dsh-ai-model-hub
用于本地 AI 模型的 DeepSeek Harness 插件和库：提供能力目录、路由器、运行时管理器和真实引擎适配器。代理指定一项能力；中心选择、启动并调用合适的模型。
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add github:Mhmd7-7/dsh-ai-model-hub#77f9bf21792e0b50f01a3808d1531b3c5e92ab41Ai Model Hub 以 dsh-ai-model-hub 发布，当前版本为 0.7.0。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
正在加载相关插件…
A modular local AI model ecosystem for DeepSeek Harness (DSH).
DSH is the agent — it reasons, plans, calls tools, and orchestrates workflows. It is not the model runtime. The specialised models (text, image, 3D, audio, video) are external and local, reached through a capability interface. DSH never learns how to launch Stable Diffusion, which Python file starts a mesh generator, or how a given engine manages its process. It only asks:
What capabilities are available?
Everything below that question is this project's job.
┌──────────────────────────────────────────────────────────────────────┐
│ DeepSeek Harness agent reasoning, planning, │
│ ───────────────── tool invocation, workflows │
│ │
│ tools: invoke_model · list_models · list_capabilities · │
│ get_model_status · start_model · stop_model · … │
└───────────────────────────────┬──────────────────────────────────────┘
│ dsh-ai-model-hub DSH plugin
│ (the ONLY DSH-aware code)
┌───────────────────────────────▼──────────────────────────────────────┐
│ ModelHub facade invokeModel({ capability }) │
│ ┌────────────────┬──────────────────┬───────────────────────────┐ │
│ │ Capability │ Model Router │ Runtime Manager │ │
│ │ Catalog │ deterministic │ lifecycle · health · │ │
│ │ registry + │ capability-first │ idle timeout · resources │ │
│ │ validation │ selection │ │ │
│ └────────────────┴──────────────────┴───────────────────────────┘ │
│ ┌──────────────────────────────┬────────────────────────────────┐ │
│ │ Model Adapters │ Artifact Store │ │
│ │ openai_compatible · http_json│ durable, typed, addressable │ │
│ │ comfyui · three_d · │ text/image/audio/video/3D/… │ │
│ │ mock (tests only) │ │ │
│ └──────────────────────────────┴────────────────────────────────┘ │
└───────────────────────────────┬──────────────────────────────────────┘
│ spawn (argv, never a shell string)
┌───────────────────────────────▼──────────────────────────────────────┐
│ Local engines: Ollama · llama.cpp · A1111/Forge · ComfyUI · │
│ TRELLIS · Hunyuan3D · Stable Fast 3D · TripoSR · … │
└──────────────────────────────────────────────────────────────────────┘
Live as a DSH profile plugin. The package installs as one unit — dsh plugin add mounts it, so the hub is composed on every boot — and it serves real local
engines only: Ollama for text, ComfyUI / A1111 / Forge for images, llama.cpp for
a second text engine, and any local image-to-3D Gradio app for meshes. There are no
mock models anywhere in the shipped catalogs and no mock fallback at runtime: a
wrong model name or a stopped engine is a loud error (or a cold start), never
fixture output.
The whole architecture — catalog → routing → runtime → adapter → artifact → chained workflow — is covered by the test suite, which uses an in-process mock adapter as a test double. That adapter is not a model: nothing ships it in a catalog, and no deployment can reach it by accident.
| Phase | Scope | Status |
|---|---|---|
| 1 | Full architecture, DSH plugin, tests | ✅ done |
| 2 | Real local text model (Ollama / llama.cpp / vLLM / LM Studio) | ✅ done — openai_compatible adapter ships |
| 3 | Real local image model (A1111 / ComfyUI) | ✅ done — http_json and comfyui adapters ship |
| 4 | Real local 3D model | ✅ done — three_d adapter ships; see docs/three-d.md |
| 5 | Runtime lifecycle + resource-aware routing | ✅ done — the machine is probed and routed against |
| 6 | Multi-model workflows | ✅ done |
Phase 2 needs no new code: the openai_compatible adapter is implemented, and
config/examples/real-models.example.json contains ready-to-copy Ollama and
llama.cpp entries. Phase 3 is implemented by two adapters, because the two
engine families are genuinely different:
http_json — engines that answer one request with a finished image:
AUTOMATIC1111, Forge, and stable-diffusion.cpp's sd-server. It POSTs the
familiar /sdapi/v1/txt2img body and decodes the base64 images that come back.comfyui — ComfyUI takes an API-format node graph, not a prompt endpoint.
Each explicitly configured workflow declares capabilities, exact input bindings
and a typed output selector. The adapter treats graph internals as opaque, edits
only declared inputs, queues on /prompt, polls /history, and retrieves the
declared output. No checkpoint or saved-workflow enumeration creates providers.Phase 4 is three_d: one adapter that speaks the shape the local 3D ecosystem
actually has — a Gradio queue API or a JSON HTTP route, one or two calls, a mesh
file at the end — so TRELLIS, Hunyuan3D, Stable Fast 3D, and TripoSR are catalog
entries rather than code. It writes a typed model_3d artifact (GLB, GLTF, OBJ,
STL, or PLY, sniffed from the bytes), and a discoverer verifies a declared engine
against the API surface it actually exposes before publishing anything.
