[teamleaderleo]

Leo: "why do the x messages from y sessions seem to always piggyback on my messages instead of going on their own"

What was wrong

• Inbox notices. When the person's prompt started a turn, inbox_prompt_written ran inbox_review, which saw a running turn that steers and steered the pending notice into it right away. In one coordinator log that happened 57 times, each notice written 1 ms after the person's prompt. So "[inbox] New messages from other agents…" landed in the person's turn, right under their words.
• Other agents' steers. Messages that aren't inbox items (peer and mesh-relay forwards, team messages, cmux agent message) steered into the person's turn at its next tool step. Claude Code's adapter batched them as "N messages from M sessions arrived during this step", the line Leo is quoting.

What changes

• TurnInfo::persons records whether the person's own prompt started the turn (inbox::persons_prompt: no agent sender, and not a notice, an inbox digest or the turn cap's notice). It is set in run_prompt_joined. Carried turns, restored host turns and test turns default to false, which is the old behaviour.
• Inbox (hub/inbox.rs): inbox_review and notice_held never steer a notice into the person's turn. Messages held back that way set Inbox::held_by_person, so they go in as one turn of their own as soon as the chat is idle, whatever their class. Before, an FYI or an action from an unrelated lane would have waited the 10-minute flush. The flag clears when a notice goes.
• Other agents' steers (hub/turns.rs, checked again as the steer is written in steer_end.rs): another agent's steer queues behind the person's turn instead of going in, and its interrupt stops nothing. A steer-only one is refused with steer.unavailable / personsTurn. The person's own steers, Send now, and an inbox digest the agent asked to drain still go in.
• The person's prompt runs exactly as written. The batch runs as its own turn after it, so the transcript draws it as its own inbox row (InboxNoticeCard), never inside the person's turn. No web change was needed. Older logs stay as they are.

Tests (tests/hub_integration/agent_inbox.rs)

• New: a_notice_never_goes_into_the_persons_turn_and_runs_on_its_own_after_it (fails on base: the notice steers in at the turn's start).
• New: another_agents_steer_waits_for_the_persons_turn (fails on base: the steer goes in).
• a_queued_prompt_never_waits_behind_agent_messages (fix(acpmux): queue automatic messages after the original steering window #27566's case) used to assert the old behaviour, that the notice goes into the queued person's turn. It now asserts that the notice runs as its own turn after that turn and that the person's words are unchanged.
• three_hundred_agent_messages_go_in_as_one_notice_at_the_next_tool_step and another_agents_steer_goes_in_when_no_prompt_waits now run under a turn started by another agent's first work, since steering into a running turn is still how a non-person turn takes notices.

Risk: this changes delivery (a state machine). A coordinator no longer reads lane messages in the middle of a long turn Leo started; it reads them right after. That is the ask.

Agent session: agent session

🤖 Generated with Claude Code

---

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

---

Summary by cubic

Other agents' messages no longer piggyback on the person's turn. Inbox notices used to steer into the person's turn the moment its prompt was written, and other agents' steers went in at its next tool step. A turn the person's own prompt started now takes neither; both wait for the turn's end and run as a turn of their own, immediately.

• TurnInfo::persons records whether the person's own prompt started the turn. A retry, a continuation, and other acpmux-owned prompts are not the person's; carried, restored, and test turns default to false.
• This applies only while the agent inbox is on; with agentInbox off, agents' sends steer in as before.
• turn_started now logs persons, so a restart and a new harness keep the turn the person's own.
• Notices held back by the person's turn are due as soon as the chat is idle, whatever their class, instead of waiting the 10-minute flush; messages already waiting as that turn starts are held back the same way. Held-back messages never steer into a turn carried on from it, and a Stop that cuts the turn still releases them.
• Another agent's steer-only prompt is refused with steer.unavailable / personsTurn; its interrupt stops nothing.
• The person's own steers, Send now, agent-requested inbox digests, and a sub-session family message from the turn's own chat still go in.
• Delivery changes: a coordinator reads lane messages right after the person's turn ends, not in its middle.

