[gezerui888]

📋 Description

doctor rust reports the compiled CodeLLDB version when CODELLDB_PATH points to an extracted VSIX. That layout stores its version in extension/package.json.

I read the selected engine's manifest after its version.json and before metadata from other installations. I added resolver and built-CLI regressions for this order and the existing fallback behavior.

Fixes #1066

🔄 Type of Change

• 🐛 Bug fix

✅ Checklist

• My code follows the style guidelines of this project
• I reviewed my code
• I commented the test setup where needed
• I added a changelog fragment
• My changes generate no new warnings
• I added tests that reproduce the defect
• New and existing unit tests pass locally with my changes

🧪 Testing

I ran these checks with Node 22.23.4, Python 3.12 and JDK 21 on macOS arm64.

pnpm run lint
pnpm run typecheck:all
pnpm run clean
pnpm run build
pnpm run test:unit --maxWorkers=4
pnpm run test:integration
pnpm run check:docs
pnpm run check:all-personal-paths
pnpm run changelog:check

The unit suite passed 8,094 tests. The existing DartAdapterFactory default io SDK-discovery test timed out on both this branch and clean main. Both runs skipped the same three tests. Main passed 8,086 tests.

Integration passed 121 tests, compared with 118 on main. Both runs skipped the same nine tests. The raw-VSIX CLI regression failed on main and passed after the fix. Lint, type checks, clean build, documentation, path and changelog checks passed.

cc @debugmcpdev

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-11-05T02:57:23.965225Z	d9e5136	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[debugmcpdev]

Reviewed line by line and hardened on the branch (the attribution rule, the manifest guard, CI-sized doctor budgets, the canary assertion); the pre-push hook and all fifteen CI checks are green on the final head.

[debugmcpdev]

Thank you, @gezerui888 — a correct fix with real regression tests on both layers, and the first PR here from you. Merged.

Rather than a review round-trip I took the branch over and built on your commit (it stays yours in the squash). What I added, and why:

• The version is now read only from the resolved binary's own root. Your candidate was the right one, but the walk still went on to the other installs' version.json (the four vendor candidates, the platform package) and then the compiled default when neither sibling file existed, so a CODELLDB_PATH binary without metadata was still reported with some other copy's number — the same misattribution doctor: getCodeLLDBVersion under CODELLDB_PATH reports the CLI's compiled default — a raw VSIX extraction has no version.json #1066 describes, one step narrower. A vendored copy with no readable metadata keeps the pin; any other source now reports no version at all rather than a guess.
• The manifest is accepted only when its name is vscode-lldb (the upstream extension's), so a project's own package.json two levels above a hand-placed binary (<myapp>/bin/codelldb) is never reported as the engine version; the path is realpath'd first so a symlinked CODELLDB_PATH reads where it points.
• A version.json that parses but has no version now falls through to the manifest instead of short-circuiting to the constant (your table pinned the old behaviour honestly; it was the one asymmetry left).
• CI-sized budgets in the integration test: the rust probe runs cargo -vV, which the canary has seen take over 30 s on a cold Windows runner, and the doctor's default 10 s probe budget would have timed the row out — so doctor runs with --timeout 60000, the kill at 90 s, each case at 120 s; the empty-stderr assertion went (a Node warning from the bundle is not a resolver fault); the bundle is copied once per file; and a fourth case proves the bare-binary answer.
• The canary's CODELLDB_PATH leg asserts the engine version for releases after 0.26.0 — the check the issue was filed to make possible.

Thanks again — the next release carries it, with the changelog line crediting you.