dsh-trustedrouter
Per-model TrustedRouter privacy and routing settings for DeepSeek Harness (dsh).
For each model, choose a privacy floor, which providers may serve it, what to optimize for, how it's billed and whether providers must be US-based. You do this from Settings → Models in the dsh web UI. The plugin adds your choices to every request as TrustedRouter's provider object.
The Privacy & routing section inside a TrustedRouter card on Settings → Models
Why a plugin
TrustedRouter reads routing and privacy policy only from the request body, for example {"provider": {"min_privacy": "confidential"}}. No header or API-key setting does the same job. dsh's model providers can add headers but not body fields, so without this plugin there's no way to ask for a privacy floor.
Requirements
- dsh web, tested on 0.2.0-rc.2 and 0.2.1-alpha.1 (see Tested with). The desktop app isn't supported.
- pnpm on your PATH. dsh installs plugins with it, and fails with "pnpm was not found" otherwise. If you don't have it, run
corepack enable pnpm or npm install -g pnpm.
- A TrustedRouter provider in dsh. Settings → Models → Add model provider → Custom model API:
- base URL
https://api.trustedrouter.com/v1;
- API protocol OpenAI Chat Completions;
- your TrustedRouter API key;
- the models you want, e.g.
z-ai/glm-5.3.
Install
From the dsh web UI:
- Sidebar → Plugins → Add plugin.
- Enter
https://github.com/AmmarByFar/dsh-trustedrouter#v0.1.0 and choose Install.
- Choose Enable now. No reload or restart is needed.
From the command line, for the web profile that dsh web uses:
dsh plugin --profile web add github:AmmarByFar/dsh-trustedrouter#v0.1.0
Then restart dsh web if it was running.
The plugin has no build step, so pnpm doesn't ask you to allow one. Its only peer dependency, @deepseek-ai/schemastery, comes from dsh itself, so the peer warning pnpm prints is expected.
Installing changes nothing yet. Requests go out unchanged until you set a policy.
Upgrading. dsh doesn't update installed plugins yet. Uninstall the plugin (Plugins → TrustedRouter privacy → Uninstall), then install the new tag. Your settings live in your profile, not in the plugin, so they're kept.
Use
Per model: Settings → Models
Open Settings → Models, find your TrustedRouter card and open Privacy & routing.
Each of the provider's models is listed with the policy it gets. Its tag says where that policy comes from:
- Custom for the model's own settings;
- Defaults when it only has the defaults;
- Uses <model> for a variant such as
z-ai/glm-5.3:nitro, which follows its base model unless you give it settings of its own.
Edit opens a model's settings:
Editing z-ai/glm-5.3: a Confidential floor, limited to two providers
| Setting | Sends | Notes |
|---|
| Privacy floor | min_privacy: no_store, zdr or confidential | Each level lists the providers that meet it for this model. TrustedRouter refuses the request rather than use a provider below the floor. |
| Prefer | sort: price, latency or throughput | |
| Only these providers | only | Empty allows every provider. |
| Never these providers | ignore | |
| Billing | usage: credits or byok | TrustedRouter credits, or your own provider key (BYOK). |
| US-headquartered providers only | jurisdiction: "us" | |
As you change settings, the editor counts the providers that still qualify. It warns you if none do, because TrustedRouter would then refuse every request for that model.
Apply saves to your profile, and the next request uses it. No restart is needed.
A few things to know:
- The section appears only on saved providers whose base URL is under a configured gateway. By default that's
https://api.trustedrouter.com/v1.
- You can only change settings from a browser on the same machine as dsh. That's a dsh rule for all settings.
- To list providers per model, your browser fetches TrustedRouter's public model list (
<base URL>/models, no API key sent) the first time you open the section. If that fails, every privacy level is offered.
Defaults and advanced settings: the Plugins page
Sidebar → Plugins → TrustedRouter privacy (under Installed), then the TrustedRouter privacy row under Components.
