[teamleaderleo]

Takes over #27787, whose owner is gone. This PR carries its commit with its authorship, adds a red/green test, and makes one change to it.

Leo asked:

0 open sub-sessions and may have at most 50 open at once

WHAT A STUPID F***ING LIMITATION

WTH IS IT COUNTING

What changes

A sub-session's width slot (maxChildren) used to be held by every child that was not closed, including hibernated children and children whose harness had died. A parent that fanned out and waited for its children kept all of its slots until it closed each child by hand.

Now only live children hold a slot: ready (harness alive between turns), running and waiting (waiting for input). A child with a message on its way to a turn also holds one: queued behind a turn or for a run slot, or starting its harness. That way a start never takes the slot a wake needs. A child that finished and hibernated (idle), whose harness died (disconnected), that failed at start (closed by close_refused), or that was closed frees its slot right away, and its history stays. A finished child keeps its harness alive for up to the idle window (2 min, IDLE_CHILD) and holds its slot until it hibernates. That idle window is also how long its process lives. The limit refuses starts. A message to a child the parent already has still wakes it, as a reopen always did.

• hub/subsession_index.rs (moved out of subsessions.rs, which was over its god-file budget on base): the index keeps open (not closed) and live children apart. Index::note_status is the only way status reaches it. It is called from every status write (write_status), on load, on the restart's lost-wait settle (log_load.rs, which used to bypass it), on offer acceptance, and when a start's reservation settles. Index::reserve counts live children, waking ones (Hub::waking_children) and reservations under one lock.
• The refusal states the live count and the counting rule. It no longer tells the agent to close finished children, because they no longer count. list_sessions gains liveChildren, and openChildren keeps its meaning. The work below a parent (work_below.rs) still looks at every open child.
• The cmux-subsessions MCP descriptions, the setting's note and help (agents.subsessions.maxChildren), and the generated Settings/MDM exports use the same definition.

My changes on top of #27787:

• In acpmux: count only live sub-sessions toward width cap #27787, the parent's subsession closed/reopened records fired on every slot change, so each hibernation logged a closed and each wake a reopened. They are back to marking a real close and reopen.
• The restart's lost-wait settle frees the slot; it used to write the status straight into the meta, bypassing the index.
• Waking children count. Without this, a parent at its limit with one hibernated child could start another while a message to the hibernated one waited for a run slot, and end up past the limit.
• The index is its own module, to stay within the god-file budget that failed cmux-next checks on the first push.

Changelog

Fixed: an agent's sub-sessions that finished, hibernated or died no longer count toward its limit of open sub-sessions.

Tests

Regression test: only_live_children_hold_a_width_slot. It walks a real child through ready, running, waiting, idle, a queued message, a harness start, running, disconnected and closed, and checks the width the parent uses, whether a start can reserve, and the parent's subsession records.

• Base (red, test-only commit c4011d1, where the test sat at hub::lifecycle::subsession_width_tests): cargo test --locked -p acpmux --lib hub::lifecycle::subsession_width_tests fails with a finished, hibernated child frees its slot (left: 1, right: 0). Fleet run https://github.com/manaflow-ai/cmux/actions/runs/41002461341
• Head (green, 775b871): cargo test --locked -p acpmux --lib -- hub::subsessions passes both width tests (the test moved to hub::subsessions::index::width_tests). Fleet run https://github.com/manaflow-ai/cmux/actions/runs/41007731238

In the same head run: width_slots_count_live_children_only (from #27787, extended), --test subsessions agents_start_children_within_their_width_and_depth_limits, --test hub_integration run_parent (the parent's subsession records), cargo clippy -p acpmux --all-targets -D warnings and cargo fmt --all --check all pass.

Observed / Not run

• Observed: the red and green fleet runs above, scripts/cmux-next/prepush-gate.py (no new violations), check-no-godfiles.sh --only rust --base (pass), gen-settings-sources.py --check, and four review rounds (BLOCK, BLOCK, BLOCK, SHIP; their findings are fixed in the later commits).
• Not run: no tagged-build check of the Settings row. Only its help text changes, and the setting's key, kind and default are unchanged.

No capture: the only native UI change is the help and note text of the existing Open Sub-sessions per Session setting; layout is unchanged.

Risk: normal. Daemon logic for sub-session admission. The worst case is a parent that can start more harness-alive children than intended. maxRunning still caps concurrent turns, and host admission (admission.rs) still caps sessions.

Agent session: agent session

🤖 Generated with Claude Code

[blacksmith-sh]

Found 9 test failures on Blacksmith runners:

Failures

Test	View Logs
--from must install a complete prebuilt tree/ --from must install a complete prebuilt tree	View Logs
/tmp/tmp92cwfl4k/summary-fail.json (live):/summary-fail.json (live):	View Logs
/tmp/tmpv7hiv1xa/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
test_every_listed_hook_is_one_glaeda_released_with_the_check (main.SigningHooksInSt epWithGlaedaTests.test_every_listed_hook_is_one_glaeda_released_with_the_check)/ test_every_listed_hook_is_one_glaeda_released_with_the_check (main.SigningHooksInSt epWithGlaedaTests.test_every_listed_hook_is_one_glaeda_released_with_the_check)	View Logs
test_the_repository_highlights_validate (main.DigestTests.test_the_repository_highl ights_validate)/ test_the_repository_highlights_validate (main.DigestTests.test_the_repository_highl ights_validate)	View Logs

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

[github-actions]

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

Review 775b871: 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=pr27931 -f ref=lane/subsession-cap-live-2 -f platform=macos -f backend_mode=local

red: cmux-next checks (god files, concurrency, crash safety, l10n); NEW: green on the base 079bc45; do not merge until green; see lane-rules

[github-actions]

Default changes

This pull request changes 1 default at 775b871bb5a3 (what everyone gets without configuring anything):

• tunable-source.CmuxNextSettings/Schema/SettingsSchema+Subsessions.generated.swift:descriptors: "c8ed4a3d7c02d932ffa48874103ee37b28c74db6a3184a47f4e4622d1e30c5d9" -> "c750ac1c552737585711d557232bd27c1dd7583a9080619c38221db139a6e00d"

Signed off on this commit by review (codex gpt-6.1-sol).

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

[teamleaderleo]

Review (codex gpt-6.1-sol): SHIP at 775b871

An independent Codex gpt-6.1-sol review (high, fast), reading the code at the exact head. Earlier rounds said BLOCK and their findings are fixed in this head's commits: the restart's lost-wait settle bypassed the index (round 1), a wake of a hibernated child was not counted (round 2), and a child hibernating between the waking read and the reservation was missed (round 3). Round 4:

No findings. The index updates are synchronized with status changes, waking reservations have no lock-order cycle, and queued/turn-lock reservations are released by their owning futures.

Red on base: https://github.com/manaflow-ai/cmux/actions/runs/41002461341. Green at this head: https://github.com/manaflow-ai/cmux/actions/runs/41007731238

[github-actions]

Shipped in:

• cmux NEXT nightly 1.0.0-nightly.4101529228901 (146a97390c)

[github-actions]

Shipped in:

• cmux NEXT nightly 1.0.0-nightly.4101585426601 (32d72d5a00)

[github-actions]

Shipped in:

• cmux NEXT nightly 1.0.0-nightly.4101495471601 (c5dd001a5d)
• cmux NEXT nightly 1.0.0-nightly.4101751355101 (71076ff59a)