🧠 Orchestrator
Opus
claude-opus-4-6 · routing & decisions
⚙️ Workhorse (default sub-agent)
Codex
gpt-5.3-codex · heavy lifting
💡 Mid-tier
Sonnet
claude-sonnet-4-6 · balanced tasks
💨 Lightweight
Haiku
claude-haiku-3-5 · cheap/fast tasks
Overview intent
This page is now strictly system snapshot (model posture, architecture context, and channel posture). Operational queueing, blockers, and agent execution state live only in Control Tower.
🧮 Model Allocation Policy — Who Uses What
| Model | Provider | Cost Tier | Role | Used For | Usage Goal |
|---|---|---|---|---|---|
| Opus 4-6 | Anthropic | 💰💰💰 High | Orchestrator | User replies, routing decisions, approval gates. Never spawned as sub-agent. | Minimal — keep lean |
| Sonnet 4-6 | Anthropic | 💰💰 Mid | Balanced sub-agent | Nuanced writing, complex reasoning needing Anthropic quality but not Opus cost. | Moderate — when Codex isn't ideal |
| Codex 5.3 | OpenAI | 💰💰 Mid | Default workhorse | Research, coding, multi-step analysis, tool-heavy tasks. Relieves Anthropic rate pressure. | Heavy — default sub-agent |
| Haiku 3.5 | Anthropic | 💰 Low | Cheap/fast | Summaries, formatting, status checks, simple cron jobs, telemetry, hygiene. | Heavy — prefer for simple tasks |
Plans: Anthropic Max ($200/mo) · OpenAI Plus ($20/mo). Primary constraint: never hit API rate limits. Opus/Sonnet/Haiku share Anthropic pool — spread load to Codex (OpenAI) to reduce 429 risk.
What is this setup?
This is a personal AI operating system running on your Mac Mini (MMM4). OpenClaw is the runtime layer: it receives messages, applies policy, routes to the right session, calls tools, and sends responses back to you.
At a systems level: Input channels → Gateway → Main session → Sub-agents/tools → Memory + external systems (Notion/GitHub) → Output channels. Your laptop can SSH into MMM4 and operate the same runtime from anywhere.
Model policy: Opus orchestrator makes routing decisions and delegates work. Codex is the default workhorse for sub-agents (research, coding, heavy tasks). Sonnet for balanced Anthropic tasks when Codex isn't ideal. Haiku for cheap/fast background jobs. Opus usage is minimized — it decides, delegates, and stays lean. Plans: Anthropic Max ($200/mo), OpenAI Plus ($20/mo).
Memory is split into two layers: session context (temporary) and durable memory files (MEMORY.md, memory/*.md). Notion is your external objective system; writes are manual-only by your explicit command.
First-principles dashboard objective: show state (what is true now), control (what you can change), risk (what can break), and next actions (what to do next). If a panel doesn't help one of those four, it should be removed.
Channels (on/off)
| Webchat | on |
| Slack | on |
| Telegram | on |
| Discord | on |
| iMessage | off |
| off |
Control Tower — single operating view
Use this page as your default cockpit: what is running, what is blocked by Nick, and what matters most now.
Ongoing Projects (live scorecard)
Launch Twitter account for AI + mindfulness content
18%
Stablecoin expertise content pipeline
12%
Autonomous background ops (low-cost)
40%
Scoring rule: 0-100 based on concrete milestones completed, not vibes. I’ll ask for progress check-ins on a cadence and update this honestly.
Agent Activity (live plan)
| Agent | Status | Workstream | Cadence / next run |
|---|---|---|---|
| Main (April) | On-demand | Orchestration, approvals, reverse prompting | Runs when Nick messages |
Mode: Planned/manual reporting for reliability. Upgrade to live telemetry after stability is proven.
Nick-blocked queue (priority ranked)
| Blocker | Score | Owner | Next action |
|---|---|---|---|
| Grant Full Disk Access for Things DB read | 10.0 | Nick | Enable access for OpenClaw runtime; rerun things today --limit 5 --format json |
Scoring method: Value ÷ Effort (higher = do first).
Top priorities (next actions)
| Priority | Task | Ball | Scheduled agent | Next action |
|---|---|---|---|---|
| P1 | Unblock Things ingestion | Nick | Main (April) | Grant disk permissions and confirm with one probe command |
Active projects overview (what April is working on)
| Project | Progress | Current focus | Next milestone |
|---|---|---|---|
| Launch Twitter account for AI + mindfulness content | 18% | Define publishing cadence + first content batch | Publish first 3 posts |
This section is generated from live control-tower telemetry and priorities, so you can see where execution time is going.
System Health
Operational telemetry snapshot for core services and recent errors.
Overall
—
Awaiting telemetry…
Services Healthy
0
Services currently passing checks
Services Total
0
Total services monitored
Open Errors
0
Unresolved incidents
Recent Errors
| Time | Service | Severity | Status | Error | Next action |
|---|---|---|---|---|---|
| No telemetry loaded. | |||||
Research Tasks
Daily research telemetry for special projects. This includes news-driven probability tracking for PayPal acquisition outcomes.
