dsh-model-toolkit

Model Toolkit is one DSH (DeepSeek Harness) plugin that does two things: it lets you declare the capabilities of your hand-added models (image input, reasoning effort, system-message role) from a settings section, and it shows a provider badge with a hover panel and quota lookup next to the model seat in the input bar. It is the merge of two previously separate plugins, dsh-model-fit and dsh-provider-info, into a single install.

中文文档：README.zh.md


Package	`dsh-model-toolkit` (also published as `@cowwo/dsh-model-toolkit`)
Version	`0.1.0`
Plugin id	`model-toolkit`
Type	DSH Web profile bundle (a regular plugin, not a dynamic debug plugin)
Runtime deps	`zod@^4` — a real dependency: `lib/typert.host.js` imports it directly, and the DSH typert loader requires a zod v4 schema (it checks the internal `_zod` marker).
Peer deps (optional)	`@deepseek-ai/cordis`, `@deepseek-ai/schemastery`, `@deepseek-ai/dsh-typert-protocol`, `@deepseek-ai/dsh-client-connection`, `@deepseek-ai/dsh-client-runtime`, `@earendil-works/pi-ai` — supplied by the DSH installation at runtime; every range is `*` so the plugin is not pinned to one DSH version.

What it does

1. Capability editing — Settings → 模型能力管理 (Model capabilities)

DSH providers come in two flavours: built-in ones ship a full catalog (with reasoning levels and image support), while hand-added ones only have a provider id, a base URL, a protocol and a model list you type yourself. Those hand-added models cannot be given reasoning levels or image support from the native Model settings page. This half writes that missing information straight into the llm-pi-ai settings namespace, where the pi-ai adapter reads it at request time, so images and reasoning actually work.

The section shows a flat list of all models of all providers (each card carries its provider badge), with:

Image input — a checkbox per model; on = input: ['text','image'], off = input: ['text'].
Reasoning effort (the UI calls it 思考等级 / thinking level) — level pills (off / high / max, plus minimal / low / medium / xhigh behind "more…"). The wire value of a selected level is editable in place, so you can map a level to the exact string your upstream expects.
Inherit from… — pick a model from the pi-ai catalog and copy its image modalities and reasoning levels (including wire values) onto the current model. compat is deliberately not inherited: catalog entries mix in fields DSH only allows the catalog itself to set, and writing one of those makes the settings layer reject the whole save. Your model's existing compat is left untouched.
Batch actions — turn image input on for all vision/omni models, give every still-unconfigured model off/high/max, or clear capabilities (see Known limitations below).
Search / filter — by model name, model id or provider name; plus "only configured".
Route-level system message role — a per-route dropdown that writes providers.<providerId>.compat.supportsDeveloperRole. Use force system when an upstream only accepts the system role and DSH would otherwise send OpenAI's developer role (symptom: intermittent HTTP 422 such as unknown variant 'developer'). "Auto" removes the field and lets pi-ai infer it. Only protocols that accept the field (openai-completions, openai-responses, azure-openai-responses, openai-codex-responses) are editable; other routes are greyed out with the reason.
One save for everything — all changed providers are committed as a single settings.mutate call with one expectedRevision, so a save either lands completely or not at all. A green/red banner reports the real outcome (a failed write never claims "saved").

2. Badge & quota — left of the model seat, and Settings → 提供商信息

The input bar has a "model seat" showing the currently selected model. This half renders a small badge to its left (via the official slot conversation.input.right, with a DOM-injection fallback for older DSH builds) showing where that model comes from — the provider's display name, falling back to the provider ID.

Hovering the badge opens a read-only popover with:

the provider block — 5 rows: display name, Provider ID, API protocol, API base URL, key environment variable name;
the current model block — 9 rows: model ID, display name, description, reasoning levels, current reasoning level, context window, max tokens, input modalities, compat info (missing values show "not provided");
the quota block — the current provider's remaining balance or plan limits, with a Refresh button (bypasses the cache; queries are read-only and never bill or consume tokens).

The settings section Settings → 提供商信息 contains the six interface settings (see below), the two feature switches, and an all-providers quota table with the columns provider | 5h | week | month | balance | expiry | action. Both the table and the popover render the same data with the same labels; a provider that has no data for a column simply leaves it empty.

Original six settings (persisted in the plugin's own JSON file):

Setting	Default	What it does
Auto refresh on hover (显示悬浮窗自动刷新)	on	Re-query the latest quota every time the popover opens, bypassing the cache. Off = use the cache.
Scheduled refresh (定时刷新)	off	Re-query at a fixed interval while the popover is open.
Refresh interval, minutes (定时刷新间隔)	5 (min 1)	Interval for the above; disabled/greyed out while scheduled refresh is off.
Show more details (显示更多信息)	on	Also show amounts (used/total) on window rows, reset countdowns in the popover, and days-left next to the expiry date.
Font size (字体大小)	Medium	Scales the popover only (small / medium / large).
Language (界面语言)	Follow system (dsh)	Language of the popover and this settings section. Forcing a language here does not change DSH's own language.

