The TUI dashboard

oe run --tui opens an ink-rendered terminal dashboard. It is the same render you'd build with tail -f on the event log, except live, structured, and per-node.

What you see

OpenExpertise run — running  · Σ in=2847 out=412

▶ fetch_diff      [collect]  · running …
▶ seed_dimensions [collect]  · running
▶ bug_review      [review]   · calling claude-sonnet-4-6  · 1024/187
▶ verify_finding  [verify]   · validating structured output  · 823/189
· score           [score]

Anatomy:

• Header line — overall run status (starting / running / finished: success) plus run-total tokens (Σ in=<input> out=<output>).
• Per-node row — status glyph, node id, phase tag, current activity string, accumulated tokens.

Status glyphs

Glyph	Status	Color
·	pending	gray
▶	running	cyan
✓	done	green
✗	failed	red
–	skipped	yellow

Activity strings

Each LLM-touching dispatcher emits a node.activity event at meaningful transitions. The TUI shows the latest activity per node:

Provider	Activities you'll see
agent (LLM)	calling <model> → validating structured output / parsing text output
skill (SKILL.md)	calling <model> (skill: <name>) → parsing text output
cli-agent (CLI process)	spawning <provider> (timeout <ms>) → parsing JSON output / parsing text output
tool	(none — tools don't emit activity events because they're synchronous deterministic code)

Tokens

node.tokens events carry { input_tokens, output_tokens, model }. The TUI accumulates per node (since one node may make multiple LLM calls via for_each or retries) and adds them to the run-level total in the header.

CLI agents (Claude Code, Codex, Gemini) don't expose token usage, so their nodes only show activity — no token counter.

Trade-offs and notes

• The TUI buffers output. While it's running, normal pino log lines do not appear in your terminal — they still flow into the JSONL event log on disk. Run oe inspect <run-id> after the fact to see the full structured log.
• The TUI flushes on completion (~100ms after run.finished) and unmounts. The final state is the last visible frame.
• For very long runs with hundreds of nodes, the TUI will scroll past your terminal height. The full event log is the source of truth.
• The TUI is implemented in @openexpertise/tui (ink + React). The state model is a pure reducer (packages/tui/src/reducer.ts) — easily unit-testable, and easy to swap out for an alternative renderer.

Programmatic use

If you're running OE from your own Node code (not through the CLI), call startTui directly:

import { startTui } from '@openexpertise/tui'
import { runExperience, EventBus } from '@openexpertise/core'

const events = new EventBus()
const tui = startTui({
  events,
  nodes: spec.graph.nodes.map((n) => ({ id: n.id, phase: n.phase })),
})

try {
  const result = await runExperience({ spec, experienceDir, dispatchers, events, args: {} })
  // wait for one render cycle so the final 'finished' state lands
  await new Promise((r) => setTimeout(r, 100))
} finally {
  tui.unmount()
}

Future enhancements

• oe inspect --tui to replay a finished run in the same dashboard.
• Pinned summary widgets (cost estimate, retry count).
• A "drill into a node" view showing that node's emitted activity timeline.
• Pause / step / drop into REPL.

Not in V1.

→ Continue with Hand-writing experience.yaml or jump to oe ultra — LLM authors for you.