DSH HUB
HomePlugin StorePlugin PacksCommunityRankingsResourcesPublish Guide
Plugin source
Back to catalog

ShawnHanXiao /

ShawnHanXiao/opendesign-reasoning-and-monitor

Verified

OpenDesign in DeepSeek Harness: fills in reasoning-effort levels for OpenDesign models and shows the 5h / 7d quota as a pill in the conversation header.

★ 0 Stars0 Forks0 IssuesN/A Community rating0 Confirmed installs
View on GitHub
READMESource: main@1af172a2

opendesign-reasoning-and-monitor

中文说明

Two things DSH is missing once you add OpenDesign as a custom model API, in one plugin and with nothing to configure:

  1. Reasoning levels that actually work. Fills in the reasoning-effort map for every OpenDesign model, so picking a level in the model menu sends a request the gateway accepts instead of a 400.
  2. A usage pill. 5-hour, weekly and billing-period usage — used percentages — in the conversation header, shown only while the session's model comes from OpenDesign.

Install it and it looks for an OpenDesign provider profile on every boot. No OpenDesign provider, no effect: it writes nothing and renders nothing.

Why the reasoning levels need repairing

DSH's llm-pi-ai provider profile carries a reasoningEfforts map that turns a chosen level into the spelling the upstream API wants. A profile typed into the "custom model API" form has none, and pi-ai's DeepSeek thinking branch tests if (options?.reasoningEffort) for truthiness, so a level missing from the map falls through to the "no reasoning" request shape — reasoning_effort: "off", which OpenDesign's gateway rejects with a 400 for most models.

What this plugin writes, per model:

reasoningEfforts:
  off: minimal     # the wire value for "off"; empty is what 400s
  low: low
  medium: medium
  high: high
compat:
  maxTokensField: max_tokens
  supportsStore: false
  supportsDeveloperRole: false
  supportsStrictMode: true
  requiresReasoningContentOnAssistantMessages: true
  thinkingFormat: deepseek

glm-5.3-flashx gets three levels, not four: its gateway does not accept medium, and offering it in the picker would only produce a 400.

Three rules keep this a repair rather than a rewrite:

  • reasoningEfforts: false is treated as your explicit refusal and left alone.
  • A model this release has no tested levels for is left without a picker rather than given invented ones.
  • compat is filled key by key; anything you set differently survives. maxTokens and contextWindow are never touched.

The repair runs once per boot, computes nothing on the second run, and every write goes through the settings seam's path edits — never a whole-section replace, which would drop redacted secrets.

The usage pill

5h 12% · 1w 34% · 1m —

Right side of the header row that carries the agent-preset chip, or nothing at all when the session is not on OpenDesign. Hover for the percentages with the used/limit counts underneath, the pay-as-you-go balance, the plan tier, the period end and the read time. Warning colour at 75%, error colour at 90%. The pill degrades as the window narrows, down to hiding itself.

The third slot is the billing period. It stays empty in practice: OpenDesign's control plane reports only the 5-hour and 7-day coding-plan windows, and the monthly maximum exists only in the published plan table with no counter behind it — every other OpenDesign plugin in the market hits the same wall. The card therefore says so instead of inventing a denominator by adding the rolling windows up, which would read as headroom nobody has. Usage is always used/limit — the direction the vendor's own console draws.

The upstream request is tracked as nexu-io/open-design#8837 (still open). The maintainers' answer is that the console already renders whatever windows the billing preflight returns, so a monthly row needs an entitlement and reset-semantics decision in the billing policy rather than a missing UI row. The day codingPlan.windows in billing/preflight carries a durationSeconds: 2592000 entry, this plugin lights the third slot up with no change at all — the windows are read, never hard-coded.

Where the numbers come from, and why you may need to connect an account

Billing lives on OpenDesign's control plane (amr-api.open-design.ai), not on the inference gateway. The sk-amr-… runtime key that DSH stores for inference is refused there (401); only a vela_ctrl_… control key can read account usage.

Two sources, in order:

  1. ~/.amr/config.json (or $AMR_HOME/config.json), which OpenDesign's own CLI and desktop app write when you sign in.
  2. This plugin's device-authorization flow — click the pill, open the returned URL, confirm the code. The resulting control key goes into the credentials store, so it survives restarts.

Nothing is read, sent or stored unless you either already have the vendor's sign-in or complete that flow yourself.

Install

dsh plugin --profile <profile> add opendesign-reasoning-and-monitor

Then refresh the browser page. No restart is needed for the host half on a running DSH that reloads its profile; the pill's stylesheet and slot appear with the page itself.

What it does not do

  • No settings page, no toggles: on when an OpenDesign provider exists.
  • No network calls to OpenDesign except the quota reads the pill needs, and those never leave your machine's credentials.
  • No writing of credentials to config files; the control key lives in the credentials store.

Tests

npm test

The pure half — the repair planner, the window maths, the payload parsers — is covered there. The seams that need a live harness (settings writes, the credentials store, the browser slot) are exercised by hand.

Licence

MIT.

—/ 5

No ratings yet

Verified DSH bundle

Commit 1af172a2ab56

Community comments

No comments yet. Be the first to write one.

DSH HUB

A community index for DSH plugins. Not an official GitHub or DeepSeek AI product.

CommunityResourcesAPIAbout