Phase 5's remaining gap is closed: ModelHub now probes RAM, VRAM, free VRAM, and
disk, re-measures on demand, and routes against measured headroom — so a model
that would not fit beside what is already loaded is refused with the shortfall in
the message instead of crashing inside the engine.
All adapters are registered by default. See
docs/adding-a-model.md, docs/three-d.md,
and the protocol templates in config/workflows/.
One command, on a machine that already has Node ≥ 22.6 and DeepSeek Harness:
# into the profile behind `dsh web` — the normal case
dsh plugin --profile web add github:Mhmd7-7/dsh-ai-model-hub
This package is the plugin: it declares dsh.bundle.patch, so dsh plugin add installs it and appends dsh-ai-model-hub to dsh.profile.bundles. The
profile composes that bundle on every boot, which is what makes the plugin
live, not merely installed. No profile file is edited by hand and no
cordis.patch.yml has to be written.
Then restart DeepSeek Harness and ask:
List the available AI model capabilities.
| What you want | Command |
|---|---|
| Another profile | dsh plugin --profile hubtest add github:Mhmd7-7/dsh-ai-model-hub |
| A pinned release | dsh plugin --profile web add github:Mhmd7-7/dsh-ai-model-hub#v0.7.0 |
| A published copy | dsh plugin --profile web add dsh-ai-model-hub |
| A checkout you are editing | node --no-deprecation scripts/install-plugin.mjs --profile web |
The last one links the working tree instead of copying it, so the plugin loads
your sources; run npm run build there first if you edited anything under
src/ or dsh-plugin/.
install.ps1 and install.sh remain one-line wrappers for the first row, and
keep their flags (-Profile, -Ref, -Repository, -SkipDoctor):
irm https://raw.githubusercontent.com/Mhmd7-7/dsh-ai-model-hub/main/install.ps1 | iex
curl -fsSL https://raw.githubusercontent.com/Mhmd7-7/dsh-ai-model-hub/main/install.sh | sh
To remove an installation:
dsh plugin --profile web remove dsh-ai-model-hub
The plugin tracks the DSH 0.x line rather than a single release:
| Supported | @deepseek-ai/dsh* >=0.1.5-rc.2 <0.3.0 |
| Developed and tested against | 0.1.5-rc.2 (this package's own tree) and 0.2.0-rc.2 (a current host) |
Two independent things decide whether that holds, and they fail in different places — which is worth knowing, because only one of them is visible in the plugin's own logs.
The manifest gate runs first, before any plugin code. DSH reads
peerDependencies and refuses the row when a @deepseek-ai/dsh* range does not
match the harness. For a 0.x version, caret pins the minor: ^0.1.5-rc.2
means >=0.1.5-rc.2 <0.2.0, so it silently refuses every 0.2.x host. That is
why the peers here are written as explicit ranges and why
tests/peer-range.test.ts exists — it asserts against the harness's own
evaluatePluginCompatibility when that package is installed, and against a
written-out range matcher when it is not. The old caret form is demonstrated
failing in the same file, so it cannot come back unnoticed.
Then the runtime surface is detected, not assumed. dsh-plugin/compat.ts is
the only module that touches a host member. It reads the context once, reports
what it found, and every registration goes through a helper that contains a
missing or differently-shaped member to one diagnostic line:
| If the host has no… | What is lost | What is not lost |
|---|---|---|
tools | everything | nothing — the load fails, by name, on purpose |
systemPrompt | the capability snapshot in the prompt | every tool; the agent can still call list_capabilities |
webServer | the Local models and Priorities pages | every tool |
effect | stopping the hub's processes on plugin unload | every tool; the effect body is not run at all, because nothing could call its disposer |
skills | the bundled skill | every tool |
sandboxPolicy | artifacts in the calling session's workspace | every tool; artifacts fall back to the hub's own store |
get | nothing | the injected-property path still resolves every optional service |
logger | host-formatted diagnostics | the diagnostics — they go to the console instead |
tools is the one hard requirement. Without a tool registry there is nothing for
this plugin to do, and a hub the agent cannot reach is worse than an honest
failure: it produces a settings page and the false impression that the capability
is available.
Two things are deliberately not wrapped. defineTool, because it is a pure
function from a closed input to a closed output whose signature has not moved
across the whole 0.1.5-rc → 0.2.0-rc line. And the import type lines, which
vanish at compile time — a host that renames a type cannot break a running plugin.
The complete runtime DSH coupling is six value imports: defineTool,
schemastery's default export, and cordis's Service.
