deepseek-harness-model-config

One-Line Description

Adds a "Model Advanced Configuration" settings page to the DeepSeek Harness Web UI, enabling creation of custom model endpoints (OpenAI/Anthropic protocols) without restarting DSH, pulling public model lists from endpoints, and configuring capabilities like context windows, max output, and reasoning effort per model.

Core Features

Add custom endpoints: Fill in name, URL, API Key and protocol, saved immediately available in session selector
Support three API protocols: openai-completions, openai-responses, anthropic-messages
One-click fetch model candidates: Pull candidate list via OpenAI-compatible GET /models, provide select all/deselect all/none
Configure capability parameters by model: context window, max output, input modalities (text/image), reasoning effort
Auto-complete from models.dev: After pulling, use https://models.dev/api.json to auto-complete missing context, max output and reasoning info
API Key only goes through Harness credential system: will not be written to settings.yaml, also won't appear in config preview

Technical Implementation

Language: JavaScript (ESM type: "module")
Key dependencies: DSH built-in client plugins @deepseek-ai/dsh-client-runtime / @deepseek-ai/dsh-client-ui-settings / @deepseek-ai/dsh-client-locale / @deepseek-ai/dsh-api-remotes (injected via window.__ModuleLoader__.load)
Architecture: Cordis dual-end plugin, host side only declares Loader name models-config-plugin, doesn't register any services, all read/write goes through host's existing settings RPC; client side registers "Advanced model config" page in settings.section slot
Entry files: src/index.js (Host, 8 lines) + src/client.js (Client, 1661 lines React) + cordis.patch.yml (bundle injection declaration)

Use Cases

When you need to connect DSH to a local OpenAI-compatible gateway (like one-api / new-api / self-developed proxy) or Anthropic-compatible service, only want to add/remove models in UI text, unwilling to manually write settings.yaml; or when you want certain models to support custom reasoning effort, custom headers, or set default context/max output for endpoints without declared capacity.

Prerequisites & Compatibility

Dependency	Min Version	Description
DeepSeek Harness	Not declared	Plugin uses `dsh.bundle.patch`, requires host to have loaded `llm-pi-ai` adapter and support `settings.section` / `settings.mutate` / `credentials.set` etc extension points
Node.js	Not declared	Plugin code only uses `export const` syntax
Platform	macOS / Windows / Linux	Cross-platform, no OS / CPU platform checks or native modules in source
Native modules	None	Only uses browser-side React and host built-in RPC

Installation

dsh plugin --profile web add github:MarvekG/deepseek-harness-model-config

Configuration Options

Config	Type	Description	Default
Endpoint Name	string	Display name for provider route and derived credential reference stem, must start with lowercase	None
Endpoint URL	string	Root address for provider requests and model discovery, requires `http` or `https` protocol, e.g., `https://gateway.example/v1`	None
API Protocol	enum	Options: `openai-completions` / `openai-responses` / `anthropic-messages`	`openai-completions`
API Key	string	Written to Harness credential system, won't enter config file	None
Custom Header	key/value	Attached when calling models, not attached during `GET /models`	None
Default Context Window	integer	Fallback value when model entry and built-in models have no context size	`262144`
Default Max Output	integer	Fallback value when model entry and built-in models have no max output	`32768`
Default Input Modality	enum	Fallback input type when model not declared (text/image)	`[text]`
Default Reasoning Effort	enum	Provider route-level default reasoning effort, from `off` to `max`	None
Reasoning Budget	key/value	Set token budget for each level for reasoning services that support budget	None
Reasoning Compatibility	object	Only valid for `openai-completions` route, controls `thinkingFormat`, `supportsReasoningEffort`	None
Cache Retention	enum	`none` / `short` / `long`	pi-ai default
Transport	enum	`sse` / `websocket` / `websocket-cached` / `auto`	pi-ai default
Request Timeout	integer(ms)	Provider SDK/HTTP request timeout	`0` (determined by underlying)
WebSocket Connection Timeout	integer(ms)	WebSocket handshake timeout	None
Stream Idle Timeout	integer(ms)	Max idle time allowed for single provider read	`300000`
Retry Strategy	object	Includes retry mode, retryable status codes, max attempts, backoff initial/ceiling, jitter ratio	`dsh-llm-retry` default
Per-Model Context	integer	Max context this model can carry	Inherited from route default or built-in model
Per-Model Max Output	integer	Model's max output capability, also used as default `maxTokens` for model requests	Inherited from route default or built-in model
Per-Model Input Modality	enum	Supported input type list (text/image)	Inherited from route default or built-in model
Per-Model Reasoning Effort	map	Select level from `off`/`minimal`/`low`/`medium`/`high`/`xhigh`/`max` and set corresponding "network spelling"	From models.dev match or user manual declaration

