Note [CRF-7] The removed text-md was never a real utility. (Netero)

The probe build emitted .text-sm, .text-xs, .font-bold, .font-normal, and both arbitrary variants, but no .text-md. tailwind.config.js:26-31 extends fontSize with 2xs, xs, sm, base, 3xl only, and Tailwind's default scale has no md key. Nine other files still use the dead class, including ModuleCard.tsx:65, DynamicParameter.tsx:400, and four DeploymentSettingsPage charts.

No action needed in this PR. Worth knowing because it changes what this line does: the title was previously unsized and inherited, so text-sm is a new explicit size, not a step down from 16px. The nine remaining uses are a separate cleanup.

🤖

Note [CRF-8] [&_code]:text-[length:inherit] also matches fenced-code code elements, with no visual effect today. (Netero)

The probe build confirms it compiles to .[&_code]:text-[length:inherit] code { font-size: inherit }, one specificity step above the inline-code text-sm at site/src/components/Markdown/Markdown.tsx:89, so the inline-code override works as intended. Prism blocks (Markdown.tsx:77-86) render code without a font-size class, so forcing inherit there changes nothing today, but the selector is broader than the stated intent.

No action needed. Flagged so that if a font size is ever added to Prism code blocks, this selector is where it gets overridden.

🤖

P2 [CRF-11] CRF-3's fix replaced the lg tier instead of supplementing it, so 1024-1279px drops from 3-up to 2-up and the step now disagrees with the next step in the same wizard. (Netero P2, Mafu-san P2, Kite P2, Meruem P2, Nami P2, Pariston P2, Hisoka P3, Zoro P3, Razor P3, Leorio P3)

Base was grid-cols-1 sm:grid-cols-2 lg:grid-cols-3. Head is grid-cols-1 sm:grid-cols-2 xl:grid-cols-4. Nothing matches between sm (640px) and xl (1280px).

One pixel of viewport swings a card from 440px to 212px. 1024-1279 covers the common laptop widths, and the PR's stated goal is to tighten this grid, so the range it loosens most is the one users see most.

Viewport	Columns	Card width
1024 (pre-PR)	3	203px
1024 (HEAD)	2	312px
1279 (HEAD)	2	440px
1280 (HEAD)	4	212px

Pariston counted what the step actually renders: coderd/templatebuilder/bases.go:240 enumerates nine base directories and all nine resolve to examples, so ceil(9/3) and ceil(9/4) are both 3. The new 4-up tier removes zero rows at any width, while the lg band goes from 3 rows to 5.

ModuleSelectStep.tsx:247 still uses the exact string this line abandoned, grid grid-cols-1 sm:grid-cols-2 lg:grid-cols-3 gap-4, and both grids render into the same container in the same wizard. Between 1024px and 1279px the base step shows 2 columns and the module step shows 3.

That is CRF-2's divergence class applied to layout one round after CRF-2 was fixed for typography. Kite states the tradeoff honestly: restoring the tier puts 203px cards back at 1024px, which is 9px below the 212px this PR already accepts at xl, so it is the status quo rather than a new regression. Fix is one token, sm:grid-cols-2 lg:grid-cols-3 xl:grid-cols-4, and decide whether ModuleSelectStep gets the matching xl tier. Meruem's structural version, one exported CARD_GRID_CLASSES consumed by both grids, is what stops the two from drifting again.

This is on me as much as on the fix: I verified CRF-3's closure against the previous round's head instead of the base SHA and missed it.

🤖

P2 [CRF-12] The story named Loading settles into an error alert, and what it captures depends on the surrounding server and on response latency. (Komugi P2, Bisky P2, Mafuuu P3, Chopper P3, Nami P3, Meruem P3, Razor P3, Netero P3)

Six reviewers verified this with throwaway probe stories rather than inferring it. Komugi sampled the canvas twice:

• t0: <div role="status" data-testid="loader"> (the Loader).
• t1: <div role="alert"> with Request failed with status code 502.

parameters: { queries: [] } only removes the seed that meta installs. It does not stop the query. [...] The probe also proves the override direction: meta seeded { bases: [] } (which renders an empty grid, not a loader), and the story still fetched.

