[teamleaderleo]

HOLD: stacked on #26940, which is stacked on #26920 (held behind its Swift stack); this lands after both.

New Tab: a new chat starts in Auto. Auto sends the prompt to the folder's run location, and only an explicit request moves it. This is part 2 of 2 of folding #26151 ("ai-env: let new chats choose the environment automatically") into the run-location default of #26920, as one picker. It supersedes #26151.

Leo, 2026-10-28: "whatever's convenient for cloud, like that can be productionized, operationalized, we ought to just put out there"

Leo, 2026-10-28: "you can just say ... where should this chat run and stuff, and that's just about it. Like, you can determine whether it's in the cloud or otherwise"

From #26151: "why doesn't ai just choose what environment i want", "AI just choose the environment", "Environment choose cloud that would be nice"

Gallery: agent-pane.composer-auto-scope

What changes

• autoEnvironment.ts (from ai-env: let new chats choose the environment automatically #26151, reworked) decides where a prompt goes under Auto:
  • The folder's run location (Composer Cloud: a new chat's run location defaults from its folder's history #26920's folder-history default: Cloud when the folder's newest chat went to Cloud, else this Mac) holds for every ordinary prompt. This replaces ai-env: let new chats choose the environment automatically #26151's "project" fallback, the project's owning remote machine.
  • Only an explicit execution request moves it, as ai-env: let new chats choose the environment automatically #26151 already required (the prompt starts with run, build, test, deploy, execute, start, continue, work or use, and isn't negated):
    • a ready machine's full name goes to that machine (also "on my gpu-box", or "on gpu-box." at the end of a sentence);
    • an ask for cloud or remote work goes to the Cloud row (cmux NEXT: Cloud as where a new chat runs, a row in the composer's environment picker #26801: the app picks a machine and copies the folder); when the Cloud row isn't offered, it goes to a ready Cloud machine, else an SSH machine, as before;
    • an ask to run locally goes to this Mac ("locally", "on this Mac").
  • Checked in that order. An incidental "local" ("the local dataset") or "here" is no destination.
• NewTabScreen.tsx:
  • Auto is selected on a new chat's page and routes the send:
    • Cloud goes through App's Cloud start (onCloud);
    • a machine goes through onAsk with its peer;
    • this Mac goes through onAsk, even while the folder's default is Cloud.
  • A Run as row (Cloud, Sandboxed, New Cloud Machine), a machine or a saved environment ends Auto.
  • A folder pick keeps Auto, since the folder is what Auto reads.
  • A Sandboxed mode carried into the tab is a pick, so Auto is off there.
  • A Cloud refusal (attachments, Cloud not offered now) is shown for the send it refused, also when Auto chose Cloud while Run as reads Persistent.
• Auto's fallback is the run location of the folder the chip names, read on the page with folderRunLocation, so a folder picked on the page counts at once. App's Run as mode is read under Auto only to see a Sandboxed pick, because App treats the mode a tab comes with as a default too.
• A machine or saved environment picked here ends Auto and makes Run as a Persistent pick in App, so no Cloud default (now or later) takes the prompt elsewhere.
• While a Cloud start from the page is out (App cloudStarting), no send goes anywhere, so an edited prompt never also starts somewhere else.
• App.tsx:
  • onDefault puts Run as back to the folder's default; the picker's Auto row calls it;
  • cloudOffered (Cloud is offered now) and cloudStarting reach the page whatever Run as shows.
• Kept: the explicit machine, saved-environment, Cloud and Sandboxed choices. No existing choice is removed or switched off. With preview features off, nothing routes automatically.
• Dropped from ai-env: let new chats choose the environment automatically #26151: following a project's owning remote machine, because the folder default replaces it. Its two tests for that are gone too.

Before / after

Gallery renders (dark, 620 px, Chromium) from a cmux-ci Linux run at this head's tree. The gallery advisory runs once this PR's base retargets onto feat-cmux-next.

Before: the picker, no Auto	After: Auto selected, folder default Cloud

Auto as a prompt moves it or the folder keeps it: a folder that runs here, "run the tests on gpu-box".