PayPal acquisition probability
—
Awaiting daily refresh…
Current thesis (why)
No thesis loaded yet.
Method: weighted signal from credible reporting, deal-rumor quality, and strategic fit indicators. Updated once per day.
Evidence log (latest signals)
| Date | Signal | Direction | Strength | Source |
|---|---|---|---|---|
| No evidence loaded. | ||||
Slack Channels (active with Claw)
| Channel | Channel ID | Purpose |
|---|---|---|
| #hq | C0AKD18KN2G | Daily driver working channel |
| #alerts | C0AKD1VH8CU | Urgent/time-sensitive notifications (failures, 429s, blocked tasks, security) |
| #dashboard | C0AK32JBKTM | Dashboard building, UI changes, automated telemetry feed |
| #april-journal | C0AL3MWCR40 | Journal prompts (6AM + 6PM), objective check-ins |
| #obj-95 | C0AL3MQV41W | Mindfulness content engine, tweet/thread drafts from Waking Up corpus |
| #obj-106 | C0AJTUZED2B | Stablecoin/PayPal/Ethereum research, wedge selection, Arbitrum L2 |
Full System Architecture — How Claw Works
How a message flows
- You send a message on Slack / Telegram / Discord / Webchat
- OpenClaw gateway receives it and routes to main session
- Claw (Opus) reads context + memory and decides what to do
- If complex → spawns a sub-agent with the right model tier
- Sub-agent runs, returns structured output
- Claw synthesizes and replies to you
- Durable decisions written to memory.md
Physical setup
| Layer | What | Notes |
|---|---|---|
| Hardware | Mac Mini MMM4 | Always-on · arm64 · macOS 26.3 |
| Network | Tailscale | 100.74.228.112 · SSH pending |
| Gateway | OpenClaw 2026.3.7 | Port 18789 · local mode |
| Your laptop | SSH into MMM4 | Via Tailscale tailnet |
| Remote dashboard | GitHub Pages | nickxma.github.io/april-dashboard |
4-Tier Model Hierarchy
| Tier | Model | Alias | When Used | Relative Cost |
|---|---|---|---|---|
| Orchestrator | claude-opus-4-6 |
/model opus |
All conversations, routing, coordination, decisions | |
| Deep Think | claude-opus-4-6 |
/model opus |
Complex strategy, hard tradeoffs, analysis | |
| Coding | gpt-5.3-codex |
/model codex |
File edits, scaffolds, tests, code review | |
| Cheap | claude-haiku-3-5 |
/model haiku |
Summaries, simple lookups, tweet reading, formatting |
Concurrency Rules
| Setting | Limit | Reason |
|---|---|---|
| Main sessions | 1 | No parallel main instances |
| Sub-agents | 2 | Rate-limit safe ceiling |
| GO PARALLEL | opt-in | Only when you explicitly say so |
Sub-agent Default Model
Sub-agents default to Codex 5.3 (gpt-5.3-codex) unless overridden at spawn time. Claw will use Sonnet for balanced Anthropic tasks and Haiku for cheap/fast jobs.
Session Types
| Type | Key | Model | Lifetime | Who talks to it |
|---|---|---|---|---|
| Main | agent:main:main | Opus 4-6 | Ongoing conversation | You directly |
| Sub-agent | agent:main:sub-XXXX | Varies (Opus/Codex/Haiku) | Single task → auto-cleanup | Main agent only |
| Isolated/Cron | agent:isolated:XXXX | Configurable | Scheduled / timed | Nobody (async) |
Context Window vs Memory
| Storage | Scope | Survives restart? |
|---|---|---|
| Session context | Current conversation | No |
| memory.md | Durable facts/decisions | ✅ Yes |
| memory/YYYY-MM-DD.md | Daily notes | ✅ Yes |
| Notion DB | Objectives/projects | ✅ Yes (external) |
Session Lifecycle
- You send message → gateway routes to a channel session
- Main agent handles directly, or spawns sub-agent
- Sub-agent runs isolated task, reports back
- Main synthesizes → sends reply to same channel/session
- Durable facts → memory.md
Channel ↔ Session Map (vernacular)
| Term | What it actually means | Example | Context sharing? |
|---|---|---|---|
| Gateway | Router/runtime process on MMM4 | openclaw-gateway :18789 | N/A |
| Main session | Your direct conversation thread with Claw | agent:main:main | Own context window |
| Channel session | Per-channel/per-thread conversation state | Telegram DM vs Discord channel | Separate by default |
| Sub-agent | Temporary worker session spawned by main | agent:main:sub-* | No direct user chat; reports back |
| Isolated session | Background/cron worker session | agent:isolated:* | Separate |
If you have 10 Telegram channels, each channel/thread can accumulate its own context. They are not 10 sub-agents. Sub-agents are spawned workers for tasks, not channels.
