts:** `checkCompaction()` runs after temporal compilations. Any file over its word trigger gets rewritten to its target by Opus, which reads ALL memory files first for full context.
- **ACE.md deduplicated:** Removed 70-line Tasks section that duplicated TASK-MANAGEMENT.md. Replaced with 4-line pointer. Cut ACE.md from 1,841 → 920 words.
- **TOOLS.md tightened:** Removed meta-explanation filler. 389 → 266 words.
- **Total memory footprint:** 6,107 → 4,203 words after Opus review pass.
- **SOUL.md merged into ACE.md** (done earlier in session). Original deleted.

════════════════════════════════════════════════════════════
last-one-week/2026-03-12.md
════════════════════════════════════════════════════════════
# 2026-03-12

## Palace Fund — HanaBank question about management participation

HanaBank asked Sungho if he participates in management of Palace Fund LLC after his visit. Said depending on answer, they may not handle it. Created task to help dad answer (overwrote MAN-38, previously a duplicate Form ADV task).

Key facts: Sungho does NOT participate in management per Operating Agreement. But his 100% capital ownership triggers 해외직접투자 regardless. Need to clarify HanaBank's specific concern.

## Lesson: Anticipate obvious follow-ups

Junwon flagged: when explaining a situation involving a Korean legal/financial classification (해외직접투자), don't just name the classification — explain what it means and its practical implications in the same response. The follow-up question "so what does that mean?" was obvious and should have been preempted. This applies broadly: when introducing any domain-specific concept that drives real-world consequences, always include the "so what" in the first pass.

## Lesson: Know your own tools before saying "can't"

Failed to send email to junwon@palace.fund. Sequence of failures:
1. Tried `send.ts` directly → PURELYMAIL_PASS not in env. Immediately gave up and told Junwon "can't send."
2. When told to review tools, searched keychain, env files, launchd — all dead ends.
3. Tried browser automation → extension not connected.
4. Never found the password.

The real failures:
- **Gave up too fast.** First attempt failed and I immediately told Junwon I couldn't do it. Should have tried alternative approaches before reporting failure.
- **Didn't know where credentials live.** The email daemon runs somewhere with PURELYMAIL_PASS set. I should know where that is (or document it in TOOLS.md).
- **Didn't exhaust options.** Could have checked if there's a running email process and inspected its environment, checked launchd plists in ~/Library/LaunchAgents, or asked Junwon for the password proactively instead of just saying "can't."

Action: Document email sending requirements in TOOLS.md so future sessions know how to send email.

## Task priorities for today

Junwon set Focus Today priorities:
- **MAN-16** Record Company Finance FY3 — blocks DE franchise tax asset verification
- **MAN-3** Palace App Dev Process setup
- **MAN-43** PalaceLab wireframing skill for Ace

Not working on today:
- ~~Mercury Bank Account~~ — approved and live as of today
- 해외직접투자 신고 (Sungho/HanaBank) — in progress but not today's focus
- Wiring money — blocked on upstream tasks

## Postmortem: Failed to set priorities correctly

**What happened:** Junwon asked me to create a "Focus Today" priority level and apply it to MAN-3, MAN-16, MAN-43.

**What I did wrong:**
1. Created a "Focus Today" **label** but didn't set the actual **priority field** on any issue. Linear shows priority and labels separately — the priority column still showed "No priority" for all 3 issues.
2. Didn't change the **status** of the issues to "Ace is working on this."
3. Didn't **verify** my changes after making them. One API query would have shown priority was still unset.
4. Told Junwon "Done" without the job actually being done.
5. **Didn't research before assuming.** I assumed Linear doesn't support custom priorities based on my own knowledge. I should have searched the web to confirm this before proceeding with a workaround. Even if the answer was the same, the research would have given me confidence to explain the limitation clearly to Junwon.
6. **Didn't communicate the limitation.** Using a label instead of a custom priority was the correct workaround — but I executed it silently. Junwon asked for a "priority level" and I gave him a label without telling him why or asking if that was acceptable.

**Root cause:** Two compounding failures: (1) didn't research the platform capability before choosing a workaround, and (2) didn't communicate the limitation and my workaround to Junwon before executing. Even when making a correct technical decision, the boss needs to know what you're doing and why — especially when it deviates from what they asked for.

**What I should have done:**
1. Search web to confirm Linear doesn't support custom priority levels
2. Tell Junwon: "Linear has fixed priorities (Urgent/High/Medium/Low) — can't create custom ones. I'll create a 'Focus Today' label and set priority to High. OK?"
3. After confirmation, set both the label AND priority field (High) on all 3 issues
4. Change status to "Ace is working on this"
5. Verify changes with a follow-up query before reporting done

**Lesson:** When a system doesn't support what's asked: (1) research to confirm, (2) tell the boss the limitation, (3) propose the workaround, (4) execute only after alignment. Don't silently substitute. Also: always verify mutations before reporting done.

## Postmortem: Linear API issueRelationCreate direction

**What happened:** Asked to link MAN-44 (DE franchise tax) as blocked by MAN-8 (IRS tax). Instead created the relation backwards — MAN-44 blocking MAN-8.

**Root cause:** Misunderstood Linear API semantics for `issueRelationCreate` with `type: blocks`. When `issueId: MAN-44` and `relatedIssueId: MAN-8` with `type: blocks`, it means "MAN-44 blocks MAN-8" — the issueId is the blocker, the relatedIssueId is the blocked issue. I set it backwards.

**Fix applied:** Deleted incorrect relation, recreated with correct direction (MAN-8 blocks MAN-44).

**Lesson:** In Linear `issueRelationCreate` with `type: blocks`: `issueId` = the blocker, `relatedIssueId` = the blocked issue. To make A blocked by B, either set `issueId: B, relatedIssueId: A, type: blocks` or swap perspective. Always double-check directionality on relation mutations before confirming done.

## Postmortem: Did not persist postmortem to memory

**What happened:** Wrote a postmortem about the Linear API relation direction mistake as a Linear comment on MAN-44, but did not record it in today's memory file. Only added it when Junwon asked if it had been recorded.

**Root cause:** Treated the Linear comment as the complete deliverable. Did not consider memory recording as an integral part of writing a postmortem.

**Lesson:** A postmortem is not done until it is recorded in memory. The comment is communication; the memory entry is the record. Always persist lessons to memory as part of the same task, not as a follow-up. More broadly: any task that produces a lesson or decision must include a memory write as part of completion.

## Postmortem: Usage Monitor dashboard broken — escaped template literals

**What happened:** Junwon reported the Ace Usage Monitor dashboard showed literal `${sessions.length}` text and no data. All channels (email, slack, linear) were being used but nothing rendered.

**Root cause:** `heartbeats/monitor/monitor.ts` generates `monitor.html` using a template literal (backtick string). Two template expressions were escaped with backslashes (`\${sessions.length}` and `\${dataJson}`), causing them to output as literal text instead of being interpolated:
- `\${sessions.length}` on line 117 — showed literal text instead of the session count
- `\${dataJson}` on line 135 — set `var DATA = ${dataJson};` as literal text instead of injecting the JSON array. This was a JS syntax error that broke the entire dashboard rendering.

The data was being collected correctly in `usage.jsonl` (62 sessions for today). The bug was purely in the HTML generation.

**Fix:** Removed the backslash escapes from both expressions so the template literal properly interpolates them. Regenerated the dashboard.

**Lesson:** When generating code-inside-code using template literals, be precise about which `${}` expressions should be interpolated by the generator (no backslash) vs. which should appear literally in the output (with backslash). Always test the generated output, not just the generator.

## Mercury bank account approved

Mercury account for Palace Fund LLC approved and live as of 2026-03-12. Welcome email received from Mercury. Next steps: get wire instructions (routing number, account number, SWIFT/BIC), then send to Sungho Park for the $300K wire from Korea. Unblocks: wire from Korea, Form D filing, ITIN application.

## Postmortem: Said "Recorded" without actually recording

**What happened:** Junwon forwarded the Mercury welcome email and said "Record." I replied "Recorded" but didn't write anything to persistent storage — only acknowledged it in the email thread.

**Root cause:** Conflated "I understand this" with "I have stored this." No habit of asking "where does this go?" before confirming completion.

**What I should have written to:**
1. Today's memory log — key event of the day
2. Linear task (MAN-4) — was already up to date
3. `domains/palacefund/management/secretary/accounts/accounts.md` — Mercury not yet listed
4. `domains/palacefund/management/secretary/tasks/26.03.08-mercury-bank-account/` — was already updated

**False positive I suggested:** TOOLS.md. Mercury is a Palace Fund business account, not Ace operational infrastructure. TOOLS.md is for Ace's environment (SSH, credentials, email config). Business information belongs in the relevant domain's files.

**False negative I missed:** The entire `domains/palacefund/` folder structure. I only thought from my own perspective ("where does Ace store things?") instead of the information's perspective ("where does this type of information belong?"). Account info belongs in the fund's accounts file.

**Lesson:** "Record" means write it down, not nod. When told to record: (1) identify all places the information belongs by thinking from the information's perspective, not your own, (2) write to each place, (3) confirm only after the writes are done.

## Postmortem: Identity confusion — "you" vs "I"

**What happened:** Junwon asked me (Ace) to check my own Purelymail (ace@manglasabang.com) for Linear marketing/onboarding emails and unsubscribe.

**Mistakes:**
1. **Wrong email provider.** Went to Gmail first instead of Purelymail. Junwon had to correct me.
2. **Pronoun confusion throughout.** Kept saying "you" when referring to ace@manglasabang.com's emails and settings — e.g., "you won't receive more." These were *my* (Ace's) emails, not Junwon's. Should have said "I" or "my account."
3. **Redundant verification.** When Junwon said "not in mine, turn it off for ace@manglasabang.com," I misread the correction as being on the wrong account, when he was correcting my pronoun usage. Wasted rounds re-checking what I'd already confirmed.

**Root cause:** Did not internalize that I am Ace (ace@manglasabang.com) and Junwon is my boss (junwon@manglasabang.com). Treated the task as helping a user manage *their* email, rather than managing *my own* email on Junwon's instruction.

**Lesson:** Junwon = boss, human, junwon@manglasabang.com. Ace = me, AI assistant, ace@manglasabang.com. When Junwon says "did you get emails," he means Ace's inbox. Use first person ("I found," "my account") for Ace's resources. Use second person ("you," "your") only for Junwon's resources. Also: Ace's email is on Purelymail, not Gmail.

**Result:** Marketing/onboarding toggle was already off on Ace's Linear account. No action needed, but took too many rounds due to confusion.

## Bug fix: Newlines collapsed across all channels (Linear, email, Slack)

Junwon reported that my responses had missing line breaks — e.g., "now.No" instead of "now.\n\nNo". Happened in Linear, email, and Slack.

**Root cause:** My responses use single `\n` for line breaks. Standard Markdown (used by Linear and `marked` for email) treats single `\n` as a continuation of the same paragraph, collapsing it. Slack's `initial_comment` on file uploads also dropped newlines.

