dsh-bailian-catalog

English | 中文

Lists every Alibaba Bailian (DashScope / Model Studio) model the configured credential can reach, inside the DSH Web GUI under Settings → 百炼模型, and adds or removes any of them from the Bailian provider route. It also shows the current session's context quota.

What it does

1. Settings panel: total model count, probed endpoint, refresh time, search, per-family collapsing, and one-click copy of a model id.
2. Add to config: every row has an Add button that writes the model into llm-pi-ai's providers.bailian.models, after which the model appears in the model picker. Clicking it again removes the model. The add is probed first — see Availability probing.
3. Availability probe: a per-row Probe button that sends one minimal chat request and reports whether the model is actually callable, so an unopened model is not mistaken for a bad key.
4. Current session quota: the context window, tokens used, remaining budget and a usage bar for the model the active session is running on.
5. Agent tool: bailian_list_models, callable from any session, with a query substring filter, an exact family filter and an exact vendor (namespace) filter.

Availability probing

A model listed by /models is not necessarily a model you may call. A Bailian workspace that has not opened a model answers 403 AccessDenied.Unpurchased, and DSH surfaces every upstream 401/403 as "invalid API key" — which points at the key instead of at the entitlement. Probing settles it.

POST /api/bailian-catalog/probe sends one max_tokens: 1 chat request and classifies the answer:

State	Meaning
ok	Callable
unpurchased	The workspace has not opened this model, or its quota is exhausted (AccessDenied.Unpurchased)
unauthorized	The key is invalid or expired
not-found	The model does not exist, or the OpenAI-compatible interface does not support it (embedding, rerank)
not-chat	A non-chat model (embedding, image, speech) rejected the messages payload — the route is reachable, but a chat request cannot verify it
rate-limited	Throttled; retry later
timeout / error	No usable answer

not-chat is deliberately a neutral result rather than a failure: probing an image or embedding model with a chat request always returns 400, which says nothing about whether the model is available.

Adding is guarded by the probe. A model in a blocking state (unpurchased, unauthorized, not-found, error, timeout) is not written to the route, and the panel shows why. Writing an unusable model into the route is what turns one bad selection into a session that fails on every turn. not-chat models are allowed through, since a chat request cannot judge them, and the probe result is returned alongside the write.

Grouping

Many endpoints advertise ids without an owner/ prefix — roughly six in ten on the author's endpoint — so grouping by that prefix alone would leave most models in a single bucket. Models are grouped by family instead, derived from id patterns (qwen* → 通义千问, glm* → 智谱, kimi* → Kimi, deepseek* → DeepSeek, minimax* → MiniMax, wan* → 通义万相, and so on), with the owner/ prefix kept as a secondary namespace value. On the author's endpoint nearly every model resolves to a real family; only gui-plus and two ops entries fall through to 其他.

How it works

• Host half (index.js)
  • Reads BAILIAN_API_KEY through the credentials service (credential store first, environment fallback). The key stays in host memory: it is never sent to the browser and never written to disk.
  • Enumerates models through the official llm.discoverModels() path, which requests GET {baseURL}/models with Authorization: Bearer.
  • Quota: reads the most recently active live session, measures used tokens with tokenMeter.measure(), and takes the model and window from session.requestContext(), falling back to llm.resolveModelInfo() when the window is unknown.
  • Add/remove writes through settings.mutate(), the structured profile-patch write path, rather than editing YAML text.
  • Probing posts one max_tokens: 1 chat request with a hard timeout and classifies the answer; error code and message are read from both the error object and the payload top level, since this endpoint uses both shapes.
  • Every route calls connection.requestRejection before doing any work.
• Browser half (client.js)
  • A hand-written ModuleLoader bundle with no build step, depending only on the shell-provided react.
  • Registers a settings.section (id bailian-catalog, order 12) and renders from the same-origin endpoints.

