PR #28297 - manaflow-ai/cmux

Problem

acpmux #28182 answers _acpmux/sessions and _acpmux/watch with no limit in byte-bounded pages when the whole list is over its largest message (~14 MiB): the first page carries a nextCursor, and the rest is read with _acpmux/sessions {cursor, ...same params}. One daemon's list was 17 MB with tools and was refused as message.too_large. The Chief's Rust client reads only the first reply, so against a paging daemon it would come up on part of the list and miss sessions on later pages (its own leftover turn and compactor sessions, a peer's session find/session fall back to the list for).

What changed

optchat-chief/src/acpmux.rs: one helper, whole_list, reads every page (the reference is acpmux's Client::whole_list in #28182): with no limit it follows nextCursor by _acpmux/sessions with the same params less the watch's enabled, and joins sessions. A cursor that comes again is refused with an error, so a bad daemon never loops it. With a limit, one plain request. A list that fits is one reply with no cursor, so older daemons work unchanged.
Every _acpmux/sessions / _acpmux/watch call in optchat-chief without a limit goes through it: the link's watch, sessions() (the list fallback for session/find and agents), and agents.rs's bounded sessions_bounded (each page keeps its 1 s bound).
Tests: tests/acpmux_session_info.rs adds a fake daemon that answers in two pages (the_link_comes_up_on_every_page_of_the_watch_list, a_session_on_a_later_page_of_the_list_is_found, a_cursor_that_comes_again_is_refused); src/acpmux.rs adds a unit test that a list with a limit is one request.

The test commit comes first; the fix follows.

Red / green

Linux fast checks (cargo test --no-fail-fast of the whole optchat-chief crate):

Red, at the test commit 93fbd17f4f47 (base plus the new tests only): https://github.com/manaflow-ai/cmux/actions/runs/41047905428. The three new tests fail as expected: the_link_comes_up_on_every_page_of_the_watch_list (left: ["s-1"], the link came up on page one only), a_session_on_a_later_page_of_the_list_is_found (listed on page two: not found), a_cursor_that_comes_again_is_refused (unwrap_err() on Ok(None)). One more fails: harness::the_chiefs_turn_and_compactor_sessions_carry_cmux_chief_and_children_do_not (harness.rs:445, left: 5).
Green, at the head 952bc6014db5: https://github.com/manaflow-ai/cmux/actions/runs/41049652619. 114 test binaries pass, 690 tests passed. All three new tests pass, and so does the whole_list_tests unit test. One binary fails, --test harness, on the same test.

Failing tests at the head, compared by name with the base:

Test	Head	Base (the merge-base 4a0b8f01e175 itself)	Verdict
`harness::the_chiefs_turn_and_compactor_sessions_carry_cmux_chief_and_children_do_not`	fails, harness.rs:445:9, `left: 5 right: 4`	fails, harness.rs:445:9, `left: 5 right: 4` (run 41050748089, tested tree 943c5e5b = the merge-base's tree; the only failure of 28 in `--test harness`)	base red: same assertion on base, not a crash. The turn session's tags gained `cmux.chief.participant` in #27935 and the test still counts 4; this PR touches neither

No crash, signal or cut-short run in either run.

No capture: no UI change.

Observed / Not run

Observed: the red and green Linux runs above, and the base run of the harness test. Fresh-home launch at the head 952bc6014db5 on a fleet Mac (build b0d14eaaee6d99eb3f453260): PASSED. Home in 6.4 s, up 30 s, no crash; the Chief thread "Chief - Main - This Mac" is listed and its intro rendered, with no error banner (run with --no-status, so no commit status was posted). An independent Codex (gpt-6-astra, high) review of 4a0b8f01e17..952bc6014db5 gave VERDICT: SHIP, with no findings.

Not run: a run against a real daemon that pages (#28182 landed after these tests ran).

Changelog

Chief reads a paged acpmux session list whole, so it sees every session on a daemon whose list is too large for one message.

Agent session: agent session

🤖 Generated with Claude Code