Add your own parameters to every model request dsh sends. OpenRouter provider routing, fallback models, reasoning effort, sampling extras, gateway headers, or any JSON field a gateway accepts. Pick one of the channels dsh already has, add parameters on a page instead of in code, see exactly what each request will carry before anything is sent, and see what the gateway answered afterwards.
dsh sends the fields its model adapters know about. Gateways such as OpenRouter take more: which providers to use and in what order, fallback models, a reasoning budget, a session ID for cache affinity. dsh does not send OpenRouter's provider routing preferences. Plugins exist for single slices of this (provider order and quantization, for example); this one is general. The same page handles any field or header for any gateway, with a ready form for OpenRouter's routing options, a live preview, and a log.
Each parameter belongs to a channel (or names its own hosts, models, and paths) and says what to add, replace, or remove in the body and headers. Parameters are applied to the request just before it leaves dsh; in the rules file each one is a rule.
Highlights
Organized by channel. The page lists the model channels dsh is set up with (the official DeepSeek channel and each one configured in llm-pi-ai) and the address each one sends to. Open a channel to see its parameters and add one in a click; it is matched to that channel for you. Parameters for any other address are under Other addresses.
Knows what each channel reads. Each channel shows which request format it reads (OpenAI or Anthropic) and which extra fields (OpenRouter's provider and reasoning, vLLM's chat_template_kwargs, …), worked out from dsh's settings and its address. The kinds of parameter it reads come first; the rest wait under Other kinds. A gateway of your own that speaks OpenRouter's protocol gets provider routing once you tick OpenRouter fields for it.
See what will be sent.Added when sending shows, as JSON, what the channel's parameters add to a request for the model you pick, including unsaved edits. Nothing is sent.
Pick providers, do not hand-write JSON.Provider routing is one list: search providers (the ones that serve the previewed model come first, with price and uptime), put them in order, decide whether other providers may be used, exclude some, sort, require zero data retention, add fallback models; price caps, quantization, an allow list, and speed preferences are under More conditions. A provider that does not serve the model is flagged, and a list that would fail with "No endpoints found" is called out before you save.
See what happened after. Each channel shows its recent requests: the status, what was added, and, when the gateway refused, what it said. Request log has every model request dsh sent.
Safe by default. Core fields (messages, tools, model, …) and credential headers (Authorization, x-api-key, …) cannot be changed. Secrets come from environment variables ({{env:NAME}}) and are never shown. If something goes wrong while applying a parameter, the request goes out unchanged instead of failing.
One hook for the usual channels. The official DeepSeek endpoint and the OpenAI-compatible and Anthropic-compatible channels of llm-pi-ai (OpenRouter, a self-hosted vLLM, an enterprise gateway) all go through the same fetch hook; see Compatibility for what was verified.
Updates in one click. The page tells you when a newer version of this plugin is out and installs it with DSH's own installer after you confirm.
Everyday comforts. A switch pauses everything at once. Unsaved edits survive leaving the page, and Ctrl/⌘ + S saves. Deleting can be undone. A rules file changed elsewhere is detected instead of overwritten. Parameters export to and import from JSON. English and Chinese, light and dark, down to phone width. If the page itself fails to draw, it shows the error and Reload instead of an empty panel.
If you run dsh from source, use pnpm dsh plugin --profile web add dsh-request-params. Restart dsh web afterwards. A Request Params entry appears in the sidebar below Plugins.
To try a checkout, use link:/path/to/dsh-request-params instead of the name.
Quick start: OpenRouter in two minutes
Route dsh to OpenRouter if you have not yet. In llm-pi-ai, openrouter is a built-in provider; it only needs a key, kept in an environment variable:
Open Request Params and pick the openrouter channel. Click Provider routing, type a provider such as deepseek and press Enter, set Other providers to Never if no other provider may serve the request, and press Save (or Ctrl/⌘ + S). Added when sending shows the provider object; put the model you use in its Model box, and if a provider is flagged as not serving it, pick another provider or model.
Send a message. The channel's Recent requests shows what was added and the gateway's status.
Choose providers and order, cap the price, require zero data retention, fall back to other models
Reasoning effort
reasoning (effort, max_tokens, exclude)
Turn thinking up or down, or keep it out of the reply
Sampling extras
top_k, min_p, top_a, repetition_penalty
Sampling fields the standard API does not have
Session stickiness
session_id in the body and the x-session-id header
Keep one conversation on one provider so caches hit
App attribution
HTTP-Referer and X-OpenRouter-Title headers
Credit your traffic to your app on OpenRouter (dsh does not send these)
Remove fields
removes fields such as store
A gateway that rejects a field dsh sends
Custom fields and headers
anything: JSON fields, headers, removals
Any other gateway
Which kinds a channel offers first follows what it reads (see What a channel reads below). An OpenRouter channel offers every kind; a channel that reads the OpenAI format offers sampling extras, custom fields, session stickiness, and remove fields; an Anthropic channel offers the last three. The examples for other gateways (turn thinking off on a self-hosted vLLM with chat_template_kwargs, Helicone headers, Cloudflare AI Gateway cache headers, and an OpenAI prompt_cache_key) come first on a channel that is that gateway. Everything else is under Other kinds, each marked with what it needs, and can still be added. Each example is an ordinary custom parameter you can edit. Other addresses offers everything.
