Repository navigation
Model menu: a started chat's harness is fixed (no rail, no Add harness…) - #22226
Conversation
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…switch Once a chat has a message, the model menu lists its harness's models alone: search, the models and More models, with no harness rail, Starred tab or Add harness…, in both the full picker and the compact menu. New Tab and chats with no message keep the whole rail. Debug Settings' "Allow switching harness in a started chat" (agent.harnessSwitchInStartedChat, off) keeps the rail and the switch in started chats. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
cmux status (one comment, edited in place; each line is updated by the job that owns it) Review App of this head (app.zip, kept 1 day)
Unclear: CI failed on
Not re-run automatically: Written by red: web / Web status; base red: the same failure on the base dde0ebb; does not block
Agents: |
|
Bead |
 [Medium impact] The PR changes two bounded user-facing workflows: agent model picking and Markdown reading settings. The PR should not merge until the webview build and width-saving failures are fixed. Findings
SummaryStarted chats hide the harness rail and Add harness option. A Debug Settings toggle restores cross-harness picks.
Reviews (3) · Last reviewed commit: "chore(next): PR autofix at 15b21d38fb92" · Reviewed by Greptile |
![greptile-apps[bot]](https://avatars.githubusercontent.com/in/867647?s=60&v=4){kind=small}
| const harnessFixed = props.harnessFixed === true && ownHarness.length > 0; | ||
| const harnesses = harnessFixed ? ownHarness : allHarnesses; |
There was a problem hiding this comment.
If a started chat's picker opens before its harness is listed, the user can select Starred in the temporary rail. When the models arrive, harnessFixed hides the rail but leaves selectedHarness on STARRED. The picker then shows only favorites, possibly an empty list, instead of all models. The user must search or reopen it to recover.
Make starredView false whenever harnessFixed is true, and add a test where models arrive while Starred is selected.
The Harnesses page browser test opens Add harness… from a started chat, so it turns on the started-chat harness switch first, through the host's composer entry point. Drop an assertion that could never fail, and reset the composer settings after unmounting. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
![cmux-autofix[bot]](https://github.githubassets.com/images/gravatars/gravatar-user-420.png){kind=avatar}
- vp fmt webviews/src/pages/markdown/settings.ts Autofix-Run: https://github.com/manaflow-ai/cmux/actions/runs/39346350706
| lineHeight: { key: "markdown.font.lineHeight", default: 1.6, min: 1, max: 2.4, step: 0.05 }, | ||
| paragraph: { key: "markdown.spacing.paragraph", default: 1, min: 0, max: 3, step: 0.1 }, | ||
| codeSize: { key: "markdown.code.blockSize", default: 14, min: 8, max: 32, step: 1 }, | ||
| codeScale: { key: "markdown.code.scale", default: 0.875, min: 0.6, max: 1.25, step: 0.025 }, |
There was a problem hiding this comment.
Renaming codeSize to codeScale leaves ReadingSettings.tsx using values.codeSize, r.codeSize, and actions.set("codeSize", value). These references no longer type-check, so the webview build fails. Without typechecking, opening Reading Settings throws when Slider reads range.min from the missing range. Update the callers and tests together with this rename.
| */ | ||
| export const MARKDOWN_READING = { | ||
| width: { key: "markdown.layout.width", default: 72, min: 40, max: 160, step: 2 }, | ||
| width: { key: "markdown.layout.width", default: 72, min: 30, max: 1000, step: 1, bigStep: 10 }, |
There was a problem hiding this comment.
The width slider now offers 30 to 1000, but MarkdownViewerSettingsSchema still accepts only 40 to 160. Choosing 300 widens the page immediately because setPreference updates it before saving. The host then refuses the write through SettingsController.setSetting, so the chosen width disappears on reload. Update the native range alongside this range.
| if (typeof code.scale === "number" && Number.isFinite(code.scale)) | ||
| put("--cmux-md-code-scale", String(numberIn(code.scale, MARKDOWN_READING.codeScale))); |
There was a problem hiding this comment.
