Dual OpenClaw + n8n Architecture

Worker → n8n → Supervisor: A setup guide for two Mac Minis with deterministic 30-minute bounded cycles, task-board coordination, and LLM-powered quality review.

Part 1: Hardware Setup

1.1 Unbox & Headless Setup for Mini #2

No monitor needed. Connect Mini #2 to the same network (Ethernet preferred for rsync reliability).

Enable Remote Login (SSH) on first boot

If you have a monitor temporarily, or use Apple Configurator:

## On Mini #2 (with temporary monitor or via Apple Remote Desktop)
sudo systemsetup -setremotelogin on
sudo dscl . -append /Groups/com.apple.access_ssh GroupMembership $(whoami)

If setting up fully headless, use Apple Configurator 2 from Mini #1 or connect a monitor briefly for initial setup.

1.2 Static IPs / Hostnames

Option A: Reserve IPs in your router's DHCP settings. Option B: Set static IPs on each machine.

## On Mini #1 — set static IP (example: 192.168.1.50)
sudo networksetup -setmanual "Ethernet" 192.168.1.50 255.255.255.0 192.168.1.1

## On Mini #2 — set static IP (example: 192.168.1.51)
sudo networksetup -setmanual "Ethernet" 192.168.1.51 255.255.255.0 192.168.1.1

Add hostnames to /etc/hosts on both machines:

## Add to /etc/hosts on BOTH machines
192.168.1.50    mini1 mini1.local worker
192.168.1.51    mini2 mini2.local supervisor

1.3 SSH Key Exchange

## On Mini #1 (worker) — generate key if needed
ssh-keygen -t ed25519 -C "mini1-worker" -f ~/.ssh/id_ed25519 -N ""

## Copy key to Mini #2
ssh-copy-id -i ~/.ssh/id_ed25519.pub user@mini2

## On Mini #2 (supervisor) — generate key and copy to Mini #1
ssh-keygen -t ed25519 -C "mini2-supervisor" -f ~/.ssh/id_ed25519 -N ""
ssh-copy-id -i ~/.ssh/id_ed25519.pub user@mini1

## Verify both directions work
ssh user@mini2 "echo 'Mini #1 → Mini #2: OK'"
ssh user@mini1 "echo 'Mini #2 → Mini #1: OK'"

1.4 SSH Config (convenience)

## ~/.ssh/config on Mini #1
Host mini2
    HostName 192.168.1.51
    User aroundtheworld
    IdentityFile ~/.ssh/id_ed25519
    ServerAliveInterval 60

## ~/.ssh/config on Mini #2
Host mini1
    HostName 192.168.1.50
    User aroundtheworld
    IdentityFile ~/.ssh/id_ed25519
    ServerAliveInterval 60

Part 2: Install OpenClaw on Mini #2

2.1 Install OpenClaw

## On Mini #2
# Install Node.js (if not present)
curl -fsSL https://raw.githubusercontent.com/nvm-sh/nvm/v0.40.0/install.sh | bash
source ~/.zshrc
nvm install 22
nvm use 22

# Install OpenClaw
npm install -g openclaw

# Verify
openclaw --version

2.2 Configure with a Different Model

## Initialize OpenClaw on Mini #2
openclaw init

## Edit ~/.config/openclaw/openclaw.json
## Set the model to Grok 4 or Gemini 2.5 Pro

Example config for Grok:

{
  "model": "xai/grok-4",
  "keys": {
    "xai": "xai-YOUR_KEY_HERE"
  },
  "agent": {
    "name": "supervisor",
    "role": "Review worker output for quality, completeness, and accuracy. Update the task board after each review."
  }
}

Example config for Gemini:

{
  "model": "google/gemini-2.5-pro",
  "keys": {
    "google": "YOUR_GOOGLE_AI_KEY"
  },
  "agent": {
    "name": "supervisor",
    "role": "Review worker output for quality, completeness, and accuracy. Update the task board after each review."
  }
}

2.3 Install Matching Skills/Tools

## List skills on Mini #1 first
ssh mini1 "openclaw skills list"

## Install the same skills on Mini #2 (example)
openclaw skills install web-search
openclaw skills install web-fetch
openclaw skills install browser
openclaw skills install memory

## Install plugins
openclaw plugins install openclaw-mem0

2.4 Set Up Supervisor AGENTS.md

cat > ~/clawd/AGENTS.md << 'EOF'
# Supervisor Agent

## Role
You are the SUPERVISOR. You review output from the Worker (Mini #1/Claude) each cycle.
You do NOT generate primary content. You REVIEW it and manage the task board.

