PR #23883 - manaflow-ai/cmux

What

The Chief Settings popover (the header's name pill) was cramped pickers, raw ids (claude-sr2-pipe, claude-opus-5-5) and a Diagnostics text dump (#89 item 2). It is now a settings page of grouped glass cards, wider (380 pt), each card only where it has something real:

Identity: avatar (menu to change it), name, where the Chief runs (display only), and the pet rows (Show Chief Pet, its character) that used to sit in a separate Appearance section.
Brain: Backend, Model, Effort and Speed, each with a one-line "what this means" under it. Every value reads by a name: official engines keep their titles, models their catalog names, and any other id is made readable by the new ChiefNames (claude-sr2-pipe → "Claude (SR2 Pipe)", claude-cr → "Claude (CodeRouter)", claude-opus-5-5 → "Claude Opus 5.5"). The raw id stays only in the popup item's tooltip.
Memory (this Mac's brain): what the Chief remembers, and Show Memory.
Diagnostics: a styled last turn: the engine by name, when it ended, then four measures (time, cost or tokens when the harness gives no price, cache %, tool calls with failures), the turn's error if it failed, Show Traces, and the recent replies behind the closed disclosure.

Behavior, Permissions and Notifications have no per-Chief setting on base, so they get no card (follow-up list below). The cards use the sidebar cards' glass (CardSurfaceView: a veil with the top-lit rim, opaque under Reduce Transparency) over the popover's overlay glass, and the shared Metrics and Typography tokens.

Data: chief.engine.get's recent already sends ts, effort, error and usage output (engine_control.rs turn_summary); ChiefEngineReport.Turn now decodes them when present, so an older brain still decodes. This Mac's trace lines give the same fields.

debug.chief_settings (DEBUG builds) shows a fixed brain answer in the open popover, since capture minis run no Chief brain; the chief-settings tour uses it.

No settings wire op here (#22639 has one). "Where it runs" stays display only.

Tests

Red, then green:

Red: fleet step c4f2516bf58298d895742b32 (chief-settings-cards-red-fdc6b7f0, the new tests on the old code): 3 of 3 suites failed. ChiefBackend.title(for: "claude-cr") gave claude-cr (want "Claude (CodeRouter)"), the Backend popup listed claude-sr2-pipe with no tooltip, and the last turn read claude-sr2-pipe · claude-opus-5-5 with no "Opus 5.5".
Green: fleet step 86e1ddd615a2e8be294c4763 (chief-settings-cards-green-5de20346 at 5de2034631): all 4 suites passed (ChiefNamesTests, HomeChiefTurnSummaryTests, HomeChiefSettingsViewTests, ChiefBackendTests; 6, 7, 9 and 9 tests). After it: e699b348 renames ChiefNames to String.chiefHarnessName / String.chiefModelName (same assertions, for the iOS package conventions gate), and a20bcef5 / 8a621b49 add acpmux's harness names and drop dashed snapshot dates, with their tests.

New: ChiefNamesTests (harness and model names, catalog name first, the Backend popup's tooltip), HomeChiefTurnSummaryTests (newer and older Turn decoding, cache %, cost, tokens, seconds, the card's last turn). Updated: ChiefBackendTests, HomeChiefSettingsViewTests (names instead of ids, the name field's edge inside the Identity card).

Capture

No capture: the media-tour run (cmux-next media, run 39706513337) is still queued and has posted no stills, and a capture from cmux16 could not lease capture-host. cmux-ci capture stops with "needs a manaflow-ai/cmuxterm-hq checkout (build-fleet/pr-capture and capture-host)". The chief-settings tour (scripts/cmux-next/media/tours/chief-settings.json, using debug.chief_settings) will post the stills when the media run gets a host.

Follow-up: settings the backend still lacks

Behavior, Permissions, Notifications: no per-Chief setting exists on base, so they get no card until the brain has one.
Where it runs: display only. Changing it needs a settings wire op (#22639 has one).

Changelog

Chief Settings is now a page of grouped cards with readable model and backend names and a last-turn summary (time, cost, cache, tools).

🤖 Generated with Claude Code

Agent session: cosmic-gray

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

Summary by cubic

Rewrites the Chief Settings popover from cramped pickers showing raw ids and a Diagnostics text dump into a settings page of grouped glass cards — Identity, Brain, Reactions, Memory, Diagnostics — each shown only where it has a per-Chief setting to offer. Every engine and model now displays a readable name (String.chiefHarnessName/chiefModelName): official engines keep their titles, acpmux's built-in harnesses read as acpmux names them (Pi, OpenCode, Gemini CLI, Aider, Muse Code), a known model keeps its catalog name, and a model id drops a trailing snapshot date (only a real date). The raw id stays only in tooltips, and Diagnostics shows a styled last turn: engine by name, when it ended, time, cost (or tokens when the harness gives no price), cache %, and tool calls, plus the error when a turn failed and the recent replies behind a disclosure.

Compatibility and testing

ChiefEngineReport.Turn decodes the newer ts, effort, error, and usage output fields only when present, so an older brain's summary still decodes.
The popover scrolls when the cards exceed the room on screen, so tall content like the pet rows and an open Diagnostics stays reachable.
New debug.chief_settings (DEBUG builds) shows a fixed brain answer in the open popover, since capture minis run no Chief brain; the body lives in HomeChiefSettingsView.handleCaptureControl in CmuxNextHomeEngine.
New tests cover naming (including acpmux's harness names and dashed snapshot dates), turn decoding from older and newer brains, and updated existing tests for the named values and scrolling.