Skip to content

Ticket #232: real live TUI workspace-wizard timeout — verification ​

  • Date: 2026-11-11
  • Branch: ticket/232-real-live-wizard
  • Base: integration/native-first at 49d2fa2
  • Issue: #232
  • Unit gate: npm run verify (strict typecheck + full test suite)

Root cause ​

The product was correct; the real harness was stale.

Agent and workspace creation open the chat column in the shared pane-selection grammar. n is a chat-interaction key (chat.new.workspace), not a pane-selection key. The harness left the wizard, saw NORMAL │ Direct conversation, and pressed n without first pressing Enter. The focused direct chat was therefore still in select mode, so the workspace wizard never opened. The failure pane showed the select-mode hint h/l column, while the script had asserted only the surface title.

The fix is in the harness: live_enter_chat_interaction presses Enter and requires the interaction hint j/k scroll before any chat key. Enter is safe when chat already owns interaction (chat.enter is a no-op). The same helper is used after the workspace wizard before the composer probes.

Follow-on real-live harness corrections ​

Running the full real path after the wizard fix exposed three more stale harness assumptions, all in the real script rather than product behavior:

  1. The 120-column terminal left the chat cell too narrow for the palette flag assertions (Skill | General was middle-truncated). The real script now uses the same 160-column shell as the fake matrix.
  2. The System Agent init report is longer than the viewport. The harness now asserts the durable report head (Ensemble init complete) and requires the visible report tail (Wrote .ensemble/init.json).
  3. The self-mention $LiveAgent review ping creates a normal assignment, but the assigned row scrolls above the viewport once the answer streams. The harness now proves the durable user row and assignment instead of a viewport-dependent line.

All three keep the real host, isolated HOME/config/project, and deterministic keys; no model output or credential is stubbed. The script still reports a missing model/catalog as an explicit skip.

Regression tests ​

  • src/v2/tui/interaction.test.ts — opening a chat target selects the column; Enter enters chat interaction; a second Enter is a safe no-op.
  • src/v2/tui/keymap.test.ts — n is absent from pane selection and chat.new.workspace is owned by chat interaction.

Results ​

CommandResult
npm run verifyPass — strict typecheck; 1811 tests passed, 0 failed, 0 skipped
npm run test:livePass — live-check passed (http://127.0.0.1:44461); no skip

Before the fix, the same command failed at live: timed out after 10s waiting for workspace wizard on this branch and on the unchanged integration/native-first baseline. After the harness fix, the real host completes agent creation, workspace creation, /init, composer palettes, mention assignment, cleanup, and the return to the native TUI.

#232 can close: the reported timeout is fixed and the real live matrix passes.

Released under the MIT License.