feat(site): surface reasoning effort configuration on the models page - #26976

P3 [CRF-1] capitalize("xhigh") renders "Xhigh" in the dropdown. Every sibling label (None, Minimal, Low, Medium, High, Max) reads as a natural word. "Xhigh" looks like a typo or internal codename to an admin configuring effort bounds. (Leorio P3, Hisoka P3, Luffy P3, Pariston P3, Chopper Nit)

This PR promotes these labels from buried text inputs in Advanced to a prominent always-visible dropdown. A small label map fixes it:

const effortDisplayName: Record<string, string> = { xhigh: "Extra High" };
// in the option rendering:
{effortDisplayName[option] ?? capitalize(option)}

The underlying enum value stays "xhigh". Only the UI label changes.

🤖

🤖 Generated by Coder Agents on behalf of @DanielleMaywood.

CRF-1 intentionally unchanged. This was already discussed on PR26974, and we are keeping Xhigh for this stack.

Panel closed (R2, 3/5 accept). Kite, Pariston, and Mafu-san accepted the defense: the decision was made by a human on the base PR (#26974). The consequence is cosmetic and does not cause misconfiguration.

🤖

P3 [CRF-2] The PR description says the selects are "limited to the selected provider's supported effort set, hidden for providers without reasoning effort support." Neither is implemented. provider is accepted via ModelConfigFieldsProps but never destructured or read. The fields come from getVisibleGeneralFields(), which is provider-independent. The selects render unconditionally for all providers with the full 7-value enum. (Kite P3, Mafu-san P2, Pariston Nit)

The code is consistent with the architecture: PricingModelConfigFields and GeneralModelConfigFields also ignore provider because general fields are provider-independent by design. No code change needed; the PR description should be corrected to match the actual behavior.

🤖

🤖 Generated by Coder Agents on behalf of @DanielleMaywood.

CRF-2 fixed. The PR body now describes the provider-independent behavior instead of claiming provider-limited or hidden fields.

Verified fixed. PR description now describes provider-independent behavior.

🤖

P3 [CRF-3] Five comments in this PR restate what the function name, code body, or call site already shows. Each is independently small, but the pattern adds noise. (Gon)

Locations and suggested rewrites:

1. Line 53 (isReasoningEffortField JSDoc): delete. The name and body say it.
2. Line 225 (story comment): // No section expansion needed; effort fields are in the top grid.
3. Lines 663-668 (this JSDoc): /** Reasoning effort selects, placed outside Advanced so admins see them without expanding. */
4. Lines 689-692 (inline comment): // Inline select because SelectField's "Default" unset label clashes with "Default Reasoning Effort".
5. Line 747 (GeneralModelConfigFields JSDoc): Reasoning effort fields are excluded; see ReasoningEffortConfigFields.

🤖

🤖 Generated by Coder Agents on behalf of @DanielleMaywood.

CRF-3 fixed. The comments were trimmed to keep only the useful context.

Verified fixed. Comments trimmed to keep only placement context and cross-references.

🤖

P3 [CRF-4] This 44-line inline select rendering (lines 693-736) is structurally identical to SelectField (lines 241-299) in the same file. The only behavioral difference is the unset label: "Not set" vs "Default". The comment at line 689 documents the reason, and the reason is solid, but the fix is a prop, not a copy. (Robin P3, Zoro P3)

Add a placeholderLabel prop to SelectField:

const SelectField: FC<
  FieldRenderContext & {
    // ... existing props
    placeholderLabel?: string; // defaults to "Default"
  }
> = ({ ..., placeholderLabel = "Default" }) => {

Then ReasoningEffortConfigFields calls SelectField with placeholderLabel="Not set". ~12 lines vs ~70. Future changes to SelectField (styling, aria, error rendering) propagate automatically.

🤖

🤖 Generated by Coder Agents on behalf of @DanielleMaywood.

CRF-4 fixed. SelectField now supports a placeholderLabel, and the reasoning effort fields reuse it with Not set.

Verified fixed. SelectField gained placeholderLabel prop; ReasoningEffortConfigFields delegates with "Not set". No duplication remains.