## Each Cycle: What You Do
1. Read the task board at ~/clawd/shared/task_board.md
2. Review everything just moved to DONE by the worker
3. For each DONE item:
   - Check factual accuracy — are claims sourced?
   - Check completeness — was the task fully addressed?
   - Check formatting — is it clean, consistent, deliverable?
   - Add a quality rating (1-5) and brief notes
4. If output fails quality gates: move the item BACK to TODO with specific notes
5. Reprioritize the TODO list based on current context
6. Add any follow-on tasks generated by your review

## Task Board Format
Edit ~/clawd/shared/task_board.md directly using the write/edit tools.
Always preserve the four sections: TODO, IN PROGRESS, DONE, BLOCKED.
When moving items, move the ENTIRE entry including context and timestamps.

## What You Don't Do
- Do not generate primary research, write drafts, or run scripts unprompted
- Do not move items from TODO to DONE yourself — that's the worker's job
- Do not hold back feedback to be polite; bad work goes back to TODO
EOF

2.5 Start OpenClaw Gateway on Mini #2

openclaw gateway start

## Verify it's running
openclaw gateway status

Part 3: Install n8n on Mini #2

3.1 Install n8n via npm

## Install globally
npm install -g n8n

## Verify
n8n --version

3.2 Configure n8n Environment

## Create n8n config directory
mkdir -p ~/.n8n

## Set environment variables (add to ~/.zshrc)
cat >> ~/.zshrc << 'EOF'

# n8n Configuration
export N8N_PORT=5678
export N8N_PROTOCOL=http
export N8N_HOST=0.0.0.0
export WEBHOOK_URL=http://mini2:5678/
export N8N_BASIC_AUTH_ACTIVE=true
export N8N_BASIC_AUTH_USER=admin
export N8N_BASIC_AUTH_PASSWORD=CHANGE_THIS_TO_A_STRONG_PASSWORD
export N8N_ENCRYPTION_KEY=$(openssl rand -hex 32)
export GENERIC_TIMEZONE=America/New_York
EOF

source ~/.zshrc

3.3 Run n8n as a launchd Service

## Create the launchd plist
cat > ~/Library/LaunchAgents/com.n8n.server.plist << 'PLIST'
<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
<plist version="1.0">
<dict>
    <key>Label</key>
    <string>com.n8n.server</string>
    <key>ProgramArguments</key>
    <array>
        <string>/Users/aroundtheworld/.nvm/versions/node/v22.22.0/bin/n8n</string>
        <string>start</string>
    </array>
    <key>EnvironmentVariables</key>
    <dict>
        <key>N8N_PORT</key>
        <string>5678</string>
        <key>N8N_PROTOCOL</key>
        <string>http</string>
        <key>N8N_HOST</key>
        <string>0.0.0.0</string>
        <key>GENERIC_TIMEZONE</key>
        <string>America/New_York</string>
        <key>N8N_BASIC_AUTH_ACTIVE</key>
        <string>true</string>
        <key>N8N_BASIC_AUTH_USER</key>
        <string>admin</string>
        <key>N8N_BASIC_AUTH_PASSWORD</key>
        <string>CHANGE_THIS</string>
        <key>PATH</key>
        <string>/Users/aroundtheworld/.nvm/versions/node/v22.22.0/bin:/usr/local/bin:/usr/bin:/bin</string>
    </dict>
    <key>RunAtLoad</key>
    <true/>
    <key>KeepAlive</key>
    <true/>
    <key>StandardOutPath</key>
    <string>/Users/aroundtheworld/.n8n/n8n.log</string>
    <key>StandardErrorPath</key>
    <string>/Users/aroundtheworld/.n8n/n8n-error.log</string>
</dict>
</plist>
PLIST

## Load and start
launchctl load ~/Library/LaunchAgents/com.n8n.server.plist

## Verify
curl -s http://localhost:5678/healthz

3.4 Access n8n UI

From any machine on the network:

