fix(jsx): a .jsx component links to the file-scope values it reads, as in .js and .tsx

[colbymchenry]

Summary

A .jsx file got no same-file value-reference edges: the references edges (metadata valueRef: true) from a reader to the file-scope const/var it reads, which codegraph impact and the impact trail in codegraph_explore / codegraph_node follow. TreeSitterExtractor.VALUE_REF_LANGS listed typescript, javascript and tsx, but not jsx, and the kernel mirrored the gap. So a component in a .jsx file had no edge to the lazy page it renders (<LoginPage /> under const LoginPage = Loadable(lazy(…))), the route table it builds or the config it maps over, while the same code in .js or .tsx had all of them.

Probe: the same React source through extractFromSource as .jsx, .tsx and .js:

	.jsx	.tsx	.js
main, wasm path	0	4	4
main, kernel path	0	4	4
this PR, both paths	4	4	4

Was leaving out jsx deliberate?

No sign of it. git log -G VALUE_REF_LANGS: the set was introduced in ec90ddf (2026-06-13) as ['typescript', 'javascript', 'tsx'], and #897 and #1186 (ArkTS) extended it. None of them mentions jsx. The design doc's coverage table says "TypeScript, JavaScript, tsx … the original languages" and never names jsx. Every other JS-family gate in the extractor includes jsx: declarators, import bindings, re-exports, store detection and the fn-ref spec. The kernel's Variant::value_refs() existed only to copy the TS set bug-for-bug, with the comment "VALUE_REF_LANGS includes typescript/tsx/javascript but NOT jsx". .jsx uses the same JavaScript grammar as .js, so nothing about it needs different handling.

Change

• src/extraction/tree-sitter.ts: add 'jsx' to VALUE_REF_LANGS.
• codegraph-kernel/src/tsjs/mod.rs: drop Variant::value_refs() and its two early returns, so every TS/JS variant runs value refs, as on the wasm path.
• Docs: the value-reference design doc and playbook now list jsx (and the playbook's list also gains arkts, already in the set).

Tests

• value-reference-edges.test.ts: a .jsx file indexed end to end: AuthRoutes → LoginPage (a lazy page rendered as a tag) and Nav → NAV_ITEMS (a table mapped inside JSX).
• kernel-tsjs-parity.test.ts: the same source as tsx, js and jsx through the kernel and the wasm path, asserting parity and the exact four value-ref edges.
• On main both jsx cases fail and the tsx/js ones pass. With the fixed TypeScript but main's kernel (CODEGRAPH_KERNEL_PATH), the parity case fails (kernel 8 edges, wasm 12), so a fix on one side only can't pass.

Validation

Before/after full indexes of the same repos, kernel-routed as users run them, dumped with scripts/dump-graph.mjs and diffed by natural keys. The pair was rebuilt and rerun on each main this branch merged: d8a7f868, ed199e60, e47cb258, and finally d0bb9be5, which includes #2420. Every run added the same edge set, line for line.

Repo	.jsx files	nodes	refs	edges removed	edges added
codedthemes/mantis-free-react-admin-template	155	same	same	0	35
uilibrary/matx-react	136	same	same	0	56
codedthemes/berry-free-react-admin-template	103	same	same	0	15
mastodon/mastodon (Rails + React, mostly .tsx now)	92	same	same	0	28
bigbluebutton/bigbluebutton, bigbluebutton-html5	201	same	same	0	266
alan2207/bulletproof-react (control, no .jsx)	0	same	same	0	0, byte-identical

Every added edge was checked against the source with a parse (the grammar the extractor uses for .jsx). Each one:

• is a references edge with {"valueRef":true}, with both ends in the same .jsx file,
• targets a file-scope const/let/var,
• has an identifier read of the target inside the reader's span,
• and its target is the only binding of that name in the file: no parameter, destructured local, import or catch binding of the same name. That matters because the shadow prune counts only const/let/var declarators, so a same-named parameter would get past it. That gap predates this PR and shows in .js and .tsx with a synthetic function F({ headCells }). An audit of all 2,201 JS/TS value-ref edges in these repos, including the ones .js/.tsx files already had, found no real case. A default value like { error = defaultError } is a read, so its edge is correct.

