Open a fresh tab when the final tab closes (APP-5242)

[warp-agent-staging]

Summary

Closing the last tab in a window has always closed the window, with no way to opt out. This adds the global preference Close window when all tabs are closed (Settings → Features → General, directly below Quit when all windows are closed, default on) and implements the keep-window behavior it unlocks.

With the preference off — or in a host that cannot close its window at all (web / link-only) — a close that would empty the window resolves to one of two outcomes instead:

• Pristine final tab → silent no-op. All four close affordances (close button, middle-click, Close tab menu item, workspace:close_active_tab) stay visible and enabled and simply do nothing. No confirmation, no toast, no animation, no undo entry, no session allocation — so repeated Cmd/Ctrl+W cannot accumulate work.
• Non-pristine final tab → close and replace. A fresh tab is seeded before the old one is detached, using ordinary New Tab semantics (configured default session mode/profile, new-tab placement, working-directory rules). The workspace never exposes or persists a zero-tab state, and the replacement is a normal tab for session restore.

The default-on path is untouched: the window still closes, pristine or not.

How it works

• Workspace::final_tab_outcome centralizes the decision (CloseWindow / NoOp / Replace) from the preference and ContextFlag::CloseWindow, and both close_tab and remove_tab route through it. Confirmation is no longer deferred to the window-close prompt when the window is going to stay open.
• Pristine is a one-way runtime marker on TabData, never inferred from visible emptiness. It fails closed: TabData::new produces a non-pristine tab and only TabData::new_pristine opts a tab in, so a creation path added later keeps its close affordances working by default. It is cleared by narrow, user-originated signals rather than the broad AppStateChanged notification: terminal input edits, command execution, every agent-prompt submission, pane-layout actions, tab/pane rename and reset, colour, pin, group membership and group pin. Restored, imported and transferred tabs start non-pristine.
• Undo Close pairs the closed final tab with the exact replacement that took its place and swaps that replacement out only while it is still pristine; otherwise it restores alongside so user work is never discarded.
• Cross-window drag: a committed handoff out of a one-tab source leaves a fresh tab behind instead of closing the source. The replacement is created only on commit, so a cancelled or returned drag never leaks a tab.
• CloseWindow-disabled hosts get the four final-tab affordances back; the workspace, not the affordance, decides what a close does there.

The approved spec is committed on this branch at .agents/specs/APP-5242-fresh-tab-on-last-close.md.

This is the alternate approach to APP-5230 (empty pane), specced and prototyped in parallel so both variants can be compared hands-on. APP-5230 is untouched by this PR.

Trying the prototype

Check out this branch and run the dogfood/internal channel build:

./script/install_channel_config    # once, puts warp-channel-config on PATH
cargo run --bin warp

The toggle ships on, which is today's behavior, so flip Close window when all tabs are closed off in Settings → Features → General to try the new behavior.

Rework changes

Addressing the code-review findings on this PR:

• Pristine tracking missed several group mutations. Only individual pin/unpin marked the tab. Group pin/unpin now marks every member (mark_group_members_non_pristine), and creating a group from a tab or selection, moving tabs into a group, removing them from one, and ungrouping all mark the affected tabs. A materially modified final tab can no longer take the silent no-op path.
• The pristine default now fails closed. TabData::new produces a non-pristine tab and TabData::new_pristine is the explicit opt-in, so any tab-creation path added later starts non-pristine and keeps its close affordances working instead of silently swallowing Ctrl+W.
• A submitted agent prompt did not break pristine, though the spec makes it a one-way mutation. The whole submission family now emits the session-mutation signal: a shared-session viewer prompt (SendAgentPrompt), an agent query from the input box (ExecuteAIQuery), a committed cloud follow-up (SubmitCloudFollowup), and CLI agent input (SubmitCLIAgentInput). Sweeping the rest of the spec's mutation list also turned up pane rename/reset and tab-name reset, which are now marked too.
  • Entering agent view is deliberately left unmarked: the runtime itself enters it when seeding a replacement under a default session mode of Agent, so marking there would make every replacement born non-pristine.
• Close outcome is now reported honestly. close_tab returns whether a close actually happened, and the local-control close handler says so rather than only blaming a cancelled warning dialog.
• Replacement identity is derived by diffing the tab list instead of trusting the active index after seeding.
• Visual proof re-recorded to cover the two paths the previous capture omitted — the touched-replacement restore-alongside undo branch and the default-on window close. The screenshot/video block below is fully replaced with the current capture.