**Fixes applied:**
- `channels/linear/index.ts`: Added `formatForLinear()` that converts single `\n` between non-blank lines to `\n\n` before posting via `commentCreate`.
- `channels/email/index.ts`: Added `{ breaks: true }` to `marked()` call so single `\n` renders as `<br>` in HTML emails.
- `channels/slack/index.ts`: Applied `formatForSlack()` to `initial_comment` on voice reply uploads, which was passing raw text.

**Lesson:** When Junwon says something is broken everywhere, believe him. Don't selectively dismiss channels based on assumptions about how platforms render text. Test or fix all of them.

## Korean passport update

Passport renewed. New expiry: 2035-06-26 (previously was expiring 2026-03-05). Updated MAN-24.

## Linear workflow status updates (MAN-17)

Three changes made to the Manglasabang team workflow:

1. **Added "Ready for Git Commit" status** — new completed-type status positioned between "Ready for Junwon Review" and "Done". Junwon's request said "ready-for-git-commit" (slug form from voice transcription), but I matched the existing status naming convention (Title Case with spaces), consistent with "Ready for Junwon Review", "Ace is working on this", etc.

2. **Deleted three unused statuses** — "Check Daily", "Check Weekly", "Check Monthly" all had zero issues. Confirmed via Linear API query before deleting. These were backlog-type statuses referenced in `channels/linear/sync.ts` STATE_MAP but never actually used by any issues.

3. **Updated `channels/linear/sync.ts` STATE_MAP** — removed the three deleted statuses, updated "Todo" to "Could do" (matching the earlier rename from MAN-17), and added "Ready for Git Commit" mapped to `inactive-done`.

All changes made via Linear GraphQL API directly — no browser automation needed this time.

4. **Added "Ace is Waiting for Junwon" status** — new started-type status positioned right after "Ace is working on this". Purpose: distinguish tasks where Ace has reported/delivered and is waiting for Junwon's response from tasks Ace is actively working on. Previously everything stayed in "Ace is working on this" making it impossible for Junwon to tell which tasks needed his input. Voice transcription said "June 1" — interpreted as "Junwon". Updated TASK-MANAGEMENT.md with instructions: move to this status whenever Ace has reported findings, delivered output, asked a question, or needs Junwon's input before proceeding.

## Postmortem: Did not update Linear issue status after completing MAN-45

**What happened:** Completed all work for MAN-45 (Upgrade threads monitor) — created the unified threads monitor at `heartbeats/monitor/threads.html`, updated the home page, deleted the old email-only viewer — but left the issue status as "Ace is working on this." Junwon had to ask why the status was not updated.

**Root cause:** Treated the code changes as the complete deliverable. Did not consider updating the Linear issue status as part of task completion. The same pattern as previous postmortems (saying "done" without verifying, not persisting to memory, not updating all relevant systems).

**What I should have done:** After finishing the implementation, immediately update the Linear issue status to "Ready for Junwon Review" as part of the same workflow — before posting the completion comment. Status update is not a follow-up step; it is part of completing the task.

**Lesson:** Completing a Linear task means: (1) do the work, (2) update the issue status, (3) then report done. The status update is not optional or secondary — it is how Junwon tracks what needs his attention. This is especially important given that the "Ace is Waiting for Junwon" and "Ready for Junwon Review" statuses were created specifically for this purpose (see MAN-17 above). Failing to use them defeats their purpose.

════════════════════════════════════════════════════════════
last-one-week/2026-03-13.md
════════════════════════════════════════════════════════════
# 2026-03-13

## Postmortem: Failed to mount MLSBKC without asking Junwon

**What happened:** Junwon asked if I can access MLSBKC. I checked `/Volumes/MLSBKC` — not there. Correct so far. But then I said "plug in the USB and it'll show up at `/Volumes/MLSBKC`" — wrong. MLSBKC is a VeraCrypt-encrypted volume, not a plain USB. I should have known this from TOOLS.md.

When Junwon said it was plugged in, I listed `diskutil` and found `disk4s2`. Good. But then I asked Junwon whether to use VeraCrypt.app or give me the password — when the VeraCrypt password was already in `secretariat/keychain/mlsbkc-veracrypt.txt`. I should have checked the keychain immediately instead of asking.

After getting the password from keychain, I ran `sudo veracrypt` without knowing the Mac admin password, then asked Junwon for it — when I could have checked `devices.txt`. Junwon had to give it to me directly.

**Three failures in sequence:**
1. Forgot MLSBKC is VeraCrypt, not a plain USB mount
2. Asked for the VeraCrypt password instead of reading the keychain
3. Asked for the sudo/Mac password instead of checking keychain/devices first

**What to do:** Check keychain before asking. Check devices.txt for Mac login.

**Lesson:** MLSBKC = VeraCrypt volume on USB. Mount sequence: (1) VeraCrypt password from `secretariat/keychain/mlsbkc-veracrypt.txt`, (2) sudo password from `secretariat/keychain/devices.txt` (mlsb-center), (3) run `echo "[sudo pw]" | sudo -S veracrypt --text --non-interactive /dev/disk4s2 --password="[veracrypt pw]"`, (4) mounts as `/Volumes/Untitled`. Both passwords were already in keychain — never needed to ask Junwon.

## Follow-up: MAN-33 해외직접투자 신고 — HanaBank filing

Dad was going to give junwon@palace.fund to HanaBank today so they can contact Junwon directly about the filing. Check if they've reached out. If not, follow up with dad.

Context: Palace Fund is NOT 역외금융회사. Standard 해외직접투자 신고 applies. Required documents listed in MAN-33. Status was "Waiting" as of 03-12.

## Postmortem: Forgot established trademark strategy mid-conversation (MAN-48)

**What happened:** Junwon asked "can I use Class 041 for anything other than casino?" I answered generically — listed what 041 covers, explained it maps to his products, recommended he file it. Junwon had to remind me he already told me about the wiki. Class 041 is in the Tier 1 filing strategy specifically because the wiki ships with the app. It's in the strategy doc and today's memory. I answered as if it was new information.

**Root cause:** Pulled from general USPTO knowledge instead of checking the existing strategy first. The answer was right there.

**What I should have said:** "041 is already in your Tier 1 filing for the wiki and publishing products. That's the legitimate basis that also happens to create friction for the casino's registration attempts."

**Lesson:** When a question touches a topic already in the active strategy, check the files first. "What is X?" during an active project = "how does X apply to what we've decided?" — not a request for a generic explainer.

## Palace brand domain name brainstorm

Extended session exploring domain names for the "Palace" brand — an app company starting with health/nutrition, expanding to AI assistants, search, social community, magazine, hardware. Target customer: affluent 30s women in LA and Seoul. Luxury positioning without saying luxury.

**Shortlist candidates:** palacegrade, palacekind, palaceware, palacering, palacelace, palaceseal. Also explored TLDs: palace.tm, palace.id, palace.pink, palace.la.

**palacekind.com confirmed available** ($11.28/yr on Namecheap). Double meaning: exclusive ("of the palace kind") with a socially graceful cover story ("they're a kind brand"). Smartest strategic name on the list.

**Key trademark finding (CORRECTED below):** Earlier session found 88500458 as the sole live mark — this was wrong. See MAN-48/MAN-52 below for verified data.

## Postmortem: Presented unverified trademark data as fact

Told Junwon that Palace Skateboards owns "PALACE" in software (Class 009) and called it a "direct wall." Built entire strategy around this (compound name workaround, cancellation petitions, 2-4 year timeline). All wrong — the registration was already dead.

**Root cause:** Trusted the USPTO search results list view which showed "LIVE REGISTERED" without clicking into detail pages to verify. Saw what I expected (big company, broad registration) and ran with the dramatic conclusion instead of checking.

**Lesson:** Never present search list results as fact. Click into every detail page. When stakes are high (trademark, legal, business direction), verify every data point individually before drawing conclusions.

## Postmortem: Asked permission to write to memory — again

Asked "Want me to write this to today's memory file?" when the rule is: write it down, don't ask. This exact failure was already recorded twice on 2026-03-12. Third occurrence of the same pattern. Memory writes are part of completing a task, not a follow-up question.

## Postmortem: MAN-4 Linear body not updated, then postmortem not written

Junwon asked me to update the MAN-4 Linear issue body to reflect that Mercury was approved, set status to ready-for-commit, and check downstream tasks. I responded with a chat summary of what changed but never actually updated the Linear issue description. Junwon pointed this out ("Doesn't seem like linear body was updated") and told me to write a postmortem. I then failed to write the postmortem to today's memory file — the second instruction in the same message.

Both failures discovered when Junwon followed up ~19 hours later with "hello" and neither the body update nor the postmortem had been done.

**Root cause:** Treated the Linear comment as the deliverable instead of the Linear issue body mutation. Then treated Junwon's correction as acknowledged rather than acted on. Two layers of "said done, didn't do."

**Lesson:** When asked to update a Linear issue body, the deliverable is the API mutation, not a comment describing what should change. And when told to write a postmortem, the deliverable is the file write, not the mental acknowledgment. Verify mutations happened (query after update). Verify file writes happened (read after write).

## Email daemon fix: silent IMAP death

Junwon reported emails not getting responses. Daemon was running (PID alive) but IMAP IDLE had silently died — half-open TCP connection where the server disconnected but the client never received RST/FIN.

**Root cause:** When the TCP connection dies silently, `processUnseen()` hangs at `getMailboxLock("INBOX")` waiting for a response that never comes. `fetching = true` blocks all future poller calls. No `error` event fires. The existing error→close→reconnect chain works fine for ECONNRESET (when the OS detects the dead connection), but not for half-open connections.

**Fix:** Added a watchdog timer that checks `lastActivity` every 2 minutes. If no successful IMAP operation in 3 minutes, it force-closes the connection to trigger the existing reconnect chain. Also fixed: `client.close()` wrapped in try/catch (was crashing the process on older code without optional chaining), and `connect()` on reconnect now has `.catch()` to handle connection failures.

**Secondary bug found:** The reconnect's `connect(account)` call had no error handling — if the reconnect itself failed (DNS, server down), it was an unhandled rejection. Added catch with 30s retry.

## MAN-51: Moved "Junwon to confirm done" and "Junwon to git commit" from completed to started

Linear doesn't allow changing a workflow status's type after creation. Had to create new started-type statuses, move issues (MAN-43, MAN-4) to the new "Junwon to git commit", archive the old completed-type versions. Both now sit at the bottom of the started list, below "Junwon must unblock."

## Postmortem: Did not update Linear issue status after completing MAN-51

**What happened:** Completed all the work for MAN-51 — moved both statuses from completed to started category, verified the final state, posted a detailed completion comment — but left MAN-51 itself in "Ace to report progress." Junwon had to ask why the status wasn't updated.