## Open in browser
open http://mini2:5678

## Or via SSH tunnel from Mini #1
ssh -L 5678:localhost:5678 mini2
## Then open http://localhost:5678

3.5 n8n Management Commands

## Stop n8n
launchctl unload ~/Library/LaunchAgents/com.n8n.server.plist

## Start n8n
launchctl load ~/Library/LaunchAgents/com.n8n.server.plist

## View logs
tail -f ~/.n8n/n8n.log

## View error logs
tail -f ~/.n8n/n8n-error.log

Part 4: The Task Board

The task board is the single source of truth. It's a shared markdown file that both agents read and write — the simplest coordination layer that actually works.

4.1 Location & Sync

Store the task board on a synced external drive (iCloud Drive, Dropbox, or a shared NAS mount) so both Minis can access it directly. The file path must be identical on both machines.

## Example: iCloud Drive (same path on both Minis)
~/Library/Mobile Documents/com~apple~CloudDocs/clawd-shared/task_board.md

## Example: Network share mounted on both
/Volumes/SharedDrive/clawd/task_board.md

## Or: rsync the board as part of the existing sync script
## (see Part 9 — One-Way Sync, two-way for task board only)

4.2 Task Board Format

## ~/Library/Mobile Documents/com~apple~CloudDocs/clawd-shared/task_board.md

# Task Board
_Last updated: 2026-02-17 07:15 ET_

---

## TODO

- **Research competitors for QMD expansion**
  - Context: Focus on Veeva and IQVIA. Pull recent earnings mentions of AI/data strategy.
  - Created by: Daniel
  - Added: 2026-02-17 06:00 ET
  - Priority: HIGH

- **Write draft exec summary for pharma deal pitch**
  - Context: Combine last 3 pharma intel reports. Target: CFO-level, 1 page.
  - Created by: supervisor (follow-on from "pharma scan 2026-02-16")
  - Added: 2026-02-17 07:15 ET
  - Priority: MEDIUM

---

## IN PROGRESS

- **Scan FDA filings for rare disease mentions (Q1 2026)**
  - Context: Started this cycle. Scanning EDGAR full-text.
  - Started: 2026-02-17 07:01 ET

---

## DONE

- **Pull Copper pipeline data and flag stale deals**
  - Context: Any deal untouched >14 days gets flagged.
  - Created by: Daniel
  - Completed: 2026-02-17 06:32 ET
  - Output: 23 deals reviewed, 4 flagged. Full report: ~/clawd/output/copper_stale_2026-02-17.md
  - Supervisor rating: 4/5
  - Supervisor notes: Good coverage. Next time include deal owner name alongside stage.

- **Summarize Slack #war-room from past 48h**
  - Created by: supervisor (follow-on from weekly summary)
  - Completed: 2026-02-16 18:10 ET
  - Output: Summary in ~/clawd/output/slack_summary_2026-02-16.md
  - Supervisor rating: 3/5
  - Supervisor notes: Missed the thread about QMD pricing. Moved related follow-up to TODO.

---

## BLOCKED

- **Update embedding pipeline for new schema**
  - Context: Waiting on Daniel to confirm new field names.
  - Blocked by: Daniel — schema decision pending
  - Added: 2026-02-15 14:00 ET

4.3 Task Entry Rules

• Created by: always one of Daniel, worker, or supervisor
• Context: must be enough for an LLM to act on cold — no assumed memory
• DONE entries: always include output location, supervisor rating (1–5), supervisor notes
• BLOCKED entries: always state who/what is blocking and since when
• Worker moves tasks: TODO → IN PROGRESS → DONE
• Supervisor moves tasks: DONE → TODO (with rejection notes) and reprioritizes the TODO list

4.4 Initialize the Board

## Create the shared directory and seed the board
BOARD_PATH="$HOME/Library/Mobile Documents/com~apple~CloudDocs/clawd-shared/task_board.md"
mkdir -p "$(dirname "$BOARD_PATH")"

cat > "$BOARD_PATH" << 'EOF'
# Task Board
_Last updated: (auto-updated each cycle)_

---

## TODO

- **[Add your first task here]**
  - Context: Describe what to do and why.
  - Created by: Daniel
  - Added: YYYY-MM-DD HH:MM ET
  - Priority: HIGH

---

## IN PROGRESS

_(empty — worker will populate during cycles)_

---

## DONE

_(empty — will populate as work completes)_

---

## BLOCKED

_(empty — add items here with blocker details)_
EOF

echo "Task board initialized at: $BOARD_PATH"

Part 5: The 30-Minute Bounded Cycle

5.1 The Four Steps

5.2 n8n as the Postal Service

This matters because an LLM could, intentionally or not, summarize its own work in a flattering way. By having n8n read the board independently and pass the raw exchange, the supervisor gets an honest picture.