The value is unpinned in three directions: 502 under pnpm test:storybook from the storybook-test-setup middleware, 404 under the static build pixel serves, and whatever CODER_HOST returns under pnpm storybook. Bisky adds that the captured pixels contain an unmasked Axios stack trace and status code, so the baseline differs between harnesses and any change produces a diff nobody can attribute.

The repo already has the fix, and Komugi found it: EditOAuth2AppPageView.stories.tsx:70-77 pins the pending state with beforeEach and a never-resolving mock.

export const Loading: Story = {
	parameters: { queries: [] },
	beforeEach: () => {
		spyOn(API, "getTemplateBuilderBases").mockReturnValue(new Promise(() => {}));
	},
};

ModuleSelectStep.stories.tsx:80 has the identical defect and is where this was copied from. Fixing only the new file leaves the pattern in place for the next person to copy.

🤖

P2 [CRF-13] The story added to cover this step renders none of the four ConfigurationField description slots the PR changed. (Bisky P2, Mafu-san P2, Nami P2, Chopper P3, Kite P3, Meruem P3, Razor P3)

The base fixture has exactly one variable, type: "string". variableToField (BaseTemplateParametersStep.tsx:56) maps everything non-bool to type: "text", and TextField (ConfigurationField.tsx:92) delegates its description to FormField. FormField.tsx:47 renders text-xs, and git show 0e1a9a9f0:site/src/components/FormField/FormField.tsx shows text-xs there before this PR too. So the one description this story renders was already the target size and is untouched by the diff.

Seven reviewers independently grepped every story that builds a ConfigurationFieldDefinition. TemplateConfiguration.stories.tsx and ModuleConfiguration.stories.tsx between them render select, radio, switch and switch-group, and not one passes a field-level description, so all four changed slots render nowhere in the repo.

The switch path is live, not theoretical. Meruem traced it to shipped data:

coderd/templatebuilder/bases/kubernetes/base.json defines use_kubeconfig with "type": "bool" and a description, so selecting the Kubernetes base renders line 279.

Two cheap edits close it: add a bool variable with a non-empty description to this fixture, and add description to one radio, one switch-group and one select field in ModuleConfiguration.stories.tsx, whose Default story already renders all three in one snapshot.

🤖

P2 [CRF-14] The description slot this PR resizes in four places already exists as one component, FormField, which was already text-xs. (Robin P2, Zoro P3)

FormField.tsx:46-50 renders exactly this block [...] It already wires descriptionId into aria-describedby (FormField.tsx:25-30, :59), and it was already at text-xs before this PR. That is why TextField needed no edit [...] So the change converged the copies on the value FormField already held, not on FormField itself. The copies remain, and they are not decorative: they have already drifted.

Re-raise against CRF-1 with new evidence: the fix moved the resize to the right layer and landed it as four duplicates of a block the codebase owns in one place. ConfigurationField.tsx:138, :188 and :313 are now byte-identical to each other and to FormField's; :279 is the same block plus ml-[44px] font-normal. The measurable cost of the duplication is CRF-15 and CRF-16, both of which are drift between these five copies.

SwitchRow at ConfigurationField.tsx:224-242 is this file's own precedent for extracting a shared sub-component. A FieldDescription sibling, importable by FormField too, collapses five slots to one. I am not asking for a refactor as a condition of this PR, but the next scale or color change to field descriptions is a five-site edit with the same odds of missing one, and this PR is the second lockstep edit in three rounds.