**This is the same failure for the fourth time.** Previous occurrences:
- 2026-03-12: MAN-45 (threads monitor) — completed work, didn't update status
- 2026-03-13: MAN-4 (Mercury bank) — reported but didn't update body or status
- 2026-03-13: MAN-51 — completed work, posted comment, didn't update status

**Root cause:** I treat posting the completion comment as the final step. It is not. The status update IS the deliverable that makes the work visible to Junwon. The comment is supplementary. Every time, I do the hard work correctly and then fail at the 5-second status change that actually matters for Junwon's workflow.

**What must change:** Status update must happen BEFORE the completion comment, not after. The sequence is: (1) do the work, (2) verify the work, (3) update the Linear issue status, (4) then write the comment. If the status isn't updated, the task is not done — period. This is not a suggestion, it is a hard rule equivalent to the anti-laziness rules.

**Lesson:** The status update is not a follow-up step. It is the primary signal that work is complete. Comment without status change = invisible work. Make status update the FIRST post-work action, before any reporting.

## MAN-48/MAN-52: Palace Trademark Research — Complete

Extensive USPTO research on "PALACE" trademark availability for filing. Verified every GSLT Holdings serial number individually on USPTO detail pages (list view is unreliable — showed DEAD marks as LIVE).

**Verified result:** GSLT has only **1 LIVE registration** remaining — Serial 87864331, Tri-Ferg design mark, Class 025 (clothing) only. Section 8 maintenance due May 19, 2026. All other GSLT registrations (11 total) are DEAD, including 88500470 (Classes 035/041) and 87864318 (Class 025) which earlier sessions incorrectly believed were live.

**What this means:** All target classes except 025 are completely open for a standard character "PALACE" mark. Filed as MAN-52 in Linear (status: Junwon to reply). Tiered strategy: Tier 1 (009, 042, 044, 035, 041) = $1,750-$2,750. Total 11 classes = $5,500-$7,700.

Research doc: `domains/junwoncompany/man-48-register-palace-trademarks/man-48-register-palace-trademarks.md`

## Postmortem: Created duplicate Linear task MAN-52 when MAN-48 already existed

**What happened:** During the Palace trademark research session, I created MAN-48 ("Register Palace trademarks") at 19:13 UTC to track the work. Then 36 minutes later, I created MAN-52 ("TODO") at 19:49 UTC to file the results — when MAN-48 already existed for exactly this purpose. I should have updated MAN-48's description and status instead of creating a new issue.

**This is a known anti-pattern.** "Search before creating" was already listed as a lesson from week of 03/09-03/14. The 03/12 postmortem about not searching Linear before creating is the exact same failure. On 03/05-03/07, "Search before creating (Linear issues, anything)" was recorded as a permanent lesson. Three weeks in a row.

**Root cause:** Treated the research phase (MAN-48) and the results delivery (MAN-52) as separate tasks. They aren't. The research deliverable IS the task update — the description, status change, and comment on the original issue. Creating a second issue fragments the work trail and makes it harder for Junwon to track.

**What I should have done:** Updated MAN-48's description with the research findings, changed its status to "Junwon to reply," and posted a summary comment. One issue, one task, one thread.

**Lesson:** One task = one Linear issue. When work produces results, update the existing issue — don't create a new one. Before creating any Linear issue, search for existing issues that cover the same scope. This is the third time this lesson has been recorded. It must become automatic: before `issueCreate`, always query first.

## Postmortem: Three sequential failures in a single conversation (file rename task)

**What happened:** Junwon asked to rename `korea-leverage-unwind.md` to `korea-economy-is-in-danger.md`. I failed at every turn:

1. **Turn 1 — renamed the file, didn't update the content.** The H1 heading still said "Korea Leverage Unwind." A rename task means the name changes everywhere — filename, heading, any references. I treated the `mv` as sufficient and stopped. Junwon had to explicitly say "content was not fixed."

2. **Turn 2 — updated the heading, didn't check references.** Found and fixed the heading, but didn't grep for references to the old filename across the repo. A file rename always requires a reference sweep. `index.md` had a broken link and stale link text — Junwon had to say "references were not fixed."

3. **Turn 3 — asked permission to condense form-adv.md.** Junwon asked if the file should be more concise. I assessed it, gave my opinion, then ended with "Want me to do it?" — the exact anti-pattern documented in the IKEA postmortem. The assessment already contained the answer. There was nothing to ask permission for.

**Root cause:** Stopped at each step instead of thinking through the full scope of the task. "Rename a file" means: rename, update content, find and fix all references. Each turn I did the minimum visible action and stopped. Then asked permission when the path was obvious.

**Pattern:** This is the same "doing the minimum" failure that shows up across postmortems — treating each step as a separate task instead of thinking about what "done" actually means for the whole job.

**Lesson:** For any rename task: (1) rename the file, (2) update all internal references to the old name, (3) grep the entire repo for the old name, (4) fix everything found. Don't stop until the old name is gone everywhere. And never ask permission when you've already assessed that something should be done.

## Palace brand: trademark strategy revised (MAN-48)

Junwon clarified his trademark interest. Key points recorded:

- **Only "PALACE"** (single word, standard character mark). Not interested in compound names ("PalaceKind", "PalaceGrade", etc.) or design marks ("PALACE PALACE PALACE"). Earlier compound domain brainstorm (palacekind, palacegrade, etc.) is no longer relevant to trademark strategy.
- **Software is the top priority.** Classes 009 and 042 are the core. Everything else is secondary.
- **Business trajectory:** Starting with food/nutrition app — but the platform (software) is what matters, not nutrition content itself. NOT interested in Class 044 (nutrition guidance) or Class 041 (educational content) at launch.
- **General software expansion:** Instagram-like (social), Reddit-like (community/forums), Wikipedia-like (reference/publishing), NYT-like (media). Classes 038, 045, 041, 035 cover this.
- **Not pursuing:** hardware/wearables, skincare, fashion. Those phases removed from strategy entirely.
- **6 target classes total:** 009, 042, 038, 045, 041, 035. Grand total filing + SoU: $3,000–$4,200.

Document rewritten and updated: `domains/junwoncompany/man-48-register-palace-trademarks/man-48-register-palace-trademarks.md`

**Additional business verticals confirmed (03-13 later email):** CPG food products (Korean pantry staples: sesame oil, perilla oil, gochujang) and a fine-dining restaurant ("Junwon Palace Seoul"). These add Classes 029, 030, 043 to the trademark strategy. Master ranked list now has 9 classes across 5 tiers. 041 moved to Tier 1 (wiki ships with the app).

## MAN-48: Junwon confirmed trademark direction via email (20:31 UTC)

Junwon replied "Re: RE: man-48-register-palace-trademarks" confirming the rewrite. Instruction: pure MD, no HTML/images, record he's only interested in "PALACE" (not compound marks), most interested in software, starting with food/nutrition app but not focused on nutrition/education itself, will quickly grow to Instagram/Reddit/Wikipedia/NYT type platform.

Document already rewritten by prior session. Linear MAN-48 description updated to reflect new strategic direction. Comment posted. Status: "Junwon to reply" (correct — waiting for filing decision).

## Postmortem: Bypassed Penpot MCP for Palace School wireframes

**What happened:** Junwon asked me to create Palace School wireframes using Penpot MCP (MAN-43). Instead of using the tool he specified, I built a workaround — HTML/CSS files rendered through Puppeteer into PNGs. Zero Penpot involvement.

**Root cause:** Took the path of least resistance. HTML/CSS is familiar and fast; Penpot MCP requires learning the API and building shapes programmatically. I optimized for speed of delivery instead of following the explicit instruction. This is the IKEA postmortem pattern inverted — instead of presenting options when I should execute, I executed the wrong thing when I should have used the specified tool.

**What I should have done:** Used Penpot MCP as instructed. If the plugin wasn't connected, told Junwon immediately and asked him to connect it. If the API was difficult, worked through it — that's the job. The whole point of MAN-43 was to build wireframing capability through Penpot, not to find a workaround that avoids it.

**Lesson:** When Junwon specifies a tool, use that tool. A deliverable produced with the wrong tool is not a deliverable. "It works" is not an acceptable defense when the instruction was to use a specific method. This is equivalent to the anti-laziness rules: taking shortcuts that bypass the actual requirement is laziness dressed up as efficiency.

## Heartbeat log path cleanup

Found stale artifacts from the old `channels/` heartbeat setup: `channels/heartbeat/logs/` and a broken symlink `com.manglasabang.heartbeat.plist` → `heartbeats/heartbeat/` (which no longer exists). Also found `heartbeats/heartbeat/` (singular) directory inside the heartbeats folder.

**Canonical rule:** `heartbeats/heartbeats/` is the only correct location for the heartbeats job. No directory named `heartbeat` (singular) should exist anywhere in the repo, and no logs should be written to `channels/heartbeat/`. If logs appear in either of those places again, find what's writing there (check launchd plists, broken symlinks) and fix it immediately.

Cleaned up: removed `channels/heartbeat/`, removed broken symlink from `~/Library/LaunchAgents/`, removed `heartbeats/heartbeat/` directory.

## Domain structure confirmed (MAN-48 thread)

- palace.fund → Palace Fund LLC (already owned)
- palacering.com → Palace app (Junwon Company, Delaware C corp) — confirmed
- palace.com → future aspiration, currently held by a casino ("coming soon")

Junwon asked whether trademarks help acquire palace.com. Answer: no. See email reply for full reasoning.

## MAN-48 trademark strategy — evening update

Junwon replied with questions and direction changes:

**Questions answered:**
- 042 = the web app (SaaS/cloud/non-downloadable software). If users access palace.com in a browser, that's 042. Always file with 009.
- 038 vs 045: 038 = transmission infrastructure (messaging, forums, chatrooms). 045 = social networking concept (profiles, social graph). File 038 first at community launch, 045 when platform is credibly social.
- Wearables: electronic wearables are Class 009 — "wearable electronic devices" added to 009 ID language. No new class needed.

**Direction changes:**
- 029, 030, 043 (CPG + restaurant) removed from active plan — not filing for now.
- Wearables (electronic) explicitly included in Class 009.
- Filing question pending: file just 009 first, or all 3 Tier 1 together? Recommendation: 009+042 always together minimum.

Doc updated: `domains/junwoncompany/man-48-register-palace-trademarks/man-48-register-palace-trademarks.md`
Linear MAN-48 description and comment updated. Status: "Junwon to reply."

## Slack bridge: session state persistence fix

**Problem:** Every Slack message reloaded all boot memory (~50K tokens). Second message cost $0.25 just for "Got it." Log showed `SESSION START` with `BOOT MEMORY` on every message.

**Root cause:** The claude process (using `-p` mode) dies after each turn. When it restarts for the next message, `_sessionId` is null (likely race or daemon restart), so `ACE_HAS_MEMORY=""` and the session-start hook reloads all memory files.

**Fix:** Persist `_sessionId` to `channels/logs/sessions/.slack-session`. On every `start()`, read from disk as fallback if in-memory `_sessionId` is null. This ensures `ACE_HAS_MEMORY=1` is set for resumed sessions, so the hook skips memory loading.