The plugin's sources are TypeScript, and Node runs them directly through type
stripping while you work in a checkout. It refuses to do that inside
node_modules:
ERR_UNSUPPORTED_NODE_MODULES_TYPE_STRIPPING: Stripping types is currently
unsupported for files under node_modules
A package installed by name — from GitHub, from a registry — lands inside the
profile's node_modules, so a .ts entry there could never be imported. That is
the entire reason this repository commits lib/: lib/ is the same sources
compiled to JavaScript, and the package's entry points (main, exports) point
at it, so dsh plugin add needs no build step on your machine.
lib/ is generated, never edited: npm run build regenerates it from src/ and
dsh-plugin/. The one setting that makes it work is
rewriteRelativeImportExtensions, so the emitted JavaScript imports ./x.js
where the source said ./x.ts.
There are two halves, and you can use either without the other.
cd dsh-ai-model-hub
npm install # typescript + the DSH packages the plugin links against
npm test # the suite, ~20 s, no engines or GPU required
npm run typecheck # strict TypeScript, no errors
npm run build # compile src/ + dsh-plugin/ into lib/
npm run demo # the vertical slice, end to end
npm run demo:workflow # a multi-step cross-model pipeline
npm run demo:3d # image → 3D → a real .glb, against a stand-in engine
The demo prints the full path — capability discovery, routing with reasons, invocation, and the artifact on disk:
=== catalog: capability discovery ===
text_to_image text → image via comfyui_z_image_turbo
text_to_text text → text via ollama_text_model
not available here: image_to_3d, image_to_image, audio_generation, …
=== routing ===
chose comfyui_z_image_turbo
rationale: comfyui_z_image_turbo (score 1): …
[rejected] ollama_text_model: does not declare capability "text_to_image"
=== artifact ===
valid PNG: yes
size: 1.2 MiB
It runs against config/models.json, so it needs the engines that catalog names:
Ollama for text, ComfyUI for images. node examples/vertical-slice.ts <catalog>
takes another catalog, and the last section deliberately asks for a capability
nobody serves, to show what a refusal looks like.
npm run demo:3d needs nothing installed. It starts a stand-in 3D engine inside
the example and runs the whole real path against it — discovery, health,
resource-aware routing, a two-step Gradio protocol, and a .glb written to disk —
then prints every HTTP request the engine saw, to make the boundary concrete. To
point it at TRELLIS instead, swap the stand-in for your catalog entry; the three
capability calls are identical. See docs/three-d.md.
# Installs into the 'web' profile (the one behind `dsh web`), then verifies.
npm run install:plugin
# ...or target another profile:
node --no-deprecation scripts/install-plugin.mjs --profile <name>
Then restart DeepSeek Harness and ask the agent:
List the available AI model capabilities.
npm run doctor # inspects: is it loadable, and if not, why not
node scripts/smoke.mjs --profile web # acts: applies the plugin and lists what it registered
doctor inspects and reports; smoke actually runs the plugin against a real
cordis context and prints the tools it registered. Use smoke when you want
proof rather than a checklist:
Registered tools:
check_model_health explain_routing get_model_status invoke_model
list_artifacts list_capabilities list_models start_model
stop_model refresh_model_discovery
Catalog: model hub ready: 2 model(s), 2 capability(ies) from …/config/models.json
Note on flags.
npm runforwards flags to the script, but npm 12 parses--profileas its own config flag and fails. Use the directnode scripts/smoke.mjs --profile <name>form for a non-default profile.
If the harness dies with:
invalid plugin, expect function or object with an "apply" method, received object
then the row cordis.patch.yml inserts is pointing at something that is not a
plugin. In this package that can only happen if the row's name was changed
away from the bare package name: dsh-ai-model-hub resolves to the package's
. export, which is the plugin (lib/dsh-plugin/index.js, exporting
name/inject/apply/Config). The hub library lives behind the
dsh-ai-model-hub/library subpath instead, precisely so the two cannot be
confused.
The row has two names, and they are easy to conflate: id is the row's name in
the composed tree (what a profile's own cordis.patch.yml uses to override or
disable it), while name is the module specifier the loader imports. Keep
name bare — a subpath such as dsh-ai-model-hub/plugin would load the host half
but silently drop the Web UI half, because DSH's client module system only
resolves a row's package manifest from a bare specifier. cordis.patch.yml
documents this in place, and npm run doctor checks it.
The plugin resolves its catalog from ordered anchors, first hit wins:
configPath — an explicit file, if configured. When set it is the only
candidate: a wrong path is an error, never a silent fallback.searchRoots, for a deployment that names its own.At every level discovery accepts models.json, model-catalog.json, or
dsh-ai-model-hub.json, directly or in a config/ subdirectory.
Anchor 4 is what makes the zero-config case work. DSH installs profile plugins as
junction links and Node resolves them to their real path, so the plugin finds the
config/models.json shipped by the checkout it was installed from — no matter
where the host was launched. Anchor 3 covers the opposite case: a host started
inside a project that carries its own catalog.