Endpoints

Method	Path	Purpose
GET	/api/bailian-catalog/list	Model list, plus configured (ids already in the route) and families
GET	/api/bailian-catalog/usage	Active session's model, tokens used and context window
POST	/api/bailian-catalog/probe	{ id }, one minimal chat request, classified into the states above
POST	/api/bailian-catalog/add	{ id }, probed first; returns already, blocked, or added
POST	/api/bailian-catalog/remove	{ id }, removes it from the route

Every route asks connection.requestRejection first, so the Host/Origin fence and browser authentication gate callers before any model list or config write is reachable. A no-cookie request gets 401; a cross-site Origin gets 403.

The first settings.mutate write materializes schema defaults (input: [], compat: { chatTemplateKwargs: {}, chatTemplateArgs: {} }) into the patch. They are semantically inert, but they add a few lines to the file.

What "quota" means here

A Bailian sk- key has no balance or quota endpoint (the usual paths — /dashboard/billing/*, /user/balance, /usage — all return 404), and account balance is only reachable through the Alibaba Cloud BSS OpenAPI with an AccessKey. So this plugin reports context quota instead:

• the context window, tokens used, remaining budget and a usage bar for the active session's model;
• states: enough budget, near the ceiling, or exhausted;
• it is measured the same way the in-session compaction decision is, so "exhausted" means that session can no longer continue and a new session restores the budget.

Install

From GitHub:

dsh plugin --profile <profile> add github:kennethfan/dsh-bailian-catalog

From a local checkout:

dsh plugin --profile <profile> add link:/path/to/dsh-bailian-catalog

The host half needs a dsh process restart before its routes exist; the browser half appears after a page refresh.

Prerequisites

You need your own Bailian (DashScope / Model Studio) API key, and — if you use a dedicated Model Studio workspace — that workspace's MaaS gateway URL. This plugin ships no credential and points at no private workspace: the default endpoint is the public DashScope compatible endpoint.

Configuration

Key	Default	Meaning
baseURL	https://dashscope.aliyuncs.com/compatible-mode/v1	Endpoint to enumerate. Point it at your workspace's MaaS gateway when you use a dedicated Bailian space.
provider	bailian	The llm-pi-ai provider key that add/remove writes to.
apiKeyEnv	BAILIAN_API_KEY	Credential reference resolved through the credentials service.
families	[]	Extra family rules (see below).

- insert:
    - id: bailian-catalog
      name: 'dsh-bailian-catalog'
      config:
        baseURL: https://llm-<your-workspace>.cn-beijing.maas.aliyuncs.com/compatible-mode/v1
        provider: bailian
        apiKeyEnv: BAILIAN_API_KEY

A workspace gateway returns the same body shape as the public endpoint; it simply sees whatever models your key is entitled to.

Custom family rules

families adds your own rules to the family table. They are matched before the built-in rules, so a rule can also reclassify something a built-in rule would otherwise claim:

- id: bailian-catalog
  name: 'dsh-bailian-catalog'
  config:
    families:
      - pattern: '^gui-plus$'   # case-insensitive regular expression
        family: GUI
      - pattern: '^sre-gpu'
        family: 运维

A rule is tested against the full model id first and against the leaf name (the part after /) second, exactly like the built-in table.

Invalid input is skipped rather than fatal — a bad regular expression, a missing pattern or family, or a families value that is not an array each produce a warning that the settings panel shows, and the built-in rules keep working.

Known limits

• The endpoint returns the models the current API key can use, not Bailian's full public catalog (180 on the author's endpoint). Listing a model does not imply entitlement — probe it.
• Probing costs one request per click. It is not run automatically for the whole list, since that would be hundreds of calls.
• A chat request cannot judge embedding, image, speech or rerank models; they report not-chat and stay unverified.
• Reads are safe; only add/remove write, and only to the configured provider's models list.
• Reachable from localhost only (the web server binds 127.0.0.1); the route fence assumes that origin model.

License

MIT