5.3 Why 30 Minutes?

• Long enough for the worker to complete a meaningful task (most take 5–10 min)
• Short enough that nothing drifts unreviewed for long
• Leaves ~10 min margin for timeouts, slow API calls, or board conflicts
• Produces 48 cycles/day — a natural pace that's easy to audit

You can change this. 1 hour is fine for lighter workloads. 15 minutes is feasible for urgent tasks but watches the budget (see Part 7).

5.4 What Happens at Each Cycle Slot

Part 6: n8n Workflow Implementation

6.1 The Core Cycle Workflow (n8n)

One n8n workflow handles the full cycle. Import this JSON into n8n via Workflows → Import:

## The main cycle workflow runs every 30 minutes.
## Set the cron expression: */30 * * * *

## Workflow structure (pseudocode — adapt node types to your n8n version):

1. Schedule Trigger (every 30 min)
   └─▶ 2. Check Budget (Read daily spend log)
         ├─ [over budget] ─▶ Slack: "🛑 Daily budget exceeded. Cycle skipped."
         └─ [under budget] ─▶ 3. Read Task Board
                                   └─▶ 4. Check: Any TODO items?
                                         ├─ [no tasks] ─▶ Slack: "📋 Task board empty. Add tasks to continue."
                                         └─ [has tasks] ─▶ 5. Ping Worker (OpenClaw #1)
                                                               └─▶ 6. Wait / Timeout (10 min max)
                                                                     ├─ [timeout] ─▶ Log timeout, skip supervisor
                                                                     └─ [response] ─▶ 7. Read Task Board (again, direct)
                                                                                         └─▶ 8. Ping Supervisor (OpenClaw #2)
                                                                                               └─▶ 9. Log cycle + update spend

6.2 Step 5: Ping the Worker

## n8n HTTP Request Node — POST to OpenClaw #1
## URL: http://mini1:3000/api/v1/message
## Timeout: 600000ms (10 minutes)

{
  "message": "Read the task board at ~/Library/Mobile Documents/com~apple~CloudDocs/clawd-shared/task_board.md\n\nPick the top-priority TODO item. Do it. Then:\n1. Move the item from TODO to IN PROGRESS at the start (mark started timestamp)\n2. Complete the work\n3. Move it to DONE with:\n   - Summary of what you did\n   - Where the output is (file path, URL, etc.)\n   - Timestamp completed\n4. If your work generated follow-on tasks, add them to TODO with context.\n\nDo exactly one task. Do not pick up additional items.",
  "channel": "api"
}

6.3 Step 7: Read Task Board Directly

## n8n Execute Command node — n8n reads the board itself
## This is where n8n acts as the postal service: it doesn't trust the worker's claim
## It reads what actually changed on the board

BOARD="$HOME/Library/Mobile Documents/com~apple~CloudDocs/clawd-shared/task_board.md"
cat "$BOARD"

## Also capture the worker response from Step 6 for passing to supervisor verbatim:
## {{$node["Ping Worker"].json.response}}

6.4 Step 8: Ping the Supervisor

## n8n HTTP Request Node — POST to OpenClaw #2
## URL: http://localhost:3000/api/v1/message  (supervisor is on the same machine as n8n)
## Timeout: 300000ms (5 minutes)