Why the host's working directory and not the session's? A plugin's
applyruns when the profile boots — before any session exists — and one host serves many sessions with different workspaces, so there is no single "session working directory" to read at that moment. Anchors 1 and 2 are the explicit way to name a workspace. Per-session resolution is a deliberate future step, not an oversight; see docs/roadmap.md.
If no anchor yields a catalog the plugin logs model hub disabled together with
every directory it tried, registers no tools, and lets DSH boot normally — a
missing or broken catalog never makes the agent unusable.
dsh plugin add <path> links the working tree — DSH installs profile plugins
as junction links — so Node resolves the plugin to its real path in the
repository instead of to a copy inside the profile. Peer dependencies then
resolve by walking up from the package: from lib/dsh-plugin/ to
<repo>/node_modules. Run npm install once in the repository and that
directory exists, because dsh plugin add forwards to pnpm, which only warns
about peer dependencies instead of installing them.
This is not theoretical: it was found by installing into a throwaway profile,
deleting the workspace's node_modules, and watching the plugin fail to resolve
@deepseek-ai/schemastery.
A linked checkout also loads lib/, exactly like an installed copy does — so run
npm run build after editing src/ or dsh-plugin/. If you would rather not
think about that, work on the hub as a library (npm test, npm run typecheck)
and only build when you want the plugin to pick the change up.
The historical obstacle on the GitHub route — Node refusing to strip types for
files under node_modules — is what committing lib/ removes; see Why the
package ships compiled JavaScript.
| Ask the agent | Tool it calls |
|---|---|
| "What AI models can you run?" | list_models |
| "What can this machine actually do?" | list_capabilities |
| "Create a futuristic city image." | invoke_model (text_to_image) → router picks the model |
| "Now make a 3D model from that image." | invoke_model (image_to_3d), chaining the artifact id |
| "Why did it pick that model?" | explain_routing — read-only |
| "Is the image model running?" | get_model_status, check_model_health |
| "Start / stop the image model." | start_model, stop_model |
The agent never names an engine, a command, or a path. That is the whole design.
The plugin also teaches the agent how to use it.
skills/dsh-ai-model-hub/SKILL.md is registered
as a skill provider on ctx.skills, the way DSH's own bundled dsh-badge skill is,
so:
~/.dsh/skills or a project skill root, and no profile
file is edited — installing or updating the plugin installs the skill;git pull changes what the agent
reads on its next load, with no reinstall step and no stale copy;The registration is deliberately not a declared inject. It goes through
ctx.inject(['skills'], …), so a composition with no skill registry — headless, or
the sdk-minimal profile — still gets the model tools and simply never runs the
skill callback. A missing or malformed skill file costs one warning and the skill,
never the tools beside it.
The file is an ordinary skill, so it also works the old-fashioned way: link or copy
skills/dsh-ai-model-hub/ into a .dsh/skills root if you would rather not run the
plugin at all. Its frontmatter is the single source of truth for the name and
description on both paths.
The agent is not the only one who needs to know where the engines are. The plugin also adds a Local models section to the DSH settings, which answers the question a user actually has: which engines are on this PC, where is each one installed, is it running, and what is inside it.
| Row | Where it comes from |
|---|---|
| Verified installation | Engine markers or executables in configured/well-known locations; a configured cwd alone is not proof |
| Configured paths | Catalog launch paths, displayed separately from verified installation paths |
| Listens on / availability | Independent bounded endpoint checks on both Scan and Refresh |
| Hub can start it | The descriptor's startable, after the deployment's allowProcessLaunch |
| Catalog models / capabilities | The same live hub the tools read, so the page cannot disagree with behaviour |
Scan is the page's main action, and it produces one list of two kinds of thing, each row labelled with what it is:
| Kind | What it is | Where it comes from |
|---|---|---|
| Ollama Model | An observed or catalog-known model | Live /api/tags, offline loopback-host manifests (OLLAMA_MODELS or ~/.ollama/models), active catalog, or last-good observations |
| ComfyUI Workflow | One complete workflow or an explained unreadable entry | Saved /userdata workflows, verified installation's workflow store, configured directories, and active catalog inline/pinned workflows |
A ComfyUI workflow is the smallest thing the list exposes. Its checkpoints, LoRAs, VAEs, text encoders, ControlNets and individual nodes are not listed — they are the workflow's own business, which is why changing what a workflow loads never changes anything an agent or a user has to know.
The two halves are discovered independently, so an Ollama instance that is down cannot hide the workflows and a ComfyUI that is down cannot hide the models; each source reports its own status. Scanning twice refreshes the list rather than growing it, because every entry's id comes from its source location.
A saved workflow is discovered without a running ComfyUI. Workflow discovery
reads the configured directories, the engine's own saved-workflow store when it
can be found on disk, the active catalog, and the workflows earlier scans
remembered — so a first launch with the engine stopped still lists the graphs that
are on the machine, and an engine that stops mid-session cannot take them off the
screen. The list says which evidence it has: ComfyUI stopped when a refused
connection proves the process is not running, ComfyUI unreachable when a timeout
or a routed address leaves that genuinely unknown, and a cached row whenever the
contract came from a previous observation rather than this pass. Only a scan that
completely read a source can conclude a workflow was deleted; a failure never can.