Feature switches

Both switches live in the Settings → 提供商信息 section and are stored in the plugin's own JSON file. They exist because the merge left a single browser plugin entry: you can no longer disable half of the plugin from the DSH plugin page, so the plugin provides its own toggles.

Key	Default	When turned off	Where to toggle	When it takes effect
`capabilityEditor`	`true`	The 模型能力管理 settings section is no longer shown, and the plugin no longer writes to your model configuration.	Checkbox in Settings → 提供商信息	After a page refresh
`badgeAndQuota`	`true`	No provider badge, no hover popover and no quota queries in the input bar. The 提供商信息 settings section still opens (that is where the switches live).	Checkbox in Settings → 提供商信息	After a page refresh

These are feature toggles, not plugin switches: the plugin itself stays loaded either way.

Install

The plugin is a profile bundle: installing it edits the profile's package.json and its dsh.profile.bundles list. Adding or removing the bundle therefore needs a restart of dsh web (bundle composition happens at startup). Changes that only touch client code need nothing more than a page refresh.

Option 1 — mount a local directory with a `link:` spec

This is what a local checkout uses. Add the dependency and the bundle entry to the web profile:

// ~/.dsh/profiles/web/package.json
{
  "dsh": {
    "profile": {
      "bundles": [
        // …
        "dsh-model-toolkit"
      ]
    }
  },
  "dependencies": {
    // …
    "dsh-model-toolkit": "link:/path/to/dsh-model-toolkit"
  }
}

Then install the link and restart:

cd ~/.dsh/profiles/web
npm install          # or: pnpm install
# restart dsh web

Option 2 — install from npm

# bare name
dsh plugin --profile web add dsh-model-toolkit

# or the scoped name (same package, same version)
dsh plugin --profile web add @cowwo/dsh-model-toolkit

# restart dsh web, then refresh the page

Uninstall / restore

dsh plugin --profile web remove dsh-model-toolkit
# restart dsh web: the plugin row and both settings sections disappear

Removing the plugin does not delete the capability data you wrote: it stays in the llm-pi-ai settings namespace and keeps working. To clear it, use "clear capabilities" and save before uninstalling (see the warning below), or edit the llm-pi-ai configuration by hand.
The badge/quota half keeps its settings in $DSH_HOME/dsh-provider-info.json; deleting that file resets this half to its defaults.

Where the data lives

The merge deliberately keeps two storage locations — no user data had to move, so there is no migration step (see docs/adr/0002-two-persistence-stores.md).

Half	Where it writes	What is stored
Capability editing	DSH's `llm-pi-ai` settings namespace	Model level: `providers.<providerId>.models[]` — `input`, `reasoningEfforts`, `compat`. Route level: `providers.<providerId>.compat.supportsDeveloperRole`. Written through the official `settings.mutate` path, so it takes effect immediately and is the same data the native Model settings page reads.
Badge & quota	The plugin's own file `$DSH_HOME/dsh-provider-info.json` (default `~/.dsh/dsh-provider-info.json`, written with mode `0600`)	This half's settings only: the six interface settings above plus the two feature switches. The filename is intentionally unchanged from the standalone `dsh-provider-info`, so existing users keep their settings. Quota responses live in memory caches, not in this file.

Network access and credentials

The badge/quota half performs outbound HTTP requests to vendor APIs to read your remaining balance or plan usage. It knows three vendors:

Vendor	What is read
DeepSeek	Account balance (multi-currency).
OpenCode Go	Rolling 5-hour / weekly / monthly usage windows (percentage, reset time, amount converted from the plan price).
Command Code	5-hour / weekly windows plus a monthly pool, and the subscription expiry from the billing cycle.

Providers that are identified but not supported still show "not supported for queries"; a provider without a configured key shows "no API key". All of these requests are read-only: they do not bill and do not consume tokens. Failures are classified (network unreachable / query timeout / 429 / server error / HTTP status) instead of collapsing into a single "query failed".

When queries happen:

Hovering the badge with auto refresh on hover (default on) issues a fresh query each time the popover opens, bypassing the cache. Turn that setting off to serve the cached value instead.
Opening Settings → 提供商信息 triggers one round of queries for the all-providers table (transient failures are retried once automatically after ~1.8 s; deterministic errors such as a missing key are not).
Scheduled refresh (default off) re-queries while the popover is open, at the configured interval.
The Refresh button in the popover always bypasses the cache.
The host caches successful per-provider results for 5 minutes; failed results are only de-duplicated for 20 seconds, so one network hiccup cannot pin an error on screen for minutes. The client-side balance cache carries a TTL as well, so turning automatic refresh off cannot leave a stale number on screen indefinitely.
Transient failures (ECONNRESET, timeouts, 429, 5xx) are retried once by the host; 401/403/404 are not retried. [provider-badge] 余量查询失败 { provider, family, error } is logged to the host console.

