blackwell-systems/agent-lsp

v0.15.1: Bug fixes and 20-48% fewer tokens on nested data

Bug fixes

Scope array double-encoding (#10): start_lsp(scope=["packages/a/src", "packages/b/src"]) was writing a pyrightconfig.json with a single-element include containing the stringified JSON array instead of the individual paths. pyright then matched zero files, causing find_references and blast_radius to silently return empty results. Reported by @nirabo.

Windows drive-letter casing (#9): C# diagnostics and code actions were silently empty on Windows because csharp-ls keys documents under lowercase drive letters (file:///c:/...) while agent-lsp queried with uppercase (file:///C:/...). Drive letters are now normalized to lowercase on both the storage and lookup sides. Reported by @SebastiaanSteenbrink.

Token savings improvement

gcf-go upgraded to v1.3.1. Nested object flattening: fixed-shape nested objects are inlined into parent rows using > path columns instead of separate attachment blocks. 20-48% fewer tokens on nested tool responses (blast_radius callers, diagnostics, explore_symbol). Graph profile output unchanged.

Before (v1.1.0):

@0 AuthMiddleware|fn|src/auth.go|42|^
.refs [3]{file,line,kind}
    src/server.go|15|calls
@1 ValidateToken|fn|src/auth.go|67|^
.refs [3]{file,line,kind}
    src/auth.go|45|calls

After (v1.3.1):

@0 AuthMiddleware|fn|src/auth.go|42|^
.refs [3]{file,line,kind}
    src/server.go|15|calls
@1 ValidateToken|fn|src/auth.go|67|^
.refs [3]
    src/auth.go|45|calls

Shared array schemas eliminate redundant field headers across records.

GCF

agent-lsp uses GCF (Graph Compact Format) as its default output encoding. GCF delivers 30-84% fewer tokens than JSON across all tool responses, with session dedup compounding to 92.7% savings by the 5th call.

GCF adoption is accelerating: it was recently merged into the Google Chrome DevTools MCP server, joining agent-lsp, knowing, and NeuroNest as production adopters.

Website: gcformat.com · Spec · Go · Python · TypeScript · Proxy · Playground