Update SK Agent + OWUI configs for OmniCoder-9B (replaces ZwZ-8B)

[jsboige]

Context

OmniCoder-9B (cyankiwi/OmniCoder-9B-AWQ-4bit) has replaced ZwZ-8B on GPU 2, port 5001 since 2026-03-28.

Same endpoint, same API key, new model_id: omnicoder-9b (was zwz-8b).

Changes Required

1. SK Agent Config (roo-extensions/mcps/internal/servers/sk-agent/sk_agent_config.json)

• Replace model entry zwz-8b → omnicoder-9b (model_id, description)
• Update agent vision-analyst description: "OmniCoder-9B — 96-107 tok/s, thinking+vision, OCR 97.5%, MME 1258.5"
• Rename sampling key zwz-8b → omnicoder-9b
• Consider adding a coder agent leveraging OmniCoder's agentic coding specialization

2. OWUI Model Wrappers (via Open WebUI API)

• Update vision-expert wrapper: backend model → omnicoder-9b
• Create OmniCoder_think wrapper (thinking ON, temp 0.7, pp 1.5, top_p 0.95, top_k 20)
• Create OmniCoder_code wrapper (thinking ON, temp 0.6, pp 0.0, top_p 0.95, top_k 20)
• Create OmniCoder_instruct wrapper (thinking OFF, temp 0.7, pp 1.5, rp 1.1, min_p 0.01)

3. Local vllm config cleanup

• Removed local sk_agent_config.json copy (commit 6c24a51f2)

OmniCoder-9B Key Specs

Metric	Value
Decode	96-107 tok/s
Tool call	1.09s
Tool parser	qwen3_coder (XML format)
Thinking	Yes (qwen3 reasoning parser)
MME	1258.5 (88.5%)
MMStar	58.5%
OCR	97.5% (+15 pts vs ZwZ)
Context	131K (configured)

Notes

• tool-call-parser is qwen3_coder (NOT hermes) — generates XML <function=name><parameter=key>value</parameter>
• torch.compile cache can corrupt after restart → fix: docker volume rm profiles_vllm-compile-cache-omnicoder
• Custom Dockerfile required: vLLM nightly + transformers >= 5.0

🤖 Generated with Claude Code