Oh My Model
English | 简体中文
Independently managed model providers, models, and prompts for DeepSeek Harness.
Definitions live in a separate models.json in the Harness home directory and are edited
in the Oh My Model panel under Settings → Models, in English or Chinese
following your Harness language preference.
Features
- Manage multiple providers and API gateways in a single configuration file.
- Fetch a provider's live upstream model list and add the advertised IDs, instead of typing every one.
- Save model definitions and typed API keys with one Save; keys go to the Harness
credentials store, never into
models.json. - Inherit context window, output limits, input types, reasoning levels, and pricing from the models.dev catalog.
- Reuse sibling model definitions within the same provider, specifying only overrides.
- Append dedicated system prompts per model, active only when that model is selected.
- Fine-grained control over upstream model IDs, request parameters, cost rates, and compaction policies.
Requires DSH >=0.1.5-rc.3 <0.3.0 and Node.js >=22.13.
Installation
Install dsh-oh-my-model@0.1.0-2 from the Plugins page. You can also install from a local directory:
- Open Plugins in the sidebar and install the package using the directory's absolute path.
- Restart DSH, then navigate to Settings → Models to access the Oh My Model panel below.
To upgrade a local installation, keep the existing link, update the files in the directory, and restart DSH. (Refreshing the browser alone does not reload host modules.)
Quick start
- Click Add, then enter the provider ID, display name, base URL, and protocol (the protocol may be left blank when it can be auto-detected).
- Enter your key in API key. It is written to the Harness credentials store
when you save;
models.jsononly records the reference name. - Click Add model and fill in the model ID. Use inherit from to pick a catalog preset, or alias a sibling to reuse an existing model from this provider. Fetch upstream models can list the IDs the endpoint advertises for you to pick.
- Unset fields inherit automatically; gray text displays the resolved values.
- Prompt appends a model-specific system prompt; Advanced settings configures custom request parameters and compaction policies.
- Click Save. It writes the typed keys to the credentials store and then saves the document; edits remain drafts until saved, and Discard drops both.
The search box at the top filters providers and models by name or ID; press Escape or
click Clear search to reset. The interface warns before closing or refreshing with unsaved changes,
and confirms before discarding edits or loading external revisions.
Fetching upstream models
A provider card offers Fetch upstream models below the manual model-ID row. The plugin asks the configured endpoint, using the stored API key when there is one, which models it advertises, and the editor lets you select which IDs to add.
- The list is fetched on demand only: it is not cached, and it is never merged into
models.jsonby itself. - Selected IDs are appended to the provider as plain
{}model entries in the draft. Nothing is written until Save, and a fetched entry behaves exactly like one typed by hand: metadata and pricing still come from models.dev through inherit from. - Keyless endpoints are fetched without authentication; a missing or unconfigured credential reference is reported instead of guessed.
Example: GPT-6 Luna with a fast variant
Add a provider, create gpt-6-luna, and select openai/gpt-6-luna in
inherit from. Add gpt-6-luna-fast as an alias of gpt-6-luna, set its
upstream model ID to gpt-6-luna, and add {"service_tier":"priority"} to its
request parameters. Click Save.
The same setup in models.json:
{
"version": 1,
"providers": {
"gateway": {
"displayName": "My gateway",
"api": "openai-responses",
"baseURL": "https://gateway.example/v1",
"apiKeyEnv": "GATEWAY_API_KEY",
"models": {
"gpt-6-luna": {
"inherit": "openai/gpt-6-luna"
},
"gpt-6-luna-fast": {
"aliasOf": "gpt-6-luna",
"upstreamId": "gpt-6-luna",
"name": "GPT-6 Luna · Fast",
"samplingParams": {
"service_tier": "priority"
}
}
}
}
}
}
Replace the example URL with your endpoint, keep apiKeyEnv as
GATEWAY_API_KEY, and type the key in API key so Save stores it. The base
model inherits its limits, input modalities, reasoning options, and pricing from the
catalog. The fast alias shares those defaults and dispatches upstream as
gpt-6-luna with priority service.
Priority service requires endpoint support. Availability, latency, and pricing
follow your provider’s priority tier. Override cost on the fast alias to match
your provider’s rates.
Saving keys and models
Save is the only save action: it writes any typed API keys to the Harness
credentials store and then saves models.json. There is no separate per-credential
save button.
| Item | API key | Model configuration |
|---|---|---|
| Written to | Harness credentials store | models.json |
| Action | Save | Save |
| Stored in file | Nothing; the file holds only the apiKeyEnv reference |
Provider and model definitions |
| Affected by Discard | Yes: a typed key is discarded | Yes |
- Discard discards typed keys together with the draft edits. Remove stored credential remains a separate, immediate action.
- The key is never written to
models.json, never enters the draft, and never appears in validation requests or logs. - After saving, a provider's credential editor shows the configured status again.
Changing
apiKeyEnvstill takes effect when the document is saved.
Inheritance
- inherit from: Copies display name, context window, output limit, input modalities, reasoning levels, and cost from a models.dev entry. Explicitly set fields override inherited values.
- alias a sibling: Reuses another model definition from the same provider, specifying only overrides and inheriting its prompt.
- Mid-conversation capabilities (dynamic prompt updates, tool changes) are inherited automatically
when the target matches pi-ai's built-in catalog. Use
inheritCompatto opt out or specify another source.
Prompts
systemPromptis appended after the session system prompt, matched by the local provider and model ID.- An alias inherits its target's prompt;
""clears it, and ↺ inherited restores it. - Disabled models do not contribute system prompts.
- A Harness
completeprompt override maintains normal precedence and supersedes appended sections.
Configuration defaults to models.json in the Harness home directory; the catalog cache and legacy
ownership records reside alongside it. Paths are configurable, and the file is watched automatically:
valid external edits apply immediately, while broken edits are safely ignored to retain the last known good configuration.
Documentation
| Document | Contents |
|---|---|
| Configuration | Grammar, fields, inheritance keywords, compaction, plugin config |
| Catalog and capabilities | models.dev cache, live upstream model list, protocol inference, capability inheritance |
| Architecture | Modules, host composition, prompt injection, route service, endpoints |
| Compatibility | Version ranges, tested combinations, updates |
License
MIT. See LICENSE.
No comments yet. Be the first to write one.