Skip to content

feat(server): dashboards call your data tools while they are built - #1678

Draft
Aditya-thesys wants to merge 15 commits into
mainfrom
ap/dashboard-create-script
Draft

Aditya-thesys wants to merge 15 commits into
mainfrom
ap/dashboard-create-script

Conversation

@Aditya-thesys

Copy link
Copy Markdown
Contributor

The server SDK side of the dashboard agent (spec: thesysdev/muse#672; server: thesysdev/coda#1110, thesysdev/muse#673).

What it does

  1. create() and edit() build on POST /v1/dashboard/chat/completions. When the model asks for data, the SDK runs your data tools and sends back only the new results with the dashboard's artifact_id, round after round, until the dashboard starts. The string they return is unchanged.
  2. create_script is handled for you: when the model saves a data script, the SDK runs it through /v1/app/execute (a draft run of the build in progress) against your tools and returns its output to the model.
  3. onProgress on create/edit reports each tool call, each script the model checks, and the dashboard starting; describeDashboardProgress turns a step into a short label.
  4. defineDashboard({ reasoningEffort }) sends reasoning_effort with every round.

Release

  • Minor (0.5.0): it needs an OpenUI Cloud with the new route, and create/edit now call your tools during the build and take longer.
  • Order: deploy coda#1110 and muse#673 first, then publish this.

Tested

  1. Unit tests for the round loop, draft runs, errors and timeouts.
  2. End to end against the local stack with coda#1110 and muse#673: create, edit, the fix flow, tool and script rounds, progress events, on DeepSeek, Sonnet 5.5, GPT-6 Luna and Haiku 5.5. Benchmarks and render checks are in muse#673.

The SDK sends the dashboard's data tools as function tools. When OpenUI Cloud
answers with tool calls, the SDK runs them on your server and posts the results
back, until generation starts. Errors, unknown tools and timeouts go back to
the model as tool results. Set probe: false to turn this off.
When the model calls create_script, the SDK runs that saved script by name
through OpenUI Cloud, as a dashboard runs its scripts, with your tools, and
sends back its output or error. The build's draft_id from the first tool-call
reply goes back with every later request. No round limit: OpenUI Cloud decides
when the model stops asking for tools.
Create and edit post to the dashboard route, OpenAI chat completions in and
out: a reply with tool_calls is run and posted back with its draft_id; a
reply that ends with stop carries the tool response for your agent's model.
…; OpenUI Cloud adds create_script; no probe option
… are internal; tool-round naming

- defineDashboard({ reasoningEffort }) sends reasoning_effort with every round
- the interpreter's draft runs (create_script) are an internal runDraft, not a
  symbol on the public execute; the public interpreter is execute only
- tools/probe.ts is tools/tool-round.ts (runToolRound); a tool-call reply with
  no artifact_id is an invalid_response
@vercel

vercel Bot commented Nov 19, 2026 •

Copy link
Copy Markdown

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment
Project Deployment Actions Updated
openui-docs Skipped Skipped Nov 19, 2026 8:11pm UTC

Request Review

The dashboard route always runs its tool loop, so requests no longer send
tools; the data tools go in the config block. Data tools run through the
interpreter's own timeout and abort handling. Removed checks for things
OpenUI Cloud always sends, and the wrapper that ignored errors thrown by
onProgress.

This branch was previously deployed

1 inactive deployment
Preview — 6fb232b2 Deployed Nov 19, 2026 by vercel[bot]
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant