React Compiler Migration

React Compiler Migration

Make a file compile. Nothing else. The marker tells you what is broken and, after the edit, whether you actually fixed it.

Rule

react-compiler-marker is the source of truth. Run it before the first edit and after every edit. A fix is not done because the reasoning is sound; it is done because the failure count dropped.

It earns that status: it loads the workspace's own babel-plugin-react-compiler and runs a real transformFromAstSync with noEmit: false, failing if codegen fails. It is not a lint pass or a heuristic. It also runs with compilationMode: "infer" and panicThreshold: "none" — the plugin defaults this repo builds with — plus enableTreatRefLikeIdentifiersAsRefs: true, which the build does not set. That last flag makes it stricter than the app build, so a marker-clean file is build-clean, not the other way round.

Know the one thing it cannot tell you: the compiled output is generated and thrown away, never executed. Jest disables the compiler (scripts/react-compiler.js gates on NODE_ENV === 'test'), so a green test run proves the source still behaves, not the compiled version. Codegen is verified; runtime behavior of the memoized output is not. Say so when reporting anything beyond a trivial fix.

Never fix a bail-out by suppressing it. "use no memo", a fresh eslint-disable, or deleting the behavior that failed are all non-fixes.

Workflow

1. Baseline

The CLI scans directories, not files, and resolves babel-plugin-react-compiler relative to the scanned path — so pass the plugin path explicitly and run from the repo root:

npx -y react-compiler-marker@latest \
  --babel-plugin-path node_modules/babel-plugin-react-compiler \
  app/path/to/containing/dir

Record the failing lines for the target file. If the file does not appear in the failures, there is nothing to do — say so and stop.

Capture the same run with --format json ... 2>/dev/null now if you intend to report before/after memo metrics (step 4). A failing file has no success entry to compare against later, and reconstructing a baseline afterwards means stashing your work in a dirty tree — not worth it.

The reported count is a lower bound, in two ways. A component skipped wholesale (see the ESLint-disable pattern) hides every other diagnostic inside it, and the compiler stops at the first failure in each function — so an identical construct three functions down stays invisible until you fix the first one. Counts often rise on the next run, and a "2 failure" file can take four rounds. That is progress, not regression; report it as such.

The converse is equally normal: plenty of files are one edit and done. Stop when the marker says zero — do not keep digging for a cascade because this section warned you one was possible.

When a fix reveals a new reason, grep the file for the construct you just fixed before re-running. Sibling copies are the common case, and knowing they exist tells you whether you are two edits from done or twenty.

2. Understand before editing

Read the failing lines and the code that made someone write them. Suppressions and defensive try/finally blocks usually encode a real bug someone hit. Preserve that behavior; change only the shape the compiler cannot digest.

Look up the failure in references/patterns.md. Each entry has a validated before/after.

Some failures should not be fixed, and saying so is a successful outcome. Check these before editing:

• Test files. *.test.tsx shows up in reports and gains nothing — the compiler is off under Jest, so migrating a test optimizes code that never runs compiled. They are also the files most likely to show pruned memo blocks. Skip them.
• False positives. Some diagnostics name a rule of React rather than a syntax the compiler cannot lower. Establish the code actually breaks the rule before restructuring anything — Hooks may not be referenced as normal values fires on any use[A-Z] identifier, including a plain boolean prop.
• Fixes that live outside the file. If the only real fix is renaming an export, changing a shared hook, or editing another team's module, stop and report it with the suggested change. A local workaround for a non-local cause is worse than the bail-out.

Report these as findings with the reason and the suggested owner, and move to the next file.

3. Edit

Follow the principles below. Make one change at a time and re-run — the marker is cheap, and it is the only way to learn which specific construct the compiler objected to. Guessing at two fixes at once wastes the signal.

4. Verify

Re-run the exact command from step 1. Confirm the file's failures are gone and no new failures appeared elsewhere in the directory.

Zero failures is necessary, not sufficient. Add --format json and read the per-function counts: memoBlocks and memoValues are what the compiler managed to memoize, and prunedMemoBlocks / prunedMemoValues are what it had to discard. A file can compile with zero failures and still be partly de-optimized, which is the actual thing the migration is for.

The CLI writes progress to stderr, so redirect it or the output will not parse:

npx -y react-compiler-marker@latest \
  --babel-plugin-path node_modules/babel-plugin-react-compiler \
  --format json app/path/to/dir 2>/dev/null

Expect prunedMemoBlocks and prunedMemoValues to be 0 on a finished file — a fix that clears the failure while pruning memo blocks is not a win.

