[colbymchenry]

Summary

• codegraph sync and the file watcher now link import { useInvocation } from './context', and every use of the name, when an edit gives the barrel context/index.ts that re-export after the importing file was indexed. Before, the import, its calls and its JSX or template tags stayed unlinked until the importing file changed or the project was re-indexed.

• Why it stayed unlinked: the barrel declares none of the names it passes on, so sync's retry by name never reached the importers' failed references. The edit adds no file and changes no definition, so neither the importer re-open through added files (fix(sync): imports through a re-exporting module follow it when it is deleted or added #2512) nor the barrel walk (fix(sync): a lazy route relinks a page that appears below its barrel #2560 / fix(sync): an import a barrel passes on under another name relinks when its module gains it #2665) runs.

• The change: importBindingTail now parks the whole name of a plain named import under its module's key (module:context), as fix(sync): a reference through an import binding links its module when the module appears later #2450 does for default, namespace and aliased bindings and fix(sync): a Python import of a name a package re-exports links when the package appears later #2507 for Python's unaliased imports. It does so only while:

  • the import resolves to a project file that carries that key (the file whose edit the key exists for), and
  • that module doesn't give the name yet.

  The by-name lookup from fix(sync): a reference through an import binding links its module when the module appears later #2450 still finds the reference by its own name.

Re-verified on current main first

The shapes originally reported (late-file hold scenario) now converge on main without this change:

• a barrel added after its importers (milo-ui useInvocation);
• a file behind a barrel added later (vben ThemeToggle, passed on as default as ThemeToggle).

#2512's importer re-open and #2560 / #2665's barrel walks cover them. A probe on ada1445 passed both add cases and failed the edit case with 2 of 8 links. This PR covers the edit.

Why the rule is narrow

Three versions were measured on the same 18 repos. All three fix the same barrel-edit cases; they differ only in how many references they park, which sync then retries whenever a file with that key changes.

	parks every plain named import	+ module file must carry the key	+ module must lack the name (this PR)
refs re-parked by a fresh index, 18 repos	11,061	7,484	2,727 (10 repos byte-identical)
vscode	—	39,082	5,074
retried per changed file vs main (mean), folo / vben / milo-ui	+15% / +19% / +17%	+5% / +3% / +17%	+3% / +1% / +3%

What the narrower rules leave out can't be satisfied through the key:

• a workspace package's own name (@vben/icons, @follow/constants) resolves to the package folder, and its entry file (src/index.ts) never has the folder's key;
• an import that resolves to nothing;
• a reference whose module already gives the name but which fails anyway, such as a constant or class passed where a function is expected. That covers 34,015 of vscode's 39,082 and almost all of ghostfolio's (Angular DI tokens).

Tests

__tests__/sync-import-binding-retry.test.ts:

• New describe, "sync links a plain named import through a barrel": the barrel added, a file behind it added, and an edit that gives the barrel the names it forwards. Each is compared edge for edge with a fresh index of the final tree in a second folder.
• New tail assertions:
  • forwarded from a barrel yet to pass it on → module:lib;
  • a missing module → name tail;
  • @main, an alias of a file with another key → name tail;
  • LIMIT, a constant its module exports, passed as a function → name tail.

Red on main ae0fc63: the barrel-edit case (2 of 8 links) and the tail test fail. The two add cases pass on main through #2512 / #2665 and stay as regression coverage.

Mutations:

• neutralizing the key check fails the @main assertion (module:@main);
• neutralizing the "module gives the name" check fails the LIMIT assertion (module:consts).

Full suite on this branch, with main 51be28d merged, the kernel built and CODEGRAPH_KERNEL_EXPECT=1: 689 files, 8,185 passed, 0 failed, 82 skipped. The related sync suites (12 files) passed as well.

A/B (main ae0fc63 vs this branch, 18 repos, each arm's sync vs its own fresh index)

Fresh index: nodes, edges and files are identical in all 18 repos. 10 repos are byte-identical (controls include gin, django-realworld and vue-realworld). The rest differ only in the tail of refs that fail on both arms.

Late sync, every 7th source file:

scenario	main	this PR
gut (emptied, then restored by an edit): vben	2,337 lines	194
gut: vue-element-admin	18	6
gut: milo-ui	1,078	1,050
gut: folo	134	126
hold (held out, then added): vue-element-admin	8	0
hold: vben	1,702	2 *
hold: infrahub	0	300 **
route (route targets held out), all 9 repos	0	0

• In every changed gut scenario, all failed-reference residue relinks: vben's 807 $t calls and its layout components, vue-element-admin's TagsView, Sidebar and Settings, milo-ui's chart utils, folo's NoopChildren. What remains on both arms is same-edge metadata and retargets, which predate this change.
• Every residue bucket went down (by 1,938 in total) except the one marked **.
• * vben hold: this is a side effect, not part of the design. Its 844 failing $t refs end up split between their name tail (462) and module:locales (382), so both by-name lookups fall under the 500-name ceiling. On main all 844 sit under one name, over it.
• ** infrahub hold: ref rows only, with identical edges. Refs into generated/types.ts (2.1 MB, indexed with no nodes) failed while the file was held out, kept their name tail, and nothing re-resolves them after it returns. A fresh index parks them under module:types. They fail on both arms; only the bookkeeping differs.

Cost

• Retry volume per changed file (mean, vs main): +1% to +3% on folo, vben, ghostfolio, infrahub, milo-ui and bulletproof; +13% on takenote.
• Hot-file syncs (median of 3, the files whose keys hold the most parked refs): the retry phase grows by +3 to +117 ms. The largest is folo constants/ui.ts, 89 → 206 ms. Whole syncs stay within noise. These were timed with the "module file must carry the key" version, which parks a superset of this PR's refs.
• vscode:
  • nodes, edges, files and refs identical apart from tails;
  • module-parked failed refs 10,541 → 15,753;
  • the retry lookup for hot files (uri.ts, lifecycle.ts) is unchanged;
  • full index: main 194.0, 194.7 and 193.1 s; this branch 183.9 and 183.1 s, both run uncontended. A third run of this branch overlapped another session's kubernetes index and took 232 s, so it is left out.

The A/B ran on ae0fc63; this branch was then synced to 51be28d (#2743, extraction), which touches none of its files. The validation kit lives in D:\cgwork\ubarrel and is deleted after the merge.

🤖 Generated with Claude Code