dsh-model-modes

Capability-aware model modes for the DeepSeek Harness web composer.

Features

• Keeps reasoning effort in DSH's official model selector. No level is hard-coded: each exact model exposes only the efforts advertised by its adapter.
• Adds a Fast switch immediately before Send.
• Fast first applies an explicit provider/model route. Without one, it can select the current model's first advertised reasoning effort.
• DSH logs the resulting request configuration in request/header.
• Addressed subagent sessions do not expose the switch.

What Fast means

DSH 0.1.0-rc.6 has no provider-neutral native Fast field. OpenAI serviceTier: priority and Anthropic speed: fast are not carried through the current adapter chain, so this plugin does not claim to enable them.

Fast is instead a verifiable policy:

1. Switch to a configured fast model route (for example a flash, turbo, or fast model).
2. Otherwise, with lowerReasoning: true, select the exact model's first adapter-advertised reasoning effort.
3. If neither capability exists, preserve the original request and log a Host warning.

Configuration

- insert:
    - id: dsh-model-modes
      name: dsh-model-modes
      config:
        lowerReasoning: true
        routes:
          - provider: myself
            model: gpt-5
            fastProvider: myself
            fastModel: gpt-5-fast
            fastEffort: minimal

provider and model accept *. Empty fast target fields preserve the source route; an empty fastEffort uses the target model's first advertised effort.

Install

dsh plugin --profile web add \
  https://github.com/DTSFO/dsh-model-modes/archive/refs/heads/main.tar.gz
dsh web

Remove with dsh plugin --profile web remove dsh-model-modes.

Development

pnpm install
pnpm run check
dsh plugin --profile web add /absolute/path/to/dsh-model-modes

Known limitations

• The switch is process-local per session and defaults off after Host restart. Request configurations already used remain durable in the session log.
• This plugin does not replace DSH's official model menu.
• Provider-native priority/speed parameters require future DSH Core and adapter support.

MIT