Feature: DeepSeek Harness Provider Integration

Purpose

Defines how the DeepSeek Harness (the dsh web GUI on this desktop) reaches the local routers through the llm-pi-ai adapter, so a dsh session uses the same models as aider and llama.vscode instead of a cloud provider.

Requirements

• DSH-R001 — A model id must match a router model id exactly (GET /v1/models) or an alias configured on the router; an unknown id fails with UNKNOWN_MODEL.
• DSH-R002 — contextWindow must match or undershoot the router preset ctx-size. A bare entry without capacities inherits the 262144/32768 defaults, which exceed every local preset and cause context overflow failures.
• DSH-R003 — maxTokens must sit well under contextWindow, typically 4096–16384.
• DSH-R004 — Prefer a curated list of the models actually used; the router may list many more.
• DSH-R005 — Local models are non-reasoning: do not request a reasoning effort when a beast model is selected; leave the effort off.
• DSH-R006 — The compat block is required for llama.cpp. pi-ai cannot recognise the endpoint URL, so without it the adapter sends OpenAI’s dialect (developer role, max_completion_tokens), which llama.cpp does not reliably accept. supportsDeveloperRole: false selects the system role and maxTokensField: max_tokens selects the max_tokens field.
• DSH-R007 — Store any non-empty credential under the BEAST_API_KEY reference through the GUI Models settings, which writes ~/.dsh/.credentials.yaml, or export it in the dsh launch environment. If --api-key is later added to the llama.service on beast, the same value must be stored.
• DSH-R008 — Any task that writes or corrects the DSH provider section must read the latest dataflow.in/endpoints/ capture for the router model ids and contexts.

Behavior

• A route declared in ~/.dsh/settings.yaml serves its configured models on the next request; no restart is needed.
• Credentials resolve from apiKeyEnv through the process environment, then ~/.dsh/.credentials.yaml, then .env files.

Dependencies

• None.

Mechanism

DeepSeek Harness ships a generic multi-provider adapter (@deepseek-ai/dsh-llm-pi-ai) keyed by provider route in the llm-pi-ai settings namespace (~/.dsh/settings.yaml). A route that pi-ai does not ship is declared outright with api, baseURL, and a models list.

Canonical Route

The beast route is the canonical local provider. The corrected example lives in sources/config/dsh-settings-beast.example.yaml.

llm-pi-ai:
      providers:
        beast:
          displayName: Beast (llama.cpp)
          apiKeyEnv: BEAST_API_KEY
          api: openai-completions
          baseURL: http://beast.lan:2001/v1
          compat:
            supportsDeveloperRole: false
            maxTokensField: max_tokens
          models:
            - id: <router model id>
              contextWindow: <preset ctx-size>
              maxTokens: <safe output cap, well under ctx-size>

Credential

llama.cpp ignores Authorization unless the router runs with --api-key.

GUI Surface

The Models settings section in the web GUI offers a custom-provider creation card (route, display name, base URL, protocol, API key, models) and per-model advanced fields (context window, max tokens). The card does not expose compat; add the compat block by editing ~/.dsh/settings.yaml directly.