LocalForge for DeepSeek Harness

A high-performance, MIT-licensed adapter that connects DeepSeek Harness (dsh) to local LLM servers — primarily LM Studio, but also any other OpenAI-compatible or Anthropic-compatible endpoint (Ollama, vLLM, llama.cpp's HTTP server, text-generation-inference, etc.).

---

Why LocalForge?

Three other LM Studio plugins for dsh exist (and we read all of them carefully — see Inspirations and the Comparison). LocalForge is the one that adds the things missing from every other adapter:

• Live discovery cache — /v1/models + native /api/v0/models fallback, with TTL and explicit invalidation. The other plugins re-probe the server on every resolveModel.
• Auto-load — if the requested model isn't resident on LM Studio, ask the server to load it (POST /api/v0/models/{id}/load) and poll until ready.
• Per-model concurrency limiter — LM Studio saturates under parallel streams on quantized models; LocalForge serializes them per model id.
• Model fallback chain — fallback: [primary, secondary]. If the primary model returns a recoverable error, the next one is tried automatically.
• Reasoning budget — caps how much internal "thinking" the model is allowed to do before producing the final answer.
• Sampling presets — code, chat, creative, precise for sane defaults that you can override per request.
• Token pre-flight — reject requests that would overflow the context window before paying for a round-trip.
• Prometheus metrics — /metrics with TTFT, TPS, queue depth, error rates, and cache hits.
• Optional health check — periodic probe with auto-reconnect when enabled; the /metrics up gauge surfaces the state.
• Anthropic-compatible adapter — second provider route using /v1/messages, useful for Claude-style system prompts and tool calls.
• Structured logging with credential redaction — every log line is safe to ship to Loki / ELK without scrubbing.
• Strict TypeScript — strict + noUncheckedIndexedAccess, full test coverage, GitHub Actions CI.

Install

Prerequisites

• DeepSeek Harness (dsh) with the web profile enabled.
• A local LLM server. LM Studio 0.4.8+ with Developer → Start Server on (default port 1234) is the most-tested target; any OpenAI- or Anthropic-compatible server works.
• Node.js ≥ 22.19 and pnpm ≥ 9.

From the DSH Plugins page

1. Open Settings → Plugins → Repository source → Add.
2. Enter github:orfeomorello/dsh-localforge#main and install it.
3. Enable LocalForge in the plugin list; if prompted, restart dsh web.

The repository installer mounts community plugins immediately. This repository also includes the compiled lib/ runtime because the repository plugin manager imports that entrypoint directly; GitHub's installer does not build TypeScript source for it. For a bundle-style installation from a terminal instead, use:

dsh plugin --profile web add github:orfeomorello/dsh-localforge#main
dsh web --dump-config # verify the dsh-localforge bundle layer
# Restart dsh web after installing the bundle

In the model picker at the top of the page you'll now see LocalForge as a provider; pick any model LM Studio has loaded. If the UI still reports failed to import, capture the complete browser console error or the dsh web terminal diagnostic: the short UI message does not expose the underlying module-resolution error.

Quick start

Minimal ~/.dsh/settings.yaml section:

llm-localforge:
  defaultConnection:
    baseURL: http://localhost:1234/v1
  models:
    - id: qwen/qwen3-8b
      preset: code
      fallback:
        - qwen/qwen3-4b
        - llama-3.1-8b

That's it. Point dsh at any conversation, select "LocalForge" as the provider, and traffic flows to LM Studio. Your DeepSeek API balance is untouched.

Configuration

See docs/CONFIGURATION.md for the full reference (every section, every field, every default).

Architecture

See docs/ARCHITECTURE.md for the design notes — the per-request resolver trick, the dual-endpoint discovery, the SSE parsing, the Prometheus metric set, and the testing strategy.

Inspirations

LocalForge is a clean-room reimplementation. We read three existing plugins to learn the patterns, then wrote everything from scratch with our own license. The inspirations are:

• Viktirr/dsh-llm-lmstudio — the cleanest TypeScript in-tree adapter, the source of the per-request resolver pattern.
• jtc268/deepseek-harness-lmstudio — the inspiration for Qwen-specific tool-calling robustness.
• starlightzfy/dsh-plugin-lmstudio — the inspiration for the plugin-bundle discovery UX.

No code was copied. See docs/COMPARISON.md for the full feature-by-feature matrix.

License

MIT © 2026 The LocalForge authors.