How does Bifrost measure up against language servers?

Immutable evidence · UsageBench v0.3.1

Short answer

Against ten reference language servers on the reviewed legacy core, Bifrost is exact on more cases than the reference server in 7 of 10 comparisons, level in 2, and behind in 1 (rust-analyzer). Each language server is compared only on the cases it and Bifrost both answer, so a reference server is never charged for an operation it does not offer.

UsageBench is BrokkAi's own benchmark: we build it to measure where Bifrost stands against established language servers, and to keep ourselves honest while doing it. Every number here is generated from digest-bound freeze evidence we cannot edit after the fact — our own misses and regressions included.

Tier: these 110 cases are the reviewed legacy core — retrospectively selected from an analyzer-informed corpus, then re-reviewed from source alone. It is the broadest comparison we publish, and deliberately not a preregistered one. The prospective slices carry that stronger claim over fewer languages.

Frozen evidence · 182 cases

• Preregistered

  36 cases

  Prospective v1

  v0.2.0

• Preregistered

  36 cases

  Prospective v2

  v0.3.0

• Retrospective

  110 cases

  Reviewed legacy core

  v0.3.1 · detailed below

Each slice keeps its own denominator and they are never pooled. The preregistered slices select and review sources before any analyzer runs; the legacy core was selected retrospectively and re-reviewed from source alone.

Ruby LSP

0.26.10

Bifrost

9/9Ruby LSP

2/9

Both exact

2

Bifrost only

7

Reference only

0

Eclipse JDT LS

1.61.0-202607142124

Bifrost

10/10Eclipse JDT LS

4/10

Both exact

4

Bifrost only

6

Reference only

0

TypeScript LS

5.3.0 (TypeScript 5.9.3)

Bifrost

16/16TypeScript LS

12/16

Both exact

12

Bifrost only

4

Reference only

0

clangd

21.0.0

Bifrost

10/10clangd

7/10

Both exact

7

Bifrost only

3

Reference only

0

Roslyn

vscode-csharp 2.140.9

Bifrost

10/10Roslyn

7/10

Both exact

7

Bifrost only

3

Reference only

0

Metals

1.6.8

Bifrost

6/7Metals

5/7

Both exact

4

Bifrost only

2

Reference only

1

Intelephense

1.18.5

Bifrost

7/7Intelephense

6/7

Both exact

6

Bifrost only

1

Reference only

0

Pyright

1.1.411

Bifrost

10/10Pyright

10/10

Both exact

10

Bifrost only

0

Reference only

0

rust-analyzer

2026-07-13

Bifrost

8/10rust-analyzer

10/10

Both exact

8

Bifrost only

0

Reference only

2

gopls

0.23.0

Bifrost

5/5gopls

5/5

Both exact

5

Bifrost only

0

Reference only

0

Strict-contract conformance across 94 shared case comparisons, from release v0.3.1 at revision 8393dafe6b9a. A case counts as exact only with complete token ranges, no unallowed extras, and the one reviewed navigation target. Disagreement is a contract result, not an automatic defect verdict — see the evidence map for every hash and denominator.