Setting markdown.code.scale has no visible effect. The new size calculations exist only in MARKDOWN_STYLE_DEFAULTS, which is not applied to the page. styles.css still sets inline code to 0.88em and blocks to 14px, so the emitted --cmux-md-code-scale is unused. Update the actual stylesheet defaults to use the scale so this setting changes the rendered text.
|
Shipped in:
|
…ollow their changes (#22516) * test(markdown): Cmd-A in the selection test is ControlOrMeta, so Linux CI selects all Meta+A on the Linux runners' Chromium and WebKit is not ProseMirror's Select All (Mod is Ctrl there), so the selection stayed empty. The other browser tests already press ControlOrMeta. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * test(changelog): the Updated card's text selection follows its headline layout e98d3a1 made the top change the card's headline, bulleted the rest and spelled the eyebrow 'Nightly · Oct 19' (the raw build moved to its tooltip). The test still dragged the first bullet for the top change, read rows 1 and 2 of a two-row list, and looked for the build. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * test(status glyphs): a wait is the attention dot in both sets 3207401 and 301bd27 replaced the talk bubble with the attention dot for a wait (in the warning color). The Codex set now draws the dot twice: the test tells the wait's from done's by position and color, not shape, and the static sheet's 2x edge probes follow the disc. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * test(narrow layouts): give the file 300 s, as its seven pages take about 195 s The default 120 s per-file limit in run-webviews-tests.sh killed it on the CI runner after six of its seven tests; the runner's own header line raises the limit for this file. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * gallery: model-menu-starred opens before the first send; the markdown host keeps table widths 75bc940 (#22226) gave a started chat's model menu no harness rail and so no Starred tab; the variant still showed a finished chat, and its play found no Starred option. It now shows a new chat, where the rail and Starred stay. 452cce0 moved a file's table column widths from the page's storage to the host (cmux.markdown.setTableWidths, back in the config as tableWidths). The gallery's markdown host had neither, so a reload forgot the widths. It now keeps them in the tab's session storage. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
…22513) * fix(next): the Starred models gallery state is a new chat (base red) #22226 fixed a started chat's harness, so its model menu has no rail and no Starred tab. The composer gallery's model-menu-starred state still drew a finished chat, so its play failed with no element for option "Starred" and composer-type-scale went red at the base for every PR. The state now opens a new chat, where the rail and its Starred tab live. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * test(markdown): table keys run as on the Mac, so Linux CI's Cmd-A reaches Select All The third Cmd-A in a table should fall through to the editor's Select All (tableKeyDown returns false once the whole table is selected). ProseMirror's keymaps decide what Mod means from navigator.platform when they load: jsdom's platform is "" and bun's is the host's, so on the Linux runners Mod was Ctrl, Cmd-A matched nothing and the CellSelection stayed. The test has been red on Linux since it landed (ade5a7d / 8ba372b) and passes on a Mac. The page now says it is a Mac before ProseMirror loads (the editor only runs in a Mac WebView), and the two ProseMirror value imports load after that. No editor source changes. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * chore(webviews): vp check passes again (formatting; one stray bun-types reference) `vp check` in webviews failed on formatting in five settings files from the Chiefs-in-Settings and host-sections changes; `vp check --fix` reformats them with no code change. Behind the formatting, the type check failed 10 times: `/// <reference types="bun-types" />` in new-tab-list-motion-render.test.tsx (0f60d16, #22397) pulled in bun's devserver ImportMeta.hot, whose `on(event, () => void)` conflicts with Vite's typed `hot.on` in every dev module. The tsconfig already loads bun's types; the line is removed. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * test(markdown): say the table keys' Mac override needs a process of its own Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * test(next): New Tab first-frame test fits the Cards page's three cards (base red) new-tab-warm-first-frame.test.ts expected a device chat on the first frame next to three acpmux sessions. Since 9215b25 (Cards New Tab template, the new default) the default page shows three chat cards, sessions first, so the device chat never had a slot and the 8 warm-open tests failed in Chromium and WebKit. The test now lists two sessions and the device chat, so it still proves a warm open paints the host's late device chats from the first frame. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(next): the webviews check gate passes its text-focus and lint steps (base red) `bun run check` in webviews stopped at check:text-focus on three Swift text fields, and `vp check` (denyWarnings) failed on 7 lint warnings. - WorkspaceTodoPageView / WorkspaceTodoRow: the TextFields get .focusEffectDisabled(); the done mark uses Palette.accent instead of Color.accentColor (check-system-blue owns that file on this PR). - AddressBarView: text-focus-allow on the read-only bar's isSelectable; AddressField is a ChromeTextField, whose init sets focusRingType = .none. - ChatDisplaySettings, NowPlayingPlayer, NewTabScreen, ThreadMinimap: line-level oxlint suppressions with reasons for the 7 warnings (controls already inside a <label>, a datalist option, role notes, and the minimap's every-commit layout effect). No behavior change. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * webviews: status-glyphs test expects the attention dot for a wait 3207401 and 301bd27 (Lawrence, 2026-10-20) replaced the talk bubble with the attention dot for waiting, in the Codex-like set and in the Static glyphs. The test still expected the bubble, so it went red on every PR. The Codex test now expects the dot for waiting, tells it from done by color, and the 2x edge probes follow the new disc. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(next settings): the Density card's 3 lint warnings get line-level suppressions (base red) The new Density card (bc4dce8) added two inputs inside a <label> and a role="radio" tile; vp check denies warnings, so the check gate went red again. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * test(markdown): Cmd-A and kept column widths pass in the Linux browser lane (base red) markdown-selection-drag: the context reports navigator.platform as MacIntel, as the Mac WebView does. ProseMirror binds Mod (Select All is Mod-a) from the platform when its keymaps load, so on Linux Mod was Ctrl and Cmd-A selected nothing. markdown-table-obsidian: the gallery's markdown host answers cmux.markdown.setTableWidths and sends the saved value back as tableWidths in the config, with the frame's localStorage standing in for the file's extended attribute. Since 452cce0 the page keeps widths through the host; the mock host had no such op, the write failed quietly and a reload lost the width. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(next): the find pill has a reduced-transparency fallback (check-layers base red) d3a9df0 gave the Liquid Glass find pill a backdrop-filter with no prefers-reduced-transparency or data-reduce-transparency fallback in its own file, so check-layers failed the webviews check gate. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * chore(next): PR autofix at 8e6179a - schemas/defaults/cmux-next.json Autofix-Run: https://github.com/manaflow-ai/cmux/actions/runs/39416502321 * chore(next): PR autofix at 6b3bea6 - schemas/defaults/cmux-next.json Autofix-Run: https://github.com/manaflow-ai/cmux/actions/runs/39418461225 * chore(next): PR autofix at e4b0d16 - schemas/defaults/cmux-next.json Autofix-Run: https://github.com/manaflow-ai/cmux/actions/runs/39419009274 * test(agent pane): the Cards New Tab template's pills are deliberate agent-pane-no-capsules has failed at the feat-cmux-next tip since 9215b25 (Cards New Tab template, u2851). That template is a pill design on purpose, from the user's scrawl of Aside's New Tab page: a round field, a round Search | Ask switch and a round selected Chats tab, all at border-radius 999px. #22285 later scoped the field and switch rules to [data-pill]. The test now lists those five Cards selectors as deliberate capsules beside the composer chips. Every other template and control is still held to small-radius rects. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * chore(next): PR autofix at 84b7643 - schemas/defaults/cmux-next.json Autofix-Run: https://github.com/manaflow-ai/cmux/actions/runs/39427646005 * test(markdown): the kept-width test reloads only after the host has the width On #22513 the resize-width test failed once in CI chromium (Received 90): the page sends cmux.markdown.setTableWidths after the drag and the gallery host stores it a task later, so a reload can beat the write. The test now waits for the stored width before it reloads, and fails loudly if it never arrives. It passes 4/4 on big-red. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * test(next): whats-new text selection follows the Updated card headline layout (base red) e98d3a1 made the first change the card headline and bullets only the rest, and the eyebrow now reads "Nightly · Oct 19" with the build in its tooltip. The drag test selects the headline, the double-click and wrap tests index the bullets from 0, and the eyebrow check looks for the day. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * test(markdown): the selection test keeps the base ControlOrMeta fix alone #22516 landed ControlOrMeta+a for this test. With this branch Mac platform init script on top, Mod became Cmd while ControlOrMeta pressed Ctrl on Linux, so Cmd-A selected nothing again. The base version passes 8/8 on big-red. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> Co-authored-by: cmux-autofix[bot] <cmux-autofix[bot]@users.noreply.github.com>
Summary
In a chat that already has a message, the composer's model menu drew a harness rail holding only that chat's harness (plus the Starred tab), then offered "Add harness…" as its last row. The chat can't switch harness (it has been locked since 2026-10-13), so a rail of one icon followed by an offer to add another read as a broken switcher.