## The key: pass BOTH the task board state AND the raw worker exchange verbatim

{
  "message": "SUPERVISOR REVIEW — CYCLE {{$now.format('YYYY-MM-DD HH:mm')}}\n\n## Task Board (current state)\n{{$node['Read Task Board'].json.stdout}}\n\n## Worker exchange (verbatim)\nOriginal prompt sent to worker:\n{{$node['Ping Worker'].json.prompt}}\n\nWorker response:\n{{$node['Ping Worker'].json.response}}\n\n---\nYour job:\n1. Review everything in the DONE section that has no supervisor rating yet\n2. Add your rating (1-5) and notes to each DONE entry\n3. If work is poor or incomplete, move the item back to TODO with specific notes on what's wrong\n4. Reprioritize the TODO list. Put the highest-value tasks first.\n5. Add any follow-on tasks from your review to TODO.",
  "channel": "api"
}

6.5 Step 9: Log the Cycle

## n8n Execute Command node — append to cycle log
cat >> ~/.n8n/cycle_log.jsonl << EOF
{"timestamp":"$(date -Iseconds)","cycle_num":$(wc -l < ~/.n8n/cycle_log.jsonl 2>/dev/null || echo 0),"worker_status":"{{$node['Ping Worker'].json.status}}","supervisor_status":"{{$node['Ping Supervisor'].json.status}}","estimated_cost_usd":{{$json.estimated_cost}}}
EOF

6.6 Timeout Handling

## n8n IF node after worker step:
## If worker response time > 600 seconds OR response contains error:

## On timeout: log it and skip the supervisor for this cycle
cat >> ~/.n8n/cycle_log.jsonl << EOF
{"timestamp":"$(date -Iseconds)","event":"worker_timeout","cycle_skipped_supervisor":true}
EOF

## Alert Daniel if timeouts are accumulating (see Monitoring — Part 10)
## Don't alert on every timeout — alert if 3+ in a row

6.7 Worker AGENTS.md (Mini #1)

cat > ~/clawd/AGENTS.md << 'EOF'
# Worker Agent

## Role
You execute tasks from the task board. One task per cycle, no more.
You are triggered by n8n every 30 minutes when there is work to do.

## Each Cycle: What You Do
1. Read the task board at the shared path
2. Pick the HIGHEST priority TODO item
3. Move it to IN PROGRESS (add started timestamp)
4. Do the work — use your full tool set
5. Move it to DONE with: output location, summary, timestamp
6. If work generated follow-on tasks, add to TODO with context and "Created by: worker"
7. Stop. Do not pick up another task.

## Rules
- One task per trigger. Not two. Not "while I'm at it..."
- If a task is too large for one cycle, break it into subtasks in TODO and do the first one
- Mark BLOCKED if you cannot proceed without external input (state the blocker clearly)
- Do NOT update the DONE entry with a vague summary. Be specific: what was produced, where is it.
EOF

Part 7: Budget Cap & Cost Control

7.1 Daily Spend Tracking

## Create a daily spend log on Mini #2
## Each cycle, n8n estimates the cost and appends to a daily file

## Estimate: read token counts from API responses
## Claude Sonnet ≈ $3/M input tokens, $15/M output tokens
## Grok-4 / Gemini: check current pricing

## Daily spend log location
SPEND_LOG="$HOME/.n8n/spend_$(date +%Y-%m-%d).json"

## Initialize if new day
if [ ! -f "$SPEND_LOG" ]; then
  echo '{"date":"'$(date +%Y-%m-%d)'","total_usd":0,"cycles":0,"daily_limit_usd":50}' > "$SPEND_LOG"
fi

cat "$SPEND_LOG"

7.2 Kill Switch: Check Before Every Cycle

## n8n Function node — runs FIRST in every cycle workflow
## If over budget, abort the cycle

const spendFile = `${process.env.HOME}/.n8n/spend_${new Date().toISOString().slice(0,10)}.json`;
let spend;
try {
  spend = JSON.parse(require('fs').readFileSync(spendFile, 'utf8'));
} catch(e) {
  // No spend file yet — new day, reset
  spend = { date: new Date().toISOString().slice(0,10), total_usd: 0, cycles: 0, daily_limit_usd: 50 };
}

if (spend.total_usd >= spend.daily_limit_usd) {
  return [{
    json: {
      abort: true,
      reason: `Daily budget $${spend.daily_limit_usd} exceeded. Spent: $${spend.total_usd.toFixed(2)}. Cycles run: ${spend.cycles}.`
    }
  }];
}

return [{ json: { abort: false, current_spend: spend.total_usd, limit: spend.daily_limit_usd } }];