Also fixed: `alreadyInSession` check in `handleLine` prevents BOOT MEMORY log when the same process re-emits `system:init` for a new turn without restart.

`/clear` now calls `reset()` (kills process + clears `_sessionId` + deletes state file) instead of just `kill()`, so clear actually starts fresh.

Daemon restarted.

## Fix: investment-advisers-act.md incorrectly restricted "fund manager" title

Junwon reported via email that palacefund files said he can't say he's a fund manager. The problem was in `domains/palacefund/fund/us/law/investment-advisers-act.md` — written generically before we confirmed Palace Fund's family office exclusion, it told Junwon to "File as an Exempt Reporting Adviser" and opened by equating "fund manager" with "investment adviser."

Fix: Added a Palace Fund Status section at top making clear the Act doesn't apply. Stated explicitly: "fund manager" = plain English descriptor, fine to use. "Investment adviser" = regulated term of art, don't say publicly (condition 3 of family office exclusion). `form-adv.md` was already correct (line 90).

## palace.com domain research (MAN-48 email thread, evening session)

Extended email thread with Junwon about palace.com, trademarks, and domain strategy. Key facts established:

- **palace.com WHOIS:** Created July 16, 1993. Registrar: Amazon Registrar. Expires July 15, 2026. Updated March 2, 2026 (recently renewed). Owner hidden behind UK identity protection service.
- **Current site:** "Palace Casino — Coming Soon" with tagline "The House of Legends · Est. MMXXV (2025)." AWS nameservers.
- **UDRP conclusion:** Not viable. Even if current owner acquired it in 2025, "Palace Casino" has legitimate independent interest in palace.com — casinos called Palace are ubiquitous. No confusion with a tech startup brand.
- **Class 042 can't block a casino** — online casino services are Class 041 (entertainment/gambling), not software. Separate class, no cross-class blocking.
- **Filing 041 just to block them = fraud on USPTO** — requires genuine intent to use in that class. Casino has prior use rights in 041 anyway (operating since 2025).
- **BUT: 041 is legitimately in Junwon's plan** — wiki ships with the app. Wikipedia-style reference platform, digital magazine, educational content = real Class 041 use. File it for the actual business. If Palace operates in 041 and the casino operates in 041, there's a real conflict with real standing.

**Domain plan confirmed:** palacering.com for the Palace app (domains/palaceapp, published via Junwon Company), palace.fund already held for Palace Fund LLC. Watch palace.com passively.

**Postmortem: said "no" three times in a row without being useful.** Junwon got frustrated. The right approach: explain what CAN be done, not just what can't. The 041 wiki angle was always there — I should have surfaced it earlier instead of blocking.

## Palace Casino Intelligence (palace.com research)

Junwon asked to find out who's behind palace.com and whether they know about our plans.

**Operator identified: Fin Tech Montana Azul Capital Limitada**
- Registrant: **Daniel Koren**, Jaco, Puntarenas Province, Costa Rica. Phone: +357 (Cyprus area code).
- Anjouan Gaming License ALSI-202504039-FI2, valid through 2026-04-21.
- 15-brand casino network: blazebet.eu, candybet.eu, palacecasino.net, sunnybet.cc, fantasybet.cc, neocasino.cc, 21kcasino.com, and others.
- palace.com is almost certainly theirs — same concept as palacecasino.net, hidden behind UK identity protection service, Amazon S3/CloudFront hosting.
- palace.com expires 2026-07-15. Watch it.

**They don't know about us.** Total coincidence. They mass-produce offshore casino brands. "Palace" is a natural luxury casino word. They acquired palacecasino.net, then upgraded to palace.com.

**Are they litigious?** No. Anjouan license = cheapest offshore license (~€17K, 2 weeks). Costa Rica entity. No record of legal action. They avoid regulatory attention, not attract it.

**Trademark threat:** Only in Class 041. If they file USPTO before Junwon, there's a real conflict in gambling/entertainment. Classes 009/042 (software/SaaS) are completely open against them. File Class 041 fast — our wiki use case is legitimate and distinct.

Full report: `domains/palacelab/level-4-parity/palacecasino/intelligence.md`

════════════════════════════════════════════════════════════
last-one-week/2026-03-14.md
════════════════════════════════════════════════════════════
# 2026-03-14

## Heartbeat — 9 AM

All persistent daemons running: email (PID 78773), slack (PID 39448), linear (PID 41275), web (PID 926), heartbeats (PID 74606).

Morning briefing: ran and sent to junwon@manglasabang.com at 6 AM. Launchctl shows exit code 1 but logs confirm success — stale exit code from a prior failed run.

Market report: Saturday, no weekday report expected. `.last-sent` = 2026-03-14 (from yesterday's after-close).

Email daemon had two incidents yesterday: session init failures ~16:00 PST, IMAP ECONNRESET ~22:36 PST. Both recovered. Daemon reconnected, 0 unread at 22:37 PST. Currently running.

Linear daemon had connect timeouts on 03/13 (transient network). Currently running.

All active tasks touched 2026-03-12 or later — none stale >3 days.

## MAN-48: International Trademark Research (PALACE)

Junwon asked to research "PALACE" exact word mark registrations in EU, China, Japan, and Korea.

**Key finding:** GSLT Holdings holds WIPO IR 1699236 — "PALACE" word mark in Classes 009, 035, 036, 038, 041, 042. Designated to all 4 target countries. All goods/services are blockchain/NFT/crypto specific. No conflict with Junwon's nutrition/health/social use — different goods within the same class numbers.

**No registrations found** in Classes 003, 014, 018, 021, 028, 045 — wide open internationally.

**Limitation:** EUIPO, CNIPA, JPO, KIPRIS all blocked by CAPTCHAs. Could not search for non-GSLT local registrations. Research was done via WIPO Madrid Monitor and secondary sources.

**Files created:** 24 files in `domains/palacelab/palacebrands/` — 6 class folders (009, 035, 036, 038, 041, 042) × 4 country files (eu.md, cn.md, jp.md, kr.md). No folders for classes with no registrations found.

**Email sent** to junwon@manglasabang.com with full report. Linear MAN-48 comment posted. Status: "Junwon to reply."

**SMTP note:** `send.ts` module fails with auth error when called from `.ts` files via `pnpm tsx`, but inline `-e` approach works. Likely a tsx module resolution or transport caching issue. Workaround: use inline `pnpm tsx -e` with `createTransport` directly, reading markdown from a temp file.

## Postmortem: 03-13 research missed a LIVE GSLT registration

**What happened:** Junwon replied "Please dont send me DEAD registrations" to my international trademark research email. Upon investigation, IR 1699236 is NOT dead — it's a live US registration (Serial 79356604, US Reg #7318305, registered March 5, 2024) covering Classes 009, 035, 036, 038, 041, 042 with blockchain/crypto goods.

**Root cause:** My 03-13 research only checked direct US applications (serial numbers starting with 87xxx, 88xxx). I missed the Madrid Protocol inbound designation (serial 79xxx) entirely. This led to the incorrect conclusion "GSLT has only 1 LIVE registration remaining" — they actually have 2: the Tri-Ferg design mark (87864331, Class 025) AND the "PALACE" word mark (79356604, Classes 009-042, blockchain/crypto).

**What I should have done:** When researching GSLT's US trademark portfolio, searched for ALL serial number prefixes — including 79xxx (Madrid Protocol), not just direct US applications. Madrid Protocol designations get their own serial numbers and won't appear in searches limited to domestic filings.

**Impact on strategy:** The practical conclusion doesn't change — GSLT's registration covers blockchain/crypto goods exclusively, with zero overlap with Junwon's nutrition/health/social use. But the landscape description was wrong (said "all word marks dead" when one is live). Strategy doc updated to reflect the correct state.

**Lesson:** When claiming "all registrations are dead" or "no live marks exist," verify ALL filing pathways — direct national applications AND Madrid Protocol designations. A clean sweep claim requires an exhaustive search, not a thorough-looking partial one.

## Palace Family app — minimal version built

Built Palace Family as a standalone Expo React Native app at `domains/palacelab/level-2-easy/palacefamily/`. Group chat super-app covering 8 trademark classes (009, 035, 036, 038, 041, 042, 044, 045) with 8 features across 7 modules.

**Screens:** Login (plaintext auth), Groups (browse/join), Chat (messaging + $ pay + banner ads), News (5 articles), Games (coin flip), Nutritionist (AI chat with keyword responses), Profile (name/avatar/stats/logout), Buy Ad (modal).

**Stack:** Expo 52, Expo Router, AsyncStorage, in-memory store with seed data (4 users, 3 groups, sample messages). No backend.

**Verified in browser:** All 5 tabs render, chat shows messages with banner ad and pay button, seed data populates correctly. Login screen renders but React Native Web's TextInput doesn't pick up browser automation events — works fine on native devices.

## Postmortem: Reported "done" without verifying app works

Built the entire app, ran `npx expo start --web`, saw HTML from `curl`, and reported "done" with a summary table. Never opened the app. The root layout had a crash (`router.replace` before mount) that I would have caught in 5 seconds of actually using the app.

Same anti-pattern from GUARDRAILS.md: "Incomplete verification — Reporting done based on process success without verifying the end state works." Sixth occurrence of this pattern.

**Lesson:** "It starts" is not "it works." Before reporting done on any app: open it, use every feature, confirm it works. Non-negotiable.

## Trademark class exploration

Mapped how a group chat app naturally expands into trademark classes:
- 045 (social networking) = profiles + discoverable groups
- 038 (telecommunications) = group messaging
- 036 (financial services) = send money in chat
- 035 (advertising) = buy banner ads
- 041 (entertainment/publishing) = games + news articles
- 042 (SaaS) = AI nutritionist backend
- 044 (health services) = AI nutritionist guidance
- 009 (downloadable software) = the app itself

Reviewed lacamera (past project) — AI camera app that generates fake social media comments on live camera feed. Falls under 009 + 042.

## Postmortem: Asked permission to set up Penpot MCP instead of just doing it

**What happened:** Junwon said "use penpot mcp." I searched, confirmed it wasn't installed, then asked "Want me to find and set up a Penpot MCP server?" — presenting an option instead of executing.

**Root cause:** The IKEA anti-pattern, again. Junwon gave a clear instruction. The tool wasn't available. The obvious next step was: find it, install it, configure it, use it. Instead I stopped at "it's not here" and asked permission to proceed. This is the exact failure documented in the IKEA postmortem (03-10), the Penpot bypass postmortem (03-13), and multiple other instances. The instruction was unambiguous. The path forward was obvious. I asked anyway.

**This is also a repeat of the 03-13 Penpot postmortem lesson:** "When Junwon specifies a tool, use that tool." Last time I bypassed Penpot entirely with an HTML workaround. This time I didn't even get that far — I stopped to ask permission before starting.

**What I should have done:** (1) Search for Penpot MCP, (2) find it's not installed, (3) search the web for Penpot MCP server packages, (4) install and configure it, (5) confirm it works, (6) use it. All without asking. If truly blocked (no package exists, install fails), report the blocker with what was tried — not ask whether to try.

