[teamleaderleo]

Leo-ask: ask-a3c16b28b7
Leo-ask: ask-8f7a35dd52

Leo, 2026-10-21: "AUDIT MY CURRENTLY RUNNING APP WHERE CAN WE MAKE THINS MORE FUCKING EFFICIENT" and "WHERE IS ALL THE CPU AND RAM COMING FROM".

We audited Leo's running cmux read-only. The hottest Chromium page renderer averaged 25% CPU with a 586 MB footprint, and the GPU process another 25%. One way a hidden page ends up like that is in BrowserHibernation. Before a hidden page freezes (30 s) or hibernates, a probe asks it, with a 2 s deadline, whether it plays media or holds form edits. If the probe threw or missed the deadline, the page was treated as playing media. A page busy enough to miss the deadline therefore:

• never froze;
• never hibernated;
• was asked again every 30 s for as long as it stayed hidden.

The busiest pages were the ones that never slept.

Change

• A probe the page does not answer is now unknown, not "playing" (PageProbe.answered).
• The tab's last media report still counts.
  • The media the tab last reported still keeps it awake: audible, playing while the tab is muted, or Picture in Picture.
  • The page's media script pushes that state (BrowserTabState.media), so reading it needs no round trip to the busy page.
• Native signals. I found:
  • WebKit reports camera and microphone use natively. That is the planner's isCapturing, which already excludes the page before any probe.
  • Chromium reports neither capture nor audible state to the app. The CEF audio callbacks are not wired.
  • So a Chromium page's Web Audio and capture are known only from its recorder's answer.
• Freeze. After 3 unanswered probes in a row, the page freezes. Any answer, a show or a close resets the count.
  • Why 3: one miss can be a load or a pause of a few seconds. Three in a row, freezeSeconds apart (about 90 s), mean a page busy the whole time.
  • A freeze loses nothing: the page thaws when shown or driven.
• Hibernation.
  • An unanswered page gets its own unanswered exemption (it used to be reported as audio).
  • It hibernates only under memory pressure, after 3 unanswered probes, because its form edits are unknown.
  • When the hibernation probe thaws a page that is past its 3 probes and keeps it, the page freezes again right away.
• Every existing guard still applies:
  • pinned tabs
  • excluded hosts
  • DevTools
  • capture
  • reported media
  • neighbor and hovered tabs
  • an agent's 30 s grace
  • freezeHiddenTabs: "off"

Tests

UnansweredPageProbeTests (committed red first, e519a0d):

• A page that does not answer is not "playing media".
• A page that does not answer but reported audible media stays awake. This is a guard and passes on both commits.
• A real hidden WebKit page that spins its main thread, so every probe misses its deadline, freezes. On the base it never does.

Changelog

• Fixed: a hidden browser tab too busy to answer now freezes instead of using CPU for as long as it stays hidden.

No capture: nothing drawn changes.

Observed / Not run

• Observed on the fleet (aws-m4pro-5), with the focused command cmux-ci run --class light --script scripts/ci/package-test-lane.sh --arg suite --arg Packages/macOS/CmuxNext --arg <suites>:
  • Red, at 6b3aea0: the red test alone on today's base 4e912c0, branch probe-unknown-freezes-red-evidence. The first red commit, e519a0d, never ran its tests: the base's web step failed until ci: the agent pane page ships the Inspector's ContextSection.css (base red since #23789) #23935 landed. UnansweredPageProbeTests failed 2 of 3:
    • aPageThatDoesNotAnswerIsNotPlayingMedia: the probe gave playingMedia: true, so the exemption was .audio.
    • aHiddenPageTooBusyToAnswerFreezes: the busy page never froze within 15 s.
    • The media guard passed, as expected on the base.
  • Green, at the head def9c19: all 5 suites passed. UnansweredPageProbeTests passed 3 of 3, and the busy page froze in about 6.8 s. HiddenPageFreezeTests, NeighborTabsWakeTests, HiddenPageHibernationTests and NeighborTabsTests also passed.
• Not run: a live CEF page in a tagged app. The unit and WebKit tests drive BrowserHibernation through the freezer seam, which is the path Chromium pages take in the app.
• Not measured: how much of the audited 25% renderer came from this path. The audit could not map renderers to tabs; cmux browser list reports no renderer PID.

