Spike: the model picker as a pure component (scene 23)

[dossettg]

Spike: the model picker as a pure component

A composition spike before the @ephemeral-ai/mayfly-ui freeze (slice 1.11): are the public basic components enough to
build a real Mayfly component? packages/mayfly/tests/design/spike/model-picker.ts holds ModelPicker and EffortPicker
in the Phase 2a style: defineMayflyComponent (with memo), plain facts shaped like catalogRows and
currentModelSelection, a translator t, and only @ephemeral-ai/mayfly-ui imports. model-picker.spec.ts compiles them
with the real renderer (createRealSurface) and compares scene 23's goldens. No source changed; nothing is wired in.
Base: e0d4d04.

1. What composed cleanly

• Scene 23 is ui.surface + ui.list with group, detail, badge, and a per-row segment with inheritedId.
  No patterns.* fits (it is a list, not a rail, split view, status page, or decision card).
• The thinking strip, (default), Delete use default, the numbered /effort list, [current · high], Provider default (high) [current], the filter header, and the 62-column ladder all come from the list painter. There is no button: with
  no acceptActionId the list still emits selection-accept with segmentId (absent = unpinned, roadmap 2a); the spec checks it.
• A strip pinned to the model's own default draws ‹ high › and accept reports high, so an explicit high and an
  unpinned selection stay distinguishable, as sameSelection needs.
• /skills and /rename already hand an overlay a surface-rooted node, so chrome: 'overlay' is enough (not mounted here).

2. Gaps and workarounds, by severity

Blocking: none. Awkward:

1. selectedIds: [] is required on every ui.list (roadmap 3.1), so every Phase 2 component repeats it.
2. The UI package has no translator type. A component may not import MayflyTranslate (frontend/), so the spike
   redeclares (key, values?) => string; every component would.
3. memo: true compares props shallowly, but interactionTranslator(ctx) returns a new closure per call, so the memo
   never hits unless the owner hoists t (the spec shows both cases).
4. formatContextWindow lives in interaction/, off limits to a component. Phase 2a moves it into components/ or puts a
   formatted contextLabel in the fact (the spike copies it).
5. Facts the commands do not pass: openModelPicker ignores the defaultEffort that catalogRows returns, and
   switchEffort never reads info.reasoning.defaultEffort. inheritedId and Provider default (high) need both.
6. Three interpolated catalog keys replace concatenation ({size} context, current · {effort}, Provider default ({level})); the zh catalog needs them.
7. Oracle gap: the kit never pins the live row's strip, so no golden shows that state. The component pins it, and an
   assertion covers it. A live-row walk (cursor on the current model) would pin it in the goldens.

Cosmetic:
8. acceptVerb is a closed enum, so the component cannot localize the hint verb (core does).
9. The kit's empty: 'No models match' is unreachable: a filter with no match shows core's No matches, and empty
shows only for an empty catalog, so the component uses the real No models advertised.

3. Before the freeze (slice 1.11)?

Nothing blocks it, and every fix is additive, so none must precede it. The smallest worth taking: (a) default
selectedIds to [] in ui.list (item 1); (b) a type-only MayflyTranslate export (item 2). Item 3 is Phase 2
guidance (hoist t per locale revision); the rest is Phase 2a work.

4. Parity per walk (golden/23-model-and-effort)

Walk	Result
initial (84), pages (Ctrl+N), narrow (62), thinking (→ →), default (Delete), filter frame 0	exact: every character and style class, no waiver
filter frames 1, 2 (/, d)	▌ in the field is Δ27 (waived, 1 cell); the hint row has a leading ←/→ thinking · the golden lacks (49 cells per frame)

The hint row cannot be composed away: the kit steps the strip with ←/→ while searching (ui-kit.mjs line 1046) but does
not hint it; the renderer hints it ("the cue is true", like Δ29). The spec pins the row's text and fails if the difference
disappears. It needs a new accepted difference (Δ33) or a kit fix.

5. Work of one key (counters of tests/perf/work-budget.spec.ts)

A warmed, framed picker; the last key of each sequence is counted. Cells are rowsPainted / stringsMeasured;
nodesValidated and unitsCompiled are 0 everywhere.

Key	6 rows	400 rows	4,000 rows
↓	2 / 41	2 / 198	2 / 198
→ (strip)	2 / 42	2 / 199	2 / 199
/	1 / 45	1 / 199	1 / 199
d (keeps all)	0 / 44	0 / 198	0 / 198
9 (narrows)	0 / 11	1 / 14	12 / 68 (11 matches)
Enter	0 / 41	0 / 198	0 / 198

Within W3's rows budget (at most 2 painted, nothing validated or compiled), whatever the catalog size. Strings exceed W3's
40 because the overlay frame re-measures its rows on every key (bare list: 8 at 6 rows, 40 at 400; framed: 26, 198). That is
constant from 400 to 4,000 rows, but no budget covers a framed list; a W3-framed row would.

Ran: vitest run packages/mayfly/tests/design packages/mayfly/tests/perf (154 tests), design:golden:check,
check:agent-docs, lint (no finding in the spike), typecheck (tsc -b does not include tests/). No full gate.

---

Draft spike, not for merge: no source changed. The report is also in docs/design/spike-model-picker.md.

🤖 Generated with Claude Code