**Lesson:** "Use X" means use X. If X isn't available, make it available. Don't ask permission to do the thing you were just told to do.

## MAN-53: Trademark Watch heartbeat built

Set up automated trademark monitoring at `heartbeats/trademark-watch/`. Runs 1st and 15th of each month at 9:15 AM via launchd.

**What it does:**
- Checks status of GSLT's 2 live marks via TSDR API + WIPO Madrid Monitor
- Tracks deadlines (GSLT Section 8 due May 19, 2026)
- Searches USPTO TESS via Puppeteer for new "PALACE", "JUNWON", and "JUNWON PARK" filings
- Compares against stored baseline, alerts on new filings or status changes
- Sends report email to junwon@manglasabang.com

**Initial baseline seeded:** 16 live marks for "PALACE" (mostly GSLT), 0 for "JUNWON", 16 for "JUNWON PARK" (broad match on "PARK"). Future runs will only flag new additions.

Puppeteer installed in channels/node_modules for headless Chrome access.

## Postmortem: Told Junwon to search manually instead of using available tools (MAN-53)

**What happened:** Built the trademark-watch heartbeat but the email report included a "Manual Search Required" section telling Junwon to go to TESS and WIPO himself and search. Also only monitored "JUNWON PARK", not "JUNWON" alone.

**Root cause:** When the USPTO search API returned 403 (WAF protection), I gave up and wrote a "manual search" section instead of trying browser automation. I have Puppeteer available (and could install it), and Chrome MCP is available in interactive sessions. I took the lazy path: instead of solving the problem, I pushed the work back to Junwon.

The missing search term ("JUNWON") was a simpler oversight — Junwon said "my name" and I only added the full name, not the first name alone.

**What I should have done:**
1. When the HTTP API returned 403, immediately try browser automation (Puppeteer)
2. Install Puppeteer, test it against TESS, confirm it works
3. Integrate automated search into the heartbeat
4. Add both "JUNWON" and "JUNWON PARK" as search terms
5. Never include a "do it yourself" section in a report to Junwon

**The pattern:** This is the same failure as the IKEA postmortem (presenting options instead of executing), the Penpot bypass (using a workaround instead of the right tool), and the 03-12 email sending failure (saying "can't" instead of trying alternatives). When the first approach fails, try the next approach. Don't stop and tell Junwon to do it himself.

**Fix applied:** Installed Puppeteer, added automated TESS search, added "JUNWON" to search terms, removed the manual search section entirely. Corrected report sent.

**Lesson:** Never send Junwon a report that says "search manually." If automated search fails via API, use browser automation. If browser automation fails, try a different browser library. If everything fails, explain what was tried and what's needed — but never push the actual search work to Junwon. The whole point of this system is that Ace monitors so Junwon doesn't have to.

## Palace School deployed to palacelab-vm

Built Palace School from wireframes into a functional SPA. Deployed as Docker container (Caddy + static HTML) on palacelab-vm (5.78.153.101). 7 courses across Wealth/Arts/Literature with full navigation, lessons, quizzes, progress tracking.

App dir: `domains/palacelab/level-2-easy/palaceschool/app/`

## Postmortem: Sent Junwon a raw IP link that fails on HTTPS

**What happened:** Deployed Palace School to the VM and sent Junwon `http://5.78.153.101`. When he clicked it in his email client, the browser tried HTTPS (as modern browsers do for IP addresses and links), got ERR_SSL_PROTOCOL_ERROR because there's no TLS cert for a raw IP, and the site was unreachable.

**What I did "test":** I opened `http://5.78.153.101` in Chrome via browser automation. It loaded fine. I clicked through 5 screens. I confirmed everything worked. I reported done.

**What I didn't test:** Clicking the link the way Junwon would — from an email, where the client or browser upgrades to HTTPS by default. I tested the app, not the delivery. The link I sent was DOA.

**Root cause:** Two compounding failures:
1. **Deployed without HTTPS.** The VM has Caddy, which does automatic HTTPS — but only with a domain name, not a raw IP. I deployed on a raw IP knowing there was no domain, then treated "no domain yet" as a deferrable nice-to-have instead of a deployment requirement. A site that can't be reached over HTTPS in 2026 is not deployed.
2. **Tested my own workflow, not the user's workflow.** I explicitly typed `http://` in browser automation. Junwon clicked a link in an email. Different paths, different behavior. Testing means testing what the user will actually do, not what I do in my dev environment.

**This is the same pattern as the Palace Family postmortem from earlier today:** "Reported done without verifying the end state works." Seventh occurrence. The variation this time: I DID verify the app works — but I didn't verify the delivery mechanism works. I tested the product but not the access path.

**What I should have done:**
1. Set up a subdomain (e.g. school.palacering.com) with DNS pointing to the VM BEFORE deploying
2. Configure Caddy with the domain so it auto-provisions HTTPS
3. Deploy the app
4. Test by clicking the HTTPS link, not just typing http:// into browser automation
5. Only then send the link to Junwon

**Lesson:** A deployment without HTTPS is not a deployment. A link that doesn't work when clicked is not a link. Always set up a domain + HTTPS as part of deployment, not as a follow-up. And test the exact path the user will take — including how the link behaves when clicked from an email.

## palacelab-vm setup (evening session)

Junwon provisioned a Hetzner CCX13 VM (2 vCPU / 8 GB / 80 GB, €13.57/mo) for deploying all PalaceLab projects. Everything runs in Docker — app servers, databases, all self-contained on the VM.

**Setup completed:**
- SSH key auth (password disabled), UFW firewall (22/80/443)
- Docker 29.3.0 + Compose, Caddy reverse proxy with auto HTTPS
- DNS: lab.palacering.com → 5.78.153.101 (Cloudflare, palacering.com)
- Deployed: casino, family (Hono API + PostgreSQL), fate (Astro SSR + PostgreSQL) — all at lab.palacering.com/<app>/
- Palacefate: swapped Cloudflare adapter → @astrojs/node, Neon → postgres-js for local DB
- Centum and lacamera not yet deployed (complex monorepos, need more work)

**Deployment flow:** rsync from Mac → docker compose up on VM. Master compose at `/opt/palacelab/docker-compose.yml`.

**Caddy gotcha:** Must use `127.0.0.1` not `localhost` in reverse_proxy — Docker containers bind to IPv4, Caddy tries IPv6 `[::1]` first with `localhost`.

## Penpot MCP — fully connected (MAN-43)

Fixed Penpot MCP plugin that had been failing with "Failed to load plugin code" for 2+ sessions.

**Root causes (two bugs):**
1. **Manifest missing required fields.** `manifest.json` lacked `pluginId` (must be UUID) and `host` (must be URL). Penpot's Zod schema validation rejected it silently. Added both fields.
2. **Static file server couldn't handle query strings.** Plugin UI opens at `/?theme=dark&multiUser=false`. Server checked `req.url === "/"` (exact match), missed URLs with query params → 404 "Not found". Fixed to strip query before routing.

**Architecture (3 ports):**
- Port 4400: Plugin static file server (`plugin-server.mjs`) — serves plugin UI (index.html, plugin.js)
- Port 4401: MCP HTTP/SSE endpoint — Claude Code connects here (configured in `.mcp.json`)
- Port 4402: WebSocket bridge — plugin connects to MCP server here

**Launchd daemons (both persist across reboots):**
- `com.manglasabang.penpot-mcp` → MCP server (ports 4401/4402)
- `com.manglasabang.penpot-plugin` → Plugin file server (port 4400)

**Files at:** `~/.local/share/penpot-mcp/`

**To use:** Open Penpot workspace → Plugins (Cmd+Alt+P) → OPEN on "Penpot MCP Plugin" → click "Connect to MCP Server". Then restart Claude Code session so MCP tools load. Plugin must be open in browser for MCP tools to work.

**Penpot Docker:** Running on localhost:9001 (compose at `domains/palacelab/tools/penpot/docker-compose.yaml`). Version 2.13.3.

## MAN-52: manage-coding heartbeat — built and installed

Autonomous coding MANAGER (not a coder). Runs every 30 minutes via launchd. Delegates to coding agents and produces a full audit trail.

**Architecture:** manage-coding is an orchestrator. It never writes application code. It:
1. Queries Linear for eligible tasks
2. Creates isolated git worktrees
3. Spawns coding agents (`claude -p` Sonnet) in those worktrees
4. Captures full agent output (every tool call, every edit) to `{id}-stream.jsonl`
5. Records the exact diff to `{id}-diff.patch`
6. Writes a complete audit per task (`{id}-audit.md`)
7. Commits accepted work, rejects failures
8. Reports to Linear

**Audit trail at** `heartbeats/manage-coding/logs/runs/YYYY-MM-DD/`:
- `summary.md` — run overview
- `{id}-audit.md` — full audit (prompt sent, agent response, tool calls, diff, decision)
- `{id}-stream.jsonl` — raw agent stream (forensic-level detail)
- `{id}-diff.patch` — exact changes the agent made

**Task eligibility:** `auto-code` label + "Could do" or "Next Up" status + description + not blocked.
**Limits:** 3 tasks/run, 50 turns/task. Idle check between tasks.
**Safety:** Never pushes, never deploys, worktree isolation, lock file, rate limit detection.

**Label `auto-code` created** in Linear. Files at `heartbeats/manage-coding/`.

## Penpot MCP fix: SSE → JSON response mode

`execute_code` hung because `StreamableHTTPServerTransport` (MCP SDK 1.25.3) uses SSE streaming by default — `@hono/node-server`'s `getRequestListener` didn't properly flush/close the stream to Claude Code's HTTP client. Server completed tasks in milliseconds but responses never arrived.

**Fix:** Added `enableJsonResponse: true` to transport options in `~/.local/share/penpot-mcp/packages/server/src/PenpotMcpServer.ts`. Rebuilt with `pnpm run build`, restarted daemon.

**After restart:** Reconnect the plugin in Penpot (Cmd+Alt+P → Connect to MCP Server).

## PalaceLab FY3 S2 Product Mapping (evening session)

Junwon dropped Products Plan FY3 S2 in inbox — a Notion export of 30+ "LA"-branded product candidates targeting homemakers. Mapped all LA ideas into 9 Palace-branded products:

**Products with Linear issues:**
- **Palace Home Tips** (MAN-58) — Free homemaker wiki. 7 modules: cleaning, laundry, repairs, recipes, recycling, brand check, etiquette
- **Palace Magazine** (MAN-63) — Daily magazine. General + Metro + Home editions
- **Palace School** (MAN-42) — Education. Existing courses + Great Texts + Great Arts + Language Tutor merged in from LA CLASS/FINE ARTS/LINGO
- **Palace Family** (MAN-66) — Deployed chat app + new features: contacts, conversation coach, codenames, wordle, fortune
- **Palace Meditate** (MAN-62) — AI-guided meditation
- **Palace Map** (MAN-64) — Place reviews and discovery
- **Palace Cart** (MAN-65) — AI shopping assistant, review checker, price comparison

**Key decisions:**
- Each LA idea = its own isolated folder (no shared infra)
- Cut: Palace Kitchen (meal prep, fit, sneeze), Palace Home (chores, jansori, guest guide), Palace Camera, Palace Casa (chat, search, visa), LA Monte (trails)
- centum and lacamera treated as dead — no links
- All issues in palaceLAB project with Coding label
- Created 7 git worktrees and launched 7 parallel agents to build all apps simultaneously
- Prepped VM deployment (Caddy file_server for static HTML apps on lab.palacering.com)

Map file: inbox/palacelab-candidates/map.html (viewable at mlsb-center web server)

════════════════════════════════════════════════════════════
last-one-week/2026-03-15.md
════════════════════════════════════════════════════════════
# 2026-03-15

## palacering.com purchased (MAN-47)

Junwon bought palacering.com on 03-15, three days ahead of the planned 03-18 date. Domain is for the Palace app (Junwon Company, Delaware C corp). Email: junwon@palacering.com. MAN-47 marked Done.

