Ollama Jev Protocol Status on chonk

The Ollama installation on chonk.lab.bitsmasher.net remains at v0.35.0. The /v1/systemone endpoint shipped in this version is unimplemented — it returns HTTP 400 with the error message "Value looks like object, but can't find closing '}' symbol" for every valid JSON payload tested across multiple field names (state, prompt) and formats.

Model Inventory

ModelSizeFamilyStatus
tev1:latest4.5 GB (Q8_0)qwen35Pulled, recognized as qwen35 family, context 262144
nimble:latest9.5 GB (Q8_0)Bespoke-Nimble-9B-merged-currentPulled but impractical — requires ~9.5 GB VRAM on 16 GB RX 9070, leaving insufficient headroom for concurrent qwen3.6 primary model operation

GPU Constraints

chonk's AMD RX 9070 (16 GB VRAM) + Granite Ridge iGPU presents a hard ceiling for local model selection. The decisionModel must stay under approximately 7 GB to leave adequate headroom for the qwen3.6 primary model and system processes. This rules out nimble:latest for concurrent use but leaves tev1:latest as a viable fast-triage option once Ollama is upgraded.

OpenClaw decisionModel — Not Wired

OpenClaw's openclaw.json has no decisionModel configuration set. The routing primitive exists in the build schema but is completely unused. Once Ollama is upgraded to v0.36+ (highest-priority fix, which should resolve the /v1/systemone parsing issue), the next step is to wire:

agents.defaults.decisionModel: "ollama/tev1:latest"

Post-Upgrade Plan

panw-steward Agent — Tools Augmentation

The panw-steward (Palo Alto firewall steward) agent's tools profile was updated during this session:

No gateway restart is required — tool profiles are evaluated per-session at OpenClaw runtime. The changes take effect on the next session initialization for panw-steward.