[lawrencecchen]

cmux-ad m14 and m18 (Lawrence's P0, "resizing splits/columns, still not animating the terminals"): in every frame of a divider or column-edge drag, and of the spring its release starts, each terminal takes that frame's pane size and its text re-wraps at that size. The program in it gets its new size when the drag lands.

Why

• The drag hold (Debug tunable terminalDrag.holdSurface, RESIZE VIDEOS A3, on by default) kept each terminal's surface at its drag-start size for the whole drag: cut where the pane was narrower, the terminal background beside it where the pane was wider.
• At the release spring's start, restAt moved the hold to the final size, so the release spring showed the final grid, not the pane's size.
• pixel-perfect's trace (STATUS 13764): 157-159 held ticks in every col-drag-2 film; the surface was up to 150 pt off its pane. That was so with motion.frameOwner.terminals on or off.

Fix

• Delete TerminalDragTunables.holdsSurfaceDuringDrag (terminalDrag.holdSurface) and its path (TerminalSurfaceView+PaneMotion.swift).
• At the drag's first frame each terminal leaves a spring hold the drag caught (a split, a close or a press still landing: endPaneMotion()).
• From there the surface takes its pane's size in every frame through the placement every resize uses (TerminalHostView.placeSurface and the frame owner's place(_:anchor:)), with the present rule of motion.frameOwner.terminals:
  • off: the bounds change at once, and Ghostty's frame of that size lands 1-4 refreshes later;
  • on: ready-or-keep, through TerminalFramePresenter (frame-owner-s4's T-KEYS changes "keep").
• The reflow preview (Ghostty's render-only re-wrap, no SIGWINCH) re-wraps the text at each size. During a drag no final grid is reported, so TerminalGridPolicy.previewsViewGrid stays true.
• The release spring of a drag holds no final size and reports no final grid at its start (paneMotionWillRest and joinPaneMotion skip restAt while the drag is followed). A final grid at the start would set geometry.final and turn the preview off for the spring.
• The PTY still gets one grid per drag, when the release lands, as before. frame-owner-s4's T-KEYS PR puts a 25 ms coalescer (Ghostty.app's rule) in the common report path; drags then use it with no change here.
• stopFollowingResizeGesture() has no drag hold to release any more.

No present-mode override for drags: the drag follows the switch, which T-KEYS turns on by default (m18).

Changelog

• Improved: While you drag a divider or a column edge, each terminal resizes with its pane and wraps its text to the new width in every frame.

Frames

• Frames: every terminal in the panes a divider or column-edge drag resizes, in every drag frame and every frame of the release spring.
  • Before: the surface kept its drag-start size, cut or beside the terminal background, until the release spring, which held the final size.
  • After: the surface has its pane's size in every frame and its text re-wraps there. With motion.frameOwner.terminals off, Ghostty's frame of each size shows 1-4 refreshes after the bounds (top-left, not stretched); with it on, the frame owner's ready-or-keep.
• Frames: a drag that catches a split's or a close's spring: the terminals in it leave the spring's final-size hold at the drag's first frame.
• No content stops following its pane.

Tests

Red first (2399b3523bf, tests only, on feat 9ac955e274c), then the fix (96ad37a03a3); a second red/green pair for Codex round 1's blocker (ca64d25e72e test, b78df0c4f55 fix). The head merges feat (216a6d14b74).

Run	Ref	Step	Result
red, non-display	2399b3523bf	7c4ee94d75b6ddb8d88e290d	DividerDragTerminalSteadyTests FAIL: p1's surface took 1 width in 30 drag steps (widths.count → 1), "drag step 1: p1's surface (600.0, 800.0) in a (592.0, 800.0) area" ...; the swap test: the surface held at 600 pt. DividerDragTerminalResizeCountTests passes.
red, display (gui-session-v1)	2399b3523bf	58d8d7065b215146f8919a23	TerminalDividerDragFrameTests FAIL, both switch values: "new sizes in drag frames [0, 0]" of 20; frames behind their pane [20, 20] (off), [21, 21] (on)
green, 36 suites (first fix)	96ad37a03a3	807d9ee21deb7bfe0713d1ff	35 suites pass; SplitSpringMotionTests.fiveRapidSplitsAreOneMotion fails on its 8.33 ms frame-time bound (28.1 ms), a timing test with no drag, the same test that failed at #35940's head (2ff367f37c3d6070185f8309, 11.66 ms)
red: a surface swapped in mid-drag (Codex round 1)	ca64d25e72e	4de6e21ebb4e7fb530560fff	aSurfaceSwappedInMidDragFollowsTheDrag FAIL: fresh.isFollowingResizeGesture → false
green, 36 suites	216a6d14b74	5cf38b4cf9c34221b74af108	ci-step m14d2-green exited 0: 36 suites, 187 tests
green, display (gui-session-v1)	216a6d14b74	dd5500c490fe9bed76d6b35e	ci-step m14d2-display-216a6d14b74 exited 0. TerminalDividerDragFrameTests: switch off, new sizes in drag frames [20, 20] of 20, 0 frames behind their pane; switch on, [17, 17] of 20, 20 frames behind (the presenter's keep path: each size shows one commit later; T-KEYS deletes "keep"); misplaced 0, unsized 0, stretched 0, PTY grids [1, 1]
app build	216a6d14b74	bdb41a9fe9c381e4682715ac (tag m14-drags-v1)	done