Tools & Capabilities
| Tool | What it does | Status | Notes |
|---|---|---|---|
web_search | Search the web (Brave) | active | API key configured |
web_fetch | Fetch + extract page content | active | |
exec | Run shell commands on MMM4 | active | Elevated mode available |
Read / Write / Edit | File system | active | ~/.openclaw/workspace |
browser | Browser automation | intermittent | Use API-first |
notion skill | Read/write Notion databases | active | Read-first; write on explicit cmd |
github / gh | Issues, PRs, CI | active | Authed as nickxma |
cron | Scheduled jobs + reminders | active | |
sessions_spawn | Spawn sub-agents | active | Max 2 concurrent |
memory_search | Semantic search over memory | active | |
pdf | Analyze PDF documents | active | |
gog | Google Workspace (Gmail/Cal/Drive/Sheets/Docs) | active | Authed via gog CLI |
tts | Text-to-speech | active | |
image | Vision model | active | |
nodes | Paired device control | active | Camera/screen blocked by policy |
Blocked by Policy
- camera.snap / camera.clip / screen.record
- calendar.add / contacts.add / reminders.add
- Infra/config changes without explicit GO
- Notion writes without explicit command
- Parallel execution without GO PARALLEL
Skills Installed
| Skill | Purpose |
|---|---|
| notion | Read Notion DBs; manual writes |
| github | gh CLI — PRs, issues, CI |
| coding-agent | Delegate complex coding tasks |
| slack | Slack messaging (Socket Mode) |
| gog | Google Workspace — Gmail/Cal/Drive/Sheets/Docs |
| obsidian | Obsidian vault read/write |
| things-mac | Things 3 task ingestion |
| weather | Weather lookups (no key needed) |
| tmux | Terminal session management |
| openai-whisper | Audio transcription |
| video-frames | Video frame extraction + analysis |
| 1password | Credential lookups (1Password CLI) |
| gemini | Google Gemini model access |
| clawhub | Install/update skills from clawhub.com |
| healthcheck | Security / host audits |
Active Config Snapshot
| Setting | Value |
|---|---|
| Main model | claude-opus-4-6 |
| Sub-agent default | gpt-5.3-codex |
| Deep think | claude-opus-4-6 |
| Coding | gpt-5.3-codex |
| Cheap/menial | claude-haiku-3-5 |
| Max concurrent sessions | 1 |
| Max sub-agents | 2 |
| Compaction mode | safeguard |
| Workspace | ~/.openclaw/workspace |
| Gateway port | 18789 (loopback) |
| Tailscale SSH | off — enable: sudo tailscale set --ssh=true |
Model Switching Cheatsheet
| Command | Switches to | Use for |
|---|---|---|
/model opus | Opus 4.6 | Default — orchestrator for all conversations |
/model codex | Codex 5.3 | Default sub-agent workhorse |
/model sonnet | Sonnet 4.6 | Balanced Anthropic tasks when Codex isn't ideal |
/model haiku | Haiku 3.5 | Quick/cheap tasks |
How to Change Config Safely
- Tell Claw what you want changed
- Claw proposes exact change + waits for GO
- Applied via
config.patch(safe merge) - Gateway restarts, pings you when back
Memory — What Gets Stored
- Decisions and preferences
- Constraints and guardrails
- Active projects and deadlines
- Open loops and commitments
- Names, recurring routines
Secrets are not stored unless explicitly requested.
Memory — Key Constraints
- Parallel blocked by default →
GO PARALLELto unlock - System changes need explicit approval
- Rate-limit guardrails on concurrency
- Notion writes manual-only
- First-install artifacts archived (2026-02-19)
Open Loops
| Loop | Status | Next Action |
|---|---|---|
| Dashboard rebuild as objectives homepage | In progress | This task — update all stale data, ship r16 |
| Research tasks JSON stale | Stale | Last updated Feb 24 — needs daily refresh fix |
Source
Notion — Objectives DB
37 objectives tracked.
Read-first. Writes only on explicit command.
Type: "read my objectives from Notion" to pull a live list.
Highlighted Objectives
| Objective | Status | Notes |
|---|---|---|
| Baddies (dating) | Active | Personal life objective |
| Mindfulness content pipeline | Active | Waking Up transcript corpus → tweets/threads/essays/newsletter |
| Stablecoin expertise | Active | PayPal BCDC work, Arbitrum L2 wedge · deadline April 20 |
| Sell Manhattan co-op | Recovering | Material financial objective |
Tasks (Clear Ownership Queue)
| Task | Status | Ball | Next step | Scheduled agent |
|---|---|---|---|---|
| Rebuild dashboard to reflect latest system state | In progress | Claw | Update all stale data, ship build r16 | Coding Agent (Codex) |
| Stablecoin wedge matrix research | Active | Claw | PayPal BCDC + Arbitrum L2 wedge analysis · deadline April 20 | Research Agent |
| Mindfulness tweet pipeline | Active | Claw | Waking Up transcript → tweets/threads/essays/newsletter | Coding Agent (Codex) |
Status values: Ready, In progress, Waiting, Blocked, Done. Ball = whose court it is right now. Agent assignment mode: Planned (manual) for reliability (no noisy live polling/rate-limit risk).