Under its address, each channel shows its protocol: the request format it reads and the extra fields it reads on top.
Worked out for you: the format comes from the channel's api setting in llm-pi-ai (openai-completions, anthropic-messages, …), from the official DeepSeek channels (Anthropic), from the hosted service, or from the paths of the requests seen. Hosted services are recognized by their address: openrouter.ai (OpenRouter fields), helicone.ai, gateway.ai.cloudflare.com, and api.openai.com.
Yours to correct: Change sets the format and ticks the fields the endpoint reads. The common case is a gateway of your own that speaks OpenRouter's protocol: tick OpenRouter fields, and provider routing, reasoning effort, and app attribution come first on that channel. The setting is saved with the parameters under the endpoint's host and base path; Use what was detected removes it.
What it changes: only which kinds come first and how sure the catalog hints are. Requests are sent as the parameters say, whatever the protocol.
Catalog hints: the provider list and whether a provider serves a model come from the public OpenRouter catalog. For a parameter that goes to openrouter.ai, a provider list none of which serves the previewed model is called out before you save. For any other gateway the same hints are only a reference: use the provider names your gateway expects, typed in if needed.
Writing your own parameter
Custom fields and headers has these parts:
Body fields: a JSON object merged into the body. Objects merge with what is there; anything else is replaced.
Headers: added to the request; a header of the same name is replaced.
Under More: Defaults (like body fields, but a field the request already has is left alone), Remove fields (dotted paths such as store or stream_options.include_usage; removal happens first, then fields are set), Remove headers, and Allow changing core fields.
Every parameter has a scope. On a channel, All models limits it to some of the channel's models (deepseek/*, *-chat), and Edit hosts and paths in its menu shows the address it matches. Under Other addresses a parameter names its own hosts (openrouter.ai, *.example.com, 127.0.0.1:8130), optional models, and optional paths (prefix). Without a path, it reaches model endpoints only (…/chat/completions, …/messages, …/responses, …:generateContent); name a path to reach another endpoint.
When several parameters apply to one request they run from top to bottom, and a later one sees the result of the earlier ones. Values may use {{model}} (the request's model), {{session}} (the dsh session ID; for channels where dsh sends none, a stable key derived from the conversation), and {{env:NAME}} (an environment variable; if it is not set the field is skipped and the log says so).
messages, system, tools, tool_choice, input, instructions, prompt, contents, stream, and model cannot be set or removed unless a custom parameter explicitly allows it.
Authorization, x-api-key, api-key, x-goog-api-key, Cookie, Host, and the headers that frame the request (Content-Length, Content-Type, …) can never be changed.
A parameter with a problem is not applied, and the page shows what is wrong. If applying parameters fails while a request is being sent, the request goes out unchanged and the log says so.
Parameters are stored on your machine in <profile>/.dsh-request-params/config.json. Nothing is sent anywhere because of them.
The request log lives in memory only (the last 50 requests; cleared on restart). It holds the model, the address, and the changes made; for a failed request it also keeps the first 2 KB of the gateway's answer, which can quote the request. It never holds the conversation itself. Values from environment variables appear as "not shown", and are replaced with [env] if the gateway's answer quotes them.
The channel list comes from dsh's own channel list and settings (with secrets redacted), to show each channel's address and models; the page changes nothing there and stores nothing.
The page reaches the network for two things. OpenRouter's public catalog (/providers, /models, /models/{id}/endpoints) is read only while the page shows a provider routing parameter: the provider list, the providers that serve the previewed model, and fallback model suggestions; the provider list also works offline from a built-in copy. The update check asks the npm registry DSH is configured with for this plugin's latest version (see Updates). Neither sends a key or any content.
The rules file
The page reads and writes <profile>/.dsh-request-params/config.json. You can also edit it by hand: the next model request uses the new rules, and an open page loads them within a few seconds (or, if it has unsaved edits, asks what to do).
revision counts saves. A page that saves over a newer revision is stopped and asked what to do.
A rule may also carry note (free text) and preset ({ "id", "form" }, the form the page reopens). A preset that no longer builds exactly this rule's body and headers is dropped, and the rule stays as a custom rule.
Fields the page does not know, and values of the wrong type, are dropped when the file is read. A file that is not valid JSON is reported on the page and left untouched; Move it aside and start fresh renames it.
protocols holds what you set each endpoint reads, keyed by host and base path: format (openai, anthropic, or other) and extensions (openrouter, vllm, openai, helicone, cloudflare), each true or false; false removes one that was detected. Only the page reads it.
Export writes the same object without revision. Import accepts that object or a bare array of rules; on Add to current parameters, an imported protocol replaces yours for the same endpoint.