A file that was failing has no success entry, so its "before" memoBlocks is undefined rather than zero. Do not go stashing changes to reconstruct it: either capture the JSON in step 1 before your first edit, or report before-metrics as N/A and give the after-figures alone. Ignore test files here; they routinely show pruning and gain nothing from being migrated. Then:

yarn prettier --write <file> && yarn eslint <file> && yarn jest <dir>

Add yarn lint:tsc when the change has type surface — a new helper signature, a changed parameter type, a Pick<...>. It checks the whole project and takes about a minute, so skip it when the diff is purely statement-level (relocating a finally, hoisting a const) and say that you did.

All three matter. A compiler-clean file with a new ESLint error is not done — this repo still enforces react-hooks rules, so keep the useMemo/useCallback the linter asks for even where the compiler makes them redundant. Stripping manual memoization is a separate codemod for after the migration, not part of it.

If failures remain, go back to step 2 — do not report a partial fix as done.

5. Report

State the diff, why the original code existed, what behavior is preserved, and the before/after failure counts. Flag any new type assertion or behavioral risk for review rather than burying it.

Give the change a risk tier, because a reviewer cannot infer it from a green checklist — every fix in this skill produces the same "0 failures, tests pass" line regardless of blast radius:

• Rename-level — a swap to an existing API, or hoisting an expression into a const. Behavior is unchanged by construction.
• Structural — control flow moved: finally relocated, a try reshaped, a helper extracted from a hook. Behavior is preserved by an argument you should state, and the argument is what the reviewer checks.
• Stateful — anything touching a cache, a ref's lifetime, or memo dependencies that decide when work re-runs. Name the screen to smoke-test, and remember Jest ran this code uncompiled.

Principles

These outrank compiler-satisfaction. A file that compiles but reads worse is a failed migration.

1. let is a smell. Reassignment means the reader has to track a variable through the function. Prefer a const with a ternary, an extracted function with early returns, or a derived value.
2. Minimum diff. Change the lines the compiler objects to. Do not reformat, rename, reorder imports, or "improve" untouched code in the same pass. Most of the compiler's restrictions are lexical — they care where a construct is written, not what runs — so the fix is usually to move that construct behind a helper, not to restructure the logic around it. New branches in the diff mean you took the wrong route.
3. Look for the existing API before writing one. The construct the compiler rejected is rarely unique to this file, and the library that supplied it often ships a hook that already hides it (useAnimatedValue for useRef(new Animated.Value(x)).current). Grep for the candidate first: with precedent in the repo it is a one-line rename and the smallest possible diff; without precedent, it is a new API in a migration pass and belongs in the report.
4. Extract freely. A closure or a plain function above/below the component is cheap and usually clearer than inlined branching. Module scope is also outside the compiler's analysis, which makes it the escape hatch for syntax the compiler rejects. Split at a seam that already exists — a helper needing ten parameters means you cut in the wrong place. Match the file's existing naming; a fresh alert or data parameter can trip lint rules the file was already avoiding. Move the body verbatim so the diff reads as a relocation, and land it inside the same ONLY_INCLUDE_IF build fence as the imports it uses.
5. Stay in the file. No new files unless the user asks. Do not edit the upstream hook or shared util to fix a local bail-out.
6. No comment slop. No JSDoc on new functions. No comment restating what the next line does. Keep a comment only when it explains a constraint the code cannot show — a non-obvious invariant, or the bug a workaround prevents. Delete comments that described a suppression you removed. A helper that exists solely to satisfy the compiler qualifies: one line leading with why it was extracted and naming the diagnostic that forced it, so the next reader can search for it and knows when it can be deleted.
7. Extraction moves code out of a restricted region; it never changes what that region covers. Shrinking the span of a try, a catch, or an effect is a behavior change wearing a refactor's clothes. If the compiler-friendly shape genuinely cannot cover the same span, say so instead of shipping it quietly. There is exactly one enumerated exception: converting a finally into a trailing statement, which does shrink a protected region and is permitted only when both preconditions in the try/catch pattern entry are cleared and stated. Anything else that narrows a protected span is a finding, not a fix.

Patterns

One entry per validated failure reason, with before/after: references/patterns.md.

Append a new entry only after a fix has been verified by the marker and accepted in review. An unvalidated pattern is worse than no pattern.