Dashboard

cue ui serves a local dashboard for the current repo. It is localhost only and has no authentication.

cue ui          # http://127.0.0.1:4224
cue ui 5000
cue ui --no-open

It is the same pipeline as the CLI. GitHub labels and .cue/runs/ stay the shared state either way.

Overview

• Total agent spend, cost per pipeline stage, cost per issue
• Cumulative spend trajectory
• The agent:* label board (active issues only — agent:done is omitted on purpose)
• A live log streamed over SSE while stages run

You can trigger process and run from the dashboard; they execute the same TypeScript as the CLI. The header Process now button is cue process (every actionable issue). Open the chevron to pick one issue by title — the same set cue run lists (agent:ready, agent:approved, agent:replan) — and the button becomes Run #n. Clear the selection to go back to processing the whole queue.

Issues in the agent:planned column carry Approve and Replan buttons — the same human gate, without leaving the board. Approve swaps the label to agent:approved and starts the dev run immediately (if a run is already in progress, the approval sticks and the next process picks it up). Replan expands an inline feedback box; the feedback is posted as an issue comment (where the replan stage reads it) before the label swaps to agent:replan.

Issues in the agent:failed column carry a Retry button (also shown on the issue's run page). Retry re-queues the right stage deterministically: back to agent:replan if a plan revision was pending, to agent:approved with a fresh worktree if a plan exists, or to agent:ready for a re-triage otherwise.

Set webhookUrl to also get a POST notification the moment a plan awaits approval or a draft PR awaits merge — useful when nobody is watching the board.

Runs

A transcript explorer over .cue/runs/, split into:

• Active — issues still on the label board
• Done — recorded runs whose issue has left the board (merged, closed, or agent:done)

Pick an issue, pick a recorded stage run, then read:

• the exact prompt that was sent
• the flattened event transcript (tool calls, thinking, results)
• the raw log entry

Denied tool calls are surfaced explicitly. A stage's --allowedTools allowlist is a common cause of odd agent behaviour.

Finished work is invisible on the label board by design. Anything listing runs reads the on-disk issue index, so archived and even deleted issues stay browsable with no gh call.

Developing the dashboard

The SPA lives in ui/ (react-router in library mode, React Compiler, shadcn base-nova). The CLI serves the production build from ui/build/client.

bun run ui:build     # required after changing ui/app/
bun run ui:dev       # Vite on :5173, proxies /api to a running cue ui
bun run fixtures     # snapshot local .cue runs so the SPA renders without the API

Contributor notes: Contributing.