[lawrencecchen]

Fixes the base red from #31936.

What

fork_through::usage_stats::the_session_list_and_info_carry_what_the_chats_transcript_says (added by #28015) fails on every feat-cmux-next push since #31936 landed:

assertion `left == right` failed: {"autocompact":"capped","autocompactCap":400000,...}
  left: String("capped")
 right: "none"

#31936 makes acpmux start every Claude Code session whose person set no window with CLAUDE_CODE_AUTO_COMPACT_WINDOW=400000, and report it as autocompact: capped, autocompactCap: 400000. That is the intended behavior. This test still expected the old none for a chat with an empty Claude home and a temp-dir folder. #31936 updated the unit test (usage_stats_tests) but not this integration test.

The product is right, so only the test changes:

• It expects autocompact: capped at harness_limits::CLAUDE_AUTOCOMPACT_WINDOW (not a copied literal).
• A second chat asks the fake Claude for CLAUDE_CODE_AUTO_COMPACT_WINDOW, so the cap the list reports is the one the harness really runs with. If the spawn rule and the report rule drift apart, the test fails.

Changelog

none

Testing

Command: cmux-ci run --class isolated --script scripts/ci/cmux-tui-rust-check.sh --os linux --arg test --arg the_session_list_and_info_carry_what_the_chats_transcript_says --ref SHA, 5 runs each side.

• Base 741fb0170111 (red): failed 5/5 (left: String("capped"), right: "none"): jobs 2c523624a919409cb0e7f492, 423936e8ef312ded3398a7df, 3ee38634d49f35d730d60c0f, a27867ef112138f5fcf1f231, 6cf826e7c58ec8e055804844
• Head (green): passed 5/5 at ecd406b80a72795f23d6fd05a6979393276672f2 (1 test run, 1 passed): jobs 53dff0a15d564535f9c0dd22, 50cd0d243239c3f73ab96ba1, 9e54562e7f885c1b0eee7fe9, 4aada263df514c5adddda909, ad5ec28a8eb0173a8f584896
• Earlier base 7233516b7cc4: job 0a95b7d3fe634947d7023898 failed the same way, also on its alone-rerun.

Observed / Not run

• Observed: the fleet runs above (Linux, Blacksmith). prepush-gate.py: no new violations.
• Not run: macOS. The test has no cfg(target_os) gate and the cap rule does not depend on the OS.
• Not run: clippy (the feat clippy red in cmux-tui-core/src/mux/tests/read_paths.rs is another agent's fix); the test mode lane does not run clippy.

risk: low (test only)

[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 ecd406b: 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=pr32312 -f ref=base-red-usage -f platform=macos -f backend_mode=local

Shipped in:

• cmux NEXT nightly 1.0.0-nightly.4145363708701 (c81ea581fe)
• cmux NEXT nightly 1.0.0-nightly.4145643176801 (627d40b6dd)
• cmux NEXT nightly 1.0.0-nightly.4145866056201 (d87662ff9d)
• cmux NEXT nightly 1.0.0-nightly.4145971823401 (507438d2a9)
• cmux NEXT nightly 1.0.0-nightly.4146140953301 (ad51b07c27)

The feat-cmux-next push run https://github.com/manaflow-ai/cmux/actions/runs/41454213833 at 986d466 failed: cmux-next checks (god files, concurrency, crash safety, l10n).
Those jobs last passed at f21d85e. This pull request is one of 5 merged since: #32305, #32312, #32182, #32265, #31743.
If the failure is in your change, fix forward on feat-cmux-next. Pull requests run only the tiers their paths reach (docs/ci/cmux-next-tiers.md); label a batch PR full-ci to run them all before merging.

[lawrencecchen]

Review (claude-opus-subagent): SHIP at ecd406b

Checked by reading only, with no local build. The test should change, not the product. #31936 meant every acpmux-started Claude session where the person set no window to run with CLAUDE_CODE_AUTO_COMPACT_WINDOW=400000, and to report capped/400000. lift_claude (spawn) and read_usage (report) both use the same rule, claude_autocompact_unset. This test has an empty CLAUDE_CONFIG_DIR, a temp-dir cwd and no flag or env, so capped at the constant is the correct expectation.

The test compiles as far as I can tell. acpmux::harness_limits is pub mod, and CLAUDE_AUTOCOMPACT_WINDOW is a (&str, &str). Value has PartialEq with &str, u64 and String. hub.session_summary(&hub.resolve(..).unwrap()) has the same shape as the existing call in transfer_guard.rs.

The probe chat runs after every earlier assertion. The usage_stats events are filtered by the first session's id, so the probe does not change tokens24h or the event count.

The preview is the raw agent_message_chunk text. Each turn resets it, it is not masked, and short_text(160) does not cut the 38-character reply. fake_claude.py streams NAME=value for env: even with FAKE_CLAUDE_TRANSCRIPT_USAGE=1, which changes only the stored transcript.

The probe compares the reported cap with the env the spawned harness actually received, so the test fails if the spawn rule and the report rule drift apart.

Non-blocking: an ambient CLAUDE_CODE_AUTO_COMPACT_WINDOW or CLAUDE_AUTOCOMPACT_PCT_OVERRIDE in the test runner's env, or a .claude/settings.json in a temp-dir ancestor, would change the result. The old none expectation was exposed to the same ambient env, so this is not a regression.