Code Intelligence — Code-world grounding

🌐 日本語

NOTE

The Language Server Protocol (LSP) lets Claude consult the symbols, types, and references that actually exist in this repository instead of guessing. Where MCP grounds the LLM in external facts, Code Intelligence grounds it in the code in front of it.

How it works

• Definition — confirms a symbol exists; a null answer is proof of non-existence before code is committed.
• Hover — returns the real type signature (RxJS operator overloads, Angular signal API, etc.) so generation is constrained to the installed version.
• References — lists exact call sites without loading every file, cutting investigation tokens dramatically.
• Diagnostics — push type and lint errors during the edit loop, so most Hallucination-class symbol errors never reach disk.

Related chapters

• Part 9 — Code Intelligence (overview)
• LSP as Grounding
• Hallucination and Symbols
• Live Type Errors
• Grep / Read / LSP — Which Tool When?

Structural problems it addresses

• Hallucination — symbol-level grounding stops the most common code-generation failures (made-up function names, wrong signatures, fake import paths) before they reach disk.
• Knowledge Boundary — project-private types and post-cutoff APIs become resolvable instead of guessable.
• Context Rot (secondary) — symbol queries return precise targets instead of forcing whole-file reads, leaving more budget for actual work.

See also

• Topic: MCP — the other grounding pillar (external-world facts).
• Topic: Hooks — test-execution Hooks catch semantic bugs LSP cannot see.
• Feature Index