Skip to main content

DSH AI Providers 0.3.0 works with DSH 0.2.1

· 4 min read
Firefox2100
Firefox2100
Maintainer of DSH AI Providers

DSH AI Providers 0.3.0 is published. It runs on both DSH 0.2.0-rc.2 and DSH 0.2.1-rc.2, narrows the DSH range it claims to support, and lets a language-model route tell DSH more about itself.

Why a release was needed​

DSH 0.2.1 checks each plugin's declared requirements when it starts, and it disabled @dsh-ai-providers/runtime from version 0.2.0 of this project. The package pinned @deepseek-ai/dsh-llm to exactly 0.2.0-rc.2, which no other DSH release can satisfy. The source itself needed no change to run: the packages type-check and pass their tests against both DSH releases.

DSH also reads a different field than the one these packages used. It looks at the top-level engines.dsh, which now decides alone when present, not at dsh.engines. 0.3.0 declares engines.dsh as >=0.2.0-rc.2 <0.2.1 || >=0.2.1-rc.2 <0.2.2. That stops before the next patch release instead of at the next minor, because DSH is pre-stable and a patch can change what a provider relies on. The @deepseek-ai/dsh-* peer ranges list the two tested releases, and the release check now enforces all of this in every package.

Two commands keep it honest. pnpm check:dsh <version> type-checks and tests a temporary copy of the workspace against another tested DSH release, and CI runs it for the older one. When a change cannot serve both releases, the older one will be dropped, not kept by luck.

What a route now reports to DSH​

The previous releases registered a route with its models and little else. The adapter now covers more of what DSH's own model adapter offers:

  • Retry policy. A route can set retry in DSH's own shape. Without it DSH uses its defaults; a wrong policy makes the route unavailable and says why.
  • One reading per call. The model facts and the request that follows come from the same reading of the routes, so an edit between the two cannot mix them.
  • Configurable providers. Every route appears in DSH's directory of configurable providers, with the address of its connection and any problem that stops it being served.
  • Model discovery. DSH can ask a settings section for models. A configured route answers with its own models, and a draft endpoint is asked with GET /models using a key that is used once and not stored.
  • System prompt in the conversation. A model whose chat template reads a changed system message anywhere in the conversation can say so with systemInHistory, which lets DSH change the system prompt without rewriting the start of a cached conversation. It is off by default, because many local chat templates insist that the system message comes first.

The route editor has a field for the retry policy and a checkbox for systemInHistory. Existing configuration keeps working without changes.

Things that behave differently on DSH 0.2.1​

  • A key stored in DSH's credential store now takes precedence over an exported environment variable with the same name, and a key can be stored while the variable is exported. On DSH 0.2.0 the exported variable wins.
  • DSH expects a model's reasoning levels from the least to the greatest effort and uses the first one for light work such as session titles. List a route's levels in that order.

What is not covered​

Images are sent exactly as stored, so the providers declare no image resizing and no image pricing, and every request carries the complete tool list. Both are deliberate: declaring more would promise behaviour the wire code does not implement. Both tested DSH releases are pre-releases, and the providers have been exercised with LocalAI. The language models guide describes each setting.