The remembered workflows live in
~/.dsh/plugins/dsh-ai-model-hub/workflow-catalog.json, beside the capability
policy and for the same reason: it is machine state that changes without anyone
editing a file. It is per host and endpoint, so repointing the catalog at another
ComfyUI does not resurrect the old engine's workflows. workflowCatalogPath
configures it; DSH_AI_MODEL_HUB_WORKFLOW_CATALOG redirects it for a test.
Scan starts the configured engines; looking never does. Opening or reloading the page reads the machine as it finds it, and says what did not answer. Pressing Scan is a deliberate act, so it checks each engine, starts any that is stopped through the lifecycle manager's own launch path, polls its API until it answers within a bounded budget, reads it, and finally releases its hold:
| Stage | What happens |
|---|---|
| Checking engines | Each engine's own readiness route is probed. An engine that answers is adopted and never restarted |
| Starting engines | A stopped engine is launched with the catalog's configured command; a missing command, a forbidden launch, or a refused spawn is reported with the setting that fixes it |
| Waiting for readiness | Bounded polling of the engine's API — a listening socket is not readiness. ComfyUI cold-starts get a longer default budget because loading checkpoints takes minutes |
| Scanning | Every readable engine is read: Ollama's models, ComfyUI's saved workflows |
| Complete / Partially complete | complete is false whenever an engine could not be started, timed out, died mid-read, or has no launch command, and the engine's own report says which and why |
While the scan reads, the engines it started are held against automatic idle
unloading; the hold is released in a finally, on success, failure and
cancellation alike, so a scan can never leave an engine permanently unsweepable. An
engine that was already running is neither started nor held — it is the user's
process, not the scan's. Repeated Scan clicks cannot start a second copy: the launch
goes through the same gate a cold invocation uses, which serializes transitions per
model, and a later scan simply reuses what is now running.
Engine states distinguish running, installed_unreachable, not_found in the
checked locations, and unknown. Connection failure alone proves neither that a
process is stopped nor that an installation is absent. The installation is the
directory whose contents are the engine — a ComfyUI whose main.py lives under
src/ is installed at src/, which is also where its saved-workflow store hangs
from. Nothing is ever installed automatically.
Resource contract status is separate from availability. Each resource reports
live, local, catalog, or cached freshness and its observation timestamp.
An answering ComfyUI server does not verify a workflow's weights or execution;
local Ollama manifests do not verify every referenced blob. Failed refreshes keep
the last useful snapshot visibly stale; Retry requires no page reload. Superseded
requests cannot replace newer results or cache/provider publications.
Workflow contract statuses include:
| Status | Meaning |
|---|---|
ready | Saved in API format — the hub queues it as it stands |
needs_conversion | Saved in the editor's UI format. It is listed and converted for execution, but it is not the same as an API export |
invalid | Readable, but produces nothing a caller can keep. The reason is shown |
unreadable | The file or the engine's copy of it could not be retrieved. The transport reason is shown, because the document was never seen |
unavailable | The engine needed to interpret it could not be reached |
Running a remembered workflow while ComfyUI is stopped goes through the same lifecycle manager as any other cold engine: the configured instance is started, its readiness is awaited within the declared budget, and the graph is queued once it answers. A startup that is refused or fails leaves the workflow listed with the reason, and never falls through to a subscription provider — a workflow the user picked is not silently swapped for a different one.
Runnable workflows are published to the hub as workflow-backed providers, so the
agent can route to them with the existing invoke_model tool by pinning the
workflow's id — the same harness, no second execution path. The public parameters
a workflow exposes (prompt, negative prompt, image, seed, steps, CFG, width,
height) are inferred from its graph and are the only fields a caller may set.
Engines the catalog does not declare are still listed — a1111, comfyui, and
ollama are known by name — with the directories that were checked, so "Forge is
installed but the hub has not been told" is visible instead of silent.
Two halves, both dependency-free:
dsh-plugin/inventory.ts builds the answer and serves it on
GET /dsh-ai-model-hub/inventory. Every probe is bounded and failures are
contained; the route is injected on demand, so a headless profile with no web
server keeps every tool and simply serves no page. The scan itself lives in
dsh-plugin/scan.ts, and the graph → public-contract inference it uses is in
src/comfy/scan.ts.dsh-plugin/client.js is the browser half. DSH loads plugin clients through
window.__ModuleLoader__ as plain side-effect scripts — no top-level
import/export, React handed in through the factory's require — so this is
hand-written JavaScript with no bundler and no build step, matching the rest
of the project. It registers one settings.section contribution.Configure a workflow directory when your workflows live beside the catalog rather than inside ComfyUI:
config:
comfyuiWorkflowDir: 'C:/workflows/comfy' # every *.json under it, recursively
scanTimeoutMs: 4000 # per-probe budget; filesystem walks are also capped
workflowCatalogPath: '' # where discovered workflows are remembered
A relative directory is resolved against the catalog that configured it, and
%VAR% and ~/ are expanded from the host's own environment, so the same
configuration works on Windows and on a POSIX host. The directory is a source of
workflows only — nothing in it enumerates weights or custom nodes.
