dsh-liquid-inference

Liquid Inference for DeepSeek Harness: a liquid model route plus read-only market tools.

Install

Not published to npm yet. Install from GitHub or from a packed tarball, then export a buyer key with the spend scope and restart the profile:

export LIQUID_API_KEY=sk-liquid-...

From GitHub. The built lib/ is committed, so no build step or allowBuilds entry is needed:

dsh plugin --profile web add github:architect-xyz/liquid-dsh-plugin

Pin a commit (github:architect-xyz/liquid-dsh-plugin#<sha>) so a later push cannot change what you install.

From a tarball. Needs no build permission:

git clone https://github.com/architect-xyz/liquid-dsh-plugin.git
cd liquid-dsh-plugin && pnpm install && pnpm pack
dsh plugin --profile web add ./dsh-liquid-inference-0.1.0.tgz

Pick Liquid Auto (liquid.auto) in the model picker.

What it adds

Model route liquid — no adapter code. The bundle patch configures the stock llm-pi-ai adapter with api: openai-completions, baseURL: https://router.inference.ai.exchange/v1 and apiKeyEnv: LIQUID_API_KEY.

• Cap. Each request carries x-liquid-cap-usd, the most it can cost, locked before the first token. The default is 1 USD; set LIQUID_CAP_USD in the launching environment to change it.
• Models. liquid.auto plus its ranking variants :cheapest, :economy, :balanced and :premium (documented in the description of GET /v1/exchange/routers), and deepseek-v4-flash as a pinned model. Add any id from GET /v1/models with a models entry in your profile's cordis.patch.yml.
• Reasoning. Levels low, medium and high are sent as reasoning_effort. The exchange passes the level through unchecked, so a level the winning provider rejects fails that request.

Tools (plugin liquid-inference)

Tool	Endpoint	Key
liquid_get_book	GET /v1/exchange/offers	no
liquid_get_markets	GET /v1/exchange/markets	no
liquid_get_routers	GET /v1/exchange/routers	no
liquid_get_job	GET /v1/jobs/{id}	yes

Chat, image, speech and video purchases are deliberately not tools: the harness's own model route is the one place that spends.

Data flow

Full prompts and tool results go to router.inference.ai.exchange, which forwards them to whichever third-party provider wins the auction. The key is read per request, from the credential store when the profile mounts one and from LIQUID_API_KEY otherwise (headless profiles disable the store). It is never written to config.

Notes

• The patch replaces the llm-pi-ai row's whole config. If another bundle also configures pi-ai providers, restate them in your profile's cordis.patch.yml.
• Develop: pnpm install && pnpm run check. Commit the rebuilt lib/ with source changes (CI fails on a stale one). pnpm run test:install packs the bundle, installs the tarball into an isolated profile of a real dsh (DSH_VERSION picks the release, default next) and boots it; it needs network access.