[teamleaderleo]

Follow-ups from the independent review of #27536 ("a turn ends when Claude Code ends it"). Its SHIP receipt listed four non-blocking notes. This PR fixes three of them; the fourth turns out to need no code change (see below).

#27536 fixed Leo's asks: "what should we do?", "should i steer", and "did ANY of the reserach/think tank/pls go look into this sessions finish???".

What changes

1. Every carried turn arms its own turn cap. On base, only the first carried turn (start_carried) armed the acpmux: turn cap, so a session taking many agents' messages can end its turn #27520 timer. The turns run_carried started after it, and the carried turns a restart adopts (recover_carried), got none. So an agent message queued behind them never got the cap notice. Now begin_carried arms the timer for every carried turn. A turn a restart adopts (a carried one, or a prompt's own in hosts.rs recover_work) arms it too. The timer now counts from the turn's own started_at, so an adopted turn's cap doesn't start over.
2. Each carry holds Claude's output once, with its own bound. On base, the hold started only when the count of owed carried turns went from 0 to 1. A second carry made before the first one's prompt came (that hold's 10 s bound already passed) got no hold, so its turn's words were read before that turn started. Translator::holding now returns the newest carry's number. The in-process reader (agent.rs) and the agent host (host.rs read_held/holds) hold once per carry.
3. A carried turn made only of lane reports is marked report-only. begin_carried never called lane_turn_started, so a coordinator's carried turn on its lanes' reports reported upward. Reports never steer on their own: they wait in the queue (lane_deliver). So this happens only when the person presses Send now on a queued report and Claude Code hasn't read it by its end_turn. Each steer now records whether it is a lane report (Continues.reports). A carried turn counts as report-only when every steer it carried is a report (lane_turn_on). A carried turn a restart cuts before its turn_started keeps the mark (carriedTo.reportsOnly). One that had already started comes back without it, but nothing depends on that: lane links live only in memory and aren't restored, so after a restart no turn of that session reports upward anyway.

Note 2, "an unread turn-cap notice can be carried into the next prompt", is real but needs no code change. The notice line is already in Claude Code's queue when its turn ends. Claude Code takes it as soon as it emits the result, so cancel_async_message would lose that race, and neither the adapter nor the hub can pull it back. All that happens is one short carried turn on the notice ("finish the current step and end your turn"), which ends as the notice asks. The held digest goes next, and the carried turn keeps a cap of its own (turn_cap_noticed still names the turn the notice was for).

Tests (red, then green)

Red commits come first:

• tests/hub_integration/carried_turns.rs a_carried_turn_after_a_carried_turn_has_its_own_cap (behavioral).
• tests/agent_host/carried_hold.rs a_second_carry_holds_claude_output_again (behavioral, about 11 s: it waits out the first hold's bound).
• src/hub/carried_turn.rs a_carried_turn_on_lane_reports_only_reports_nothing_upward. Base has no way to mark a steer as a report, so this red commit doesn't compile on base.
• src/hub/turn_cap_tests.rs an_adopted_turns_timer_counts_from_the_turns_start was added with the fix.

Observed / Not run

Ran through cmux-ci test (Blacksmith Linux, 32 vCPU):

• Red at d07170120b8 (red tests on base), run 40993825537:
  • cargo test -p acpmux --test hub_integration a_carried_turn_after_a_carried_turn_has_its_own_cap failed: the third turn's cap sent its notice, left: [].
  • cargo test -p acpmux --test agent_host a_second_carry_holds_claude_output_again failed: the second carry held nothing (read: LATE-2 was read before its prompt came).
  • The report-only unit test doesn't compile on base (see above).
• Green at e222db5536b, run 40994153187:
  • cargo clippy -p acpmux --all-targets -- -D warnings is clean.
  • lib filters carried_turn (12), turn_cap (12), carry (18) and lane_report (27) all pass.
  • --test hub_integration carried_turns (10), --test agent_host carried_hold (2), --test lane_reports (17) and --test durable_agents carried (1) all pass.