^{Written for commit ccd05c9. Summary will update on new commits.}

[github-actions]

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

Review ccd05c9: 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=pr30157 -f ref=lane/ao-inbox-own-turn -f platform=macos -f backend_mode=local

red: cmux-next checks (god files, concurrency, crash safety, l10n); NEW: red check: Companion workflows is not red on the base 39cc5ac; do not merge until green; see lane-rules

Shipped in:

• cmux NEXT nightly 1.0.0-nightly.4130052256501 (28aa638015)

[cubic-dev-ai]

1 issue found and verified against the latest diff

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. When an issue isn't valid or won't be fixed in this PR, reply in its thread with the reason and then resolve the thread. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="cmux-tui/crates/acpmux/src/hub/steer_end.rs">

<violation number="1" location="cmux-tui/crates/acpmux/src/hub/steer_end.rs:258">
P2: A steer-only request that races with the person's turn start returns the generic no-turn error because this hold is not recorded for the later error mapping. Track the person-turn hold and return `persons_turn_steer_refused()` in that path.</violation>
</file>

<sub>Reply with feedback, questions, or to request a fix.

View guided diff | Turn on auto-fix | Re-trigger cubic</sub>

[cubic-dev-ai]

P2: A steer-only request that races with the person's turn start returns the generic no-turn error because this hold is not recorded for the later error mapping. Track the person-turn hold and return persons_turn_steer_refused() in that path.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. When an issue isn't valid or won't be fixed in this PR, reply in its thread with the reason and then resolve the thread. At cmux-tui/crates/acpmux/src/hub/steer_end.rs, line 258:

<comment>A steer-only request that races with the person's turn start returns the generic no-turn error because this hold is not recorded for the later error mapping. Track the person-turn hold and return `persons_turn_steer_refused()` in that path.</comment>

<file context>
@@ -255,6 +255,9 @@ impl Hub {
             } else if !digest && super::inbox::persons_prompt_waits(session) {
                 prompt_held.store(true, std::sync::atomic::Ordering::SeqCst);
                 true
+            } else if !digest && super::inbox::persons_turn(session) {
+                // The person's own turn (`inbox.rs`): it goes in after it.
+                true
</file context>

[cubic-dev-ai]

4 issues found across 8 files

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. When an issue isn't valid or won't be fixed in this PR, reply in its thread with the reason and then resolve the thread. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="cmux-tui/crates/acpmux/src/hub/steer_end.rs">

<violation number="1" location="cmux-tui/crates/acpmux/src/hub/steer_end.rs:258">
P2: A steer-only agent message refused by this new `persons_turn` check at write time gets the generic 'steer: no turn is running' error, not the `steer.unavailable` / `personsTurn` refusal that `turns.rs` returns. It also skips the `steer_withdrawn` log that the `promptWaits` path writes for a logged steer that never went. Track the person's-turn hold with its own flag, like `prompt_held`, and return `persons_turn_steer_refused()` and log `steer_withdrawn` when it is set.</violation>
</file>

<file name="cmux-tui/crates/acpmux/src/hub/turns.rs">

<violation number="1" location="cmux-tui/crates/acpmux/src/hub/turns.rs:495">
P1: Agent-pair delivery still cancels the captured running turn after accepting a queued message, and it skips only review boxes. Add the person-turn guard to that cancellation path so another agent cannot stop the person’s turn.</violation>

<violation number="2" location="cmux-tui/crates/acpmux/src/hub/turns.rs:495">
P1: This interrupt decision uses a stale turn snapshot, so an agent message can still cancel a person’s turn if that turn starts while queue admission waits. Recheck `persons_turn` at the cancellation boundary before honoring `interrupt`.</violation>

<violation number="3" location="cmux-tui/crates/acpmux/src/hub/turns.rs:597">
P2: An auto-resume continuation of a turn another agent started is classed as the person's own turn. `agent_sender` ignores `auto_resume` provenance, so `persons_prompt` returns true and notices and other agents' steers wait through it. That changes existing behavior for agent-started turns. Derive the turn's agent from the cut turn's sender (`CutTurn.from`) so `persons_prompt` sees it, or confirm that auto-resume and subagent retry turns are meant to count as the person's.</violation>
</file>

<sub>Reply with feedback, questions, or to request a fix.

View guided diff | Turn on auto-fix | Re-trigger cubic</sub>

[cubic-dev-ai]

P1: This interrupt decision uses a stale turn snapshot, so an agent message can still cancel a person’s turn if that turn starts while queue admission waits. Recheck persons_turn at the cancellation boundary before honoring interrupt.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. When an issue isn't valid or won't be fixed in this PR, reply in its thread with the reason and then resolve the thread. At cmux-tui/crates/acpmux/src/hub/turns.rs, line 495:

<comment>This interrupt decision uses a stale turn snapshot, so an agent message can still cancel a person’s turn if that turn starts while queue admission waits. Recheck `persons_turn` at the cancellation boundary before honoring `interrupt`.</comment>

