Windows Local Setup for DeepSeek Harness (DSH) with NVIDIA NIM

[collectivewinca]

Windows Local Setup for DeepSeek Harness (DSH) with NVIDIA NIM

Summary

Successfully ran DSH v0.1.1-rc.2 on Windows 11 (PowerShell 5.1) as a self-hosted coding-agent runtime pointing at NVIDIA NIM (integrate.api.nvidia.com/v1). The setup works with the free-tier NVIDIA key and the deepseek-ai/deepseek-v4-flash-0731 model.

Prerequisites

• Node.js v22.x, npm v11.x, pnpm v11.x
• Windows 11 / PowerShell 5.1

Key Steps & Issues Encountered

1. npm i -g @deepseek-ai/dsh → installs v0.1.1-rc.2
2. YAML encoding gotcha — PowerShell's > writes UTF-16 (BOM null bytes); DSH parser rejects this. Fix: always use Set-Content -Encoding ASCII for .yaml/.yml files.
3. Built-in deepseek-official provider — DSH has a hardcoded provider route deepseek-official that ignores arbitrary settings.yaml keys. Must override via cordis.patch.yml.
4. Cordis patch override — The default model is set by plugin agent-default-model (provider: deepseek-official, model: deepseek-v4-flash). Override in profiles/<web|headless>/cordis.patch.yml:

   - id: agent-default-model
     config:
       provider: nvidia
       model: deepseek-ai/deepseek-v4-flash-0731

5. NVIDIA key & models — Key nvapi-... authenticated successfully (410 not 401). Only deepseek-ai/deepseek-v4-flash-0731 is live on this tier; other live models: moonshotai/kimi-k3, qwen/qwen3-next-80b, etc.
6. Web UI launch — dsh web --host 127.0.0.1 --port 3080 --no-open → HTTP 200 at http://127.0.0.1:3080
7. Headless end-to-end — dsh --profile headless "Reply with exactly: DSH NVIDIA WORKS." → Output: DSH NVIDIA WORKS.

Commands That Work

# Web UI
dsh web --host 127.0.0.1 --port 3080 --no-open

# Headless one-shot
dsh --profile headless "Say hello in one sentence"

# Interactive REPL
dsh --profile headless

Issues Summary

#	Issue	Fix
1	UTF-16 BOM in YAML files	Set-Content -Encoding ASCII
2	deepseek-official provider ignored	Override via cordis.patch.yml → agent-default-model
3	deepseek-v4-flash EOL Aug 7 2026	Use deepseek-ai/deepseek-v4-flash-0731
4	NVIDIA key auth + model discovery	Query /v1/models, pick live model id

Verified Working

• ✅ NVIDIA API key authenticated
• ✅ deepseek-ai/deepseek-v4-flash-0731 replies "DSH NVIDIA WORKS."
• ✅ Web UI serves at http://127.0.0.1:3080 (HTTP 200)
• ✅ Headless task prints result and exits

Contact / Notes

• NVIDIA key: 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
• This setup routes DSH through NVIDIA NIM, which supports ~15+ models including DeepSeek, Kimi, Qwen, Gemini, and Grok via the TeamoRouter gateway.
• The full setup report (with all process/IO details) is at D:\Internship\Projects\edgecli\deepseek-harness-dsh-setup.md

---

Contributed by: VE LAB / Edge CLI experiment • Aug 2026