dsh-llm-ollama

English | 中文

Ollama Cloud integration for DeepSeek Harness. Chat uses Ollama's OpenAI-compatible Chat Completions endpoint through the shared pi-ai-backed adapter. Model discovery and the Web Search/Fetch providers remain on Ollama-native APIs because those independent capabilities are not part of the chat protocol.

The package root exposes the Cordis plugin contract and OllamaAdapter. The same artifact exports ./client, which contributes the Ollama Cloud card under Settings → Plugins → Plugin configuration. The protocol and capability split is recorded in ADR 0001.

Installation

DeepSeek Harness 0.1.0-rc.6 or later is required. Install directly from GitHub:

dsh plugin --profile web add github:NOirBRight/dsh-llm-ollama
dsh web

The repository tracks release-ready lib artifacts, so GitHub installation needs no build-script allowlist. A source checkout can use a link installation after running pnpm run build.

Web configuration

Open Settings → Plugins → Plugin configuration → Ollama Cloud. The card stores the API key through the Harness credentials API under OLLAMA_API_KEY; the Host never returns the stored literal. It saves the native base URL and model catalog together as one revision-fenced llm-ollama settings mutation.

Fetch available models opens the picker immediately and calls the package's loopback-only RPC with the unsaved endpoint and one-shot key. The Host reads /api/tags, deduplicates native ids, and enriches up to six models concurrently through /api/show. The native metadata supplies context windows plus vision, thinking, and tools flags that /v1/models does not expose. The picker starts from the current draft selection, preserves current-only models, and replaces the draft catalog when applied.

The card's Cloud usage section mirrors ollama.com/settings: the Host reads GET /usage with the stored (or one-shot) key and renders the session and weekly windows as consumed-percentage meters plus the week's per-model request counts. The credential never crosses to the browser. A self-hosted endpoint without the usage surface shows an unsupported note instead of an error.

The model catalog starts collapsed and lists one row per model: a drag handle reorders rows (the order persists with the catalog), the chevron opens that row's context window, Default thinking, and capability flags, and the trash button removes it.

Plugin configuration screenshots

Cloud usage and the complete weekly model activity list:

Sortable model catalog:

The Models page lists saved ollama-cloud models and can select them. Current Harness releases do not expose a third-party editor slot inside that page, so this package owns its editor under Plugin configuration.

Capability and protocol split

Chat uses:

POST <openai-base>/chat/completions

The configured baseURL remains the native Ollama API address. The plugin maps chat to its /v1 sibling:

https://ollama.com/api  ->  https://ollama.com/v1
http://localhost:11434/api  ->  http://localhost:11434/v1

The Ollama-native independent capabilities remain:

model discovery  ->  GET /api/tags + POST /api/show
web search       ->  POST /api/web_search
web fetch        ->  POST /api/web_fetch

Search and Fetch are ctx.web providers, so they work with any selected chat model. A DeepSeek, Codex, Kimi, or OpenAI-compatible chat model can still call the Ollama-backed web_search tool when the profile selects ollama-cloud.

OpenAI Responses is not the default because Ollama supports only the non-stateful flavor. Anthropic Messages is not the default because Ollama Cloud needs an extra Bearer header and that compatibility surface has no model listing or prompt caching.

Web search and fetch

The Host plugin registers both Web providers under ollama-cloud. Registration alone does not change deployment policy; pin the desired providers in the profile patch:

- id: web
  config:
    searchProvider: ollama-cloud
    fetchProvider: ollama-cloud

Omit fetchProvider to keep the built-in HTTP fetcher while moving only search. Both providers reject redirects before following them. Each attempt has a 15-second default budget and one transient timeout or pre-response transport failure is retried. HTTP errors, malformed replies, missing credentials, redirects, and caller cancellation are not retried.

Config

- id: llm-ollama
  name: 'dsh-llm-ollama'
  config:
    apiKeyEnv: OLLAMA_API_KEY
    baseURL: https://ollama.com/api
    defaultContextWindow: 262144
    streamIdleTimeoutMs: 300000
    webRequestTimeoutMs: 15000
    retryPolicy:
      mode: normal
      backoff:
        initialDelayMs: 500
        maxDelayMs: 10000
        jitterRatio: 0.1
    models:
      - id: gpt-oss:20b
        name: GPT-OSS 20B
        contextWindow: 131072
        thinking: true
      - id: llava
        name: LLaVA
        contextWindow: 4096
        vision: true

The provider route remains ollama-cloud and the settings namespace remains llm-ollama. Only configured catalog models are accepted for chat. The adapter does not install a request-level maxTokens default; output is not capped from the catalog. Per-row contextWindow is the DSH compaction budget.

The fallback context window is 262,144 tokens. Discovery should normally provide an exact model value; the fallback also leaves room for pi-ai's context-safety reserve when metadata is unavailable.

Model capabilities

vision controls text/image input modalities. thinking enables selectable reasoning efforts. Known Ollama Cloud families expose only vendor-real levels and pin a plugin defaultEffort used when the session has not picked one (GLM-5.2 and Kimi K3 default to max; DeepSeek V4 and MiniMax M3 to high; GPT-OSS to medium; Nemotron Super/Nano to low). Unknown thinking models keep off, low, medium, high, and max with no plugin default. tools records discovery metadata; the actual request carries the current DSH tool definitions.

The OpenAI Chat Completions profile is pinned for Ollama: it sends max_tokens, reasoning_effort, and streaming usage, preserves system-role messages, and does not send store, max_completion_tokens, or prompt_cache fields.

Model Experience

Prompt effects

The system prompt and all provider-neutral messages are translated by PiAiAdapter into OpenAI Chat Completions messages. Tool calls retain provider-issued ids and tool results return with the matching tool_call_id. Images are encoded as base64 data URLs only for models marked vision-capable.

Token effects

Usage maps to Harness input/output counts. maxTokens is clamped against the configured context capacity by pi-ai, leaving a safety reserve. Ollama does not currently expose cache-read/cache-write accounting through this endpoint.

KV-cache effects

Stable model, system prompt, history, tool definitions, and request options preserve a stable serialized prefix. Tool-call ids are provider-issued protocol fields and are replayed unchanged. Changing earlier messages, tools, images, model id, or reasoning/output options can invalidate provider-side reuse.

Known limitations and deferred work

• GenerateOptions.stop is not supported by the shared PiAiAdapter.
• Models absent from the saved catalog are rejected; the old native adapter's pass-through behavior is removed.
• /api/show reports thinking capability but not the exact accepted effort set, so the plugin applies Ollama's general rule and the GPT-OSS exception.
• Ollama does not publish per-model output limits.
• Logs written by v0.2.2 and earlier can contain duplicate ollama-call-0 values; existing logs are not migrated.
• Structured-output format configuration is not exposed by this package.