dsh-custom-provider

YAML configuration

The same provider and model catalog can be managed directly in settings.yaml.

Configuration

Web settings

Open Settings → Custom Providers in DSH Web, then:

1. Add a provider ID, display name, API base URL, and credential reference.
2. Enter the API key. The field is write-only and stores the value through the DSH credentials service.
3. Add one or more models with their model IDs, display names, context windows, and maximum outputs. Capacity inputs accept integers and K/M suffixes such as 200K or 1m.
4. Expand Advanced only when the endpoint requires compatibility overrides or custom reasoning-effort mappings.
5. Save the provider. Its models become available to the model picker immediately after the settings change is applied.

settings.yaml

llm-custom:
  providers:
    example:
      displayName: Example Provider
      apiKeyEnv: EXAMPLE_API_KEY
      api: openai-completions
      baseURL: https://api.example.com/v1
      compat:
        supportsStore: false
        supportsDeveloperRole: false
        thinkingFormat: deepseek
        supportsReasoningEffort: true
        maxTokensField: max_tokens
        requiresReasoningContentOnAssistantMessages: true
      models:
        - id: example-model
          name: Example Model
          contextWindow: 262144
          maxTokens: 32768
          reasoningEfforts:
            off:
            high: high
            max: max

apiKeyEnv is a credential reference, not the secret itself. Configure its value through DSH's credentials service or the Web settings page.

Configuration reference

Each key below is relative to llm-custom.providers. .

In the Web UI, a provider ID must start with a lowercase letter and contain only lowercase letters, digits, and dashes.

Provider fields

Field	Required	Description
displayName	No	Name shown in the model picker. Defaults to the provider ID.
apiKeyEnv	Yes	DSH credential reference resolved before every request.
api	Yes	Wire protocol. The supported value is openai-completions.
baseURL	Yes	Absolute HTTP(S) base URL of the OpenAI-compatible endpoint. Trailing slashes are normalized.
compat	No	Compatibility defaults inherited by every model on the route.
models	Yes	Static model catalog containing at least one model. Model IDs must be unique within the provider.

Model fields

Field	Required	Description
id	Yes	Model ID sent to the provider.
name	No	Name shown in the model picker. Defaults to id.
contextWindow	Yes	Positive integer context-window size in tokens.
maxTokens	Yes	Positive integer maximum output size in tokens.
reasoningEfforts	No	Selectable DSH reasoning levels mapped to provider wire values. Set to false to disable the reasoning control.
compat	No	Per-model compatibility values. Each declared field overrides the provider-level value.

Compatibility fields

compat accepts the same fields at provider and model level:

Field	Values	Effect
supportsStore	true / false	Whether the request may send the OpenAI store parameter.
supportsDeveloperRole	true / false	Whether system prompts may use the developer role.
thinkingFormat	openai, deepseek, openrouter, together, zai, qwen, string-thinking, ant-ling	Reasoning-content format understood by the endpoint.
supportsReasoningEffort	true / false	Whether the request may send a reasoning-effort parameter.
maxTokensField	max_completion_tokens / max_tokens	Request field used for the maximum output limit.
requiresReasoningContentOnAssistantMessages	true / false	Whether replayed assistant messages retain reasoning_content.

Only declare compatibility fields required by the endpoint; omitted model fields inherit the provider-level values.

Reasoning-effort mappings

The supported DSH levels are off, minimal, low, medium, high, xhigh, and max. A mapping value is the exact string expected by the provider. Only off may be empty (null in YAML), and a mapping must expose at least one non-off level. Omitted levels are unavailable in the picker.

reasoningEfforts:
  off:
  medium: medium
  high: high

Use reasoningEfforts: false when the model does not provide a reasoning control.