Now a started chat's harness is fixed, as in Claude Code:
turnCount > 0or any row): the full picker is search, that harness's models and More models. No harness rail, no Starred tab, no Add harness…. The compact menu (agent.modelMenu) gets no rail either; its More models opens the same rail-less picker. Row stars still toggle, since they don't need the rail and feed the New Tab's Starred tab.agent.harnessSwitchInStartedChat, default off). On, a started chat keeps the whole rail, never just its own harness, plus Add harness…, and a pick in another harness starts a new chat there through the existingchat.newpath. It reaches the page live as thecomposerevent'sharnessSwitch, besideplainFooterandmodelMenu. It's a Debug Settings tunable, not a cmux.json key, so the settings lists don't change.Mechanism:
ModelPickertakesharnessFixed;ComposerPickerspasses its existinglockedvalue, which the toggle now overrides. If acpmux hasn't listed the chat's own harness yet (first open after launch), the picker keeps the old full rail rather than show an empty menu.Testing
Ran on Huge Blue (
bun test, per file):test/model-picker-picks.test.tsx22/22,test/model-menu-compact.test.tsx19/19,test/signed-out-rail.test.tsx4/4;model-picker-composition,model-picker-height,effort-picker-slider,gallery-coverageandgallery-manifestpass.CMUX_BROWSER_TESTS=1 bun test test/agent-pane-harness-catalog.test.ts(Chromium) passes at 15b21d3. It opens Add harness… from a started chat, so it turns the started-chat harness switch on throughwindow.cmuxAcpmuxComposerfirst, which tests the toggle end to end in the page.bun run typecheck: no type errors in 1629 files.vp check(fmt + lint) is clean on the changed files.test/ui-rules.test.tsfails the same way at the base ca13780 (+4 received); it isn't from this change.bun testrun in one process fails about 148 tests at the base from cross-file state; the only new failures at the head were the three new tests, which pass when run per file.verify-local.py --affected: Swift syntax, test wiring, xcstrings and the other checks pass.localizationfails with 35k parity errors in catalogs this PR doesn't touch (a base finding).Observed / Not run
Not run: the Swift side (
AgentPaneTunables.harnessSwitchInStartedChat, thecomposerevent'sharnessSwitchand the newAgentPaneHarnessSwitchTunableTests) is only syntax-parsed here. It follows themodelMenuwiring line for line. PRs into feat-cmux-next skipcmux-next swift test, and no native build ran on this Mac. I haven't checked the toggle in a tagged build.Changelog
Changed: In a chat that has started, the model menu lists only that chat's harness's models, with no harness rail or Add harness….
Proof
Gallery fixtures
agent-pane.model-pickerstarted-chat(full picker) andstarted-chat-quick-menu(compact menu) let the Linux gallery advisory run capture both states. Screenshots will be added here from that run.Checklist
🤖 Generated with Claude Code