🤖 dsh-local-ai
- 1024 store channel:
npm i -g dsh1024 once, then dsh1024 plugin --profile web add dsh-local-ai (counts toward the deepseek1024.com install ranking).
Local-model (Ollama) integration for DeepSeek Harness.
Discover, pull, remove, and inspect local models, route requests to them by task type or keyword with automatic fallback to the cloud, and get a one-shot status overview via /ollama.
Official repository. This is the only official repository of dsh-local-ai, maintained by PerryLink. Same-name repositories under other accounts are not affiliated.
English · 简体中文 · Español · Português · हिन्दी
Compatibility
| Surface | Status |
|---|
| Harness | DeepSeek Harness dsh-v0.1.6-alpha.2 (verified 2026-09-18: dual typecheck rulers + 133 tests + self-contained/artifacts gates). The peer range admits every supported line: >=0.1.2-rc.1 <0.2.0 || >=0.1.5-alpha.1 <0.2.0 || >=0.1.6-0 <0.2.0; dev/test pins are 0.1.6-alpha.2. |
| Node | ^22.19.0 || >=24.0.0 |
| Backend | Ollama (local HTTP API + CLI probe) |
| Model | Text-only route (inputModalities: ['text']); tool calls and tool results are supported |
What you get
dsh-local-ai makes Ollama a first-class local provider in DeepSeek Harness:
- Discovery & management —
ollama_list (installed models, running models, disk usage), ollama_show (parameter size, quantization, context length), ollama_pull, and ollama_remove.
- Health check — process liveness (via the
ollama CLI) and API responsiveness (via /api/version), reported as two independent signals.
- Official adapter — the
ollama provider route is registered through ctx.llm.registerAdapter (LlmAdapter), with configurable model mapping and temperature / max-tokens / stop translation.
- OpenAI-compatible backends — LM Studio, vLLM, and llama.cpp
--server each register as their own openai:<name> provider through the same LlmAdapter seam, reusing one OpenAI /v1/chat/completions adapter (text-only route).
- Local routing —
model_route rules route requests to a local model by task type (purpose), case-insensitive keyword, or always, with automatic fallback to the cloud when the local route fails before producing content.
/ollama command — a one-shot status overview: models, disk usage, health, and suggestions.- Zero dependencies, HTTP first — everything talks to Ollama's HTTP API (the CLI is used only for the process probe); no model files are bundled.
request (loop)
│ llm/stream waterfall
├─ rule matches? ──▶ route to ollama ──▶ Ollama /api/chat (NDJSON stream)
│ └─▶ route to openai:<name> ─▶ /v1/chat/completions (SSE)
│ └─ fails first ─▶ fall back to cloud (next())
└─ no match ──▶ cloud provider
tools ──▶ /api/tags · /api/ps · /api/show · /api/pull · /api/delete
health ──▶ /api/version (API) + ollama list (process)
Quick start
# 1. install the bundle into your profile
dsh plugin --profile web add "github:PerryLink/dsh-local-ai#main"
# or from npm (published releases)
dsh plugin --profile web add dsh-local-ai
# 2. configure routing in your profile patch (cordis.yml) and restart
dsh --profile web
Minimal routing configuration (the rule ships commented out in cordis.patch.yml):
- insert:
- id: dsh-local-ai
name: dsh-local-ai
config:
route:
- model: llama3.2
keywords: ["confidential", "offline"]
Then verify the row mounts:
dsh --profile web --dump-config | grep -A2 'id: dsh-local-ai'
Install & uninstall
- git channel (latest
main): dsh plugin --profile web add "github:PerryLink/dsh-local-ai#main" — the prepare script builds with production dependencies only.
- npm channel (published releases):
dsh plugin --profile web add dsh-local-ai.
- tarball channel:
pnpm pack in this repo, then dsh plugin --profile web add ./dsh-local-ai-<version>.tgz.
- uninstall:
dsh plugin --profile web remove dsh-local-ai (or remove the row from the profile patch).
If pnpm reports ERR_PNPM_IGNORED_BUILDS for this package, add allowBuilds: { esbuild: true } to your pnpm-workspace.yaml — the dsh CLI prints the exact snippet.
Configuration
All tunables are Schemastery Config fields (changeable from cordis.yml). An id-targeted override replaces the whole row — restate every key you need. cordis.patch.yml documents each key inline.
| Key | Default | Meaning |
|---|
baseURL | http://127.0.0.1:11434 | Ollama HTTP API base URL; /api/* paths are appended |
requestTimeoutMs | 30000 | Per-request HTTP timeout (milliseconds) |
graceMs | 15000 | Subprocess terminate grace for the health-check CLI probe |
defaultContextWindow | 8192 | Context capacity used when a model has no exact value |
maxTokens | 4096 | Per-request output cap used when a model has no exact value |
temperature | (none) | Default sampling temperature (0..2); omitted leaves the provider default |
vision | true | Declare and serialize image support when the model reports vision; false keeps the route text-only |
visionCacheTtlMs | 30000 | Milliseconds a /api/show capability probe stays cached (0 disables caching; a pull or remove invalidates that model) |
models | [] | Harness-visible → Ollama model mappings |
models[].name | (required) | Harness-visible model name (GenerateOptions.model) |
models[].model | = name | Ollama model id |
models[].contextWindow | (none) | Per-model context capacity |
models[].maxTokens | (none) | Per-model output cap |
models[].temperature | (none) | Per-model sampling temperature |
backends | [] | OpenAI-compatible local backends (LM Studio / vLLM / llama.cpp) |
backends[].name |
Tools & surfaces
| Surface | Kind | What it does |
|---|
ollama_list | tool | List installed models, running models, and disk usage |
ollama_show | tool | Show parameter size, quantization, context length, family, format |
ollama_pull | tool | Pull (download) a model |
ollama_remove | tool | Remove a model |
ollama_health | tool | Process liveness + API responsiveness |
/ollama | command | One-shot status overview (models + health + suggestions) |
Consumes the public host services ctx.llm (registerAdapter), ctx.tools, ctx.subprocess (CLI probe), and ctx.commands. It registers no llm/stream short-circuit by default — the routing listener passes through (next()) unless a rule matches.
