Highlighting: Make the worker the default and warm it during startup

[masonmcelvain]

• close Highlighting: make --fast the default with warm workers #1065
• built on top of fix(ui): bound and surface large-diff highlight retries #785 in spirit: it does not include that change, but landing it first or alongside is still the plan from the issue

Syntax highlighting now runs in the worker by default, and the worker starts compiling the review's grammars during startup, before the OpenTUI import. On a small diff the default path used to block keyboard input for roughly a quarter second after the review was already on screen: the first file of each language paid one-time Oniguruma grammar compilation on the main thread, and source-backed highlighting tokenizes whole files. That work now happens in the worker while the renderer loads.

Changes

• Warm during startup. New preload worker request (protocol v5) resolves one theme and grammar and tokenizes a short sample so the grammar's root, comment, and string rules compile. main.tsx sends one preload per distinct changeset language (review order, capped at 8) after bootstrap and before importing OpenTUI. Render requests queue ahead of waiting preloads, so a visible file waits behind at most one grammar. Warm-up is best-effort and skipped where the worker is not used.
• Worker by default. Interactive highlighting offloads wherever supportsHighlightWorkerOffload() holds and the theme has no syntax_scopes overrides. Inline stays for Windows compiled binaries, scope-override themes, and static piped output. The offloadLargeDiff prop plumbing from App down to the highlight hooks is removed; loadHighlightedDiff keeps an offload option for the static pager, and the document service consults eligibility alone.
• HIGHLIGHT_WORKER_MIN_LINES removed. It was checked against source-backed metadata so nearly every real file already qualified, and with the worker as the default any file below it would have paid the full grammar compile on the main thread for a handful of lines.
• --fast is a no-op. Still parsed so aliases and existing invocations do not error; hidden from --help and the public reference; the changeset notes the deprecation for removal in a later release.
• Permanent worker refusals fall through to inline. Non-retryable worker errors (unsupported-language and friends) now take the inline path, which re-renders as plain text and caches the result, instead of being marked retryable and re-requested on every scroll while the file sits in the prefetch halo.
• Tests and benchmark. A PTY test launches a three-grammar review, sends page-down just after first paint, and requires the answer within max(60ms, 4× the settled answer) for the same key once color has landed. Cherry-picks the spawn-inclusive startup benchmark from Startup: begin loading OpenTUI once the plan is interactive, overlapping git bootstrap #1062's investigation and extends it with startup_first_color_ms and startup_post_paint_key_max_ms.

Measurements

Source mode, Linux x64, benchmarks/startup-first-frame.ts, 7 launches per side, same machine:

Metric	main	this branch
startup_first_frame_ms	435	462
startup_first_color_ms	639	561
startup_post_paint_key_max_ms	105	34

The +28ms first frame is the worker thread starting alongside the OpenTUI import, the same order as the issue's prototype. The PTY latency test measured the first page-down at 72–105ms on main and 6–28ms here across repeated runs.

Validation

• bun run typecheck, bun run lint, bun run deps:check, bun run check:docs
• bun run test: 2245 pass, 4 fail. All four reproduce on unmodified main in this environment: test/cli/entrypoint.test.ts and test/cli/log.test.ts pick up my user config.toml (watch = true), packages/hunk-jj/src/source.test.ts needs jj on PATH, and brokerClient.test.ts > authenticates after a successor becomes healthy is a 5s timeout flake.
• bun run test:integration: 184 pass, 1 fail. The failure is watch-start-integration.test.ts's teardown assertion (child.exitCode still null 500ms after SIGKILL), which fails 2 of 4 runs on main as well.
• PTY smoke of hunk diff and hunk show HEAD~1 on this repository at 140×30, unified layout, github-dark-default; navigation and color land as before.
• Static hunk show HEAD~1 | head, hunk --fast show HEAD~1 | head, and hunk diff --help from source.
• Not tested: macOS, Windows. The Windows compiled-binary path is unchanged by design (supportsHighlightWorkerOffload still gates it, warm-up skips it).

Follow-ups

• fix(ui): bound and surface large-diff highlight retries #785 (bounded worker retries) should land first or alongside.
• hunk log's interactive history surface does not warm the worker; it has no file list when the plan is selected, so its first review still starts a cold worker.
• Remove --fast in a later release.

🤖 Generated with Claude Code

[vercel]

@masonmcelvain is attempting to deploy a commit to the Modem Team on Vercel.

A member of the Team first needs to authorize it.

[greptile-apps]

PR author is not in the allowed authors list.