How it works
dsh's model adapters send requests through the global fetch. This plugin wraps fetch once. For a POST request with a JSON body that looks like a model call (or that a rule names), it reads the body, applies the matching rules, and sends the changed request; every other request passes through untouched. Responses and streams are never read or changed (only the status, and a failed response's first 2 KB, are noted for the log). Uninstalling the plugin or pausing it puts everything back to what dsh sends by itself.
When a parameter does not seem to apply
Open the channel. Added when sending shows what its parameters add for the model in the Model box. If it says no parameter applies to this model, check the parameter's model scope (the All models button).
Look at the channel's Recent requests, or the Request log. If the request is not there, the model request did not go through fetch (or no message has been sent since the last restart).
Open the request. No parameters means the host, model, or path differs from every parameter. A host without a port matches any port; write 127.0.0.1:8130 to match one.
Nothing to change means the request already had those values.
A note that a parameter has problems or uses an environment variable that is not set says what to fix.
If the gateway answers 4xx, open the request: the fields that were added are listed with the gateway's own message. Pause all parameters shows whether they are the cause. For OpenRouter, "No endpoints found" usually means the providers you allowed do not serve that model; the routing form flags them for the model in the preview.
Compatibility and limits
Needs dsh with the web client and Node ^22.19.0 || >=24. Tested with dsh 0.2.1-rc.2.
Verified against real dsh runs on three kinds of channel: the official DeepSeek endpoint (/v1/messages), an OpenAI-compatible llm-pi-ai channel (/chat/completions), and an Anthropic-compatible llm-pi-ai channel (/v1/messages).
Applies to requests sent through the global fetch. A provider that captured its own fetch earlier would not be seen; the request log shows what the plugin saw.
A gateway may ignore or reject fields it does not know. OpenRouter ignores unsupported parameters unless require_parameters is on.
Limits: 100 parameters; per parameter, 64 KB of JSON each for the body fields and the defaults, and 40 headers.
Uninstall
dsh plugin --profile web remove dsh-request-params
Requests go back to what dsh sends by itself. The parameters stay in <profile>/.dsh-request-params/config.json, so installing again brings them back; delete the folder to clean up.
Updates
The page checks for a newer version of this plugin and can install it.
Check: when you open the page, the plugin asks the npm registry DSH is configured with (a mirror works) for the latest version of dsh-request-params, at most once every 12 hours. The request names only the package. Turn it off with Check for updates automatically at the bottom of the page; Check for updates there and in the ⋯ menu still works. Development installs (link:) are never checked.
Upgrade: a banner at the top shows a newer release with See what changed and Upgrade…. After you confirm, DSH's own installer downloads that release's tarball from the registry and installs it, with progress and Cancel install. The installer restores the profile's files when an install fails or is cancelled, and the banner then shows the installer's message and the command to upgrade by hand. Only the latest release is offered, never a prerelease, and the tarball address must be this package's file on the registry that answered.
After the upgrade: depending on the DSH version, the new version takes over at once (this page closes; open it again) or loads when you restart dsh (web: run dsh web again; desktop app: quit and reopen). The next time you open the page it says Upgraded to the new version. Your parameters stay, and running sessions are not affected. The banner shows the command that goes back to the old version.
Installed from GitHub: the banner shows dsh plugin --profile web add github:xiaohai-78/dsh-request-params instead of a button.
By hand: run dsh plugin --profile web add dsh-request-params@latest, or on the Plugins page use Add plugin with https://registry.npmjs.org/dsh-request-params/-/dsh-request-params-<version>.tgz. Without @latest the command does nothing, because DSH keeps the version it installed.
From 0.1.0: 0.1.0 has no update check; upgrade it by hand once.
Host plugin: wraps fetch, serves POST /api/dsh-request-params (get, save, check, dryRun, preset, recent, clearRecent, resetDamaged, channels, orProviders, orModels, orEndpoints, and the update actions updateStatus, setUpdateCheck, dismissUpgraded, upgrade, upgradeStatus, cancelUpgrade), reloads the rules when the file changes
rules.js
Rule engine: validation, host/model/path matching, templates, applying rules to a body and headers
interceptor.js
The fetch wrapper and the request log
presets.js
Preset forms and the rules they build
config-store.js
Rules file with atomic writes and revision checks
channels.js
Reads the configured channels from dsh's llm and settings services and works out each one's endpoint and protocol
protocols.js
What an endpoint reads: the formats and extensions, working them out from a channel's settings and address, and reading the protocols set in the rules file
update.js
Version check helpers without DSH services: version comparison, how the profile depends on the package, the registries to ask, and the check that the registry's tarball address is this package's file on that registry
self-update.js
The update actions: the check with its 12-hour cache, the switch and the note an upgrade leaves in <profile>/.dsh-request-params/settings.json, and the install through DSH's plugin manager (installBundle, cancelInstall, the plugin-manager/install-state progress)
openrouter.js
OpenRouter catalog client with cache and offline fallback; openrouter-providers.json is the built-in provider list
Tests
npm test
runs the node --test suites in test/. What each feature must do, and how it was verified on real dsh runs, is in docs/requirements.md. Screenshots in this README come from those runs; the OpenRouter catalog in them is OpenRouter's public one, and no request was sent to OpenRouter.