Observed / Not run

• Observed: the fleet runs above, the prepush gate, check-no-godfiles (TerminalHostView.swift kept at its base line count), check-motion, the package conventions lint.
• Codex round 1 at b72c7e2d3fa: FIX. (1) blocker: a surface swapped in mid-drag did not follow the drag; fixed red-first (ca64d25e72e, b78df0c4f55): the host follows the drag and a surface joins its host's drag. (5) should-fix: the display test pins the switch before each frame; fixed. (4) The PTY count is asserted by DividerDragTerminalResizeCountTests (one grid per drag, none mid-drag; unchanged and green); this PR keeps that rule until T-KEYS' 25 ms coalescer. (2) A forced report (a terminal shown mid-drag) still sends at once, as before this PR. (3) Not verified: a drag that catches a running key-resize hold (FRAME OWNER S3 followedHold). With the switch off the host's follow rule may keep the key hold's size until that hold ends; no test here covers a drag during a key spring.
• Known effect, not this PR: with motion.frameOwner.terminals on (the default from perf: resize perf paths on by default (cmux-ad m19, no switches) #36311), a drag shows each size one commit late, through the presenter's "keep" (display step dd5500c490fe9bed76d6b35e: 17 of 20 drag frames take a new size, 20 frames one commit behind the pane). frame-owner-s4's T-KEYS PR deletes "keep". With the switch off: 20 of 20, 0 behind.
• Observed (pixel-perfect film, reported to the RESIZE OWNER; tables in STATUS 14353): px-col-drag-2-sw-terminals at 96ad37a03a3 against base feat 9ac955e: held terminal frames 155-161 at the base, 0 at the head. One terminal edge frame per film at the release (p0's high side cut by 102 px or more). No other new red.
• Observed at head 9d5718797f84a02bed13ebd4b57d527e84e449d5 (after the feat merge): launch smoke PASS on macOS 14.8.7 and 15.7.7 (os-matrix step 5d88e49aad86d5579008326d, build dc809e30886d6b7643c3ce94, 0 crashes, 0 hangs); focused suites 8b429b07ee626b091b4c0cfc exit 0 (10 suites, 51 tests).
• Not run: films (after merge), a tagged dogfood build.

[github-actions]

cmux status _{(one comment, edited in place; each line is updated by the job that owns it)}

Review 9d57187: HQ feed card · gallery before/after runs · checks

App of this head (app.zip, kept 1 day)

gh workflow run reload-build.yml -R manaflow-ai/cmux --ref feat-cmux-next -f tag=pr36328 -f ref=fix/m14-drags-live-terminals -f platform=macos -f backend_mode=local

Default changes

This pull request changes 3 defaults at 9d5718797f84 (what everyone gets without configuring anything):

• tunable-source.CmuxNextDesign/Tunables/TerminalDragTunables.swift:all: "cab3d3153edd27ef2c43eb3abd1a3819800e7847845fef5e2100863692ffbbd4" -> "060460dd240b8038982e2f9778e3393893e00fed4957d1f2cd3a7b281f72ada6"
• tunable-source.CmuxNextDesign/Tunables/TerminalDragTunables.swift:holdsSurfaceDuringDrag: "7f939fb347a62f071f4887444a385f8e40c6898667371343891e8dceb5fef395" -> (none)
• tunable.terminalDrag.holdSurface: false -> (none)

Signed off on this commit by review (claude-review-subagent, started by lead).

Merge receipt: merged as 058a07e4ae11 with these defaults changed, signed off by review (claude-review-subagent, started by lead).

Regenerate the snapshot with python3 scripts/ci/defaults_guard.py snapshot.