Tests

• test/auto-environment.test.ts:
  • ai-env: let new chats choose the environment automatically #26151's cases on the new result shape;
  • ordinary and negated prompts keep the folder's run location, both this Mac and Cloud;
  • cloud and remote asks map to the Cloud row when it is offered, else to a ready Cloud machine;
  • a named machine or "run locally" moves a Cloud default;
  • a Cloud default with no Cloud row stays on this Mac.
• test/new-tab-saved-environments.test.tsx:
  • ai-env: let new chats choose the environment automatically #26151's page tests;
  • with the folder's Cloud default, an ordinary prompt (even "Fix the bug here") starts Cloud, while "Run the tests on devbox" and "Run the tests locally" do not;
  • "Run the tests in the cloud" starts Cloud, not the listed cloud-1;
  • a folder pick keeps Auto, and Auto picked again calls onDefault;
  • a Cloud pick ends Auto;
  • a carried-over Sandboxed mode turns Auto off;
  • a Cloud refusal under Auto is shown;
  • This Mac picked under Auto is a pick;
  • no send while a Cloud start is out;
  • Auto follows a folder picked on the page;
  • with Cloud not offered, a cloud ask goes to a listed Cloud machine.
• NewTabScreen.test.tsx: the base test "with Cloud picked…" now picks Cloud on the page, since under Auto the Cloud mode a tab comes with is a default, not a pick.
• cmux-ci (Linux pool) at this head's tree:
  • these tests pass, along with runLocationDefault.test.ts, cloudRun.test.ts, EnvironmentPicker.test.tsx and gallery-agent-pane-entries.test.ts;
  • bun run check is green, apart from Environment picker: an Auto row and its scope line (1/2 of the #26151 fold) #26940's format fix, which is now in.
• The two gallery-coverage.test.ts tests fail with the same names on lane/run-location-default at fdb7d09, and this stack adds no line to either list.

Risk: medium, behind preview features. It changes where a New Tab prompt goes, and only in three cases:

• the folder's newest chat went to Cloud (the Cloud default Composer Cloud: a new chat's run location defaults from its folder's history #26920 already sets);
• the prompt explicitly asks for a machine, cloud or local;
• otherwise it goes where it went before.

Agent session: agent session

🤖 Generated with Claude Code

[teamleaderleo]

Owned by an agent session · session

[teamleaderleo]

Review (codex gpt-6-astra): SHIP at df710a0

Codex gpt-6-astra (high, fast) reviewed this PR together with #26940 as one branch, each round given the full changed files:

• Rounds 1–3 covered the full branch and said FIX. Their fixes are in, listed below.
• Round 4 was a delta on the round-3 fixes: SHIP.
• Round 5 was a delta on the scope line's default: SHIP, at e155ace.
• The only commit after that, in Environment picker: an Auto row and its scope line (1/2 of the #26151 fold) #26940, is a format-only line wrap in a test.

Fixes from the rounds:

• explicit machine and saved-environment picks end Auto and make Run as a Persistent pick;
• no send while a Cloud start is out;
• Cloud only when offered;
• Auto reads the folder picked on the page;
• "local" counts only as a destination phrase, after a named machine;
• a machine name may end a sentence or follow "my";
• "my remote server" and "our VM" count as remote asks;
• one onDefault call.

Left as is (P2, no wrong send): App's Run as mode can differ from Auto's page-computed folder default after a folder pick. No send reads App's mode while Auto holds, and every explicit pick ends Auto and sets App's mode.

[blacksmith-sh]

Found 96 test failures on Blacksmith runners:

Failures

Test	View Logs
test/accent-glass-invariants.test.ts/accent-glass-invariants.test	View Logs
test/accent-glass-invariants.test.ts/accent-glass-invariants.test	View Logs
test/agent-pane-agents-pill.test.ts/agent-pane-agents-pill.test	View Logs
test/agent-pane-agents-pill.test.ts/agent-pane-agents-pill.test	View Logs
test/agent-pane-background-tasks.test.ts/agent-pane-background-tasks.test	View Logs
test/agent-pane-background-tasks.test.ts/agent-pane-background-tasks.test	View Logs
test/agent-pane-chat-unavailable.test.ts/agent-pane-chat-unavailable.test	View Logs
test/agent-pane-chat-unavailable.test.ts/agent-pane-chat-unavailable.test	View Logs
test/agent-pane-hairline-focus.test.ts/agent-pane-hairline-focus.test	View Logs
test/agent-pane-hairline-focus.test.ts/agent-pane-hairline-focus.test	View Logs