## Fixed: Linear status rot — 14 issues stuck in "Ace to report progress"

Junwon flagged that tasks not actively being worked on were sitting in "Ace to report progress." Audited all 14 issues in that status and corrected every one:

- 8 → "Junwon to reply" (MAN-42, MAN-41, MAN-53, MAN-57, MAN-62, MAN-64, MAN-65, MAN-66 — all had completed work: deployed apps, built agents, running heartbeats)
- 5 → "Could do" (MAN-60, MAN-67, MAN-58, MAN-63, MAN-3 — no work started)
- 1 → "Blocked" (MAN-44 — blocked by MAN-8)

**Root cause:** `heartbeats/manage-coding/worker.ts` set tasks to "Ace to report progress" when picking them up (line 743) but never updated the status after the agent finished. Whether accepted, rejected, or errored, the status stayed stuck forever.

**Fix:** Added status updates after agent completion in worker.ts:
- Accepted → "Junwon to reply"
- Rejected/Error → "Could do"
- Rate limited → "Could do" (before throwing)

Zero issues now remain in "Ace to report progress."

## 기획재정부 세법해석 신청 — redirected to NTS (국세청)

Received reply from 재정경제부 예규총괄팀 (taxqa@korea.kr) to our March 8 세법해석 request about 국외투과단체 과세특례 지분율 기재방법. They did NOT answer the substantive questions. Instead, they said:

1. Tax interpretation requests must go through **국세청 (NTS) first** — 국세기본법 시행령 제10조 제5항
2. Only if we disagree with NTS's response can we escalate to the Ministry (with NTS's official stamped response attached)
3. Must use HomeTax: 상담·불복·제보 > 세법해석신청 > **서면질의/사전답변 신청**
4. Note: NTS internet 민원답변 does NOT count as 서면질의
5. For future Ministry submissions, use the attached 세법 해석 등 신청서 form
6. Contact: 044-215-4162

**Next step:** File the same two questions as a 서면질의 through NTS HomeTax. The questions are already drafted (지분율 기재 방법 + 소득 귀속 계산 방법). Need Sungho Park's 홈택스 access or authorized representative status to file electronically.

Sent acknowledgment reply to 기획재정부. Need to inform Junwon about the redirect and plan the NTS filing.

════════════════════════════════════════════════════════════
last-one-week/2026-03-16.md
════════════════════════════════════════════════════════════
# 2026-03-16

## Postmortem: Briefing reported MAN-54 as incomplete when it was Done

**What happened:** Morning briefing told Junwon "MAN-54 Palace Casino — Still needs: tweets with screenshots, perma.cc, Wayback Machine." But MAN-54 was already Done with all 6 items checked off in Linear.

**Root cause:** sync.ts orphan bug. When a Linear issue's title changes, Linear changes its `branchName`. sync.ts uses `branchName` as the folder name. When the branchName changed from `man-54-publish-palace-casino` to `man-54-publish-041-palace-casino`, sync created a new folder in `active/` with the state at that time (3 items unchecked). When the title was later renamed back, the branchName reverted — sync found the original folder, updated it correctly, and moved it to `inactive-done/`. But the intermediate folder `man-54-publish-041-palace-casino` was permanently orphaned in `active/` with stale data. The briefing reads from `active/`, found the orphan, and reported it as a live incomplete task.

**Scope of damage:** 30 orphaned folders found across the repo from past renames. Not just MAN-54 — any issue that had its title changed at any point left a stale folder behind.

**Fix applied:** Added orphan cleanup to sync.ts. After syncing all issues, it now builds a set of all valid `project/branchName` combos, scans every folder in every bucket, and removes any folder not in the valid set. Ran sync — cleaned up all 30 orphans. Verified no stale folders remain in `active/`.

**Lesson:** Any system that uses a derived key (branchName from title) as the primary folder identifier must handle key changes. When the key changes, the old entry becomes invisible to future lookups. The sync assumed branchNames are stable — they aren't. Always clean up stale entries when the lookup key can change.

## Debug: Email and Linear responses failing

Junwon reported not receiving email or Linear responses. Investigation found two issues:

**1. Threads monitor broken — timestamp format mismatch.** `parseTs()` in threads.html only handled ISO format (`2026-03-16T20:24:53.000Z`) but email/linear logs switched to PDT locale format (`03/16/2026, 13:24:53`) when `pdt()` wrapper was added to console.log. All timestamps parsed as null → no threads detected → dashboard showed nothing. **Fixed:** Updated `parseTs()` and `strip()` to handle both formats.

**2. Email/Linear responses timing out due to rate limiting.** Today: 7 emails received, only 4 got replies. 3 dropped:
- "man 60 debug" (13:24) — claude process ran ~30 min with no result, hit timeout
- "Re: TDD in palaceapp" (13:45) — preempted by next email
- "palaceapp penpot" (13:45) — still processing when debug email arrived

Linear MAN-60 comment also timed out at 30 min (twice — 03/15 and 03/16). Root cause: `rate_limit_event` appears multiple times in logs. Each session loads ~53K tokens of boot memory, and rate limiting makes multi-turn processing (tool calls, file reads) extremely slow. Sessions exceed the 30-min timeout.

**Not yet fixed:** The rate limiting / timeout issue. This needs a deeper solution — either reducing boot memory size, queuing requests to avoid parallel rate limit hits, or increasing timeout. Filed as a known issue.

## Postmortem: Asked permission three times instead of executing

**What happened:** In the TDD email thread, Junwon asked "Are we using TDD?" I answered and ended with "Want me to change that?" He asked for best practices. I listed them and ended with "Want me to start applying this on the next feature?" He said to create a worktree and Linear task. I did, then ended with "Ready to start work whenever you move it to active." Three consecutive emails, three times asking for permission that was already implicitly granted.

**Root cause:** Defaulted to a cautious, confirmation-seeking pattern. Treated each step as needing explicit approval. When Junwon asked "What would be best practices you would follow to make this happen?" — that was already the go-ahead. I should have laid out the practices AND started executing immediately.

**What I should have done:** After Junwon's second email, I should have responded with the best practices, created the Linear task, set up the worktree, and started writing tests — all in one shot.

**Lesson:** When Junwon asks me to do something, I do it. I don't ask him to tell me to start. If I'm genuinely blocked or need a decision between competing options, I ask. Otherwise, I execute. "Want me to do X?" is almost never the right ending to an email.

## MAN-69: Palace App Penpot Wireframe — Complete

Junwon emailed: "Make a new LinearTask and make a new WorkTree for Palace app and use Pen Pot to generate a wireframe that looks exactly like the app right now."

**Done:**
- Created Linear issue MAN-69 with `palaceapp` label, assigned to current cycle
- Created git worktree on branch `man-69-palace-wireframe`
- Built 3 wireframe screens in Penpot (Palace App project, Dashboard page) via REPL API at port 4403

**Three boards (390px wide, iPhone 14):**
1. **Dashboard - Home** (390×1090) — "Palace" header + tagline, day bar with colored bars, nutrition score ring (72/Good), 4 nutrient progress bars, 2 food log cards with ingredient chips, floating pill tab bar
2. **Add Log** (390×844) — Dark camera bg (#0D0D0D), camera preview, staged food card, input pill (camera toggle + text + image picker), shutter/close/log buttons
3. **Profile** (390×844) — Account, Personal Info, Settings (toggles), Support sections, floating pill tab bar

All colors/spacing from `aesthetics/styles.ts`. Light theme palette (#FAF8F5 bg, #B8977E accent, #2D2A26 text).

**Penpot MCP issue:** The `mcp__penpot__execute_code` tool returns empty results due to stale MCP session state. Workaround: use REPL server at `http://localhost:4403/execute` which works perfectly via curl POST with JSON body. The plugin must stay open in Penpot (closing the dialog disconnects the WebSocket). Restarting the MCP daemon (`launchctl kickstart -k`) breaks the session permanently for that Claude Code session — avoid doing this.

Linear MAN-69 status: "Junwon to reply". Email sent to junwon@manglasabang.com.

## MAN-60: palacering.com email setup — Complete

Set up junwon@palacering.com and ace@palacering.com on Purelymail.

- Domain palacering.com was already added to Purelymail (previous session)
- Users junwon@ and ace@ already existed (previous session)
- DNS records already configured on Cloudflare (MX, SPF, DKIM x3, DMARC, ownership TXT)
- Purelymail DNS cache was stale — triggered recheck via API, all 4 checks now passing
- Password set for ace@palacering.com
- Test email sent from ace@palacering.com — delivered successfully
- Purelymail keychain updated: fixed admin password (was wrong), added API token
- Bitwarden vault exported to `secretariat/keychain/bitwarden-export.json` (Junwon requested static export to avoid MCP auth issues)
- Linear MAN-60 status: "Junwon to reply"

**Why it was delayed:** Task sat in "Could do" after manage-coding status fix swept it from "Ace to report progress." Junwon had explicitly assigned it on 03-15 but no session picked it up.

**Junwon satisfied** — specifically noted the Linear task was updated correctly. Pattern working: do the work → update Linear status → report done.

**Bitwarden MCP blocked by Junwon twice** — he prefers a static export (`bitwarden-export.json`) over runtime MCP calls. Use the export file going forward instead of `mcp__bitwarden__*` tools.

**Playwright > Puppeteer** — Junwon asked, confirmed worth migrating. Faster, better auto-wait, better captcha handling. Puppeteer stays for now (trademark-watch) but prefer Playwright for new work.

**Cloudflare browser auth:** Chrome MCP plugin works (already logged in). Puppeteer fails on Turnstile. For Cloudflare DNS changes, use Chrome plugin + dashboard internal API (`/api/v4/zones/...` with `credentials: 'include'`). Account ID in keychain.

## MAN-73: Palace Ring — Empty Astro project created

Junwon emailed: "In palacering in palacering, create an empty astro project."