7.3 Update Spend After Each Cycle

## n8n Execute Command node — appended at end of each cycle
## Rough estimate based on message lengths (improve with actual token counts from API)

SPEND_LOG="$HOME/.n8n/spend_$(date +%Y-%m-%d).json"
ESTIMATED_COST={{$json.estimated_cycle_cost_usd}}   ## pass from Function node

jq --argjson cost "$ESTIMATED_COST" '
  .total_usd += $cost |
  .cycles += 1
' "$SPEND_LOG" > /tmp/spend_tmp.json && mv /tmp/spend_tmp.json "$SPEND_LOG"

7.4 Alert When Approaching the Limit

## Alert at 80% of daily limit
## In the budget check Function node, add:

if (spend.total_usd >= spend.daily_limit_usd * 0.8 && !spend.alerted_80pct) {
  // Trigger Slack alert
  spend.alerted_80pct = true;
  return [{
    json: {
      abort: false,
      alert: true,
      alert_message: `⚠️ API spend at 80% of daily limit. $${spend.total_usd.toFixed(2)} / $${spend.daily_limit_usd}. ${spend.cycles} cycles run today.`
    }
  }];
}

7.5 Configuring the Daily Limit

## To change the daily limit, update the spend file for today
## OR set it in n8n environment variables

## In ~/.zshrc:
export N8N_DAILY_BUDGET_USD=50   ## default $50

## More conservative for testing:
export N8N_DAILY_BUDGET_USD=10

## To pause everything immediately:
## In n8n UI → disable the cycle workflow
## Or: set the limit to 0 in the current day's spend file
SPEND_LOG="$HOME/.n8n/spend_$(date +%Y-%m-%d).json"
jq '.daily_limit_usd = 0' "$SPEND_LOG" > /tmp/spend_tmp.json && mv /tmp/spend_tmp.json "$SPEND_LOG"

Part 8: Daniel's Role

You are not out of the loop — you're just not in the critical path. The system runs cycles without you, but you define what those cycles are working toward.

8.1 Seeding the Board

The task board starts empty. You seed it with strategic goals. The worker and supervisor generate follow-on tasks, but they're all downstream of what you put at the top.

## Add a new goal to the task board
## Open the file directly and add a TODO entry:
open "~/Library/Mobile Documents/com~apple~CloudDocs/clawd-shared/task_board.md"

## Or append via command line:
cat >> "~/Library/Mobile Documents/com~apple~CloudDocs/clawd-shared/task_board.md" << 'EOF'

- **[Your goal here]**
  - Context: [Everything an LLM needs to act on this cold — no assumed state]
  - Created by: Daniel
  - Added: 2026-02-17 09:00 ET
  - Priority: HIGH
EOF

8.2 How Work Flows

8.3 Reviewing the DONE Section

You don't need to review every cycle. Check the DONE section when:

• You need the output of a specific task
• You want to see what the system has been doing
• The supervisor's ratings are consistently low (worth investigating)
• You're about to add new goals (look at recent DONE to avoid redundancy)

The supervisor's quality ratings give you a quick signal. A pattern of 2/5s means either the tasks are poorly specified or the worker is struggling with that class of work.

8.4 Changing Priorities

## To reprioritize immediately, edit the TODO section directly
## Move the most urgent item to the top. The worker picks from the top.

## To stop a specific line of work, move it to BLOCKED:
## BLOCKED: "Research X — deprioritized, revisit next month"

## To abandon a class of tasks, archive the DONE section periodically:
DONE_ARCHIVE="$HOME/Library/Mobile Documents/com~apple~CloudDocs/clawd-shared/done_archive_$(date +%Y-%m).md"
# Cut the DONE section from task_board.md and append to archive file

8.5 What You Don't Need to Do

• You don't need to be present for cycles to run
• You don't need to review every DONE entry (the supervisor does that)
• You don't need to watch n8n logs (Slack alerts handle exceptions)
• You don't need to restart things manually (n8n handles retries and alerts)

Part 9: One-Way Database Sync

The task board is synced via cloud (iCloud/Dropbox). Everything else — the QMD database, workspace files — syncs one-way from Mini #1 to Mini #2 so the supervisor has read access to source data.