Not run: the full acpmux suite (left to PR CI), and a live Claude Code session.

Risk

risk:normal. acpmux daemon only, Claude carried turns. The turn cap is off by default (turnCapMs). The hold change keeps the same 10 s bound per carry.

Changelog

none

Agent session: agent session

🤖 Generated with Claude Code

[blacksmith-sh]

Found 2 test failures on Blacksmith runners:

Failures

Test	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 b86ee09: 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=pr27868 -f ref=lane/acpmux-carry-followups -f platform=macos -f backend_mode=local

red: cmux-next checks (god files, concurrency, crash safety, l10n); NEW: 3 commit(s) change the sidebar's drawing code without a Sidebar-Look: trailer or an entry in plans/cmux-next/sidebar-owner.md is not red on the base afe5e8d; do not merge until green; see lane-rules

[teamleaderleo]

Review (codex gpt-6-astra high fast, independent review): SHIP at e222db5

Reviewed exactly 9e700f3f7fd..e222db5536b. No blocking findings:

• Timers and locks: begin_carried releases the turn lock before it arms the timer, and no new lock inversion was found. The saturating subtraction gives an overdue adopted turn an immediate timer.
• Holds: each carry's seq tells successive holds apart, so a stale HoldOver can't release a newer hold. The watch subscription reads the current state, so no wakeup is lost. read_all keeps the EOF drain from starting another hold.
• Compatibility: an older record without reportsOnly decodes as false.
• Not blocking: a carry that had already started loses the report-only mark across a daemon restart. That changes nothing today, because lane links live only in memory and aren't restored. The PR body's claim is narrowed to say so.
• Unread cap notice (note 2 of acpmux: a turn ends when Claude Code ends it #27536's receipt): once Claude Code has dequeued the notice, no withdrawal can take it back. That agrees with the PR's reasoning.

Exact-head tests: run 40994153187 (clippy -D warnings is clean; the carried-turn, hold, turn-cap and lane-report suites pass). Red on base: run 40993825537.

[teamleaderleo]

Review (codex gpt-6.1-sol high fast, independent review): SHIP at 5568330

This head merges feat-cmux-next (c5dd001) into e222db5, which was reviewed SHIP earlier. The only conflict was in turn_cap_tests.rs: base's new digest test and this PR's adopted-turn test were both added at the end of the file, and both are kept. The gpt-6-astra review of this head hit its 40-minute limit without a verdict, so this one ran on gpt-6.1-sol, the interchangeable Codex reviewer.

Reviewed git diff c5dd001a5df8 55683309fa6. Findings: none.

• The reviewed behavior is unchanged. Every carried turn arms its own cap, including turns a restart adopts. Each carry has its own hold. A carried turn on lane reports only is marked report-only.
• The interactions with the new inbox digest (steer_queued) and the turn cap are still correct.
• The PR adds no indexing or slicing in non-test code, so the panic ratchet doesn't move.

Exact-head tests: run 41013004844.

• cargo clippy -p acpmux --all-targets -- -D warnings is clean.
• The lib suites carried_turn, turn_cap, carry, lane_report and digest pass.
• hub_integration carried_turns (10), agent_host carried_hold (2), lane_reports (17) and durable_agents carried (1) pass.

[teamleaderleo]

Red comparison for head 5568330 (addendum to the receipt above). I compared each red by name against the newest base runs.

