Foreword

When developing agents with DeepSeek Harness (DSH), the model catalog is unavoidable: to select a specific model directly in the model selector, a prerequisite is that a plugin has connected its description, endpoint, and credentials. If you hold a Zhipu GLM Coding Plan membership or a BigModel account, these models are not in the default DSH selection list, particularly lacking glm-5.3 with 1M context.

The opensaozi/dsh-bigmodel-catalog introduced below solves this problem by integrating Zhipu BigModel / GLM Coding Plan as a model catalog plugin into DSH.

What is this

opensaozi/dsh-bigmodel-catalog is a DSH model catalog plugin maintained by OpenSaozi, under the MIT license, with the npm package name @deepseek-ai/dsh-model-catalog-bigmodel (version 0.1.0-rc.5, private package).

After installation, it unlocks glm-5.3 (1,000,000 token context limit, 128K max output), while also supporting glm-4.5, glm-4.6, and glm-5. It connects to Zhipu’s developer-facing Coding PaaS endpoint https://open.bigmodel.cn/api/coding/paas/v4/models.

DSH’s philosophy is “everything is a plugin,” and the model catalog is no exception. Please note: this is a community integration and has no affiliation with, nor endorsement from, Zhipu AI, BigModel, DeepSeek, or pi-ai.

Core Features

  • Unlock 1M Context: Directly select glm-5.3 in the model selector, with a context limit of 1,000,000 tokens and a maximum output of 128K.
  • Coding Plan Focused: Connects to the Coding PaaS endpoint, supporting models including glm-5.3, glm-4.5, glm-4.6, and glm-5.
  • Live Reconciled Catalog: Combines real-time discovery from the endpoint while preserving verified exceptions to maintain support for glm-5.3.
  • Independent Provider Routing: The plugin registers its own zai adapter, reusing only the exported PiAiAdapter for message and stream transformation, without relying on the llm-pi-ai provider configuration or directory seams. The catalog, real-time filtering, credential parsing, and adapter registration for the zai route are all handled by this package.
  • Credential Security Isolation: BIGMODEL_API_KEY is parsed via the Harness credential service for each request; keys are not written to Cordis configuration or logs.
  • No-Build Installation: The repository includes dsh.bundle patches and pre-built runtime artifacts; no local build steps are required after installation.

How the Catalog is Reconciled

Two points to know:

  1. The Coding Plan Models endpoint currently does not return a directly callable glm-5.3; this ID is a clearly marked verified exception.
  2. The endpoint only provides model IDs and does not contain complete capability metadata. Context and output capacity come from an audited description table and are locked by package tests.

Refresh timing: Run once on mount; run again when the hosted credential corresponding to BIGMODEL_API_KEY changes. On refresh failure, the maintained catalog is preserved and the error is logged.

Additionally, the plugin itself does not add any request content—it only controls which BigModel descriptions (e.g., glm-5.3) are available for selection; request transformation is handled by the exported PiAiAdapter. Therefore, it has no direct impact on tokenization or KV Cache; switching models may affect provider-side cache eligibility.

Installation and Enablement

First, prepare a credential named BIGMODEL_API_KEY in the Harness credential service, then install the reviewed revision to the specified profile:

dsh plugin --profile <profile> add github:OpenSaozi/dsh-bigmodel-catalog#<commit-sha>

The repository comes with dsh.bundle patches and pre-built runtime artifacts; no local build is needed after this step.

Post-Installation Configuration

After installation, the bundle contributes the following Cordis configuration:

- id: model-catalog-bigmodel
  name: '@deepseek-ai/dsh-model-catalog-bigmodel'
  config:
    apiKeyEnv: BIGMODEL_API_KEY
    baseURL: https://open.bigmodel.cn/api/coding/paas/v4
    displayName: BigModel

During deployment, you can override the two items as needed:

  • baseURL: Points to a different Coding Plan endpoint;
  • apiKeyEnv: Points to a different credential reference.

Suitable Scenarios and Notes

Suitable scenarios: Developers holding a Zhipu GLM Coding Plan membership or BigModel account who need to analyze large codebases, process long documents, or perform high-intensity code generation.

Notes:

  • The plugin runs with the current dsh process permissions; check the source code and license (MIT) before installation.
  • Refresh failures will not clear the catalog; the maintained catalog is preserved and the error is logged.
  • The plugin has no direct impact on tokenization or KV Cache, but switching models may change provider-side cache eligibility.

Summary

This plugin does one thing very focused: connecting Zhipu Coding Plan models into the DSH model selector, with a focus on unlocking glm-5.3 with 1M context, while handling credential isolation, catalog reconciliation, and independent routing properly. If you are already using Zhipu’s Coding Plan, you can use these models in DSH immediately after installation.

Repository URL: https://github.com/OpenSaozi/dsh-bigmodel-catalog

Community directory listing page: https://www.skillhub.cn/plugins/OpenSaozi/dsh-bigmodel-catalog (This address is from a lead, not directly verified in repository materials)