blackwell-systems/agent-lsp

Performance: get_change_impact 100x faster on large files

The primary bottleneck for the /inspect skill and /lsp-impact analysis has been eliminated. get_change_impact on a 2,295-line file with 80+ exports previously hung for 20+ minutes. Now completes in under 30 seconds on cold workspaces, under 5 seconds on warm ones.

What changed

• Parallel worker pool (8 concurrent goroutines) replaces sequential reference query loop
• GetReferencesRaw: batch callers skip per-file WaitForFileIndexed, which was the 15s-per-symbol bottleneck
• workspaceLoaded flag: once all $/progress tokens complete, WaitForFileIndexed becomes a no-op for all subsequent calls. This is the proper readiness signal, not per-file diagnostic heuristics.
• Warmup query: one blocking GetReferences call absorbs the cold-start cost; result is reused (not discarded)
• Struct fields excluded: avoids 50%+ unnecessary reference queries on Go codebases where exported structs have many exported fields
• Context cancellation in workers for early exit
• Transitive references parallelized (was sequential)
• Test function deduplication in output

Impact

The /inspect skill now works reliably on external repositories. Large codebases (2,000+ line files, 80+ exports) that previously caused multi-minute hangs now complete in seconds.

Also in this release

• mcp-assert lint added to CI (fails on schema errors)
• Nullable array schema fix for Gemini 2.5 Flash compatibility (from v0.5.3)