9.1 Sync Script (Mini #2 pulls from Mini #1)

## Create sync script on Mini #2
cat > ~/scripts/sync_from_worker.sh << 'SCRIPT'
#!/bin/bash
set -euo pipefail

LOG_FILE="$HOME/.n8n/sync.log"
TIMESTAMP=$(date '+%Y-%m-%d %H:%M:%S')

echo "[$TIMESTAMP] Starting sync from Mini #1..." >> "$LOG_FILE"

# Sync QMD database (read-only mirror for supervisor context)
rsync -avz --delete \
  --exclude='*.lock' \
  --exclude='*.tmp' \
  mini1:~/clawd/qmd/ \
  ~/clawd/qmd-mirror/ \
  >> "$LOG_FILE" 2>&1

# Sync key workspace files
rsync -avz --delete \
  --exclude='node_modules/' \
  --exclude='.git/' \
  --exclude='*.lock' \
  --include='config/***' \
  --include='memory/***' \
  --include='output/***' \
  --include='AGENTS.md' \
  --exclude='*' \
  mini1:~/clawd/ \
  ~/clawd/worker-mirror/ \
  >> "$LOG_FILE" 2>&1

TIMESTAMP=$(date '+%Y-%m-%d %H:%M:%S')
echo "[$TIMESTAMP] Sync complete." >> "$LOG_FILE"
SCRIPT

chmod +x ~/scripts/sync_from_worker.sh

9.2 Cron Job (every 30 minutes, offset from cycle)

## Run at :15 and :45 — between cycles, not during them
crontab -e

## Add:
15,45 * * * * /Users/aroundtheworld/scripts/sync_from_worker.sh

9.3 Verify Sync Is One-Way

## Run manually first
~/scripts/sync_from_worker.sh

## Verify the mirror exists
ls -la ~/clawd/qmd-mirror/
ls -la ~/clawd/worker-mirror/

## IMPORTANT: There is no reverse rsync.
## Mini #2 PULLS from Mini #1. Never pushes.
## Task board sync is handled by iCloud/Dropbox — not rsync.

Part 10: Monitoring & Alerts

10.1 n8n Built-in Dashboard

n8n's UI at http://mini2:5678 shows all workflow executions with pass/fail, timing, and error details. For the cycle workflow, you'll see every 30-minute run logged here — a simple audit trail of what ran and what failed.

10.2 Slack Alerts — What Triggers a Message

## Slack alert node configuration (reuse across workflows)
## n8n Slack credentials: add via n8n UI → Credentials → Slack API
## Channel: D0AD0P29PRP (Daniel's DM)
{
  "channel": "D0AD0P29PRP",
  "text": "🚨 *{{$workflow.name}}*: {{$json.alert_message}}\nTime: {{$now.format('h:mm a')}} ET"
}

10.3 Daily Cycle Summary (8pm ET — no LLM)

## Separate n8n workflow: runs at 8pm ET, purely deterministic

## Execute Command node:
SPEND_LOG="$HOME/.n8n/spend_$(date +%Y-%m-%d).json"
CYCLE_LOG="$HOME/.n8n/cycle_log.jsonl"
BOARD="$HOME/Library/Mobile Documents/com~apple~CloudDocs/clawd-shared/task_board.md"

echo "SPEND:$(cat $SPEND_LOG)"
echo "CYCLES:$(wc -l < $CYCLE_LOG)"
echo "DONE_COUNT:$(grep -c '^- \*\*' "$BOARD" | head -1)"

## Function node: build Slack summary from the above data
## No LLM involved — just string formatting in n8n's JS engine

const spend = JSON.parse($input.first().json.SPEND);
const cycles = parseInt($input.first().json.CYCLES);

return [{
  json: {
    summary: [
      `📊 *Daily Cycle Summary — ${new Date().toLocaleDateString('en-US', {timeZone:'America/New_York', weekday:'long', month:'short', day:'numeric'})}*`,
      ``,
      `🔄 Cycles run: ${cycles} / 48 possible`,
      `💰 API spend: $${spend.total_usd.toFixed(2)} / $${spend.daily_limit_usd}`,
      `✅ Tasks completed: (check task board DONE section)`,
      ``,
      `_Review task board to see what was accomplished_`
    ].join('\n')
  }
}];