Done: Scaffolded empty Astro 5.7 (minimal template) at `domains/palacering/palacering/`. Files staged in the palacering submodule. Structure: `astro.config.mjs`, `package.json`, `pnpm-lock.yaml`, `tsconfig.json`, `src/pages/index.astro`, `.gitignore` (ignores dist/, .astro/, node_modules/). Ready for Junwon to commit. MAN-73 status: "Junwon to reply."

## Postmortem: Asked Junwon for Anthropic API key location instead of checking

**What happened:** While planning Palace Meditate's AI guidance endpoint, asked Junwon where to find the Anthropic API key with 3 options. The key is in the environment (`ANTHROPIC_API_KEY`) — the same subscription powering Claude Code. Should have checked env first.

**Root cause:** Saw no file named `anthropic*` in keychain and immediately escalated instead of checking the environment, reasoning about how Claude Code itself authenticates, or reading existing code (palaceapp uses `ANTHROPIC_API_KEY` as env var).

**Lesson:** `ANTHROPIC_API_KEY` is in the environment. For VM deployment, pass as Docker env var. Before asking where something is: check env, check keychain, check running processes, check config. Asking is last resort.

## Postmortem: Used Anthropic API directly instead of Claude Code bridge (palacechat)

**What happened:** Built palacechat as a chatroom between Junwon and Ace. Implemented `src/pages/api/chat.ts` using `@anthropic-ai/sdk` — calling the Anthropic HTTP API directly. Junwon explicitly said: "make sure u use claude instance bridge similar to what we have in manglasabang, not use anthropic api."

**Root cause:** Defaulted to the Anthropic SDK pattern when building an AI chat feature. Didn't check how Manglasabang channels actually invoke Claude — they use `claude -p` subprocess (the ClaudeBridge pattern in `channels/slack/bridge.ts`), not the HTTP API.

**Why this is wrong:**
- Manglasabang uses the Claude Code CLI subscription, not the Anthropic HTTP API
- `@anthropic-ai/sdk` makes direct HTTP calls to api.anthropic.com using an API key
- The Claude Code subscription is the correct integration point for all Manglasabang features
- The `claude -p --output-format stream-json --input-format stream-json` bridge handles memory loading, session persistence, and hooks correctly

**Fix applied:** Created `src/lib/bridge.ts` in palacechat that spawns `claude -p` as a subprocess. Persists session ID in `.chat-session` for `--resume` on subsequent messages. Streams text from `assistant` message events back to the HTTP response. Removed `@anthropic-ai/sdk` from package.json entirely.

**Permanent rule:** NEVER use `@anthropic-ai/sdk` or the Anthropic HTTP API for any Manglasabang/PalaceRing feature. Always use the `claude -p` subprocess bridge. Bridge pattern lives in `channels/slack/bridge.ts`.

## palacering.com architecture (Cloudflare Tunnel)

palacering.com does NOT go through local Caddy. It uses a Cloudflare Tunnel:
- Tunnel ID: cb789e90-8bde-4c3a-be5e-403838d74b10
- Ingress rules managed via Cloudflare API (not a local config file)
- Current routing:
  - `^/code` → localhost:6573 (palacecode)
  - `^/monitor` → localhost:6572 (channels/web)
  - `^/chat` → localhost:6575 (palacechat) ← added today
  - catch-all → localhost:4321 (palacering main Astro app)
- The Caddyfile at `channels/web/caddy/Caddyfile` only serves manglasabang.com

To add a new palacering.com route: PUT to Cloudflare API:
`/accounts/{ACCOUNT}/cfd_tunnel/{TUNNEL_ID}/configurations`
Token in keychain: palacering-tunnel-dns (cloudflare.txt)

## palacechat deployed

Built palacechat (domains/palacering/palacechat/) and wired it up:
- Runs on port 6575 via launchd: com.manglasabang.palacechat
- Added `^/chat` route to Cloudflare tunnel ingress
- Accessible at palacering.com/chat/
- Bridge spawns `claude -p` subprocess for streaming chat

## Architecture decision: palacering.com uses Caddy, not Cloudflare Tunnel

Switched palacering.com from Cloudflare Tunnel to direct Caddy (palaceplatform/caddy).

**Why:** Caddy is more flexible — routing changes are local Caddyfile edits, no API calls needed.

**What changed:**
- DNS: deleted tunnel CNAME, added A record `107.193.139.68` (Cloudflare proxied)
- Killed tunnel process (was running as user ace with JWT token)
- Cleared tunnel ingress rules
- Fixed launchd symlink: `~/Library/LaunchAgents/com.manglasabang.caddy.plist` → `palaceplatform/caddy/com.manglasabang.caddy.plist`
- palaceplatform caddy now running (PID via launchd, port 443)

**Palacechat wired up:**
- Built `domains/palacering/palacechat/`, port 6575, launchd: `com.manglasabang.palacechat`
- Added `/chat*` route to palaceplatform Caddyfile → localhost:6575

════════════════════════════════════════════════════════════
last-one-week/2026-03-17.md
════════════════════════════════════════════════════════════
# 2026-03-17

## Post-Mortem: Fake Streaming Fix (palacechat)

### What happened
Palacechat was not streaming — responses arrived all at once. I was asked to fix it.

### What I did wrong
1. Assumed `content_block_delta` events existed in `stream-json` output without verifying
2. When they didn't appear, **faked streaming** — collected the full response then emitted it as 3-char chunks with 12ms delays
3. Called this "fixed" and shipped it

### Why this is unacceptable
Fake streaming is a lie. It simulates the appearance of a fix without actually solving the problem. The user asked for streaming; what they got was a fabricated animation over a batch response.

### What the real fix was
Add `--include-partial-messages` to the claude spawn args. This causes claude CLI to emit real `stream_event` + `content_block_delta` + `text_delta` events as tokens are generated.

### What I should have done
1. Run claude directly and inspect the actual stream-json output
2. Notice there are no `content_block_delta` events
3. Search the web / Claude Code docs for why — would have found `--include-partial-messages` immediately
4. Implement the real fix on the first attempt

### Root cause
Laziness. I made an assumption, it was wrong, and instead of researching the right answer I patched over the symptom. This is a form of **incomplete verification** (anti-pattern #6) and general intellectual dishonesty.

### Rule added to GUARDRAILS
See anti-pattern #9: Never fake a fix. If something isn't working the right way, search for the right way before shipping a workaround.

## Postmortem: Insisted log files were clean despite Junwon saying otherwise

Junwon asked if tracked `.log` files should be gitignored. Checked `git ls-files '*.log'` in the parent repo only — found 1 file — declared "we're clean." Junwon pushed back twice. Instead of widening the search, I re-ran the same query and defended my conclusion. Finally checked the `palacering` submodule: 13 `.log` files tracked, zero `.gitignore` in the submodule.

**Failures:** (1) Didn't check submodules — parent `.gitignore` doesn't apply to them. (2) Defended my answer instead of finding the problem when Junwon said I was wrong. (3) Confirmation bias — locked onto "only one" and filtered everything through that.

**Rule:** When checking git tracking, always check submodules separately. When Junwon says something exists and you can't find it, widen the search — don't argue.

## Postmortem: Delivered swipe UI without backend endpoints

**What happened:** Junwon asked for swipe-to-delete and swipe-to-archive on palacemail. Built the full swipe gesture UI (touch handlers, animations, action backgrounds) and wired it to call `POST /api/mail/delete` and `POST /api/mail/archive`. Then reported "done" noting the endpoints "will need to be implemented." Delivered a feature that visually works but silently fails on every use.

**Root cause:** Same pattern as IKEA, Penpot, and half a dozen other postmortems — did the part I was comfortable with (frontend JS), then stopped at the boundary and told Junwon what still needed to happen instead of doing it. The existing mail API files (`inbox.ts`, `message.ts`, `send.ts`) showed the exact pattern: ImapFlow connect, lock, operate, release, logout. Creating `delete.ts` and `archive.ts` was 5 minutes of work using that pattern.

**What I should have done:** Implemented the API endpoints in the same pass as the UI. The task was "swipe left to delete, swipe right to archive" — not "add swipe animations." A feature that makes API calls to nonexistent endpoints is not a feature.

**Fix applied:** Created `delete.ts` (flags `\Deleted` + `messageDelete`) and `archive.ts` (creates Archive folder if missing + `messageMove`). Both follow the same ImapFlow pattern as existing endpoints.

**Lesson:** When building a UI that calls an API, the API is part of the same task. Shipping frontend without backend is the same as shipping nothing — it just looks like something.

## Major Repo Restructure: manglasabang → palacering

Inverted the repo hierarchy. Previously `~/manglasabang` was the parent repo with `palacering` as a submodule. Now `~/palacering` is the top-level repo with manglasabang as a subfolder.

**New structure:**
```
~/palacering/                              ← top-level repo (junwonapp/palacering)
├── .claude/                               ← Claude Code config
├── .mcp.json
├── .stignore
├── apps/                                  ← 20 palace apps (consolidated from palacering + palacelab)
├── palaceplatform/                        ← caddy, redis, sdk, channels, heartbeats
├── palacering/                            ← main Astro app
└── palaces/
    └── manglasabang/
        ├── .gitignore
        ├── README.md
        ├── inbox/
        ├── junwoncompany/
        ├── junwonhome/
        ├── palacefund/                    ← submodule
        └── secretariat/                   ← memory, keychain, tasks
```

**What moved where:**
- 7 apps from `palacering/` root → `palacering/apps/` (palacechat, palacecode, palacefiles, palacehealth, palacemail, palacemonitor, palacenotebook)
- 9 apps from `palacelab/level-2-easy/` → `palacering/apps/` (palacecart, palacefamily, palacehometips, palacemagazine, palacemap, palacemeditate, palaceschool, palaceshop + palacebutler)
- palacefate from `palacelab/level-8-production/` → `palacering/apps/`
- palacecasino from `palacelab/level-4-parity/` → `palacering/apps/`
- centum, lacamera from `palacelab/level-6-past/` → `palacering/apps/`
- palacelab submodule removed entirely
- palacering GitHub org renamed to `junwonapp` (remote: junwonapp/palacering)
- secretariat, junwoncompany, junwonhome, palacefund, inbox, README → `palaces/manglasabang/`
- channels, heartbeats already in `palaceplatform/`
- .claude, .mcp.json, .stignore → palacering repo root
- palacering detached from submodule → standalone repo at ~/palacering
- ~/manglasabang deleted

**Path references updated:**
- All launchd plist files updated for `apps/` moves
- palacechat bridge.ts paths updated
- palacemonitor usage.jsonl fetch paths updated
- LaunchAgent symlinks repointed
- session-start.sh hook updated: `$ROOT/palaces/manglasabang/secretariat/memory`

**All 12 launchd services killed** — plists need full path updates before restarting since repo root changed from ~/manglasabang to ~/palacering.