To see what a stopped engine still yields on this machine, with every HTTP probe refusing:
node scripts/verify-live-offline.ts # writes nothing
node scripts/verify-live-offline.ts ./scratch-catalog.json
Open Settings → Local models, then use the Edit priorities button at the top left or the Priorities tab. The Inventory tab stays mounted alongside the editor, so switching views or rescanning does not discard an unsaved draft.
The editor uses responsive cards, human-readable capability names, provider search, and All / With options / Enabled filters. Search matches capabilities and their providers; it keeps each matching capability's full ordered list visible so filtering never changes the meaning of a rank. Technical status and the policy file are available in expandable details, and Save / Cancel stay reachable in a sticky action bar. Light and dark colors follow the host theme.
Each capability contains every option that genuinely declares it — local models, ComfyUI workflows, and subscription backends in the same list, because to the hub they are the same kind of thing: a way to serve a capability.
On hosts that only report effective enablement while globally disabled, individual editing is paused rather than guessing at saved permissions. Enable and save the global switch to edit those options. Simply opening this page, switching tabs, or changing filters never changes the policy or starts a model.
Three facts about an option are deliberately kept apart, and confusing them is the failure this section exists to prevent:
| Fact | Where it lives | What it means |
|---|---|---|
| Enabled / Disabled | The policy you save | May this option serve this capability |
| Position | The policy you save | When it is tried, relative to the others |
| State | The live runtime | Is it usable right now — ready, stopped, starting, unhealthy, unavailable, unsupported |
So "disabled" is not "stopped", and "lower priority" is not "unavailable":
Disabled, keeps its position, and is
never run — not as a first choice and not as a fallback — and its engine is
never started for it. Only you can change that.Reordering is real drag-and-drop — and every row also has Move up, Move down, Move to top and Move to bottom buttons with accessible names, so the list is not mouse-only. Nothing is written until Save; Cancel reverts to the saved policy. That matters for a list whose whole point is an arrangement: a drag that saved immediately would make a mistake irreversible.
The policy is persisted as its own versioned document, keyed by stable ids
(never display names), at ~/.dsh/plugins/dsh-ai-model-hub/capability-policy.json —
override it with policyPath. It survives reloads, restarts, and catalog refreshes,
and an option that is temporarily undiscovered keeps its saved settings rather
than being forgotten.
Nothing new activates itself. A newly discovered option — a pulled Ollama model,
a workflow saved in ComfyUI, a subscription backend a plugin just registered —
appears in the list immediately, disabled, and stays that way until you enable
it. Catalog presence, a descriptor's own enabled flag, and visibility in a picker
are not authorization. Options that were already configured when the plugin loaded
keep working exactly as they did, so the feature is a no-op on an existing
deployment until you touch it.
For every request, in order:
There is no attempt cap by default: a saved order of ten options really is ten
attempts. A deployment that needs a smaller budget sets attemptBudget, and the
failure message then says how many options went untried rather than implying they
failed.
A bare 2xx proves that something is listening. It does not prove the engine can
serve work — a half-loaded engine or any local web page answers 200 too, and this
project has been bitten by exactly that. A health check may therefore demand the
shape of a real answer:
"health": {
"kind": "http",
"path": "/system_stats",
"expect": "json-key",
"expectKeys": ["system", "devices"]
}
any (the default) keeps the historical permissive behaviour; json requires a JSON
body; json-key additionally requires one of expectKeys at the top level. A body
that fails the expectation is an unhealthy model with the observed content in the
reason. The shipped catalog uses this for ComfyUI (/system_stats) and Ollama
(/api/tags).
Automatic local-engine startup is the deployment-level allowProcessLaunch switch,
and it has always been the one place to decide it — once, not per request:
config:
allowProcessLaunch: true # the hub may start an engine its catalog declares startable
Both halves must agree: the descriptor declares lifecycle.startable: true with a
start command, and the deployment opts in. With launching off, an engine that is
merely stopped is reported as not startable by the hub with that as the reason —
never launched anyway, and never confused with a failed launch. See
docs/security.md before enabling it.
invoke_model reports the requested capability, the policy revision that decided
it, the ordered list of permitted options, everything that was skipped with the
switch or fact that skipped it, and how each attempt turned out — distinguishing
launch forbidden, stopped, startup failed, and generation failed by
the stage it failed at. That is reported on success too: "why did this run on the
second option?" is answered by the first option's reason, not by a log hunt.