See docs/llm-pi-ai-parameters.md for complete field table of endpoint advanced parameters and per-model parameters.

FAQ

Q: Does this plugin conflict with DSH's built-in Models settings page?

A: Under standard Web profile, both will be visible simultaneously. DSH design explicitly makes them mutually exclusive. This plugin can run independently after disabling the ui-settings-models line at profile level, otherwise two UIs will modify the same config (src/client.js:1232, docs/model-config-plugin-design.md:188).

Q: Where is the API Key written, will it leak to settings.yaml?

A: No. The plugin only writes credential references (like MY_ENDPOINT_API_KEY) through Harness's credentials service. The real key is only submitted once in memory and won't appear in plain text in settings.yaml or config preview (package.json:5-7, README.md:18).

Q: Which API protocols are supported?

A: Currently provides three dropdown options: openai-completions, openai-responses, anthropic-messages. When fetching model candidates, anthropic-messages will automatically be switched to OpenAI-compatible GET /models because model lists are generally in OpenAI format (src/client.js:1375-1377, 1181).

Q: Can it work normally when models.dev is unreachable?

A: Yes. Endpoint-returned models can still be manually checked and edited, but context, output, input modality, and reasoning capabilities won't be auto-completed. After submission, it will prompt "Endpoint results still editable" (src/client.js:136, 270, 1184, README.md:37).

Q: When editing a saved endpoint, if the API Key field is left blank, will it overwrite the original Key?

A: No. Leaving blank means "continue using existing credentials". Only when a new value is filled will credentials.set be called to update (docs/model-config-plugin-design.md:140).

Q: How to uninstall and update?

A: Both require first remove then add:

dsh plugin --profile web remove dsh-models-config-plugin
dsh plugin --profile web add github:MarvekG/deepseek-harness-model-config
dsh web

(README.md:62-77)

Q: When re-fetching models, will already checked ones be cleared?

A: No. Already checked models keep their check status, newly appearing models are unchecked by default, models no longer returned by endpoint will also remain displayed to avoid accidental deletion during refresh (src/client.js:1171-1174, README.md:13).

Q: Can I configure global custom Headers for an endpoint? Will they also be included when fetching models?

A: Custom Headers can be added in endpoint advanced parameters, but Headers are only attached when calling models, not in GET /models requests (README.md:30, docs/llm-pi-ai-parameters.md:37).

Difficulty

Beginner — Just fill in the form in Web UI, no manual YAML writing needed; understanding DSH's profile / plugin concepts and having a ready OpenAI/Anthropic-compatible gateway address is enough to get started.

Known Issues & Limitations

Model fetching only covers OpenAI-compatible GET /models; model lists for other protocols (like pure Anthropic protocol endpoints) require manual editing (docs/model-config-plugin-design.md:225)
When models.dev doesn't provide reasoning info for a model, the UI won't guess capabilities for users; users need to add reasoning levels themselves (docs/model-config-plugin-design.md:226)
Independent plugin form is mutually exclusive with DSH's built-in ui-settings-models: when enabling this plugin in profile, should disable the original Models settings line, otherwise both UIs will modify the same config simultaneously (docs/model-config-plugin-design.md:188, 227)
Model discovery has a 4MB receive limit; 401/403 are attributed to credential issues, unreachable/non-JSON are categorized as "fetch failed" (docs/model-config-plugin-design.md:198)