Permissions & data
- Permissions:
network:outbound to the Ollama endpoint you configure; no native code, no filesystem access, no storage.
- Data: every model list/detail, health fact, and error message shown to the model or the user is sanitized (endpoint userinfo and secret query params dropped, control characters stripped, lengths bounded) before display. Tool and command results are logged by the harness's own tool/command seams.
- Credentials: the plugin stores and reads no credentials. It only issues HTTP requests to the endpoint you configure, plus the local
ollama list process probe.
Security boundaries
- No re-routing by default — the
route list is empty unless you opt in; a request reaches a local model only through an explicit rule or an explicit ollama provider selection.
- Sanitize before display — endpoint addresses and local paths are sanitized before they reach tool output, the
/ollama command, or error messages.
- Zero bundled models — downloads and storage are Ollama's own responsibility; nothing is shipped in the package.
- Failure loud, failure contained — invalid config fails the mount; a local route that fails before producing content falls back to the cloud (
next()), so a down Ollama never bricks a conversation. One carve-out: a failure carrying IMAGE_OFFLOAD_REQUIRED is rethrown instead of retried on the cloud — that code is the official offload circuit asking this same local route to drop retained images, and falling back would skip the circuit while silently sending a local-only request to a remote provider.
- Model-visible ⟺ logged — routing only changes which provider serves a request (the assistant message is logged with its
ollama provenance); no new model-visible input is invented.
Known limitations
- npm 0.1.5-rc.2 — developed and tested against
@deepseek-ai/dsh@0.1.5-rc.2; newer harness baselines are expected to work but are verified by the monthly compat workflow.
- Vision when the model reports it — models whose
/api/show capabilities include vision declare inputModalities: ["text","image"] and carry base64 image payloads on user messages (opt out with vision: false); text-only models still reject image content (UNSUPPORTED_CONTENT).
- Mid-stream fallback — once a local route has started producing content, a later failure is forwarded (not retracted); only a failure before the first token falls back to the cloud.
Development
pnpm install # node ^22.19 || >=24
pnpm run typecheck # tsc: src + tests against the published 0.1.5-rc.2 types
pnpm run typecheck:ci # strict tsc against published rc.2 types (skipLibCheck off)
pnpm test # vitest: real Context/LlmRuntime/ToolRuntime/CommandRuntime/subprocess seams
pnpm run test:coverage # coverage gate (90/80/90/90)
pnpm run build # tsdown bundle + tsc declarations (lib/)
pnpm run verify:self-contained # dependency specs resolve from the registry
pnpm run verify:artifacts # built ESM face + bundle patch present
node scripts/check-readme-sync.mjs # five-language README sync gate
node scripts/check-endpoints.mjs # M3 endpoint-liveness probe (Ollama /api/version)
pnpm pack # the published tarball
Topics
dsh, dsh-plugin, deepseek-harness, deepseek, cordis, ollama, local-llm, local-models, offline, privacy, model-routing
Contributors
- @PerryLink — creator and maintainer: adapter, routing, tools, health check, sanitization, and the five-language docs.
- @LABEST-IA — tool-call CallId fix (PR #2), and the tool-call slot and vision-support reports (issues #1, #3, #5).
PerryLink DSH Plugin Family
This project is one of the 40 DeepSeek Harness plugins maintained by PerryLink. If this one helps you, the others likely will too:
| Plugin | One-liner |
|---|
| dsh-auto-review | Second-model auto-review on the approval chain, fail-closed by default |
| dsh-background-agents | Durable background child agents with a Web UI sidebar, messaging and interrupt |
| dsh-budget | Cost governance for DeepSeek Harness: budgets, carbon, and latency in one panel. |
| dsh-checkpoint-rewind | Claude Code /rewind-equivalent: snapshots, session forks, one-shot restore |
| dsh-claude-move | Migrate Claude Code sessions, memory, skills and CLAUDE.md into DSH |
| dsh-click | Cross-platform native desktop control for DeepSeek Harness — Windows first. |
| dsh-composer-history | Terminal-style input history for the web composer: arrows, Ctrl+R search |
| dsh-data-quality | Dataset quality checks and citation cross-checks (the optional numeric bridge consumed here) |
| dsh-defend | Prompt-injection, jailbreak, and secret-leak defense for DeepSeek Harness. |
| dsh-doublecheck | Engineering-discipline guard: requirements grill, test gates, adversary review |
| dsh-draw | Unified static-image generation routing for DeepSeek Harness. |
Install from the DSH Desktop Market
All PerryLink plugins are browsable in the built-in DSH Desktop Market: Market → Sources → add source → paste https://perrylink-dsh-catalog.perrylink.workers.dev/catalog-source.json → select it. Installation still goes through the Market's npm-identity verification and your confirmation.
License
Apache License 2.0 © 2026 dsh-local-ai contributors