<file context>
@@ -481,7 +492,7 @@ impl Hub {
             // agent's message stop a review box's review, or go ahead of the person's prompt that
             // waits (`inbox.rs`).
-            interrupt: opts.interrupt && !capped && !review_held && !prompt_held,
+            interrupt: opts.interrupt && !capped && !review_held && !prompt_held && !persons_held,
             interrupt_turn: opts.interrupt_turn.as_deref(),
             provenance: opts.provenance.as_ref(),
</file context>

[cubic-dev-ai]

P1: Agent-pair delivery still cancels the captured running turn after accepting a queued message, and it skips only review boxes. Add the person-turn guard to that cancellation path so another agent cannot stop the person’s turn.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. When an issue isn't valid or won't be fixed in this PR, reply in its thread with the reason and then resolve the thread. At cmux-tui/crates/acpmux/src/hub/turns.rs, line 495:

<comment>Agent-pair delivery still cancels the captured running turn after accepting a queued message, and it skips only review boxes. Add the person-turn guard to that cancellation path so another agent cannot stop the person’s turn.</comment>

<file context>
@@ -481,7 +492,7 @@ impl Hub {
             // agent's message stop a review box's review, or go ahead of the person's prompt that
             // waits (`inbox.rs`).
-            interrupt: opts.interrupt && !capped && !review_held && !prompt_held,
+            interrupt: opts.interrupt && !capped && !review_held && !prompt_held && !persons_held,
             interrupt_turn: opts.interrupt_turn.as_deref(),
             provenance: opts.provenance.as_ref(),
</file context>

[cubic-dev-ai]

P2: A steer-only agent message refused by this new persons_turn check at write time gets the generic 'steer: no turn is running' error, not the steer.unavailable / personsTurn refusal that turns.rs returns. It also skips the steer_withdrawn log that the promptWaits path writes for a logged steer that never went. Track the person's-turn hold with its own flag, like prompt_held, and return persons_turn_steer_refused() and log steer_withdrawn when it is set.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. When an issue isn't valid or won't be fixed in this PR, reply in its thread with the reason and then resolve the thread. At cmux-tui/crates/acpmux/src/hub/steer_end.rs, line 258:

<comment>A steer-only agent message refused by this new `persons_turn` check at write time gets the generic 'steer: no turn is running' error, not the `steer.unavailable` / `personsTurn` refusal that `turns.rs` returns. It also skips the `steer_withdrawn` log that the `promptWaits` path writes for a logged steer that never went. Track the person's-turn hold with its own flag, like `prompt_held`, and return `persons_turn_steer_refused()` and log `steer_withdrawn` when it is set.</comment>

<file context>
@@ -255,6 +255,9 @@ impl Hub {
             } else if !digest && super::inbox::persons_prompt_waits(session) {
                 prompt_held.store(true, std::sync::atomic::Ordering::SeqCst);
                 true
+            } else if !digest && super::inbox::persons_turn(session) {
+                // The person's own turn (`inbox.rs`): it goes in after it.
+                true
</file context>

[blacksmith-sh]

Found 6 test failures on Blacksmith runners:

Failures

Test	View Logs
main.DeveloperIdAppTests.test_download_prints_the_active_profile_only/ test_download_prints_the_active_profile_only	View Logs
nightly must build and hand off the macOS 26 helper and Xcode 26.5 app before publishin g/ nightly must build and hand off the macOS 26 helper and Xcode 26.5 app before publishin g	View Logs
stubs: /System/Library/Frameworks/Security.framework/Versions/A/Security has no stub/ Security has no stub	View Logs
stubs: CoreFoundation.tbd: install-name /System/Library/Frameworks/ CoreFoundation.framework/Versions/A/CoreFoundation, missing 1: ['_CFRetain']/ CoreFoundation, missing 1: ['_CFRetain']	View Logs
stubs: CoreFoundation.tbd: install-name /System/Library/Frameworks/ CoreFoundation.framework/Versions/A/CoreFoundation, missing 1: ['_log2']/ CoreFoundation, missing 1: ['_log2']	View Logs
test_the_nightly_is_covered (main.NightlyStepCoverageTests.test_the_nightly_is_cove red)/ test_the_nightly_is_covered (main.NightlyStepCoverageTests.test_the_nightly_is_cove red)	View Logs

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

[teamleaderleo]

Review (codex-review-box/gpt-6-astra high, lane-place): BLOCK at 51ef316

BLOCK at 51ef316

Visual verdict: no UI changes; no images needed.