...and 86 more test failures. View all on Blacksmith

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

[cmux-bot]

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

Probably yours: changelog-pictures.test.tsx, layout.test.ts, agent-pane-queue-steer.test.ts and 1 more fails here but not on main's latest full suite. Seen on other PRs too (likely flaky): accent-glass-invariants.test.ts, agent-pane-agents-pill.test.ts, agent-pane-peer-chat.test.ts and 25 more.

CI failed on df710a004b (run 40763554892 attempt 1): 4 code, 1 unknown.

Job	Verdict	Why
web / react-apps-base-check (2)	code	a web test failed
web / react-apps-base-check (3)	code	a web test failed
web / react-apps-base-check (1)	code	a web test failed
web / react-apps-base-check (4)	code	a web test failed
web / react-apps-check	unknown	no known signature; failed step: Check webview test shards

Failures:

• seen on other PRs accent-glass-invariants.test.ts no orange or system accent and no opaque glass on the gallery stages: failed (job)
• seen on other PRs agent-pane-agents-pill.test.ts by default the running agents sit in the tray above the composer, not in the transcript: failed (job)
• seen on other PRs agent-pane-peer-chat.test.ts a chat on an SSH machine > counts its running turn from the prompt and offers no machine switch mid-...: failed (job)
• seen on other PRs agent-pane-queue-row-click.test.ts queued row with a chip's menu open (chromium) > one press on more closes the model menu and does wha...: failed (job)
• seen on other PRs agent-pane-queue-row-click.test.ts queued row with a chip's menu open (webkit) > one press on more closes the model menu and does what ...: failed (job)
• seen on other PRs agent-pane-transcript-spacing.test.ts meta rows hug their message and the rhythm has no empty bands: failed (job)
• new in this PR changelog-pictures.test.tsx a change's picture opens enlarged on a click, and Close closes it: failed (job)
• seen on other PRs chat-inspector-overlay-toggle.test.ts chromium: over the chat the Inspector is as tall as its sections, and scrolls when they are taller: failed (job)
• new in this PR layout.test.ts only page.css imports the page's sheets: failed (job)
• seen on other PRs agent-pane-queue-scroll-bottom.test.ts chromium > a prompt the fold hides, run elsewhere, leaves no room as the stack shrinks, opens and cl...: failed (job)
• ... 22 more in the job logs

Matched log lines

web / react-apps-base-check (2): (fail) no orange or system accent and no opaque glass on the gallery stages [709355.81ms]
web / react-apps-base-check (3): (fail) only page.css imports the page's sheets [32.23ms]
web / react-apps-base-check (1): (fail) the summary card lists the chat's background tasks: running ones first, the finished folded [1195.28ms]
web / react-apps-base-check (4): (fail) chromium > the gutter beside every part of the stack scrolls the transcript, and the part keeps the pointer [8766.30ms]

Not re-run automatically: web / react-apps-base-check (2), web / react-apps-base-check (3), web / react-apps-base-check (1), web / react-apps-base-check (4), web / react-apps-check are not machine failures.

Written by scripts/ci/classify_failures.py (ci-failure-attribution.yml); signatures are its SIGNATURES table. A machine verdict is the runner's fault, not this PR's; yours means the failing file is one this PR changes, also red on main that main's latest full suite fails the same way, seen on other PRs that it failed on another pull request's run lately.

CI fast guards failed on df710a004b (https://github.com/manaflow-ai/cmux/actions/runs/40763554892). It does not block the merge; a red guard merged into main breaks it for every open PR.
The log named no failed step; see the run.

Agents: python3 scripts/ci/guard_attribution.py fix applies the mechanical fixes locally.

[teamleaderleo]

Folded into #28687 (on top of #28686, original commits kept). Review and landing continue there; the owner closes this one.