dsh-bigmodel-catalog
开发工具 活跃维护

dsh-bigmodel-catalog

OpenSaozi/dsh-bigmodel-catalog

原生支持接入GLM Coding Plan及glm-5.3等多款智谱大模型,为智谱BigModel官方模型清单插件,无需额外配置即可扩展对应框架可调用的模型资源池,部署使用简单。

1
Stars 标星
0
Forks 分支
1
Watchers 关注
0
Open Issues
TypeScript
主要语言
MIT
开源协议
31 KB
仓库大小
1 个月前
最后推送
一键安装扩展 / 插件指令
dsh plugin --profile web add github:OpenSaozi/dsh-bigmodel-catalog
git clone https://github.com/OpenSaozi/dsh-bigmodel-catalog.git
git clone git@github.com:OpenSaozi/dsh-bigmodel-catalog.git
README.md main

DSH Plugin: LLM Provider - BigModel GLM

中文 | English

Allows connecting to the Zhipu AI BigModel / GLM Coding Plan as an LLM Provider, enabling DeepSeek Harness to utilize the dedicated Coding Plan route and unlocking the verified 1M-token context model glm-5.3 (along with glm-4.5, glm-4.6, and glm-5).

Target Audience and Use Cases

Designed for developers with a Zhipu GLM Coding Plan membership or BigModel account who need to analyze large codebases, process long documents, or perform intensive code generation.

Key Features

  • Unlocks 1M Context Window: Directly select glm-5.3 in the model selector, featuring a 1,000,000-token context limit and 128K max output capacity.
  • Tailored for Coding Plan: Connects to Zhipu AI's developer-focused Coding PaaS endpoint (https://open.bigmodel.cn/api/coding/paas/v4/models).
  • Live Reconciled with Verified Exceptions: Combines live endpoint discovery while maintaining verified support for glm-5.3.
  • Independent Provider Route: The plugin registers its own zai adapter. It reuses the exported PiAiAdapter only for message and stream conversion, without depending on llm-pi-ai provider configuration or catalog seams.
  • Secure Credential Isolation: The plugin resolves BIGMODEL_API_KEY through the Harness credential service for each request and never places the key in Cordis configuration or logs.

Installation and Quick Start

Install a reviewed revision into a DeepSeek Harness profile. The repository includes a dsh.bundle patch and prebuilt runtime artifacts so no local build steps are required:

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

The installed 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

The plugin owns the complete zai route: catalog, live filtering, credential resolution, and adapter registration all live in this package. Deployments may override baseURL for another Coding Plan endpoint and apiKeyEnv for another credential reference.

Contributing

Please attach provider evidence for every catalog change: exact model ID, protocol, context/output capacity, reasoning behavior, and direct request verification where applicable. Update both README languages and run package tests in a matching DeepSeek Harness checkout.

License and Disclaimer

MIT License. This community integration is not affiliated with or endorsed by Zhipu AI, BigModel, DeepSeek, or pi-ai.

Model Experience

Catalog Selection

What the model sees

No package-owned text. The plugin controls which BigModel descriptor, such as glm-5.3, can be selected; the exported PiAiAdapter owns request conversion.

Token effect

No direct token effect. Tokenization and context use remain properties of the selected model and pi-ai's z.ai-compatible request implementation.

KV Cache effect

No direct effect. Selecting a different model may change provider-side cache eligibility, but this plugin adds no request content.

Known Limitations and Deferred Work

  • Refresh runs on mount and when the managed credential named by BIGMODEL_API_KEY changes. Failure keeps the maintained catalog and logs the error.
  • The Coding Plan Models endpoint currently omits a directly callable glm-5.3; that exact id is an explicit verified exception.
  • The endpoint supplies ids, not complete capability metadata. Context and output capacities therefore come from the reviewed descriptor table and remain pinned by package tests.