git:20260923.1ec71d0 to git:20261007.212a13f

8 added, 1 removed. Audit A to A.

# Terminal UI (`src/ui/`)

- OpenTUI React (`@opentui/react`, `@opentui/core`), not Ink: lowercase intrinsics (`<box>`, `<text>`,

`<scrollbox>`), and the JSX runtime comes from `jsxImportSource`, so no `import React` for JSX.

- Colours only through `theme.ts` (the approved web palette with xterm-256 fallbacks); no raw hex in

components. No gradients, shadows or glows. Every status is a glyph plus a colour, never colour alone.

- - The approved `frontend/` UI is the visual reference, and `docs/ui/DESIGN-SYSTEM.md` records how it

+ - The approved website UI (`Shelra/Shelracode-frontend`) is the visual reference, and `docs/ui/DESIGN-SYSTEM.md` records how it

translates to terminal cells; keep it in step with `theme.ts` and `acceptance.test.ts`.

- The log is the only permanent surface. A plan, a load or a turn summary appears when it exists, and the

detail opens on demand (`/plan`, `/diff`, `/checks`, `/context`). Something with nothing to show never draws.

- Show only what the runtime knows: read progress, tokens, agents and verification from `Agent` state, never

from model text. UI text is English.

- Known traps: OpenTUI's `<textarea>` keeps the `onSubmit` it was created with, so use

`components/text-area.tsx`. Markdown and code blocks use the local renderers (`markdown.tsx`,

`code-highlight.ts`) because OpenTUI's built-in rendering broke list indents, block margins and code colours.

- Render tests (`testRender`) run only under `bun test` (see the Bun-only rule in the root CLAUDE.md); pure

logic such as `activity.ts` and `observability.ts` stays in Vitest.

- Verify the rendered result, not the source: `SHELRA_DEMO_SCENARIO=<name> bun run scripts/ui-demo/main.tsx`

mounts the real `App` and `Agent` with real tools on a throw-away fixture; only the model is scripted

(scenarios in `scripts/ui-demo/scenarios.ts`; point `HOME`/`USERPROFILE` at a scratch directory). Check

80x24 and 120x40. The owner uses Windows Terminal and Warp.

- UI deliverables get three rendered passes, each improving on the last (owner's rule). Load the

`terminal-ui` skill for TUI work.

+ - Responsiveness (measured 2026-10-06, `docs/audits/2026-10-ui-responsiveness.md`): every mounted renderable costs layout

+ on every frame, so the log mounts only its last 150 items and reaching the top mounts 100 more

+ (`transcript-window.ts`, `hooks/use-transcript-window.ts`). Nothing on the render path may call into the agent for

+ something that grows with the session (`getContextStats` is cached per message); polled snapshots go through

+ `sameSnapshot` so an unchanged poll does not redraw; streamed text must not re-render its finished blocks

+ (`markdown.tsx` memoizes each block). `scrollToBottom` follows the end only while the reader is there; sending a

+ message uses `jumpToBottom`. Re-measure with `bun run scripts/perf/ui-stress.tsx` (`SHELRA_PERF_WARM=<turns>`).