The TrustedRouter privacy page under Plugins
- Gateways: the base URLs the policy applies to, one per line. Add the EU gateway,
https://api-europe-west4.quillrouter.com/v1, if you use it.
- Defaults for every model: the same settings as a model has. A model's own settings override them field by field.
- Advanced defaults:
- preferred provider order (
order);
- whether to fall back to other providers (
allow_fallbacks);
- maximum prompt and completion prices in USD per million tokens (
max_price).
- Per-model overrides (JSON): every model's settings, including any
provider field the Models page doesn't show.
- Raw body rules (JSON): fields deep-merged into every POST under a base URL, for anything that isn't
provider.
- Log changes to the terminal: see Check that it works.
Config file
Both pages write the trustedrouter row in your profile's cordis.patch.yml. For dsh web that's ~/.dsh/profiles/web/cordis.patch.yml, or under $DSH_HOME if you set it. You can also edit it by hand:
- id: trustedrouter
config:
gateways: [https://api.trustedrouter.com/v1] # the default
defaults: { min_privacy: zdr } # for every model on those gateways
models: # keyed by model id
z-ai/glm-5.3: { min_privacy: confidential, only: [tinfoil, privatemode] }
rules: [] # optional [{ baseURL, body }], deep-merged after the policy
log: false
How a request's provider is built:
- A model's settings are its entry in
models, merged over defaults. A :suffix id such as z-ai/glm-5.3:nitro uses its base model's entry unless it has its own.
- The result is merged over any
provider the request already has, so your policy wins. Lists replace rather than combine.
- Values are type-checked, so a misspelled privacy level or a malformed URL is rejected. Fields the plugin doesn't know are passed through; TrustedRouter rejects unknown options with a 400, so a typo fails loudly rather than being ignored.
Check that it works
- Turn on Log changes to the terminal (or set
log: true).
- Send a chat with a TrustedRouter model. The terminal running
dsh web prints a line like this:
[trustedrouter] added provider to POST https://api.trustedrouter.com/v1/chat/completions for z-ai/glm-5.3: {"min_privacy":"confidential"}
- Optionally, prove the floor is enforced. Set Confidential on a model no confidential provider serves, such as
anthropic/claude-haiku-4.5. The chat should fail with TrustedRouter's 400: No route candidates match the requested provider filters.
Repeat step 2 after every dsh upgrade (see Limitations). The log never includes request bodies, only the URL, the model id and the fields added.
How it works and limitations
- It wraps the dsh server's
fetch. POSTs under a gateway get provider merged into their JSON body. GETs, such as model discovery, and requests to other URLs pass through untouched. A request whose model has no policy is sent byte for byte.
- It fails closed. If a POST to a gateway needs a policy but its body can't be read, the plugin refuses it rather than send it without the policy. The refusal is always printed to the terminal.
- It relies on dsh internals. It works because dsh's model clients use the global
fetch. If a dsh release changes that, requests would go out without your policy, and nothing would error. Check the log line after upgrading dsh. This repo's CI also runs weekly against dsh's latest and alpha releases to catch it.
- Settings are per model, not per chat.
- Web only:
dsh web in a browser, not the desktop app.
Tested with
| dsh | Node |
|---|
0.2.0-rc.2 (npm latest) | 22 |
0.2.1-alpha.1 (npm alpha) | 22 |
dsh is pre-release, so newer versions may move the settings pages this plugin hooks into.
Development
npm install # the schemastery devDependency, for tests
npm test # unit tests, host and client
test/e2e/run-all.sh # headless dsh against a mock gateway; DSH_VERSION=0.2.0-rc.2|alpha|latest
test/e2e/clean-install.sh # install the packed plugin into a fresh dsh home and check both halves
test/ui/scratch-web.sh start <dir> # a scratch dsh web with this checkout linked in, for the browser
There's no build step: index.js (the server half) and client.js (the browser half) are what ships. PLAN.md has the design notes and the dsh internals they depend on.
License
MIT