New regression coverage for the rework: group pinning marks every member, grouping a lone tab stops the no-op, tabs the runtime did not seed start non-pristine, typing reaches the workspace through the real Input → TerminalView → PaneGroup → Workspace chain, type-then-delete stays non-pristine, navigational interaction preserves pristine state, every agent-prompt submission path marks the tab, and pane rename / tab-name reset mark the tab.

Verification

Regression tests in app/src/workspace/view_tests.rs, covering both preference states and each branch separately:

• Setting defaults to on and round-trips through the settings store.
• The default path still closes the window and seeds no replacement, pristine tab included.
• Three repeated closes on a pristine final tab leave tab, pane group, pristine state, confirmation state and the undo stack untouched.
• A non-pristine final tab is replaced, leaving exactly one tab whose pane group differs and which starts pristine.
• A multi-tab close that drains to one tab lands on the same outcome with no stale active index.
• Both Undo Close branches: swap out a still-pristine replacement, restore alongside a touched one.
• CloseWindow-disabled hosts: all four affordances available (including the workspace:close_active_tab binding being in context) and both outcomes reachable.
• Both cross-window handoff branches.
• The pristine contract itself, driven through the production event chain rather than injected events.

ContextFlag gained a thread-local, RAII test override (mirroring FeatureFlag::override_enabled) so the CloseWindow-disabled host can be exercised without mutating process-wide state.

Repository checks, all from the repo root with the current master merged into this branch:

• ./script/format --check — clean.
• cargo clippy --workspace --exclude warp_completer --all-targets --tests -- -D warnings and cargo clippy -p warp --all-targets --tests -- -D warnings — clean.
• cargo nextest run --no-fail-fast -p warp -p warp_core (the packages this diff touches) — 6288 passed / 3 failed. The three failures (ambient_agent_headers_for_task_overrides_existing_cloud_agent_header, test_decorations_with_multibyte_chars, test_histignorespace_support_in_zsh) reproduce identically on the unmodified base branch and are environment-dependent, not caused by this change.

Visual proof below was captured against this branch's running dogfood build. The current recording walks the setting off, runs echo FIRST-TAB and closes that tab so the window stays open with a fresh replacement, runs echo SECOND-TAB in that replacement, then runs Reopen closed session — which restores the closed tab alongside the worked-in replacement, leaving both FIRST-TAB and SECOND-TAB intact — and finally switches the setting back on, where closing the last tab closes the whole window.

CHANGELOG-NEW-FEATURE: Added a "Close window when all tabs are closed" setting — turn it off and closing the last tab opens a fresh tab instead of closing the window.

Originating thread: https://warpdev.slack.com/archives/C0BDQDW8V5E/p1786112332604639

Computer-use video recordings (3)

Verifying close-window-when-all-tabs-closed setting: Locating the new General setting in Warp Settings → Features, toggling it off, testing close-button behavior on tabs (untouched vs typed-in), undoing a close, then toggling the setting back on and closing the last tab.

Warp close-on-last-tab setting walkthrough: Recording of finding the "close window when all tabs closed" setting in Warp Settings → Features → General, toggling it off, closing an untouched tab three times, running echo APP-5242-MARKER and closing that tab, reopening the closed session, then toggling the setting back on and closing the last tab.

Warp tab close & reopen-session verification: Walkthrough of Warp: turning off the "close window when all tabs closed" setting, closing/reopening tabs via Reopen closed session, then turning the setting back on and closing the last tab so the window closes.

Computer-use screenshots (10)

The Warp window still open with exactly one "New session" tab after clicking the untouched tab's × close button three times (each click replaced it with a fresh tab rather than closing the window).

After typing text into the terminal and clicking the tab's × close button, the Warp window stayed open with one fresh replacement tab titled "bash" and the typed text no longer visible.

Warp Settings → Features → General showing the "Close window when all tabs are closed" toggle (ON), its supporting text, and neighbouring rows.