1. [P1] [blocking] Pending inbox messages can remain undelivered after a Stop aborts a newly started person turn. inbox_turn_started disarms the timer and marks them held (inbox.rs:1231), but the early cancellation return in fail_unstarted_turn (turns.rs:1511) skips inbox_turn_ended; no review or wake is scheduled afterward.

2. [P1] [follow-up] A host restart during fallback clears the person-turn marker: host_started resets turn_persons (hosts.rs:868), while the same logical turn has no new turn_started record. Adoption then restores persons: false (hosts.rs:1197), allowing agent messages to steer into the person’s turn after restart.

3. [P1] [follow-up] Messages held by a person turn can steer into a carried successor. After the original turn ends, inbox_review sees the successor and selects Next::Steer (inbox.rs:1001) without honoring held_by_person, so the notice is piggybacked instead of getting its own turn.

BLOCK at 51ef316

[cubic-dev-ai]

1 issue found across 3 files (changes from recent commits).

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. When an issue isn't valid or won't be fixed in this PR, reply in its thread with the reason and then resolve the thread. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="cmux-tui/crates/acpmux/src/hub/inbox.rs">

<violation number="1" location="cmux-tui/crates/acpmux/src/hub/inbox.rs:1017">
P2: `held_by_person` is only cleared when a notice is delivered. If the held messages are withdrawn before that, the flag stays set with nothing held. A later message in a carried turn then waits for that turn to end instead of steering in. Clear the flag when no new message remains, or check it only when a held message is still new.</violation>
</file>

<sub>Reply with feedback, questions, or to request a fix.

View guided diff | Turn on auto-fix | Re-trigger cubic</sub>

[cubic-dev-ai]

P2: held_by_person is only cleared when a notice is delivered. If the held messages are withdrawn before that, the flag stays set with nothing held. A later message in a carried turn then waits for that turn to end instead of steering in. Clear the flag when no new message remains, or check it only when a held message is still new.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. When an issue isn't valid or won't be fixed in this PR, reply in its thread with the reason and then resolve the thread. At cmux-tui/crates/acpmux/src/hub/inbox.rs, line 1017:

<comment>`held_by_person` is only cleared when a notice is delivered. If the held messages are withdrawn before that, the flag stays set with nothing held. A later message in a carried turn then waits for that turn to end instead of steering in. Clear the flag when no new message remains, or check it only when a held message is still new.</comment>

<file context>
@@ -1009,10 +1009,12 @@ impl Hub {
                             held_by_person = turn.persons;
                             if !steers
                                 || turn.persons
+                                || inbox.held_by_person
                                 || capped(&meta, turn.started_at, now)
                                 || prompt_waits(session)
</file context>

[teamleaderleo]

Review (codex gpt-6-astra high, review box): SHIP at bee457c

Round 3, a delta review of 51ef316..bee457c after the round-2 BLOCK. Reviewer's words:

SHIP at bee457c

Visual verdict: no images; Leo tries the build

The three previous findings are fixed. The configuration gate preserves inbox-disabled steering. Static review found no new lock or borrow problems; tests could not run because Rust tooling is unavailable.

1. [P2] [follow-up] inbox.rs:1017: Withdrawing every held message leaves held_by_person set: neither inbox_unsend nor the empty-inbox branch clears it. After the person’s turn ends, unrelated messages arriving during a later agent turn are consequently held until that turn ends. Clear the flag when withdrawal removes the last pending message, and cover this sequence with a regression test.

Follow-ups:

• Clear stale held_by_person after all held messages are withdrawn.

SHIP at bee457c

[teamleaderleo]

Review (codex gpt-6-astra high, review box): SHIP at ccd05c9

Delta review of bee457c..ccd05c9 (sub-session family messages still go into the person's turn; the turn_cap tests run with agentInbox off), after the SHIP at bee457c. Reviewer's words:

SHIP at ccd05c9
Visual verdict: no images; non-UI delta.

1. No findings.

Follow-ups: none

SHIP at ccd05c9

[teamleaderleo]

Red classification at ccd05c9: cmux-next checks (god files, concurrency, crash safety, l10n) fails on Checks job structure / Companion workflows. The failing test is tests/test_car_guards.py::test_cmux_next_follows_the_committed_target_graph (Lists differ: [] != ['CmuxNextAgentIntegrationsTests', ...]).

It is a base red. The same test fails the same way at the feat-cmux-next tip 2a2aa0c and passes at this PR's head. This PR changes only cmux-tui/crates/acpmux (no Swift, no workflow, no test map). The god-file lines in that job's log are warn-only and are not what fails.