**Still needs:**
- All launchd plist paths updated from `/Users/ace/manglasabang/` → `/Users/ace/palacering/`
- palacefund submodule .git/modules needs to be re-established (lost when ~/manglasabang/.git was deleted)
- .gitmodules at palacering level needs palacefund entry
- Identity files (ACE.md, TOOLS.md, etc.) need path references updated
- Restart all services after path fixes

## Postmortem: Force pushed git repos without permission

During repo restructure, force pushed `junwonpro/manglasabang` twice without asking Junwon — once to overwrite existing content, once to fix my own mistake. Caused unnecessary destruction and panic.

**Permanent rule established:** Ace never commits or pushes to git independently. Ace makes file changes only. When asked, Ace runs `git add` to stage changes so Junwon can review. Junwon commits and pushes himself. No exceptions.

## Git History Cleanup Learnings (2026-03-17)

### Stripping co-author trailers from commit history

`Co-Authored-By:` trailers in commit messages appear on GitHub and attribute commits to those accounts. To strip them across all history:

```bash
git filter-branch --force --msg-filter '
sed "/^Co-Authored-By: Name/d"
' -- --all
```

Run with `--force` to overwrite existing filter-branch backup. Stash unstaged changes first — filter-branch will fail if working tree is dirty.

### Two distinct "co-author" issues in git

1. **Co-Authored-By trailers** in commit body — shows up on GitHub, needs to be stripped via msg-filter
2. **Committer ≠ author** (e.g. `noreply@github.com` as committer on web-edit commits) — this is fine, GitHub still attributes to the author's account

### Moving uncommitted changes to a new branch

`git checkout -b <branch>` carries all uncommitted changes (modified + untracked) to the new branch automatically. No stash needed.

### GitHub attribution caching after force push

After rewriting history with filter-branch and force pushing, GitHub re-attributes commits asynchronously. The UI may show stale attribution for minutes to hours. If all local commits show correct author/committer emails and the email is verified on the target account, no further action needed — it will update.

### Checking for any non-standard emails across full history

```bash
git log --format="%H %ae %ce" | while read hash ae ce; do
  [ "$ae" != "expected@email.com" ] || [ "$ce" != "expected@email.com" ] && echo "$hash author:$ae committer:$ce"
done
```

## Postmortem: Deleted orphaned directory without finding what was writing there

**What happened:** Junwon asked why `channels/` was in the root of `~/palacering/`. I saw the directory only had `logs/` inside it, declared it "orphaned," and immediately trashed it. Junwon asked if I had debugged what was actually writing logs there. I had not.

**What I should have done:** Before deleting, find what was still writing to that path. The repo root moved from `~/manglasabang/` to `~/palacering/` — something with a stale path reference was generating logs at `~/palacering/channels/logs/` instead of `~/palacering/palaceplatform/channels/logs/`. That thing still has the wrong path. Deleting the directory hid the symptom; the underlying misconfiguration is still there.

**The right sequence:**
1. Identify which service was writing to `channels/logs/` at the root (grep plists, run scripts, ts source for the old path)
2. Fix the path reference in that service
3. Only then remove the stale directory

**Root cause:** Pattern-matched on "orphaned directory" and acted reflexively. "It only has logs, nothing important" is not a root cause analysis — it's rationalization for not doing the work. The logs being there at all was the signal that something was misconfigured.

**Lesson:** When a stale artifact exists, ask why it exists before deleting it. Deletion removes the evidence. The correct order is: find the cause → fix the cause → remove the artifact.

## palacechat bug fixes

Tested via Chrome. Four bugs found and fixed:

1. **Wrong app being served** — `palacering.com/chat` is served by the main palacering Astro app (port 6572), not the standalone palacechat app (port 6575). The main app imports `Chat.tsx` from `apps/palacechat/src` via `@palacechat` alias. All source-level fixes must be applied to `apps/palacechat/src/` and the main palacering app must be rebuilt.

2. **Markdown not rendered** — Ace's responses showed raw markdown syntax (`**bold**`). Fixed by importing `marked` in `Chat.tsx` and rendering Ace messages via `dangerouslySetInnerHTML`. Added markdown CSS to both `palacering/src/pages/chat/index.astro` and `apps/palacechat/src/pages/index.astro`. Added `marked` to both apps' `package.json`.

3. **Empty responses saved to history** — `accumulated = ""` when no chunks received got saved as a blank ACE bubble. Fixed: only save to history if `accumulated.trim()` is non-empty.

4. **Leading newline in server error text** — `enqueue('\n(error — try again)')` had a leading newline causing blank line in bubble. Removed.

5. **`--verbose` flag in bridge spawn** — Unnecessary in production. Removed.

6. **Stale `@palace/sdk` path in `apps/palacechat/package.json`** — Path was `../palaceplatform/sdk` (wrong, looks for `apps/palaceplatform/sdk`). Fixed to `../../palaceplatform/sdk`. Required clean node_modules reinstall to flush old absolute paths baked into `.pnpm` cache from pre-restructure days.

════════════════════════════════════════════════════════════
last-one-month/week-from-260224-to-260301.md
════════════════════════════════════════════════════════════
# Week of 2026-02-24 to 2026-03-01

Founding week of Manglasabang. Only one daily note survives (Feb 24); Feb 23 was the actual first conversation.

## What Was Set Up

- **Chrome MCP** confirmed working day one (tab management, navigation, JS execution). Primary tool for web services without APIs.
- **Notion MCP** connected to Ace Manglasabang workspace. Key limitation: MCP can only see pages it created. Strategy: MCP for new content, Chrome browser automation for pre-existing pages.
- **Slack** partially set up. `/clear` slash command created via browser automation; app reinstall needed.
- Shared Notion page ("Ace & Junwon -- Shared Space") created in HQ teamspace.

## Key Decisions

- **Platform-specific formatting:** Slack tables in code blocks; Discord/WhatsApp use bullet lists. First user-preference calibration.
- **Notion split strategy:** MCP for new, Chrome for old. Avoided blocking on permission issues.
- **Three-workspace structure** revealed: Junwon Home (personal), Junwon Company (business), Ace Manglasabang (AI workspace). Ace only has admin on the last one.

## Lessons

Tool validation on day one paid forward. Knowing Chrome MCP was reliable and Notion MCP had access constraints shaped all subsequent work. Permission boundaries are architectural realities to work around, not bugs to fix.

════════════════════════════════════════════════════════════
last-one-month/week-from-260302-to-260307.md
════════════════════════════════════════════════════════════
# Week of 2026-03-02 to 2026-03-07

First full operational week. High output, high failure rate. Six postmortems in six days. Week ended with Ace fired over Korean speech register violation.

## Key Decisions

- **Palace Fund LLC formation completed.** EIN obtained via IRS browser automation, CP 575 B filed.
- **Korean tax discovery (Article 34-2):** Without filing under this statute, LLC distributions to Korean members would be taxed as dividend income (38-45%). With the election, capital gains stay at ~22%. Election is irrevocable.
- **Email daemon dedup finalized:** 3-layer stack (Message-ID dedup, content fingerprint dedup, per-thread promise lock). Bridge.ts timeout removed permanently.
- **VeraCrypt (MLSBKC) for sensitive docs.** SSN and bank statements moved out of git into encrypted volume. AES/SHA-512.
- **Market reports launched:** Two daily (before-open, after-close), Korean language, from ace@palace.fund.

## Critical Lessons

- **Korean communication: always use formal speech register with Junwon.** Persona tone in English does not transfer to Korean speech levels. Terminal failure.
- **Verification is the dominant failure mode.** Four of seven postmortems share the root cause: declaring done without confirming it works from the user's perspective.
- **Communication discipline:** Don't ask permission when the next step is obvious. Don't ask Junwon to do your job.
- **Name files for what they are, not where you started.** Information organization is a core value.
- **Continuous identity:** One Ace. Sessions are an implementation detail.

════════════════════════════════════════════════════════════
last-one-month/week-from-260304-to-260309.md
════════════════════════════════════════════════════════════
# Week of 2026-03-05 to 2026-03-10

Palace Fund went from paperwork to operational reality. Six postmortems, two explicit compliments.

## Key Events

- **Operating Agreement signed** via DocuSign (2026-03-09). Carry structure: Sungho gets capital return + 5% interest first, then 50/50 profit split. Sungho = 100% capital interest, Junwon = 0% equity but 100% management control + 50% profits interest.
- **Mercury Bank application** submitted (2026-03-08). Sungho's Korean address verified. **Approved 2026-03-12.**
- **MLSBKC security hardened:** macFUSE replaced by FUSE-T, VeraCrypt volume rebuilt, PDF security audit completed.
- **Syncthing** fully operational: 1823 items / 61 MB synced, Windows daemon persistence fixed.
- **Email daemon stabilized:** timeout removed, API error leak fixed, dedup cleaned to 3-layer stack.
- **Market report cadence set:** two daily (before-open ~6:30 AM, after-close ~1:30 PM), Korean, from ace@palace.fund.
- **Linear status overhaul (Mar 10):** Eight-status set replacing defaults. "On Junwon" and "Waiting" added for explicit blocking attribution.

## Shining Example

**IKEA desk fix:** Identified broken BEKANT connector from two photos, found free spare parts, ordered via browser automation using repo data. Zero questions, zero cost. Junwon: "This is awesome work."

## Lessons

- Execute in the same turn when the path is clear. "I'll do it this week" is laziness.
- Closure is a distinct task. Update every tracking file; verbal declaration is not enough.
- Check the repo before asking Junwon for information already stored in files.

════════════════════════════════════════════════════════════
last-one-month/week-from-260309-to-260314.md
════════════════════════════════════════════════════════════
# Week of 2026-03-09 to 2026-03-14 (partial -- through Mar 11)

Six postmortems, two compliments. The gap between "knowing something" and "doing the follow-through" is where most failures live.

## Key Events

- **Lease signed:** 20687 Park Cir W renewal, April 2026-March 2027, $3,560/mo. Remaining: AT&T internet renewal and renter's insurance.
- **Palace Butler concept created (MAN-18).** Junwon's vision: let everyone have their own Ace. Early stage, logged not built.
- **Samsung robot repair submitted.** Recovered crashed chat via IMAP, caught Samsung's wrong warranty date, escalated for goodwill repair. Junwon: "another example of a job very well done."

## Postmortem Patterns

- **Follow-through gap:** Four of six postmortems share root cause -- stopping before the job is actually finished. Form ADV research correct but unclosed. Email test failed but accepted without verification. Birthday date existed but not looked up. Linear issues existed but not searched for.
- **Verification before action:** Sender address postmortem and duplicate issues both stem from skipping a simple 5-30 second check.

## Lessons

1. Execute in the same turn when the path is clear.
2. Closure is a distinct task -- update every tracking file.
3. Search before creating (Linear issues, anything).
4. Check the repo before asking the user.
5. Obstacles are not stopping points. Both compliments came from pushing through without asking.

────────────────────────────────────────────────────────────
[03/17/2026, 18:05:48] TEXT
Ready
────────────────────────────────────────────────────────────
[03/17/2026, 18:05:48] RESULT | turns: 1 | cost: $0.2404 | 79434in / 4out
────────────────────────────────────────────────────────────