dsh-plugin-h3-hip

DeepSeek Harness Host bundle for h3-hip.c (h3 --serve, protocol v1alpha). Independent from the HIP tree so the plugin can track DSH preview breakage without an engine release.

Requires h3-hip.c ≥ v0.12-exp with --serve, and a DSH web profile that already mounts tools + jobs (dsh-jobs-local + dsh-tool-jobs). Pin the DSH CLI you tested; preview APIs break.

What M1 provides

• h3_generate — semantic submit to POST /v1/jobs, then ctx.jobs tails SSE
• h3-prompt skill — MiniMax-H3 three-part prompts and reference tags
• GET /v1/info protocol check (X-H3-Protocol: v1alpha; 426 → tell the user to upgrade)
• Fake daemon for tests (npm test)

Not in M1: prompt lint, Web MP4 card, GPU scrape, SQLite job ledger, CLI Plan B.

Install

Start the engine (loopback only):

./h3 -d "$H3_MODEL" --serve

Install this bundle into the web profile (from a clone of this repo):

dsh plugin --profile web add github:alexhegit/dsh-plugin-h3-hip#<40-char-sha>
dsh --profile web --dump-config   # look for "# == dsh-plugin-h3-hip"
dsh --profile web

Local link while developing:

dsh plugin --profile web add ./

Config is environment-first: H3D_BASE_URL or H3D_BIND + H3D_PORT (default http://127.0.0.1:8571). Do not bind the daemon to the public internet; DSH is the only network surface.

Agent usage

1. skill → h3-prompt when writing the prompt.
2. h3_generate with prompt and optional preset / mode / geometry.
3. job_output / job_kill on the returned DSH jobId.

Develop

npm test
node tests/fake-daemon.mjs --listen   # H3D_PORT=8571 stand-in

License

MIT. MiniMax-H3 weights are under the model community license — review before IM/public sharing (h3-hip.c design Q1).