[tulsi-builder]

TL;DR

In Create agent, a long Goose model name no longer squeezes the Model label to one letter per line. Goose also gets its own Thinking row there, as Claude Code and Codex already have (#2409). The model name and its thinking level now share space by content instead of splitting it 50/50.

What

• Model title stays whole. The design system caps a compact choice at 60% of its PreferenceRow, but only when the choice is the trailing's direct child. Create agent nests Goose's picker in ModelDiscoveryRecovery, so the cap never applied and the picker took the whole row. The cap now also applies when a lone wrapper holds the compact choice (components.css).
• Name gets what the level leaves. In ModelFamilyMenu.module.css, .model and .effort were both flex: 1 at a zero basis, so they always split the trigger evenly. That cut long names early and left short levels like "Low" sitting in empty space. Now the effort is flex: none and the name is flex: 0 1 auto, and the name truncates first. Visual direction came from Xela. There's no extra · before the level, because · Default already uses the dot.
• Goose Thinking row in Create agent. Goose now supplies CreateThinking, which reuses its existing Settings ThinkingField in the compact row treatment. When a type supplies a Thinking row, the dialog passes showEffort: false to its model field. Changing the model still resets effort to that model's default.
• Failed level reads are visible. Compact Select doesn't render descriptions, so the compact field wraps itself in ModelDiscoveryRecovery. It shows the error and Try again, the same pattern the model field uses. Failed reads are already evicted from GooseSetup's cache, so a retry really reads again.

Non-goals: Goose agent cards still pick the level from the model menu. Shortening data_workflow_tools. model prefixes is a possible follow-up.

Screenshots

Create agent with Goose on a long model (data_workflow_tools.goose.goose-deepseek-v4-flash-0731), rendered from the real dialog in Chromium at 640px with a fake goose.

Before

After

Tests

• Browser, goose-model-row.spec.mjs, two cases added. Both are layout geometry, which jsdom can't measure.
  • Create agent's Model title whole: renders CreateModel in a PreferenceRow at 520px. Before the CSS fix it failed with titleLines: 4, and it passes after.
  • Name gets what the level leaves: covers a short inherited name, which stays whole with the level right after it and no gap before the caret, and a long inherited name, where · Default stays attached and the name truncates. On the prior commit it failed with truncated: true for the short name, and it passes after.
  • Both pass in Chromium and WebKit, and the existing Settings case still passes.
• Vitest, three tests added.
  • Goose CreateThinking reads per-model levels and saves a choice.
  • A failed level read shows the alert and Try again, then retry recovers. This failed before the recovery fix.
  • The dialog passes showEffort: false to the model field when a Thinking row exists.
• vitest run src/bundled/goose src/bundled/agents src/features/agents → 1,804 passed at e66e471a3.
• Before the rebase, the full pnpm test:browser run showed 7 failures (conversation, layout Alt+ArrowUp, typing composer ×2, page-placement). Those same specs also fail on base 9c82d1f0a without this change.

Review

Crack Catcher reviewed the change and found one medium issue: a failed Thinking read was silent. That's fixed in the last commit. tulsi tried the dev app before that fix.

Deferred: native-app check in both themes and with enlarged text. The new error state hasn't been seen in the native app because it needs Goose to fail.

🤖 Generated with Claude Code

[kalvinnchau]

🤖 One recovery-state issue found.

[kalvinnchau]

🤖 [P3] Refresh Thinking after Model recovery

If Model and Thinking discovery both fail when Create agent opens, restoring the upstream and clicking Model’s Try again recovers Model and returns setup readiness to ready, but Thinking retains its old error. CreateThinking's provider/model/revision read key remains unchanged across setup.check(), so the user must click a second retry despite successful setup recovery. Reproduced with a targeted unit test and in the actual app on Chromium and WebKit with modeled upstreams. Make Thinking follow setup re-check generations or coordinate invalidation through the recovery owner, while preserving Thinking-only retry.

[tulsi-builder]

🤖 Fixed in c382500. CreateThinking now subscribes to setup's readiness and bumps ThinkingField's revision each time a check settles. A Model Try again (setup.check() clears the levels cache) therefore re-reads Thinking without a second retry. Thinking-only retry is unchanged.

Regression test: reads Create agent's Thinking levels again when a setup check recovers in src/bundled/goose/tabs.test.tsx. Both reads fail, then setup.check() recovers, and Thinking must show its levels with no alert. It fails on e66e471 and passes on c382500. The existing Thinking-only retry test still passes.