msclo
qwen2.5:7b · qwen3:8b
qwen2.5:32b · phi4
nomic-embed-text
pcdev
qwen2.5:14b · llama3.1:8b
nomic-embed-text
MAL :9420 dead
msi01
Ollama via pcdev tunnel
MAL SSH routing issue
JESEOF · H100
NC40ads_H100_v5 available
24 vCPU quota free
canadacentral — spinnable now
Anthropic
Claude Sonnet 4.6 (current)
Claude Opus 4 · Haiku 4.5
OpenAI / Google
o3 / GPT-4o available
Gemini 2.5 available
API keys not configured
AUTO GEAR ROUTING CHAIN
T0 · INSTANT
Local Fast Caps
$0 / task · <1s
qwen2.5:7b · msclo
qwen3:8b · msclo
qwen2.5:14b · pcdev
llama3.1:8b · local
USE FOR: heartbeat checks · quick lookups · classification · CDNET pattern matching · simple Q&A · routing decisions
msclo ✓ LIVE
pcdev ~ MAL dead
↓ if T0 quality score < 0.7 or task complexity HIGH
T1 · ANCHOR
Fleet Heavy Caps
$0 / task · 3-8s
qwen2.5:32b · msclo
gpt-oss:20b · msclo
phi4 · msclo
JESEOF qwen:72b FP16 ⚡
USE FOR: DESFACTOR runs · CDNET analysis · SORRY chain reasoning · architecture decisions · code generation · multi-step tasks
32b ✓ msclo
JESEOF not spun
↓ if T1 quality score < 0.75 or task needs external knowledge
T2 · PAYG
Anthropic Escalation
$0.003–$0.015 / 1K tokens
claude-haiku-4.5
claude-sonnet-4.6
claude-opus-4
USE FOR: main sessions (current default) · external comms drafts · legal language (ATMOS) · final floor decisions · anything T1 cannot resolve
Sonnet 4.6 ✓ LIVE (current)
Opus 4 ✓ LIVE
MANUAL MODE — always available. Use session_status(model=...) to pin any model anytime.
HOW FAR AWAY — GAP ASSESSMENT
DONE
T0 Local Fleet — msclo
running now
msclo has 7 models live. qwen2.5:7b, qwen3:8b, qwen2.5:32b, phi4, nomic-embed-text. MAL :9420 live. Ollama healthy. This tier IS available right now — just not wired into OC auto routing yet.
30 MIN
Fix pcdev MAL — restore T0 full capacity
SSH + 1 command
pcdev Ollama is healthy at :11434. MAL bootstrap.js died after Ollama rebind. Just needs a restart via SSH. SSH has been flaky — might need Kay to manually restart on pcdev directly.
- SSH to 192.168.2.16
- cd ~/mal && node bootstrap.js &
- Verify :9420 responds
2–3 HRS
JESEOF H100 — spin + load Qwen 72B FP16
Azure CLI, ~$3–5/hr
NC40ads_H100_v5 ready in canadacentral. 24 vCPU quota free. Spin VM → install Ollama → load qwen2.5:72b FP16 (full quality, no quantization loss). Wire into MAL as T1 anchor. Expected: 67–69% ARC-AGI (+3-5pp over Q4_K_M).
- az vm create NC40ads_H100_v5 canadacentral
- curl https://ollama.ai/install.sh | sh
- ollama pull qwen2.5:72b (FP16 — ~150GB)
- Add JESEOF_ENDPOINT to MAL config
4–8 HRS
Auto-Gear Router Service
New Node.js service
A lightweight service that sits between OC and the model tiers. Task arrives → classify (simple/medium/complex/critical) → route to lowest available tier that can handle it → escalate if quality threshold not met. Default: T0. PAYG only as last resort.
- Task classifier (regex + embedding similarity)
- Tier health checker (polls each endpoint)
- Quality scorer (length + coherence heuristic)
- Escalation logic + audit log
- Deploy as master1-system service :9430
1 DAY
OpenClaw Model Override Hook
openclaw.json + OC config
OC already supports per-session model override via session_status(model=...). The auto-gear router can use this API to dynamically switch models. Default mode: auto (router decides). Manual mode: pin any model. Main session stays Anthropic by default — auto gear routes sub-agent and heartbeat tasks to OSS.
- Add auto-gear endpoint to openclaw.json
- Define task routing rules in HEARTBEAT.md
- Sub-agent spawns default to T0/T1 (OSS)
- Main session stays T2 (Anthropic) — manual override anytime
ONGOING
External Provider Wiring — OAI + Google
API keys + OC provider config
o3, GPT-4o, Gemini 2.5 Pro — all available via API but not currently configured in OC. Adding them makes them available as T3 escalation above Anthropic (for ARC-AGI tasks specifically). Not urgent — Anthropic handles T2 well. Worth adding when o3 drops in price.
- Add OPENAI_API_KEY to OC secrets
- Add GOOGLE_AI_API_KEY to OC secrets
- Configure as T3 in auto-gear router
AUTO GEAR READINESS — WHERE WE ARE NOW