[All Platforms][Agent&Skills] TUI Ctrl+C re-launches openclaw startup banner instead of dropping to sandbox shell

[mercl-lau]

Description

NemoClaw v0.0.70 regression: after sending two Ctrl+C signals in the openclaw TUI chat session, the TUI re-shows the startup banner and config-warnings screen instead of exiting to the sandbox shell prompt. The first two sub-tests (TUI opens and responds to prompts, session continuity) pass — only the exit path is broken. This is a 7/7 runner failure (all platforms), regression from v0.0.69 where the same test passed on all runners. The openclaw version in sandbox is 2026.5.27 (27ae826).

Environment

Device:        All CI runners
OS:            Ubuntu 24.04 / Ubuntu 22.04 / macOS 26.5 / Windows 11 WSL2
OpenShell CLI: v0.0.70 (openshell 0.0.44)
NemoClaw:      v0.0.70
OpenClaw:      2026.5.27 (27ae826)

Steps to Reproduce

1. Install NemoClaw v0.0.70 and onboard a sandbox
2. Open TUI chat: nemoclaw my-assistant chat
3. Send a prompt — agent responds correctly
4. Press Ctrl+C twice to exit TUI

Expected Result

TUI exits and drops to the sandbox shell prompt (sandbox@...[$#]).

Actual Result

TUI re-launches — startup banner and config-warnings screen re-appears instead of exiting to shell. 7/7 runners fail identically.

Test assertion: step[9] TIMEOUT waiting for /sandbox@.*[$#]/i
Error: "Clean exit flow did not complete."

Regression

• v0.0.69 pipeline #56134184: T5882255 PASSED on all 7 runners
• v0.0.70 pipeline #56299521: T5882255 FAILED on all 7 runners

---

NVB#6395790

[cjagwani]

Thanks for the detailed repro and the clean v0.0.69→v0.0.70 bisect — that made this quick to run down.

I checked the NemoClaw side thoroughly, and nothing in our OpenClaw path changed between the two tags: the pinned OpenClaw build is byte-identical (OPENCLAW_VERSION=2026.5.27, same integrity hash — no version bump), and Dockerfile, generate-openclaw-config.mts, sandbox-init.sh, nemoclaw-start.sh, blueprint.yaml, and the OpenClaw manifest are all unchanged (same baked skills/plugins). The connect/exit code is identical too. So there's no NemoClaw config, startup-latency, or exit-path change that would flip this.

What's actually happening: the check exits via OpenClaw's two-Ctrl+C path, which only completes when both Ctrl+C land within a 1000ms window (resolveCtrlCAction) — the first press arms and re-renders the "press ctrl+c again to exit" banner, and if the second lands outside that window it reads as a re-launch, exactly what you captured. Since OpenClaw and our sandbox setup are unchanged, the flip has to be a timing change in the pipeline harness or the CI runner environment between the two runs, which we can't see from the repo.

If it keeps flaking: OpenClaw already ships a deterministic, non-racing exit in the build we pin (2026.5.27 / 27ae826) — SIGTERM/SIGHUP → clean exit (src/tui/tui.ts:1631 / :1634), plus a documented Ctrl+D. Having the exit step use that instead of the two Ctrl+C is the reliable route, independent of timing. (Confirmed with upstream — openclaw/openclaw#98641, closed as already-implemented via PR #75381.)

Flag us if it comes up again and we'll dig in with you — and thanks again for the clean report.