Agent session: Air-Blue

🤖 Generated with Claude Code

[github-actions]

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

Review def9c19: 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=pr23980 -f ref=probe-unknown-freezes -f platform=macos -f backend_mode=local

red: cmux-next checks (god files, concurrency, crash safety, l10n); base red: the same failure on the base 4e912c0; does not block

[blacksmith-sh]

Found 12 test failures on Blacksmith runners:

Failures

Test	View Logs
/tmp/tmpjdpeo8w9/summary-fail.json (live):/summary-fail.json (live):	View Logs
/tmp/tmpw4mjm3t5/summary-missing.json (live):/summary-missing.json (live):	View Logs
[new_workspaces] debug.new_workspace timed 0 of 50 new workspace rows/ [new_workspaces] debug.new_workspace timed 0 of 50 new workspace rows	View Logs
[new_workspaces] the window draws no frames/[new_workspaces] the window draws no frames	View Logs
[tabs] debug.tab_switch timed 0 of 40 switches/ [tabs] debug.tab_switch timed 0 of 40 switches	View Logs
[tabs] the window draws no frames/[tabs] the window draws no frames	View Logs
convention violations found (ERROR lines above)./ convention violations found (ERROR lines above).	View Logs
nightly cmux-tui resolve/nightly cmux-tui resolve	View Logs
test_every_submodule_free_checkout_on_an_owned_runner_resets_submodules (main.Reuse dWorkspaceSubmodules.test_every_submodule_free_checkout_on_an_owned_runner_resets_submo dules)/ test_every_submodule_free_checkout_on_an_owned_runner_resets_submodules (main.Reuse dWorkspaceSubmodules.test_every_submodule_free_checkout_on_an_owned_runner_resets_submo dules)	View Logs
test_every_submodule_free_checkout_on_an_owned_runner_resets_submodules (main.Reuse dWorkspaceSubmodules.test_every_submodule_free_checkout_on_an_owned_runner_resets_submo dules) (job='web-bundles')/ test_every_submodule_free_checkout_on_an_owned_runner_resets_submodules (main.Reuse dWorkspaceSubmodules.test_every_submodule_free_checkout_on_an_owned_runner_resets_submo dules) (job='web-bundles')	View Logs

...and 2 more test failures. View all on Blacksmith

^{Need help on this PR? Tag @codesmith-bot with what you need.}

[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.

[teamleaderleo]

cmux-next crash-class rules / cmux-next crash-class rules is red on 11 of 11 fresh feat-cmux-next PR heads in the last 2:00:00, so it is most likely a base red, not this PR's. The base-reds lane has it; a base merge picks up the fix once it lands.

[teamleaderleo]

Review (subagent, correctness first): SHIP at 6158387

An independent acpmux Claude reviewer read three rounds. It found two problems, both since fixed:

• 2a85c0e: HOLD. The refreeze in prepare skipped the neighbor, hover and agent guards and the freeze-off setting. Fixed in 2c50bfe.
• 3a5a4d9: HOLD. Its !probing filter stopped refreezing pages that prepare keeps. Reverted in 6158387; only the freeze probe now counts misses.

It gave SHIP at 6158387 with two non-blocking nits:

• Hibernate Tab clears answeredBusy, which needs a rare sequence of events.
• One branch in doubts is unreachable.

[teamleaderleo]

Review (subagent, correctness first): SHIP at def9c19

def9c19 merges the base (it includes #23935) into the reviewed 6158387. The PR diff is unchanged: git diff origin/feat-cmux-next...def9c19 is byte-identical to the one passed at 6158387. A fresh reviewer confirmed this by reading both diffs and the merged BrowserHibernation.swift. Its reds that are not this PR's:

• Namespace-type errors in ChiefWorkspaceName.swift and AttentionHighlightLook.swift.
• The test_cmux_next_checks_workflow submodule-reset test.
• Guard status and release-notary.

The BrowserHibernation.swift god-file warning is warn-only. The file was already over the ceiling on the base, at 530 lines.

[teamleaderleo]

Bead cx-3lsfx created for this PR by the cmux coordinator.

[github-actions]

Shipped in:

• cmux NEXT nightly 1.0.0-nightly.3973137732601 (15186f150b)
• cmux NEXT nightly 1.0.0-nightly.3973389583201 (48a69e9f66)