There is no global master switch for the network. If you want the plugin to stop talking to vendors altogether, turn off badgeAndQuota: that disables the badge, the hover popover and the quota queries as a whole.

How the credential value is handled (accurate statement):

The browser never receives the key value. The frontend only ever learns the name of the environment variable (the "key env var" row).
The host process does read the key value — it calls ctx.credentials.resolve("<ENV_VAR_NAME>") — and uses it only in the outbound HTTP request to the matching vendor.
The key value is never written to any file by this plugin, and never returned to the frontend.

This corrects an earlier claim in the standalone dsh-provider-info README that the key value was never read at all. The host has to read it to talk to the vendor.

Known limitations

"Clear capabilities" deletes globally and cannot be undone. The button removes input, reasoningEfforts and compat for every model of every route at once — including fields you wrote by hand, such as thinkingFormat or maxTokensField. There is no confirmation prompt, and once you save, the change is irreversible. This is a known risk; the plan for phase 2 is the minimal fix of adding a confirmation step.
The two halves are still two separate settings sections. The merge put them in one package and one plugin entry, but 模型能力管理 and 提供商信息 have not been merged into a single UI.
One browser plugin entry means half of the plugin cannot be disabled on its own. Use the capabilityEditor and badgeAndQuota switches above instead. Note that they are plugin-internal settings, not DSH plugin switches — the plugin page still shows one plugin and one switch.

Relationship to the old plugins and migration

dsh-model-toolkit 0.1.0 is the merge of:

Old plugin	Version merged	Half it became
`dsh-model-fit` ("模型能力管理")	`0.2.3`	Capability editing
`dsh-provider-info` ("输入框显示提供商")	`0.17.0`	Badge & quota

The merge is zero migration — no data is moved:

Remove the two old plugins from the profile (dsh plugin --profile web remove dsh-model-fit and … remove dsh-provider-info, or edit dsh.profile.bundles by hand). Do not run them side by side with this package.
Install dsh-model-toolkit as described above and restart dsh web.
Your capability data keeps living in the llm-pi-ai settings namespace, and the badge/quota half keeps reading ~/.dsh/dsh-provider-info.json (the file name did not change), so its six interface settings are carried over untouched. Both feature switches default to true, which matches having both old plugins installed and enabled.

Development

There is no build step. client/client.js is a hand-written, pre-built bundle committed to the repository (in DSH's __ModuleLoader__ format), and lib/index.js is plain ESM. Edit the source and it takes effect — there is no bundler to run and the package declares no npm scripts.

dsh-model-toolkit/
├── package.json           # package metadata + dsh.bundle.patch + dsh.client declaration
├── cordis.patch.yml       # inserts the plugin row `model-toolkit` into a profile
├── lib/
│   ├── index.js           # host: ModelCapabilityService + ProviderBadgeService, one entry point
│   ├── catalog.js         # catalog helpers
│   └── typert.host.js     # the merged typert host manifest (5 endpoints)
├── client/
│   └── client.js          # one pre-built browser bundle, two internal scopes
├── tools/                 # development-time verification scripts (not shipped)
├── docs/adr/              # architecture decision records for the merge
└── GLOSSARY.md            # the domain vocabulary used by this project's docs

One package, one plugin entry. One npm package, one plugin row (model-toolkit), one browser entry. lib/index.js mounts both services: ModelCapabilityService and ProviderBadgeService.
One typert manifest, five endpoints: modelCapability/source, providerBadge/modelInfo, providerBadge/balance, providerBadge/settings, providerBadge/providers.
One client bundle, two independent scopes (module 1 = capability settings page, module 2 = badge/quota). The type symbols in the manifest keep their original dsh-model-toolkit#… prefixes.
UI integration points: settings sections model-fit (order 12) and provider-info (order 13), and the badge slot conversation.input.right (order 15, with a DOM-injection fallback when the slot is not available).
Verification scripts live in tools/ and run with plain Node. They are development-time only, are not listed in the package's files, and are therefore not published. They never touch the network and never write your real $DSH_HOME. tools/README.md is the authoritative list of what each script checks; run the whole offline suite from the package root with:
```
node tools/verify-all.mjs
```
Individual scripts can be run on their own, for example node tools/verify-format.mjs (rendering strings and the hover state machine, extracted from the real client/client.js) or node tools/check-typert-manifest.mjs . dsh-model-toolkit (the typert manifest contract, 5 endpoints). Outbound verification is deliberately kept outside this directory: the old repository's verify-quota.mjs would call the vendor APIs with a real API key and needs explicit consent, so it is not part of the package.

dsh-model-toolkit

설치

README