Note: Placeholder / candidate. Not scheduled, not a commitment. Capturing the pattern while it's fresh from #208. This is a developer-facing foundation concern, not a user-facing workflow — the closest template fit is below.

What task are you trying to do?

Ship UI work on the composer, chips, and popovers without paying a CSS-override tax every PR.

PawWork's design language has permanently diverged from opencode's (pill shapes, brand orange #FF5910 / #E56A2E, rounded-full / radius-xl popovers, menu-below composer, custom WorkspaceChip / SendButton), but the underlying primitives still come from @opencode-ai/ui (Kobalte wrapper + popover.css + shared Button variants). Every feature PR that touches chips or popovers pays a tax to override upstream styling, and most of the iteration rounds on #208 were spent fighting that tax rather than moving the product forward.

What do you do today?

Patch case by case. From #208 alone:

• Three popovers (dialog-select-model, prompt-input variant, workspace-chip) use inline style={{ "border-radius": "16px" }} to beat [data-component="popover-content"] { border-radius: var(--radius-md) }. Tailwind ! modifier and class both lose to source order.
• SendButton uses custom #E56A2E + inline color: #ffffff on the Icon to beat [data-component="icon"] { color: var(--icon-base) }.
• as any cast on Kobalte triggerProps to inject data-action attrs (Popover types don't accept data-*).
• Model chip gets hover via Button component; Variant and Workspace chips rebuild hover manually. One round used invalid hover:bg-background-base-hover token (silently dropped by Tailwind) — only visual review caught it.
• Starter cards iterated rounded-xl → rounded-2xl → rounded-full across three rounds because no shared chip primitive pinned the shape.

None of these are bugs. They're all the same shape: brand-layer design colliding with a primitive layer that assumes a different design.

What would a good result look like?

Three focused replacements, each a self-contained sub-PR:

1. PawWork Popover primitive — thin wrapper over Kobalte.Popover that does NOT pull popover.css. Radius / background / border baked into Tailwind at the primitive layer. Migrate model / variant / workspace popovers off the shared @opencode-ai/ui/popover wrapper.
2. <Chip> primitive — unified trigger for Model / Variant / Workspace chips. Hover, focus, disabled, truncation in one place. Removes piggyback on Button and removes the "why does this chip not hover like that chip" asymmetry.
3. Brand-layer design tokens — explicit --pawwork-accent-brand, --pawwork-radius-chip, --pawwork-radius-popover, --pawwork-shadow-chip. Stop co-opting @opencode-ai/ui tokens (--radius-md, --accent-brand) whose semantics live elsewhere.

Acceptance criteria:

• Zero inline style={{ "border-radius": ... }} in prompt-input.tsx, workspace-chip.tsx, dialog-select-model.tsx
• Zero as any casts on Popover trigger props in packages/app/src/components
• A new feature PR adding a chip-shaped trigger requires zero CSS override to match design
• Brand orange / chip radius / popover radius have a single source of truth in the PawWork layer

Audience

Both — this is developer-facing, but it directly improves the speed and quality of every UI iteration that reaches users. Visible payoff is faster, more consistent visual updates for everyone.

Extra context

Non-goals. Not rewriting sidebar / session list / command palette / tool-execution UI / permission dialogs / chat stream rendering. Not replacing @opencode-ai/ui/button wholesale. Not a frontend rewrite. Not changing the opencode-底座 commitment (session / SDK / providers / tools stay upstream).

When (not now). Proposed after #208 lands and the visual language settles ~2 weeks, and after the dynamic shell tabs series (Task 1-13 Codex handoff) lands. Batch with the next deliberate brand iteration so the swap piggybacks on a visible change instead of landing invisibly. Prevents thrashing the foundation layer while the design it supports is still settling.

Alternatives considered.

• Keep patching. Sustainable, but the tax compounds with every new chip surface. Current state.
• Full UI rewrite. 2-3 months, blocked by the opencode-底座 commitment (≤10K users, no 自研底座). Rejected.
• Wait for opencode upstream to match. Design has permanently diverged. Not coming back.

Open questions.

• Keep @opencode-ai/ui/popover for non-chip popovers (sidebar context menus etc.) and only introduce a second primitive for chips, or replace the shared wrapper entirely? Probably the former — narrower scope.
• <Chip> primitive in packages/ui/src/components/chip-pw.tsx (shared) or packages/app/src/components/prompt-input/chip.tsx (app-local)? Probably app-local first, promote if reused.