A timeout or a dropped connection after the request reached the backend leaves a
question this process cannot answer: did it generate? The hub refuses to guess. It
reports an explicit unresolved outcome, with any identifiers that can be used to
look the work up, and does not fall back — because generating again could
double-charge a subscription or repeat a three-minute render. A refusal the backend
stated (quota, a login problem, a 4xx) is a definitive "no", and the plan moves
on as usual.
The shipped catalog (config/models.json) declares two real
engines — Ollama for text and ComfyUI for images — and includes a specific ComfyUI
workflow provider. Ensure Ollama has the configured model and ComfyUI has the
workflow's required weights; installing a new checkpoint or saving a workflow does
not, on its own, register a new ComfyUI provider.
To change them, or to add more, edit that file or point the plugin at your own
with configPath / searchRoots. Everything else is copy-and-adjust from
config/examples/real-models.example.json:
an A1111/Forge/sd.cpp server uses http_json, llama.cpp uses
openai_compatible, and ComfyUI uses its own comfyui adapter driven by a graph
template — an example is in
config/workflows/z-image-turbo.api.json,
an API-format {client_id, prompt} graph. The provider additionally declares exact
bindings for caller inputs and an outputs node/type selector; the graph's
internals are not inferred. To add a ComfyUI provider, configure the host's
adapterConfig.workflows with id, name, workflowPath, capabilities,
bindings, outputs and optional routing/resource settings. The static model
alternative uses adapterConfig.providerKind: "workflow" and workflowId.
A catalog names those files (workflowPath for ComfyUI, stepsPath for 3D) with
paths relative to the catalog itself, because a catalog can live anywhere and
the process that reads it can be started from anywhere. The shipped
config/models.json therefore says
workflows/z-image-turbo.api.json — config/ is that file's own directory —
while the same template from inside config/examples/ is
../workflows/z-image-turbo.api.json. Absolute paths also work. A path that does
not resolve names itself and its base in the error rather than guessing.
A wrong model name is a loud failure — the engine answers 404 and the invocation fails — never fixture text: there are no mock models in any shipped catalog. See docs/adding-a-model.md.
For supported enumerating engines such as Ollama, set discoverModels: true on
the plugin to register pulled models. ComfyUI remains workflow-only and
configuration-driven: discovery does not publish its loose checkpoints or saved
workflows. Add a ComfyUI workflow descriptor to the host config (or a static
workflow model) and reload the catalog:
- insert:
- id: dsh-ai-model-hub
name: 'dsh-ai-model-hub'
config:
discoverModels: true
It is off by default, and off means no engine is ever contacted for introspection. With it on:
config/models.json still wins. A discovered model whose id a hand-written
entry claims is dropped before the catalog is built, so the file stays the
place to set exact resources or fix priority. ComfyUI workflow providers are
always explicitly configured, independently of this discovery switch.discoveryTtlMs (60 s). The
refresh_model_discovery tool bypasses the cache — that is the path right
after ollama pull.Discovery never bypasses resolveDescriptor(): it produces descriptors in the
same shape as a models.json entry, and the router, runtime manager, and
adapters are unchanged by it. See
docs/architecture.md.
Nothing is launched by default: allowProcessLaunch is false, so the hub talks
to engines you run but will not start any. See
docs/security.md before turning that on.
Artifacts land in <workspace>/artifacts, where "workspace" means the
calling session's, not the host process's working directory. That distinction is
the whole point: dsh web is one long-lived process serving many sessions with
different workspaces, and it is built — and used to resolve this path — before any
session exists. So the root is resolved per tool call, from DSH's own
ctx.sandboxPolicy.resolve({ session }).workspaceRoot, which is the same value the
file tools treat as the workspace boundary.
The consequence a user notices: switch workspace in the GUI and output follows,
with no configuration. list_artifacts reads the same per-call root, so chaining a
workflow still finds what the previous step produced.
Set artifactRoot in the plugin row to pin one absolute directory for every
session instead — the right choice when collecting artifacts centrally, the wrong
one when you want each conversation's images beside its code. The hub refuses a
relative root rather than guessing at a base.
Turn it on and engines start themselves. A descriptor that marks its engine
startable — "startable": true plus a start command, as the a1111 host in
config/examples/real-models.example.json
does — lets start_model bring the engine up, and lets invoke_model do it
implicitly for a cold model, waiting for the health check before the request is
sent. A stopped ComfyUI then costs one cold start, not a failed task:
invoke_model({ capability: 'text_to_image', prompt: 'a red fox in deep snow' })
→ text_to_image served by comfyui_z_image_turbo in 22620 ms (cold start).
Two switches have to agree — the descriptor's startable and the deployment's
allowProcessLaunch — so an 8 GB model cannot be started by a catalog file alone.
Starting an engine that is already healthy is a no-op rather than a duplicate
process, and only processes the hub started can be stopped by stop_model.
The agent asks for a capability; the router picks a model.
