blackwell-systems/agent-lsp

Three features that make agent-lsp work on large Python and TypeScript codebases.

Persistent daemon mode for Python and TypeScript

Language servers like pyright need minutes of background indexing before get_references works on large repos. agent-lsp now automatically spawns a persistent daemon broker that survives between sessions:

• First session: daemon starts, indexes in the background (~10s for FastAPI)
• Subsequent sessions: instant connection to the warm daemon
• Auto-exits after 30 minutes of inactivity
• Go, Rust, and other fast-indexing languages bypass this entirely (zero overhead)

Validated on FastAPI (1,119 Python files, 80K stars): get_references returns 1,214 results across 556 files instantly. Previously timed out at 5 minutes on every attempt.

New CLI commands: agent-lsp daemon-status, agent-lsp daemon-stop

/lsp-inspect skill (21st skill)

Full code quality audit for a file or package. Combines get_change_impact (batch dead symbol + test coverage analysis) with LLM-driven heuristic checks:

• dead_symbol, test_coverage (LSP-verified)
• silent_failure, error_wrapping, coverage_gap, doc_drift, panic_not_recovered, context_propagation (heuristic)

Runs inline with zero setup. Replaces the external inspector agent (no background agent spawning, no permission gates, no warmup flag files).

Workspace scoping via scope parameter

start_lsp now accepts a scope parameter that generates a temporary language-server config limiting analysis to specific subdirectories:

start_lsp(root_dir="/repo", language_id="python", scope="src/mypackage")

Generates pyrightconfig.json (Python) or tsconfig.json (TypeScript) with restricted include paths. Auto-cleanup on shutdown with backup/restore of existing configs.

Multi-signal warmup gate

For servers that don't emit $/progress tokens (pyright, jedi-language-server), a new readiness gate waits for diagnostics or hover confirmation before issuing reference queries. Extended first-query timeout (5 min) with guidance message on failure. Zero overhead for gopls/rust-analyzer.

Performance

• get_change_impact: 100x faster on large files. Parallel worker pool (8 goroutines), GetReferencesRaw skips per-file waits, warmup query reused, struct fields excluded, context cancellation in workers.