dsh-plugin-ollama-bridge

English | 简体中文

Hand the bulk low-difficulty long-text work to the Ollama models on this machine, and keep the leading model on scheduling and review.

A DeepSeek Harness plugin that registers four tools covering the whole loop: see what is installed → try one call → run a batch over long text → fill metadata gaps over the web when needed. Long source text never enters the conversation: only counts, failure distribution and a review sample come back, so the saving is on main-model tokens.

Tools

Tool	What it does
ollama_models	Lists installed models with parameter count, quantization, context length, capabilities, and roles hints (code / reasoning / vision / embedding / long-context …)
ollama_ask	One question, unverified answer — for probes or a few high-value judgements
ollama_batch	Chunking + concurrency + retries + a deterministic check on every item + review sample + optional result file
ollama_lookup	Web lookup of model capability, after the user approves the exact queries

Recommended flow: look before you choose

1. Call ollama_models for the inventory and metadata hints.
2. Summarise what each model is good for. roles are metadata hints, not measured quality.
3. When metadata is not enough (benchmarks, known weaknesses), call ollama_lookup — it hands the exact queries to the user for approval, and returns approved: false without searching when they decline.
4. Run ollama_batch with the chosen tag. An unknown tag fails immediately and names the installed models.

ollama_batch returns only: item count, passed/failed, failure-reason counts, at most 20 failure snippets, and a verbatim review sample (2% of the items by default, at least one and at most eight). The source text stays inside the plugin process.

Install

From GitHub (recommended, updatable)

In DSH, open Plugins → + Add plugin and enter:

github:wth373737/dsh-plugin-ollama-bridge

From a local directory

pnpm add file:/path/to/dsh-plugin-ollama-bridge

Restart DSH after installing. The plugin appears under Plugins → Installed plugins as “本地 Ollama 桥”.

Configuration

The config block of cordis.patch.yml:

baseURL: http://127.0.0.1:11434
model: qwen2.5:7b
numCtx: 8192
concurrency: 2
timeoutMs: 180000

• numCtx: context available per chunk, and the main driver of memory use.
• concurrency: parallel requests. 1–2 is sensible on one machine; lower it for larger models.
• If Ollama keeps its models outside the default directory, set OLLAMA_MODELS to the model root.

How far to trust the output

Every item carries one of three states, in the written file and in the returned sample:

Label	Meaning
ok	Passed every deterministic check, but not read one by one
*抽样待复核*	Sampled and returned; the leading model read it
!!校验失败(原因)	Malformed: empty, too long, not JSON, outside the allowed choices, missing a required substring, regex miss, or refusing

Deterministic checks verify form, never meaning. Report “all items are well-formed” separately from “the content is correct”, and raise sample_rate for small batches.

Boundaries

• Reads: any absolute path the host process can read.
• Writes: only inside the session workspace. A path outside it is refused with the workspace root in the message; an unresolvable workspace refuses the write rather than guessing.
• Network: only the configured baseURL; ollama_lookup reaches the host's configured search provider only after the user approves.
• Approval: with no answerer available, ollama_lookup refuses to look up instead of silently skipping approval.

Development

npm test     # offline smoke test: stubs fetch/web/userQuestions and checks
             # verification, sampling, labels, write confinement, approval gate

No Ollama needed, and nothing is written outside the workspace.

License

MIT © 2026 wth373737