Concurrent callers raced past the tree-sitter load and got regex

[nedonatelli]

The reason tree-sitter still was not running after #48 shipped it.

The race

if (!treeSitterLoadAttempted && extensionGrammarsPath) {
  treeSitterLoadAttempted = true;      // flips immediately
  ...
  treeSitterAnalyzer = await ...        // assigned much later
}
if (treeSitterAnalyzer?.supportedExtensions.has(ext)) return treeSitterAnalyzer;
return regexAnalyzer;

The symbol indexer calls getAnalyzer concurrently for every file in the
workspace (Promise.allSettled over ~1000 files). All but the first saw the
flag already set, skipped the load block, found the analyzer still null, and
took the regex fallback.

The grammars loaded correctly the whole time — into a variable nobody awaited.

Why every previous test passed

A sequential call always worked. Every direct test of
createTreeSitterAnalyzer, including the ones I added in #45 and #48, called it
once and got a working analyzer. Only concurrency exposes this, and nothing
tested that.

Diagnosed from a real install

The SideCar output channel showed grammars loading at 13:51:15, and the graph
written that same second was regex-shaped:

kind	tree-sitter	regex	live graph
method	1487	0	6
function	2275	4468	4433
class	201	214	214
interface	611	639	639

method is the categorical discriminator — the regex analyzer emits none at
all, so their near-absence is proof rather than a judgement call about two
similar function counts. I initially mis-read this myself with a method > 0
threshold that called 6 a success.

Cost

Memoizing the load means every caller awaits it. Measured: 43ms for 28
extensions, so it is free.

Worth flagging that this contradicts the comment on getGrammarsPath, which
claims loading all grammars took "3m20s cold in the extension host". Either
that figure is stale or the host is far slower than Node; if the latter, this
delays first index rather than blocking the UI, and correctness beats a fast
wrong answer. Someone should re-measure in the host and fix whichever comment is
lying.

Testing

50 concurrent calls — one or two do not reproduce it. Mutation-checked:
restoring the fire-and-forget shape reports "50 of 50 concurrent callers fell
back to regex".

Relationship to #50

#50 bumps GRAPH_VERSION so existing caches repopulate. Still needed and still
correct — but it was not sufficient on its own, because a fresh rebuild also
used regex. Both are required for a user to actually get a tree-sitter graph.

npm run check: green.

🤖 Generated with Claude Code

[github-actions]

✅ All checks passed

Check	Status
Type check	✅ success
Lint	✅ success
Tests	✅ success

Posted by SideCarAI-Bot