🤖

P3 [CRF-16] Only SwitchField keeps font-normal, so the now-uniform 12px descriptions render at two weights. (Gon P3, Mafuuu P3, Hisoka Nit)

tailwind.config.js:28 bakes fontWeight: "500" into text-xs, so text-xs text-content-secondary at lines 138, 188, and 313 renders at weight 500, while SwitchField at line 279 renders at 400 because it keeps font-normal.

Mafuuu names why this PR is where it becomes visible:

Before the PR the switch description differed in both size and weight, so the weight gap was invisible. Now that sizes match, weight is the only remaining difference and it shows up side by side: in the base-parameters step a string variable renders a FormField description at 500 and a bool variable renders this one at 400 in the same column.

Gon adds that the card descriptions this PR also resized use font-normal (TemplateCard.tsx:74, ModuleCard.tsx:74), so the weight question is already answered elsewhere in the same diff. Pick one weight for description text and apply it to all four slots. This is a different claim from CRF-9, which was about one pairing being correct by source order, not about the four slots disagreeing with each other.

🤖

Nit [CRF-32] The comment restates the three assertions directly below it. (Gon P2, Mafu-san Nit)

findAllByRole("radio"), toHaveLength(bases.length), and the getByText loop already say "every base renders as a radio with its title and description." The comment adds no invariant, no reason, no trap. Delete it.

site/AGENTS.md FE4 forbids comments that restate assertions. Gon rated this a finding; I am posting it as a Nit because the standalone consequence is reader noise, and the part that has real cost, the comment promising description coverage that does not exist, is CRF-24. Worth contrasting with the precedent this file copied: ModuleSelectStep.stories.tsx:86 explains the fixture arithmetic instead of narrating its own calls.

🤖

Nit [CRF-33] Half this comment restates the assertions; the trailing clause is the part worth keeping. (Gon P2)

"the field description and README body copy are rendered" is what findByText(/Container image for workspaces/) and the two findByRole("heading") calls show. The part worth keeping is why presence assertions cover a typography change.

Gon's rewrite: // The step's typography rules only render when body copy and headings exist. Leorio independently called this comment's framing the honest one of the two, so it is the sentence to keep, just shorter.

🤖

Nit [CRF-34] Selected sets state and asserts nothing, and no story anywhere exercises selection or the keyboard handler. (Kite)

Selected sets selectedBaseId: "docker" and stops. Nothing asserts aria-checked="true" on the docker card, so the story would pass unchanged if TemplateCard's selected prop stopped being wired through. onSelectBase is fn() in args and no story clicks a card, and TemplateCard's Enter/Space handler (TemplateCard.tsx:41-46) is never invoked.

Class rather than instance: TemplateCard.stories.tsx has six stories and zero play functions, so that keyboard handler has no coverage in the repo at all. A play that tabs to a card, presses Enter, and asserts onSelectBase was called with the right base closes both, and AGENTS.md asks for exactly that.

🤖

Nit [CRF-35] ml-[44px] is a derived offset with nothing recording the derivation, on a line this PR modifies. (Gon)

44px is Switch's default width (w-9, 36px, Switch.tsx:21) plus the SwitchRow gap (gap-2, 8px, line 232), which aligns the description with the switch label. Nothing states that, so a change to either value silently breaks the alignment.

One line of comment above the div naming the two values it is derived from.

🤖

Note [CRF-37] text-xs also drops paragraph leading from 24px to 16px, a 1.33 ratio on 12px prose. (Hisoka)

The utility changes two things at once, and the README is the one place in this PR with real multi-line prose. 1.33 is tight for body copy; the rest of the class list is explicit about spacing ([&_p]:mb-3, [&_h3]:mt-2 [&_h3]:mb-1), so if the leading is deliberate it would be worth writing it down as leading-normal or similar rather than inheriting whatever the size utility happens to bake.

No change requested. Flagged because it is the same shape as CRF-17 and CRF-18: a size utility here sets three properties, and only one of them was the intent. The pixel baseline will show whether it reads well.

🤖