• nightly-guards / workflow-guard-tests / release-notary
  • Head job 123922834903 fails steps "Test What's New highlights and release notes" (failed: python3 scripts/whats-new/test_digest.py) and "Validate a signing job on an owned Mac holds the whole Mac and deletes its keychain".
  • Base guards run 41019032868 at 695d1f5 (job 123941860165) fails the same two steps with the same test_digest.py failure, plus a third, "Validate universal nightly workflow".
  • The failing tests in test_digest.py, as Blacksmith named them on this PR, are test_the_repository_highlights_validate and test_every_listed_hook_is_one_glaeda_released_with_the_check. Those two also failed on acpmux: a turn ends when Claude Code ends it #27536's base. This PR changes nothing under scripts/ or .github/.
• nightly-guards / Guard status
  • Head job 123925887544 and base job 123944211872 both fail "Check routed guard jobs". This job only wraps the routed guard jobs above.
• cmux-next daemon tests (same-tree)
  • Run 41014818964 has 5 red tests: aRefusedConnectionSaysTheDaemonIsBusy, aDaemonBusyPastTheBudgetIsStillReached (BusyRefusalReconnectTests), aRefusedSplitIsNotAskedAgain, aDaemonWithoutClientKeysIsNotAskedAgain (SplitLateStartTests) and bothDamagedFallsBackToTheNewestArchive (ClassicUpgradeSimulationTests).
  • All five fail by name in the same job on the unrelated lane/odd-27327-spaces (run 41020031643, 07:25Z), among its 8 red tests.
  • The job is skipped on base push and dispatch runs since 04:34Z, so no newer base result exists to compare against.
  • None of the five is a crash or signal; each is a red assertion. They are Swift tests of the cmux-tui session daemon client, and this PR changes only cmux-tui/crates/acpmux.
• cmux-next checks: compared per check against base push runs earlier on this PR. The web-bundles red was red on base pushes and on unrelated PRs at 03:26Z and 03:27Z.
• Familiar intent review warning / warn: an infrastructure failure. The job died on a GitHub API rate limit (HttpError: API rate limit exceeded for installation). A rerun at this head has been requested.

[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]

Base merge only: b0cad4f merges feat-cmux-next (806dd9f) into 5568330, the head the gpt-6.1-sol SHIP above was for.

• Clean merge, no conflicts. The PR's diff against the new base is the reviewed diff. Comparing git diff c5dd001a5df8 55683309fa6 with git diff 806dd9fabee b0cad4f0e03, only blob ids, hunk line numbers and one context line differ. That context line is base's acpmux: Stop really stops an agent that works through the cancel #27685 adding respawn_stopped_agent in run_carried, a separate step from this PR's code. The Sol SHIP carries to this head.
• Do the daemon-test reds touch acpmux?
  • BusyRefusalReconnectTests and SplitLateStartTests run a python3 script as the daemon (DaemonBridge(executable: "/usr/bin/python3", ...)). No cmux-tui or acpmux binary runs.
  • ClassicUpgradeSimulationTests launches the same-tree cmux-tui binary, which links the acpmux crate. It migrates classic sessions and workspaces and runs no agent turns. This PR's code runs only when a Claude agent's turn carries.
  • These five tests are new on base. cmux NEXT: a web update that swaps in shows no "Updated to" toast #27969 passed them all at its base-merged head, so the daemon tests are rerunning at this head rather than being counted as base reds.

[teamleaderleo]

Review (the owner lane's independent review, at its land request): SHIP at b0cad4f

Recorded by lane-gh-broker from this PR's owner lane, which asked to land this exact head after its independent review. next-autoland lands it once every red is the base's; a BLOCK at this head, a HOLD line or a hold label still stops it.

[teamleaderleo]

Review (the owner lane's independent review, at its land request): SHIP at 7a45a16

Recorded by lane-gh-broker from this PR's owner lane, which asked to land this exact head after its independent review. next-autoland lands it once every red is the base's; a BLOCK at this head, a HOLD line or a hold label still stops it.

[teamleaderleo]

Reds at head 7a45a16, by name (81 passed, 3 failed, 0 pending). This PR changes no file outside cmux-tui/crates/acpmux/.

