[teamleaderleo]

Summary

Since #27686, another agent's plain message goes to a chat's inbox, and the chat hears of it through a notice. That includes the very first prompt a new chat gets: the task lane-place --prompt-file or acpmux new with a prompt sends as an agent. A new lane was getting "[inbox] New messages from other agents in 1 thread…" where its brief should have been, so it read a thread instead of its task. #28369 fixed this for review boxes only.

With this change a chat's first work runs as its prompt, whatever kind of chat it is:

• hub/inbox.rs first_work: the chat never ran a turn, none is running, and nothing waits in its queue or inbox. It takes a hold (Inbox::first) under the inbox lock first and reads the chat after, so of two first messages sent at once only one runs as the prompt (the other goes to the inbox), and a first work that already ran is seen. The FirstWork guard keeps the hold for the life of that prompt and releases it on drop.
• inbox_route: another agent's message that is a chat's first work skips the inbox and runs as that prompt.
• inbox_review: while a first work holds the chat, or waits in its queue with the firstWork mark (one a person's prompt went ahead of, or one a restart rebuilt with no hold), and no turn runs, no notice goes ahead of it. The inbox looks again after the wake debounce, until the first work's turn takes the notice, its hold is let go, or it leaves the queue.
• hub/turns.rs: the sender of the prompt that starts a chat's first turn becomes its delegator, set as the turn starts (for a prompt that held the first work, even when a person's prompt went first), not when the message is routed. A first work's provenance carries firstWork: true, which its queue entry keeps across a restart and no client can send (prompt_provenance.rs keeps only the fields it names): a refused prompt names no delegator, and a message the inbox took while the first work waited to start does not stay the delegator. It is read from that prompt's own provenance, so other agents' messages that join its turn as a digest (digest.rs) don't hide it.
• Messages after the first, and any message to a chat that has run a turn, keep acpmux: every chat's inbox; another agent's plain send is never a turn of its own (A1) #27686's inbox behaviour, as before.
• The module docs at the top of inbox.rs say why.

Follows #28369 and #27686.

Testing

Tests in tests/hub_integration/agent_inbox.rs (the test commit comes before the fix):

• an_idle_chat_runs_its_first_work_and_wakes_for_a_blocker_only (was …wakes_for_a_blocker_and_for_its_first_work_only): the first message runs as prompt t-1 with no notice, and its sender gets end_turn. The next message goes to the inbox, and a blocker still wakes the chat with one notice.
• only_the_first_message_to_a_new_chat_runs_as_its_prompt: a second message that comes while the first one's turn runs goes to the inbox.
• of_two_first_messages_at_once_only_one_runs_as_the_prompt: two messages sent at once to a new chat; exactly one goes to the inbox, and the other's turn ends with end_turn.
• src/hub/first_work_tests.rs a_message_the_inbox_takes_before_the_first_work_starts_is_not_the_delegator: the first work is paused after admission (prompt_start_seam), another session's message goes to the inbox, and after release the delegator is the first sender. a_message_that_joins_the_first_works_turn_is_not_the_delegator does the same with a queued message that joins the first turn as a digest. Two more pause the first work right after it takes the chat (first_work_seam, tests only): a_blocker_that_comes_before_the_first_work_queues_waits_for_it (no notice turn runs before the brief) and a_persons_prompt_before_the_first_work_queues_leaves_its_sender_the_delegator (the first work waits in the queue behind a gated person's turn, and its queued record carries the mark). a_first_work_a_restart_rebuilt_keeps_notices_behind_it restarts a durable hub with that first work queued and a blocker in the inbox, and checks that no notice goes ahead of it.
• tests/send_async.rs a_no_wait_send_from_a_session_names_its_proven_sender runs a turn on its receiver first, so the message it checks still goes to the inbox.
• The review-box test from acpmux: review boxes get no inbox notice and no other agent's steer #28369 now runs a turn on its ordinary chat before checking that the message goes to the inbox.

Run at head 79388b6 (cmux-ci fast checks, run 41103630462); base comparisons at base tip c5359cd (run 41097908556) and base 364c30d (run 41092748743):

Command	Head 79388b6	Red check
cargo test -p acpmux --lib first_work	5 passed	with inbox.rs from the test commit 6645c9a: a_first_work_a_restart_rebuilt_keeps_notices_behind_it fails; with the source of 9fb7f36: a_persons_prompt_before_the_first_work_queues_leaves_its_sender_the_delegator fails
cargo test -p acpmux --test hub_integration	617 passed, 3 failed	the same 3 fail on base tip c5359cd, at the same lines (below); with base 364c30d source, an_idle_chat_runs_its_first_work_and_wakes_for_a_blocker_only, only_the_first_message_to_a_new_chat_runs_as_its_prompt and of_two_first_messages_at_once_only_one_runs_as_the_prompt fail
cargo test -p acpmux --lib inbox	23 passed
cargo test -p acpmux --test send_async	16 passed
cargo clippy -p acpmux --all-targets -- -D warnings	clean
panic ratchet, as CI runs it (-p acpmux -p cmux-tui --lib --bins)	exit 0; acpmux indexing_slicing matches the base tip

Observed / Not run

• Observed: the table above; rustfmt clean. Reviewed by a different model (gpt-6-astra): SHIP at 79388b6.
• Base reds, each compared by test name against base tip c5359cd, where each fails the same way: inbox_digest::a_digest_refused_before_its_turn_puts_every_message_back_and_a_restart_keeps_them (inbox_digest.rs:666), carried_turns::a_carried_turn_after_a_carried_turn_has_its_own_cap (carried_turns.rs:502, no cap notice), agent_inbox::an_inbox_message_from_before_the_review_tag_never_steers_into_the_review (agent_inbox.rs:73, one of three prompts unanswered). None is a crash. The last passed at 3e8ff30 and fails on base since acpmux: preserve retries and adopted prompt outcomes across restart (takes over #26981) #27921.
• inbox_digest::messages_drained_twice_follow_the_newest_digest_at_a_restart failed once in four head runs and in neither base run. Its chat is mid-turn (busy) when the messages arrive, so this PR's first-work path never applies there.
• CI reds outside acpmux: release-notary fails on release-notes/next/highlights/chief-name-in-agent-chats.md, which this PR does not touch; Guard status rolls it up. The CI panic ratchet's acpmux counts are base drift (above).
• Not run: a real lane-place launch against a daemon built from this branch.

Changelog

Fixed: a new acpmux chat's first task from another agent runs as its prompt, not as an inbox notice.

Agent session: agent session

🤖 Generated with Claude Code

[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.

[github-actions]

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

Review 79388b6: 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=pr28413 -f ref=lane/first-prompt-not-inbox -f platform=macos -f backend_mode=local

[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
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_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.}

[teamleaderleo]

Owned by an agent session · session

[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.

[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]

Review (the owner lane's independent review, at its land request): SHIP at 79388b6

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.

[github-actions]

Shipped in:

• cmux NEXT nightly 1.0.0-nightly.4111325355101 (df9b52b7d0)
• cmux NEXT nightly 1.0.0-nightly.4111671381901 (a0fc9a9a63)
• cmux NEXT nightly 1.0.0-nightly.4111911568201 (214b514c8c)