Pick the provider shape before the model name

Use the official provider when you want the documented DeepSeek path. Use a catalog provider when DSH can fetch known models for that service. Use a custom provider when the endpoint is OpenAI-compatible but not represented by the catalog. Record which layer owns the model name; a model visible at the upstream service may still be unknown to the local provider configuration.

  • Official DeepSeek provider: simplest DeepSeek-specific route.
  • Catalog provider: credential plus provider-managed model discovery.
  • Custom provider: base URL, credential, and explicit model identity.
  • Image-capable tasks: verify the model input modality before diagnosing the prompt.

Keep credentials out of examples and analytics

Enter credentials only in the DSH configuration surface intended for them. HarnessBase examples use environment-variable names or placeholders and never ask you to paste a live value into this site. Copy events record only that an example was copied—not the key, endpoint, model, or generated configuration.

Diagnose by the error boundary

A missing-credential error means configuration cannot resolve the required secret. UNKNOWN_MODEL means the selected provider cannot resolve that model identity. A 401 comes from the upstream authentication boundary and can indicate a missing, invalid, or wrong-provider key. A successful model fetch does not prove that a later multimodal request is supported.

  • MISSING_CREDENTIAL → confirm the expected credential field or environment reference.
  • UNKNOWN_MODEL → fetch models again or use the provider’s exact model identifier.
  • 401 → confirm the credential belongs to the endpoint and has not been copied with whitespace.
  • Image/input error → confirm the selected model advertises the required modality.

Custom endpoint checklist

Before adding a custom provider, confirm that its request and response behavior is compatible with the interface DSH expects. Record the base URL without accidentally appending a duplicate path segment, select an exact model ID, and test a minimal text task before tools or images.

Safe configuration notes · text
provider: custom OpenAI-compatible
base_url: https://provider.example/v1
credential: reference an environment variable
model: use the upstream model identifier
first_test: minimal text-only prompt

HarnessBase test matrix still required

Before indexing, HarnessBase will retain successful and failed runs for the official provider, a catalog provider, a custom OpenAI-compatible endpoint, and OpenRouter; reproduce missing credential, unknown model, and 401 states; and document where DSH displays each failure.

Sources and evidence boundary

Primary sources establish product facts. HarnessBase labels direct testing separately and does not convert community observations into verified results.

Target query cluster: DeepSeek Harness models · DeepSeek Harness API key · DeepSeek Harness OpenRouter · DeepSeek Harness custom provider