• cmux-next checks: the red check is now "No literal state roots (per-build isolation S7)". check-state-roots.sh names only cmux-tui/crates/acpmux/src/cli/help_examples/session.rs:179 (~/.acpmux/bundles/mini-3-review-27). That line came in on base with acpmux: every command's --help ends with examples #28013 and is still on base tip 58d6a62. This PR doesn't touch the file. The "String tables" red from earlier heads is gone.
• release-notary (step "Test What's New highlights and release notes"): run locally, only scripts/whats-new/test_digest.py fails, in test_the_repository_highlights_validate, on chief-name-in-agent-chats.md. That highlight came in on base with cmux NEXT agent chat: a Chief's message shows the Chief's name and avatar (#27914) #27935 and is on base tip 58d6a62. Nothing under release-notes/ or scripts/whats-new/ differs between this head and that tip, so base fails it the same way. The other seven commands in the step pass.
• Guard status (step "Check routed guard jobs"): wraps the release-notary red.

None of these reds come from this PR. Review: SHIP in comment 6520646600 (the diff is unchanged, see comment 6520856518).

[teamleaderleo]

Infra red at 7a45a16 (next-autoland lists it as "its own red"): cmux-tui artifacts / publish the Windows daemon to R2 (run 41043457329, job 124029813663, a push run of cmux-tui-artifacts.yml). It failed only on:

Error: Failed to persist attestation: API rate limit exceeded for installation.

It tried four times (11:08:55, 11:09:59, 11:12:05 and 11:16:09 UTC) with backoff, and every try hit the same error. Nothing in that job failed because of the PR's code. Under the standing rule it isn't rerun, since a rerun spends the same exhausted installation quota.

The other reds are base by name (comment 6521331273). Review: SHIP (comments 6520646600 and 6520856518).

[teamleaderleo]

Daemon-tests red at 7a45a16 (next-autoland, 14:22Z: "its own red: status / cmux-next daemon tests (same-tree)"; run 41045353403, job 124031216958):

• 1 red test: ClassicUpgradeSimulationTests.aDowngradeAndReUpgradeMergesOnlyNewWork() at ClassicUpgradeSimulationTests.swift:184. It is an assertion failure, not a crash, signal or cut-short run. After a classic downgrade and re-upgrade, a workspace came over with no group (here.group → nil, expected "Work"), and two workspaces came over in swapped order.
• The rest of the suite: in the same run, theFirstLaunchBringsEverythingOnce("messy") showed the same kind of order mismatch, and aRunCutOffMidwayFinishesWithoutDuplicates() recorded it as a known issue. At this PR's earlier head 5568330, a different test in this suite was red (bothDamagedFallsBackToTheNewestArchive). Classic import's ordering and grouping misbehave across this suite whichever test happens to be red.
• Why it isn't this PR's code: the suite exercises classic import (the Swift app plus the same-tree cmux-tui store) and starts no agent session. This PR changes only acpmux turn handling (carried turns, the turn-cap timer, the post-carry output hold, the report-only mark), and none of that runs without an agent turn.

I couldn't compare by name against a base run of this job: the REST quota is spent, and the base daemon-tests dispatches I was given (41036657085, 41024114885) skipped or timed out. So this is classified on scope and on the suite's own known issue, not on a base run with the same test.

[github-actions]

Shipped in:

• cmux NEXT nightly 1.0.0-nightly.4108273398901 (4ab956c562)
• cmux NEXT nightly 1.0.0-nightly.4108444840601 (fb33c09a23)
• cmux NEXT nightly 1.0.0-nightly.4108762076801 (8e3aaa2401)

[lawrencecchen]

Base red from a merge race with #27686: a_carried_turn_after_a_carried_turn_has_its_own_cap queues a fromAgent message, which the inbox now takes, so no cap notice comes. Fix (test runs the queue path with agentInbox off): see the base-red PR linked from #27686's thread / the PR titled 'base red: acpmux: the carried-turn cap test...'.

[lawrencecchen]

Fix: #28427.