All 400 pass. By what the target is:

Target	Edges	Example
intl message table (defineMessages)	149	bbb's intlMessages, read by render and handlers
scalar constant	62	DEBOUNCE_TIME, ANIMATION_NONE
object table / sx style object	56	mantis avatarSX, matx initialAuthState
array table	47	mantis headCells, chart data like monthlyLabels
lazy page (Loadable(lazy(() => import(…))))	35	matx materialRoutes → AppButton, berry MainRoutes → DashboardDefault
context (createContext)	14	ConfigProvider → ConfigContext (<ConfigContext.Provider>)
module-level let assigned later	6	mastodon let EmojiPicker, Emoji;, set in onShowDropdown, rendered in render
other (regexes, selectors, a Yup schema, new DOMParser())	31

In 201 of the 400, the reader uses the value inside JSX, as a tag (<LoginPage />) or in an expression ({rows.map(…)}). In the other 199 it's plain code, such as a handler reading a message table.

What a user sees (codegraph impact, before → after):

• matx AppButton (const AppButton = Loadable(lazy(() => import("./buttons/AppButton")))): 1 → 3 affected. The materialRoutes table that renders <AppButton /> is now affected, and so is routes.jsx, which imports that table.
• mantis headCells (a column table): 1 → 3 affected. OrderTableHead, which maps over it, is now affected, and so is OrderTable, which renders <OrderTableHead>.

scripts/kernel-parity.mjs --lang jsx,javascript,tsx,typescript over the five .jsx repos with the fixed build: 0 files with diffs. 2,556 files are byte-identical and 37 defer to wasm on parse errors, the same 37 as with the unmodified build. Nodes and refs are unchanged; only edge totals grow.

Test suite. On the final branch (main d0bb9be5 + this change), with the kernel staged and CODEGRAPH_KERNEL_EXPECT=1:

• full suite: 7,091 passed, 0 failed, 80 skipped;
• the build's tsc is clean.

An earlier full run, on a box saturated by other sessions, had 5 failures in daemon/MCP/process tests (mcp-daemon, mcp-status-freshness, mcp-writer-lock, daemon-pid-reuse, and function-ref's 60 s #1820 case). None of those tests indexes a .jsx file. Each failure reproduced on main under the same load in interleaved runs, or passed on rerun.

Overlap

• fix(react): a component a file declares itself as a value is what its JSX renders #2420 (merged as d0bb9be5) routes every PascalCase reference from a .tsx/.jsx file through declaredComponent in the React resolver's Pattern 1. Value refs never reach it: they are finished edges emitted during extraction, not references for a resolver. Accordingly, every A/B above leaves refs unchanged, including the one on d0bb9be5, and jsx-child-disambiguation.test.ts (fix(react): a component a file declares itself as a value is what its JSX renders #2420's cases) passes 20/20 on this branch.
• fix(react): a JSX tag renders what the file imports under that name #2451 (open) changes which component a JSX tag renders in the jsx-render synthesizer. Those are synthesized edges, and the synthesis merge dedupes them only among themselves, so the two changes don't interact. They share only CHANGELOG.md.
• Main's other React/JS work since the original base (fix(react-router): a JSX index route is the page at its parent's address, inside the layout around it #2423, fix(js,ts): a default import is what export default declares, not the first exported function #2432, fix(react-router): a lazy route links the page a barrel module forwards #2436, fix(sync): a route rendering a same-file lazy value follows its module, and a new component gets its JSX edges #2452, fix(react,vue): a tag imported from a package renders no project namesake #2475, fix(react): a tag that passes type arguments renders its generic component #2476, fix(react-router): a lazy route links the page its loader picks #2480, fix(react-router): a lazy route follows a barrel that re-exports an import #2558) is resolution or synthesis. fix(react): a typed styled component is a component, not a constant #2429 is extraction: it makes styled components component nodes, so in every JS dialect they are neither value-ref targets nor readers. All of it is included in the final A/B.

🤖 Generated with Claude Code