The Warp window still showing a single "New session" tab and empty terminal pane after clicking the tab's close control three times.

The Warp window immediately after closing the tab that ran echo APP-5242-MARKER: a single fresh "bash" tab with an empty terminal pane, chrome intact.

The Warp window after running "Reopen closed session": a single tab titled "echo APP-5242-MARKER" with the restored command and its APP-5242-MARKER output visible in the pane.

Two-tab state right after running "Reopen Closed Session": tabs "echo FIRST-TAB" (selected, showing FIRST-TAB output) and "echo SECOND-TAB".

Tab "echo FIRST-TAB" selected, its pane showing the command echo FIRST-TAB and its output line FIRST-TAB.

Tab "echo SECOND-TAB" selected, its pane showing the command echo SECOND-TAB and its output line SECOND-TAB.

Screen immediately after closing the final tab with the setting ON: the entire Warp window is gone, leaving only the desktop wallpaper and taskbar.

Conversation: https://staging.warp.dev/conversation/0855efb3-a979-4bc5-94e4-3579ed65bcef
Run: https://oz.staging.warp.dev/runs/019fdccf-a5b0-79cf-9a19-e543ce82c104

This PR was generated with Oz.

[warp-agent-staging]

Overview

Adds a "Close window when all tabs are closed" preference that keeps the window alive on final-tab close, no-oping a pristine tab and replacing a non-pristine one. The state machine is sound, but the visual proof does not cover the branches the spec asks for, and that needs your call.

Concerns

• The recording proves the non-pristine replace branch and the undo swap-out, but not spec visual criterion 2: it ends at 0:48 with the window still open and the cursor merely hovering the close control, on a tab that had two commands run in it. The PR description nevertheless claims "the setting back on, where closing the last tab closes the whole window" — that claim is not supported by the artifact and should be corrected regardless of what you decide about coverage.
• Criterion 3, the pristine no-op, cannot be established by this capture at all: a silent no-op and a close-and-replace are visually identical, since both leave one empty "New session" tab. Distinguishing them needs a visible signal such as session count, undo depth, or a pane identifier. The PR's own screenshot caption for the earlier artifact in fact reads the opposite way, describing the clicks as having "replaced it with a fresh tab".
• Criteria 6 and 7 have no capture at all, and criterion 5's second half (mutate the replacement, then undo, to show restore-alongside) is also uncaptured. A Dogfood prototype may not need all seven captures — that is your judgment to make — but criterion 6 covers the CloseWindow-disabled host path, which is where the critical defect now in revision actually bites, so it should not be waived until that fix lands.

Verdict

Checks: build pass, tests pass, CI green, visual proof mismatched

Found: 1 critical, 3 important, 2 suggestions, 2 nits

The critical and the other two important findings are unambiguous and are already going back through a revision pass; they are not reproduced here. This review carries only the finding that needs a human decision.

[warp-agent-staging]

Review overview

Cumulative re-review of c64caa0e27e21b936281bea358265d4d0df8c8f1 verified that the prior code findings are resolved: tab-group mutations mark affected tabs non-pristine, prompt submission reaches the pristine marker, and the new conservative default preserves explicit pristine seeding for ordinary replacement paths. Focused pristine-state, agent-prompt, and pane-rename regressions pass locally.

Blocking finding

⚠️ [IMPORTANT] Visual proof does not demonstrate the default-on final-close acceptance criterion. The refreshed recording shows the setting and the touched-replacement Undo Close branch, but its final default-on sequence ends with the last tab/window still visible; it never shows the window closing after the close action. The committed spec requires proof that closing a pristine only tab with the setting on closes the window. Please re-record that sequence through the visible window-close result and refresh the linked evidence.

Verification

• cargo nextest run -p warp --lib pristine --no-fail-fast: 14 passed
• cargo nextest run -p warp --lib agent_prompt --no-fail-fast: 1 passed
• cargo nextest run -p warp --lib pane_marks --no-fail-fast: 1 passed
• git diff --check origin/master...HEAD: passed
• Linux CI and all formatting/clippy, compile, and CodeQL checks are green; macOS and Windows test jobs were still in progress at review time.

Verdict

Request changes: the implementation is code-clean on this pass, but the required user-facing default-on window-close proof is incomplete.