[elhoim]

BLUF

• Priority: medium.
• What: widget modules no longer import ink or React at runtime. Widget.renderEditor now returns a plain descriptor ({ kind, props }), and the editor components move unchanged into src/tui/components/widget-editors/. That folder is only reached through the lazy import('./tui').
• Why: perf: keep the TUI off the status line render path #575 took the TUI off the render path, but eleven widget modules still imported ink/React for their editor components. The registry builds every widget at load, so every status line render still evaluated ink, react, react-reconciler and yoga-layout, and yoga-layout instantiates its WASM when the module loads.
• Measured: the chunk statically imported by the entry shrinks from 2.22 MB to 1.27 MB and no longer contains ink, react or yoga. Median CPU per render drops 14% on a heavy config, 18% on the default config and 26% for --version. Rendered output is byte-identical.
• Update after merging main. Main has since added two React modules under src/widgets/shared that only editors use: the useTextCursor hook (text-cursor.ts, also used by the new label editor) and searchable-option-editor.tsx (now behind the locale and timezone editors). The merge moves both, with their tests, into src/tui/components/widget-editors/ so the no-ink/React guard still holds. It also ports main's later changes to the split widgets (custom-command display-width truncation, Link URL sanitizing, the cwd label and segment fixes, Number.parseInt) into the new files. Two new upstream glyph tests read slots from the editor descriptor instead of a React element's props. The entry's static closure is still ink/React/yoga-free (1.28 MB shared chunk vs 2.23 MB on main).

Details

• src/types/Widget.ts: adds WidgetEditorDescriptor. renderEditor?(props) now returns WidgetEditorDescriptor | null instead of React.ReactElement | null. SymbolSlot moves here, and symbol-override re-exports it.
• CustomText, CustomSymbol, CustomCommand, Link, Skills and CurrentWorkingDir go from .tsx to .ts. So do the shared symbol-override, max-width, speed-widget, locale-editor and timezone-editor. Each one keeps its render logic and returns a descriptor from renderEditor. Helper names and signatures are unchanged (renderSymbolSlotsEditor, renderMaxWidthEditor, renderSpeedWidgetEditor, renderUsageLocaleEditor, renderUsageTimezoneEditor), so the ~30 widgets that call them needed no edits. The only other changes are the return type annotation in BlockResetTimer and WeeklyResetTimer.
• The 11 editor components move verbatim to src/tui/components/widget-editors/*Editor.tsx. renderWidgetEditor(descriptor) in index.tsx maps each kind to its component. ItemsEditor is the only caller: it renders renderWidgetEditor(editor), and still renders nothing when a widget returns null, as before. The few model helpers these editors use (Link's buildMetadata etc., Skills' list-limit helpers, the speed WINDOW_EDITOR_ACTION) are now exported from their widget modules.
• The widget registry still builds every widget eagerly. Now that widget modules are ink-free, that costs nothing measurable, so it is left alone.
• I kept a data descriptor instead of a registry that the TUI fills at startup, so there is no import-order dependency. Widget tests that call renderEditor directly (e.g. GitRootDir) pass unchanged.
• Tests: the three editor tests move next to their components. The new widget-editors/__tests__/index.test.ts checks two things. First, every editor a widget opens from its keybinds maps to a React element. Second, a guard fails if any module under src/widgets imports ink or react as a value. The guard fails on main, where 11 modules do.
• Overlap with my other open PRs: [low] fix(custom-command): truncate maxWidth output by display width, not string length #626 edits src/widgets/CustomCommand.tsx, which this PR splits into src/widgets/CustomCommand.ts and src/tui/components/widget-editors/CustomCommandEditor.tsx. Whichever PR merges second needs a small rebase. feat: Add RClone Queue widget + code-review fixes #523 adds a .tsx widget; if it keeps an ink editor it would need the same descriptor pattern, and the new guard test would flag it.

Measurements

Built with bun run build (--splitting). The static import closure of dist/ccstatusline.js:

	shared chunk	ink / react / react-reconciler / yoga-layout modules in the static closure
main (35440e4)	2,219,391 B	55 / 4 / 6 / 4
this PR	1,271,391 B	0 / 0 / 0 / 0 (they now sit only in the lazy TUI chunk)

The runs alternated base and patched builds, each process fork/exec'd with CPU = user+sys including reaped children. 25 passes on node 24. Load1 during the run was 54.9 to 71.1 (median 62.2) on 6 cores. The host was heavily loaded, so compare the ratios rather than the absolute numbers.

arm	CPU median (ms)	CPU p90 (ms)	wall median (ms)
control node -e 0	108.6	143.7	998
main --version	1086.1	1194.9	8527
this PR --version	806.1 (-26%)	900.0	6637
main heavy config, 5 MB transcript	2009.1	2312.7	14413
this PR heavy config, 5 MB transcript	1723.8 (-14%)	1949.6	13425
main default config, 5 MB transcript	1531.0	1728.3	11383
this PR default config, 5 MB transcript	1259.7 (-18%)	1390.4	9843

Output: stdout+stderr are byte-identical between the two builds on the heavy, heavy-ttl0, default, default-ttl0, livelike-ttl0, git17-ttl0 and block configs, each with 1 MB, 5 MB and no transcript (21/21). --version is identical too.

Checks

• bun run lint (tsc + eslint): clean.
• bun run build: OK (as above).
• bun test: 2289 pass / 69 fail on this loaded host. Every failure is a timing-sensitive test: ink input tests in the TUI menus and editors, the custom-command capture tests and fetchUsageData. The same TUI/editor/widget subset fails on unmodified main under the same load: 15 failures there vs 14 here, with the same test names. The usage and custom-command files fail too: 36 on main vs 33 here. This full run was taken before the two new tests in widget-editors/__tests__/index.test.ts were added. Those pass (2/2) when run on their own.

🤖 Generated with Claude Code