Introduction

DeepSeek Harness extends Web Settings via plugins. For users who have connected multiple configured model services, the context window size usually needs to be confirmed separately; if set too large, it might also require the model to carry content beyond its actual capacity.

dsh-operating-context adds a “Working Window” page to Harness, limiting each configured model service to a working context window and keeping it within the model’s carrying capacity.

What is it?

dsh-operating-context is a DeepSeek Harness plugin maintained by AIMFllyYS, licensed under MIT, current version v0.1.0.

It is a Harness bundle and cannot run independently; the official entry point is:

npx @deepseek-ai/dsh

Its positioning is to provide a “Working Window” settings page for DeepSeek Harness, reusing the adapter, ContextMeter, and official compaction which have already read contextWindow.

Core Features

The following describes the main behaviors verified in v0.1.0.

  • Provides the “Working Window” settings page for DeepSeek Harness.
  • Limits each configured model service to a working context window, not exceeding the model’s carrying capacity.
  • Writes via settings.mutate at the user layer; does not mount a second compression engine, does not wrap resolveModel, and does not use a bundle patch that overwrites the entire line.
  • Writes different fields based on route shape: modelOverrides.<id>.contextWindow, models[].contextWindow, or defaultContextWindow.
  • Reads the ceiling of the installed catalog via llm.discoverModels; does not guess for unknown ceilings.
  • Removes the override when the selected window is not higher than the model ceiling, making apply idempotent and restoring native capacity.
  • When updating a route with an explicit model list, preserves endpoint, protocol, credential, names, output limits, and unknown metadata.
  • Provides prebuilt lib/ artifacts when installing v0.1.0; the installation does not execute package code.

Installation and Usage

First-time use requires Node.js ^22.19.0 or >=24, and pnpm available in the PATH; dsh plugin forwards to pnpm.

First install the recommended version v0.1.0:

npx @deepseek-ai/dsh plugin --profile web add github:AIMFllyYS/dsh-operating-context#v0.1.0

If operating within the DeepSeek Harness source repository, use pnpm dsh:

pnpm dsh plugin --profile web add github:AIMFllyYS/dsh-operating-context#v0.1.0

After installation, check if the configuration layer appears:

npx @deepseek-ai/dsh --profile web --dump-config

In the output, look for # == dsh-operating-context layer.

Start the Web UI:

npx @deepseek-ai/dsh web

Go to Settings -> Working Window, select size and apply.

If you no longer need this plugin:

npx @deepseek-ai/dsh plugin --profile web remove dsh-operating-context

Note that written values are retained in ~/.dsh/settings.yaml; removing the plugin does not roll back these values.

Typical Usage

After the above steps, a common usage workflow is:

  1. Install v0.1.0 and confirm the # == dsh-operating-context layer.
  2. Run npx @deepseek-ai/dsh web to open Web UI.
  3. Open Settings -> Working Window, select the desired context window size.
  4. After apply, run npx @deepseek-ai/dsh --profile web --dump-config again to verify the actual modelOverrides, models, or defaultContextWindow fields written.
  5. If no longer in use, execute the remove command; written values are still retained in ~/.dsh/settings.yaml.

Applicable Scenarios and Notes

Suitable for developers managing multiple configured model services under the DSH Web profile. It is suitable for scenarios where context windows need to be centralized into a settings page for adjustment, while ensuring written values do not exceed known model ceilings.

Precautions before use:

  • This plugin runs with the current dsh process permissions; check the source code and MIT license before installing.
  • The plugin has been compatibility reviewed with DeepSeek Harness 0.1.0-rc.5 (commit 47f943859bef60e4160492346772ded9b24f765a, 2026-08-14); the Harness plugin surface is still a release-candidate, and migration to newer versions should be re-confirmed.
  • unknown ceilings and explicit models lists cannot be reset by reaching known native ceilings.
  • If the catalog no longer contains models that still have modelOverrides entries, the adapter will reject the stale entry and may affect the entire route; the page will report the number of entries to be removed, and does not attempt removal when the catalog is unknown.
  • Current v0.1.0 contains compiled lib/ files and defines no Git build triggers; old Git revisions may fail due to frozen-lockfile, autoInstallPeers, or allowBuilds.
  • PeerDependencies include @deepseek-ai/cordis, @deepseek-ai/dsh-api-remotes, @deepseek-ai/dsh-client-locale, @deepseek-ai/dsh-client-runtime, @deepseek-ai/dsh-client-schema-form, @deepseek-ai/dsh-client-ui-primitives, @deepseek-ai/dsh-client-ui-settings, @deepseek-ai/dsh-client-ui-slots, @deepseek-ai/dsh-client-web-react and react ^18.2.0.

Conclusion

dsh-operating-context provides a centralized working window configuration page: it limits each configured model service to a carrying capacity range and writes the corresponding fields according to the route shape.

GitHub: https://github.com/AIMFllyYS/dsh-operating-context