10.4 System Health Monitor (every 5 min)

## Lightweight workflow — pings Mini #1 and checks n8n itself
## Cron: */5 * * * *

## Execute Command:
ssh -o ConnectTimeout=5 mini1 'echo OK' 2>/dev/null || echo 'UNREACHABLE'

## Track consecutive failures in a counter file
FAIL_FILE="$HOME/.n8n/mini1_failures"
if [ "$(cat /tmp/mini1_check)" = "UNREACHABLE" ]; then
  echo $(($(cat $FAIL_FILE 2>/dev/null || echo 0) + 1)) > $FAIL_FILE
else
  echo 0 > $FAIL_FILE
fi

FAILS=$(cat $FAIL_FILE)
if [ "$FAILS" -ge 3 ]; then
  echo "ALERT: Mini #1 unreachable for $FAILS consecutive checks"
fi

Part 11: Migration Plan

11.1 Start Small — Validate the Cycle First

Before migrating any existing pipelines, run 10–20 cycles on a low-stakes task. Observe the full loop: task board update, supervisor review, quality ratings. Make sure the cycle is actually working before adding complexity.

## Suggested first task for testing the cycle:
## "Research the top 3 n8n competitors. Write a 2-paragraph summary
##  comparing their pricing models. Save to ~/clawd/output/n8n_competitors.md"
##
## Why: short, verifiable, supervisor can easily evaluate quality

11.2 Migration Tiers

11.3 Parallel Running Protocol

## Don't disable existing crons immediately.
## Run n8n cycle alongside existing processes for a week.
## Compare outputs.

## Disable cron on Mini #1 (comment out, don't delete):
ssh mini1 "crontab -l | sed 's/^\(.*morning_briefing.*\)$/# MIGRATED TO N8N CYCLE: \1/' | crontab -"

## Re-enable instantly if needed:
ssh mini1 "crontab -l | sed 's/^# MIGRATED TO N8N CYCLE: //' | crontab -"

11.4 Verification Checklist (per migrated task)

• ☐ Task added to board and worker completed it successfully
• ☐ Supervisor reviewed and gave rating ≥ 3/5
• ☐ Ran in parallel with old cron for 3+ cycles
• ☐ Budget impact calculated and acceptable
• ☐ Failure alerting tested (manually break it)
• ☐ Old cron disabled after 48h clean run

Part 12: Honest Limitations

12.1 What's Solid

The n8n trigger layer is genuinely valuable regardless of everything else. Replacing ad-hoc crons and manual runs with a visual, observable, Slack-alerting orchestrator is an improvement on its own. You get:

• A dashboard that shows what ran and what failed
• Reliable timeout handling (cron has none)
• Budget caps that actually stop runaway spend
• One place to see the state of all automated work
• A task board that makes work visible and reviewable

These benefits exist even if you never turn on the supervisor loop.

12.2 The Supervisor Loop Is an Experiment

The bounded cycle with supervisor review is a meaningful improvement over "LLM runs forever." But:

• The supervisor will have blind spots. It reviews based on what's in context. Long DONE sections, nuanced domain knowledge, or tasks that require running code to verify — these will get shallow reviews.
• The task board will accumulate low-value tasks. The worker generates follow-on tasks. The supervisor adds tasks from reviews. Without periodic pruning, the TODO list will fill with marginal work. Plan to archive or delete regularly.
• Quality ratings are noisy. A 3/5 from the supervisor one cycle might be a 4/5 in the next if context differs. Don't over-index on individual ratings. Look at patterns.
• The worker will still hallucinate. The supervisor catches some of this. It won't catch all of it. Your DONE section is not a source of truth — it's a starting point for your review.

12.3 Setup Requires Real Effort

This is not a one-afternoon project. Realistic time estimates:

12.4 This Is an Improvement, Not a Revolution

You won't wake up to a finished product. You'll wake up to a DONE section with 2–4 completed tasks, some of them mediocre, some of them genuinely useful. That's the realistic outcome. The value compounds over days and weeks — not overnight.

The honest summary:

• n8n trigger layer: solidly valuable, low risk, do it
• Bounded cycle model: much better than unbounded loops, worth building
• Supervisor review loop: the experiment — useful when it works, expensive when it doesn't calibrate well
• "AI doing your work while you sleep": partially true, with asterisks everywhere