[MotionDevLab]

Summary

• what changed: the --resume session picker list now renders inside Ink's <AlternateScreen> (same component the REPL uses), and onCancel unmounts via useApp().exit() before exiting 1 in a stdout-drain callback (500 ms failsafe). Two commits, single file (src/screens/ResumeConversation.tsx, +41/−3).
• why it changed: pressing Esc in the picker exited the process but left the whole picker frame painted behind the shell prompt, looking like a frozen CLI. Root cause: Ink deliberately persists the last inline frame on exit (getRenderOpsForDone in src/ink/log-update.ts returns an empty diff when the cursor is visible), so no flush-based cleanup can ever erase it — only leaving the alternate buffer restores the main screen. An earlier flush-only attempt was verified to fail on a real terminal, which is what motivated this approach.

Impact

• user-facing impact: Esc (and double-Ctrl+C) from the resume picker now returns to a clean shell prompt. Exit code 1 (cancel) is preserved, and no resume hint or SessionEnd hooks run on cancel — same observable behavior as before, minus the residue. Under tmux -CC (where alt-screen corrupts terminal state) the picker stays inline, exactly as before — no regression there.
• developer/maintainer impact: none outside the picker cancel path. CrossProjectMessage / NoConversationsMessage exits intentionally untouched (their post-exit text must persist).

Testing

• I ran the required local preflight (see results/notes below; full check could not complete on this machine).
• exact commands and results:
  • bun install --frozen-lockfile — ok
  • bunx eslint src/screens/ResumeConversation.tsx — clean
  • bun run build — green (dist/cli.mjs, SDK bundle, externals valid)
  • bun run typecheck:type-tests — pass (10 files)
  • node bin/openclaude --version — 0.31.0 (OpenClaude); compile-cache-disabled launcher check — pass
  • bun run test:provider-recommendation — 165 pass / 0 fail
  • live test (built CLI, PTY): --resume + Esc emits ?1049h on mount and ?1049l + full terminal-mode cleanup + EXIT:1, fast, no hint noise
  • local LLM gate (laya, english ckpt): ship plurality, regression risk low (1.2), evidence sufficient (0.86) — advisory only
• focused tests: user-confirmed in Windows Terminal (screenshot in conversation): Esc restores main buffer, clean prompt, $LASTEXITCODE 1
• documented skipped checks, platform limitations, or verified pre-existing failures:
  • bun run check → test:full does not complete on this machine: process-spawning tests stall (5 s timeouts) — launcher compile-cache/heap suites and bughunter git-context tests. Reproduced the launcher files (21 pass / 9 fail) identically on a pristine origin/main worktree (C:\Users\reini\openclaude-base, detached 02ac9901, same node_modules tree) — not PR-owned (PR touches no launcher/git/bughunter surface).
  • bun run typecheck — 12 errors, reproduced byte-identically on pristine origin/main (react-reconciler type skew in untouched src/ink/{ink,reconciler,render-to-screen}.tsx; zero errors in the changed file).
  • bun run test:provider — 1748 pass / 3 fail; all 3 are network/timing environmental (stream-watchdog temp-file race, API timeout clamp, Copilot error-type mismatch), none on the PR surface.
  • bun run security:pr-scan -- --base FETCH_HEAD --head HEAD — 1 medium finding in scripts/prepare.mjs; git log origin/main..HEAD -- scripts/prepare.mjs is empty (PR never touches it — upstream rebrand churn between our base and upstream tip). Not PR-owned.

Notes

• provider/model path tested: local OpenCode Zen Router lane; UI change verified visually (Windows Terminal) plus PTY byte trace (1049h/1049l).
• screenshots attached (if UI changed): before/after screenshots are in the linked conversation; after-shot shows clean prompt + exit code 1.
• follow-up work or known limitations: none — tmux -CC keeps the old inline behavior by design.