// What the agent does. Note what it does NOT say: no model, no engine, no command.
await tools.invoke_model({ capability: 'text_to_image', prompt: 'a futuristic city' })
// What the hub does, in order:
// 1. resolve the request's artifacts
// 2. filter models by capability, input kinds, tags, resources, availability
// 3. sort deterministically (priority, then id)
// 4. start the chosen model if it is cold
// 5. call its adapter
// 6. persist the outputs as artifacts and return them
A model is data, not code. Adding one means adding a JSON entry:
{
"id": "a1111_sdxl",
"name": "SDXL (WebUI)",
"type": "image_generation",
"host": "a1111",
"capabilities": ["text_to_image", "image_to_image"],
"adapterConfig": { "model": "sd_xl_base_1.0.safetensors", "steps": 30 },
"priority": 10,
"tags": ["local", "gpu", "image"]
}
No router change. No prompt change. No DSH change.
Artifacts are the only thing that crosses a model boundary. That is what
makes text → image → 3D compose without anything knowing about PNG or GLB:
const image = await hub.invokeModel({ capability: 'text_to_image', prompt: 'a spaceship' })
const mesh = await hub.invokeModel({ capability: 'image_to_3d', inputs: [image.outputs[0].id] })
dsh-ai-model-hub/
├── src/ the hub — imports NOTHING from DeepSeek Harness
│ ├── catalog/ capability vocabulary, descriptor schema, registry,
│ │ published JSON Schema
│ ├── router/ deterministic capability-first selection
│ ├── runtime/ process lifecycle, health, idle timeout
│ ├── adapters/ adapter contract + the mock test double,
│ │ PNG/STL/WAV writers, openai_compatible,
│ │ http_json, comfyui, three_d
│ ├── artifacts/ artifact contract, the local filesystem store,
│ │ and 3D container sniffing/measurement
│ ├── discovery/ engine introspection: ollama · comfyui · a1111 ·
│ │ three_d, and the merge that keeps static first
│ ├── comfy/ the workflow contract (bindings + typed outputs)
│ │ and the scan that infers one from a saved graph
│ ├── config/ catalog discovery and loading
│ ├── util/ process guardrails, primitive validators
│ ├── machine.ts RAM/VRAM/GPU/disk probing, free figures included
│ ├── hub.ts the facade: the public API
│ └── index.ts the public surface
├── dsh-plugin/ the ONLY DSH-aware code (~1 file per concern)
│ ├── index.ts plugin entry: name / inject / apply / Config
│ ├── service.ts the hub as a cordis service
│ ├── config.ts plugin configuration schema
│ ├── types.ts the DSH API bridge — where a breaking change lands
│ ├── skills.ts the bundled agent skill: one provider, one file
│ ├── inventory.ts engine discovery + the route the settings page reads
│ ├── scan.ts the Scan pass: Ollama models + complete ComfyUI workflows
│ ├── client.js the browser half: the "Local models" settings page
│ └── tools/ discovery · lifecycle · invoke
├── skills/dsh-ai-model-hub/
│ └── SKILL.md the agent-facing skill the plugin registers
├── lib/ src/ and dsh-plugin/ compiled to JavaScript:
│ the entry points `dsh plugin add` loads, and
│ the reason a GitHub install needs no build
├── cordis.patch.yml the bundle patch: the row the profile mounts
├── install.ps1 one-line install from GitHub (Windows/PowerShell)
├── install.sh one-line install from GitHub (macOS/Linux)
├── scripts/ install-plugin · doctor · smoke — install and verification
├── config/
│ ├── models.json the catalog the package ships: real engines
│ ├── examples/ more real engines, ready to copy
│ └── workflows/ ComfyUI graph templates and 3D step declarations
├── tests/ the suite; real local HTTP servers stand in for
│ engines, so no GPU or model download is needed
├── examples/ vertical-slice · workflow · three-d-slice demos
└── docs/ architecture, components, guides
The dependency rule that makes this survivable: src/ does not import
@deepseek-ai/*. Only dsh-plugin/ does. DSH is a 0.1.x-rc line, so breaking
changes are expected; when one lands, the blast radius is the plugin layer, and
the entire hub keeps working and keeps passing its tests.
| Document | Read it when |
|---|---|
| docs/architecture.md | You want the design, the data flow, and why each boundary is where it is |
| docs/components.md | You are working on one component and want its contract in detail |
| docs/adding-a-model.md | You want to add a model — the common case needs no code |
| docs/three-d.md | You want local image-to-3D working: engine, catalog, discovery, troubleshooting |
| docs/security.md | You are about to enable process launching or widen the command allowlist |
| docs/roadmap.md | You want to know what each phase required and what remains |
npm test # the whole suite; real HTTP servers stand in for engines
npm run demo:3d # discovery → routing → a real .glb → the boundary transcript
npm run demo # capability discovery and routing against your own catalog
lib/
instead, because Node refuses to strip types under node_modules;
npm run build regenerates it.dsh plugin, which forwards to it.node_modules must resolve inside a DSH profile is a plugin that breaks
on the next release. The DSH packages it imports are peers, supplied by the
profile that loads it.MIT