Shipped in:

• cmux NEXT nightly 1.0.0-nightly.4203360930601 (d36db2bf1d)

[lawrencecchen]

Base finding, not this PR: SplitSpringMotionTests.fiveRapidSplitsAreOneMotion (an 8.33 ms frame-time bound, no drag, no terminal) failed in step 807d9ee21deb7bfe0713d1ff at 96ad37a03a3 (28.1 ms). The same test failed at #35940's head 588c9c01929 (step 2ff367f37c3d6070185f8309, 11.66 ms), a PR that does not touch it. It passed at this PR's head 216a6d14b74 (step 5cf38b4cf9c34221b74af108, 36 suites green).

[lawrencecchen]

Review (claude-review-subagent): SHIP at 216a6d1

Independent review subagent (correctness first, read-only, no build), relayed by the lead at 11:00Z. No blocker. Nits for a follow-up PR (not pushed here, a push voids this receipt):

1. DividerDragTerminalSteadyTests and TerminalDividerDragFrameTests send .ended at the last .moved point, so the new guard !isFollowingResizeGesture in paneMotionWillRest (TerminalSurfaceView+PaneMotion.swift:25) has no red/green test of its own; release a few points away or on a snapping column edge.
2. Rename …AndResizesItsProgramOnce or assert resizes == [1, 1].
3. RestingPlacement.swift:64: the doc of PaneResizeGestureFollowing still says a follower keeps its drag-start size.
4. bench-terminal-resize.py:22, :240 and its test :20 still say holdSurface is on by default.
5. Clear followsResizeGesture in viewWillMove(toWindow:) when newWindow !== window.

[lawrencecchen]

Review (claude-review-subagent): SHIP at 216a6d1

Independent review subagent, posted by the perf lead because the owner stopped on an API overload. No blocker. Follow-up nits (a later PR, not this head): the release-spring guard in paneMotionWillRest needs a red/green test (release away from the last move); rename or assert the TerminalDividerDragFrameTests resize count; update the PaneResizeGestureFollowing doc (RestingPlacement.swift:64) and the stale holdSurface text in bench-terminal-resize.py; clear followsResizeGesture on any window change.

[lawrencecchen]

Check classification at head 216a6d14b74 (none from this PR):

• cmux-next swift test (run 41996591952): WindowCleanupCoverageTests.everySuiteThatOpensAWindowClosesWhatItLeaves names AutomationsSignInTests, WindowViewGeometryTests, WindowFrameBrightnessGridTests, AgentPageFlickerLogTests, UpdatedCardFirstFrameTests, WebProjectionTests (none touched here), and HoverCardMachineColorTests 2 color asserts (not touched here). Base findings.
• build / release-compile (hq run 41996596581): infrastructure, xcodebuild -resolvePackageDependencies got HTTP 504 downloading Sparkle 2.10.0. Rerun the same SHA.

[cursor]

Bugbot needs on-demand usage enabled

Bugbot uses usage-based billing for this team and requires on-demand usage to be enabled.

A team admin can enable on-demand usage in the Cursor dashboard.

[lawrencecchen]

Review (claude-review-subagent, started by lead): SHIP at 9d57187

Scope: only the merge of feat fe07c0ab50a (#36311) into this PR. The SHIP at 216a6d14b74 covers the rest.

Method

• I compared the PR-only diff at the old head (git diff 16afd28be6f4 216a6d14b74) with the PR-only diff at the new head (git diff fe07c0ab50a 9d5718797f84). The two diffs touch the same 8 files. Their added lines are identical, byte for byte. Only the removed lines and the hunk context are different, because the base is now perf: resize perf paths on by default (cmux-ad m19, no switches) #36311.
• For the 6 files with no conflict, the feat delta (16afd28..fe07c0a) and the merge delta (216a6d1..9d57187) are identical.
• No other file differs between feat fe07c0ab50a and the head.

#36311 lines: none lost

• motion.frameOwner.terminals keeps default: true and the m19 doc text from perf: resize perf paths on by default (cmux-ad m19, no switches) #36311 (TerminalDragTunables.swift:9-10, :22).
• The only perf: resize perf paths on by default (cmux-ad m19, no switches) #36311 lines that the merge removes are the holdsSurfaceDuringDrag tunable and its "Off (the default since m19)" doc. This PR deletes that tunable.
• perf: resize perf paths on by default (cmux-ad m19, no switches) #36311's other 10 files (BrowserResizeTunables, the CEF tests, PaneContentPolicyTests, PaneGroundCoversPaneTests, TerminalFrameOwnerMainThreadTests, TerminalLaunchFrameKeeperTests, the two bench scripts, test_bench_terminal_resize.py) are the same at the head as on feat.
• The defaults guard job agrees. Its only default change is tunable.terminalDrag.holdSurface: false -> (none), plus the two hashes of the source file. frameOwner.terminals does not show as a change.