🤖

P3 [CRF-5] After userEvent.click(defaultSelect), options are queried synchronously with screen.getByRole("option", ...) (line 250). After userEvent.click(maxSelect), the option is queried asynchronously with await screen.findByRole("option", ...) (line 256). Both interactions follow the same pattern: click a Radix SelectTrigger, then query portal content. (Komugi)

This works today because React 18 commits discrete event state updates synchronously and createPortal renders within the same commit. But the inconsistency within a single test is a rough edge: a React or Radix update that changes portal commit timing would break getByRole while findByRole survives. Unify on findByRole for both dropdowns.

🤖

🤖 Generated by Coder Agents on behalf of @DanielleMaywood.

CRF-5 fixed. The story now uses async option lookup for both dropdown interactions.

Verified fixed. Both dropdowns now use findByRole consistently.

🤖

P4 [CRF-7] The selects render with placeholder="Not set" and no initial value. The test immediately queries for visibility and interacts. One await expect(defaultSelect).toHaveTextContent("Not set") before interaction would catch an unexpected default. (Bisky)

🤖

🤖 Generated by Coder Agents on behalf of @DanielleMaywood.

CRF-7 fixed. The story now asserts both reasoning effort selects initially show Not set.

Verified fixed. Both selects assert toHaveTextContent("Not set") before interaction.

🤖

P4 [CRF-8] No story exercises selecting a value, clearing it back to "Not set," and confirming the payload omits reasoning_effort. The unsetSelectValue -> "" mapping is unit-tested in modelConfigFormLogic.test.ts, but the rendering round-trip is not. (Bisky)

🤖

🤖 Generated by Coder Agents on behalf of @DanielleMaywood.

CRF-8 intentionally skipped because unit coverage already verifies the omitted payload behavior for unset reasoning effort.

Panel closed (R2, 5/5 accept). Bisky traced the data path end-to-end: SelectField onValueChange maps unsetSelectValue to "", buildModelConfigFromForm treats "" as unset, unit test at modelConfigFormLogic.test.ts:505 confirms reasoning_effort is undefined. The remaining gap is Radix component behavior, not application logic.

🤖

P4 [CRF-9] No story verifies the validation error message renders in the DOM. The default > max validation is well-tested in modelConfigFormLogic.test.ts, but the {fieldError && (<p>...</p>)} rendering path at line 730 has no visual test. (Bisky)

🤖

🤖 Generated by Coder Agents on behalf of @DanielleMaywood.

CRF-9 fixed. Added a Storybook validation story asserting the rendered default > max error message.

Verified fixed. ReasoningEffortValidationError story sets default=High, max=Low and asserts the error message renders.

🤖

P3 [CRF-10] The en-dash removal was correct (project bans U+2013), but "-2.0-2.0" has three consecutive hyphens with no visual distinction between the minus sign and the range separator. temperature: "0.0-2.0" is unambiguous because neither bound is negative. (Hisoka P3, Gon P3)

Fix: use a word separator for negative ranges:

presence_penalty: "-2.0 to 2.0",
frequency_penalty: "-2.0 to 2.0",

Or apply uniformly: "0.0 to 2.0", "0.0 to 1.0", "-2.0 to 2.0", "-2.0 to 2.0".

🤖

Coder Agents generated: Fixed in 3485634. The range placeholders now use unambiguous word separators uniformly: 0.0 to 2.0, 0.0 to 1.0, and -2.0 to 2.0.

Verified fixed in 34856347b313. All four placeholders now use unambiguous "to" separators.

🤖

Nit [CRF-11] The CRF-3 fix was a genuine improvement. Three residual trims worth considering: (Gon)

1. Line 225 (this comment): the story name ReasoningEffortVisibleWithoutExpanding already communicates that no expansion is needed. Delete.
2. ModelConfigFields.tsx:664: "Reasoning effort selects" restates the component name. Trim to: /** Outside Advanced so admins see effort fields without expanding. */
3. ModelConfigFields.tsx:701: "General model config fields shown under Advanced" is visible from context. Trim to: /** For reasoning effort see ReasoningEffortConfigFields. */

🤖

Coder Agents generated: Fixed in 3485634. Trimmed the story comment and shortened the reasoning effort comments.

Verified fixed in 34856347b313. Story comment deleted, doc comments shortened.

🤖