Configure DSH with Local LiteLLM Proxy Provider

Set up DeepSeek Harness to use a local LiteLLM proxy as a custom OpenAI-compatible provider. This writes the provider configuration to DSH’s settings.yaml so the Claude models served by LiteLLM (via Vertex AI) are available as model routes in the DSH WebUI.

For DSH installation, see the dsh-setup skill first.

Prerequisites

• DSH installed and launchable (see skill dsh-setup)
• LiteLLM proxy running locally at http://127.0.0.1:4000 (see skill litellm-vertex-ai-proxy to set one up)
• Use skill litellm-proxy-status to verify the proxy is healthy before proceeding

Architecture

DSH stores settings in ~/.dsh/settings.yaml (or $DSH_HOME/settings.yaml). The llm-pi-ai plugin manages model providers. A custom provider needs:

Field	Purpose
baseURL	LiteLLM endpoint: http://127.0.0.1:4000/v1
api	Protocol: openai-completions
apiKeyEnv	Env var holding a dummy API key (DSH requires non-empty)
models	List of model IDs matching LiteLLM’s model_name entries
compat	Protocol compatibility flags for LiteLLM

Reference Configuration

File: ~/.dsh/settings.yaml (relevant section)

⚠️ Use this exact schema. Do NOT write from memory or improvise field names. Copy this template and substitute only the marked placeholders.

llm-pi-ai:
  providers:
    litellm-vertex-ai:
      apiKeyEnv: LITELLM_VERTEX_AI_API_KEY
      api: openai-completions
      baseURL: http://127.0.0.1:4000/v1
      compat:
        supportsDeveloperRole: false
        maxTokensField: max_tokens
      models:
        - id: claude-opus-4-6
          input: [text, image]
        - id: claude-sonnet-4-6
          input: [text, image]
        - id: claude-haiku-4-5
          input: [text, image]

Compatibility Flags

Flag	Value	Why
supportsDeveloperRole	false	LiteLLM proxies role: "developer" as-is but Vertex AI Claude doesn’t accept it. Forcing false makes DSH send role: "system" instead.
maxTokensField	max_tokens	LiteLLM uses the OpenAI max_tokens field, not max_completion_tokens.

Image Input

All three Claude models support vision. The input: [text, image] declaration on each model tells DSH to allow image attachments in the WebUI for those models.

Steps

Setup

1. Run the setup script

   bash ~/.agents/skills/dsh-litellm-provider/scripts/setup.sh
   

   This script:

   • Verifies LiteLLM proxy is reachable at http://127.0.0.1:4000
   • Queries LiteLLM’s /v1/models endpoint to discover available models
   • Writes (or merges) the provider block into ~/.dsh/settings.yaml
   • Backs up the existing settings file before modifying
2. Verify the configuration

   bash ~/.agents/skills/dsh-litellm-provider/scripts/verify.sh
   

3. Test in DSH — launch DSH (via dsh-launcher or the desktop entry), open Settings → Models, and confirm the litellm-vertex-ai provider appears with the expected models.

Restore (after DSH update)

DSH updates may reset settings.yaml. Re-run setup:

bash ~/.agents/skills/dsh-litellm-provider/scripts/setup.sh

The script is idempotent — it detects an existing provider block and updates it in place.

Gotchas

• ~/.dsh/ is created on first DSH launch. If you run this skill before ever launching DSH, the directory won’t exist. The setup script creates it if missing.
• Model IDs must match exactly between LiteLLM’s config.yaml (model_name) and this provider’s models[].id. The setup script auto-discovers them from the running proxy.
• Dummy API key required. The local LiteLLM proxy has no authentication, but DSH rejects providers with an empty API key. The setup script exports LITELLM_VERTEX_AI_API_KEY=sk-litellm-local-no-auth as a dummy value and sets apiKeyEnv: LITELLM_VERTEX_AI_API_KEY. The env var must be set in the shell environment where DSH launches (the dsh-launcher script handles this automatically).
• supportsDeveloperRole: false is critical. Without this flag, DSH sends role: "developer" for system prompts, which Vertex AI Claude rejects through LiteLLM.
• Settings format may change. DSH is in developer preview. The llm-pi-ai plugin’s settings schema could change across versions. The setup script includes a version check and warns if the expected structure is missing.

Specification

ID	Capability	Verifiable By
S1	Write LiteLLM provider to ~/.dsh/settings.yaml	verify.sh checks provider block exists
S2	Auto-discover models from running LiteLLM proxy	setup.sh queries /v1/models and populates model list
S3	Idempotent — re-running setup updates in place	Run setup.sh twice, settings.yaml has no duplicates
S4	Backup existing settings before modification	.dsh/settings.yaml.bak created

Tests

Test	Spec	Command	Expected Result
T1	S1	bash scripts/verify.sh	Provider block found with correct baseURL
T2	S2	bash scripts/setup.sh && grep 'claude-' ~/.dsh/settings.yaml	Model IDs match LiteLLM output
T3	S3	bash scripts/setup.sh && bash scripts/setup.sh && grep -c 'litellm-vertex-ai' ~/.dsh/settings.yaml	Count is 1 (no duplicates)
T4	S4	bash scripts/setup.sh && ls ~/.dsh/settings.yaml.bak	Backup exists

Changelog

See CHANGELOG.md for version history.