#36328 lines: none lost

• See above: the PR's added lines are the same as at 216a6d14b74.
• TerminalHostView+PaneMotion.swift: the new PaneResizeGestureFollowing extension is first in the file. The S3 doc comment, with Frame owner S4 (switch on): a key resize draws each terminal at its pane's size with its prompt in view (m14, no low-prompt hold) #36199's "m6 and m14" text, stays above PaneMotionHoldFollowing.
• Frame owner S4 (switch on): a key resize draws each terminal at its pane's size with its prompt in view (m14, no low-prompt hold) #36199's surfaceFollowsArea and its followedSurfaceFrame change are kept.

Tests

• The merge replaces perf: resize perf paths on by default (cmux-ad m19, no switches) #36311's change to DividerDragTerminalSteadyTests with this PR's version. This does not weaken the tests:
  • widths.count >= 25 stays.
  • perf: resize perf paths on by default (cmux-ad m19, no switches) #36311's single mid-drag check abs(width - hostWidth) < 40 becomes a per-frame check with a 0.5 pt tolerance, for both panes, in the drag, the release and at rest.
  • The swap test now checks every frame after the swap. Before, it checked only the landing.
• No test of perf: resize perf paths on by default (cmux-ad m19, no switches) #36311 is deleted.

terminalDrag.holdSurface references at the head

• Code: none. The tunable, its descriptor in all, and its reader in TerminalSurfaceView+PaneMotion.swift are gone.
• TerminalHostView.holdSurface(size:anchor:) and the calls to it in TerminalSurfaceCommitSizeTests remain. This is the spring hold, not the tunable.
• Script: bench-terminal-resize.py:244-249 drag_holds_terminals() still asks for the key. With the key gone, debug.tunables returns an error, and the function returns False, as its docstring says. Its docstring at :245-246 still says "on by default since P0 RESIZE VIDEOS A3b: a divider drag keeps terminal text in place and reflows once when it lands #32006". That text comes from feat and is outdated, but it does not block.
• Comments: terminal_resize_verdict.py:11, test_bench_terminal_resize.py:20,165, and the m14 history comments in the PR's own files. These are history only.
• schemas/defaults/cmux-next.json:1598 still has holdSurface: true. Feat also still has frameOwner.terminals: false there. The bot generates this file, so I do not ask for a hand edit.

One PTY grid per drag with framesOwnedByCommit = true

• In the drag, place() goes to requestWantedFrame(), then to updateSurfaceSize(). updateSurfaceSize() holds the grid while isFollowingResizeGesture is true (TerminalSurfaceView.swift:478).
• A frame that lands goes through landing. It does not send a report (:363).
• paneMotionWillRest and joinPaneMotion skip restAt during a drag, so the release spring does not report a final grid at its start. The grid goes once, at the landing (followResizeGesture(false), or the spring settles).
• Frame owner S4 (switch on): a key resize draws each terminal at its pane's size with its prompt in view (m14, no low-prompt hold) #36199's surfaceFollowsArea needs followedHold (a key hold only), so a drag does not use it.
• Proof:
  • The display run at 216a6d14b74 (step dd5500c490fe9bed76d6b35e, switch on) gave PTY grids [1, 1], with each newest grid equal to the grid the view fits.
  • At this head, DividerDragTerminalResizeCountTests passed with the switch default now true (step 8b429b07ee626b091b4c0cfc, suite 3/10).

Observed / Not run

• Observed:
  • The diffs above.
  • The log of step 8b429b07ee626b091b4c0cfc at this SHA: 10 suites passed, 0 failed. These include DividerDragTerminalSteadyTests, DividerDragTerminalResizeCountTests, PaneContentPolicyTests and PaneGroundCoversPaneTests.
  • The defaults guard log. That job is red only because it waits for this receipt.
• Not run: the display suite TerminalDividerDragFrameTests at this head. Its last green run is at 216a6d14b74, before Frame owner S4 (switch on): a key resize draws each terminal at its pane's size with its prompt in view (m14, no low-prompt hold) #36199 was in the base. I think the risk is low, because Frame owner S4 (switch on): a key resize draws each terminal at its pane's size with its prompt in view (m14, no low-prompt hold) #36199 changes only the key-hold path (followedHold != nil), and a drag does not set followedHold. I did not compile or run tests myself.