feat(lsp): add lsp AI tool for code navigation (TS/JS)

[whitelonng]

Summary

Adds an lsp AI tool that lets the agent query a language server for code navigation. Currently supports TypeScript / JavaScript via typescript-language-server. This is the AI-tool-only version (no editor integration); a follow-up can add it to the Monaco/CodeMirror editor once that lands.

Supported operations

Operation	Input	Returns
goToDefinition	filePath, line, character	Location[]
findReferences	filePath, line, character	Location[]
hover	filePath, line, character	Hover (type info / docs)
documentSymbol	filePath	Symbols in file
workspaceSymbol	query	Symbols across project
goToImplementation	filePath, line, character	Location[]

Positions are 1-based (as editors display them); the tool converts to 0-based internally and back to 1-based in output.

Architecture

Agent calls `lsp` tool
  → builtin-lsp-tool.ts (LocalTool definition)
    → lsp-client.ts (JSON-RPC 2.0 over stdio)
      → spawn('typescript-language-server', ['--stdio'])

No external LSP client library (vscode-languageserver-protocol / vscode-jsonrpc). The client implements minimal Content-Length framing and request/response matching in ~200 lines — this avoids adding a new dependency to kun/ and keeps the PR self-contained.

Files

File	Lines	Description
kun/src/adapters/tool/lsp-client.ts	~330	Session management (reference-counted per workspace, 30s idle cleanup), JSON-RPC framing, initialize handshake, didOpen/didClose, 6 query methods
kun/src/adapters/tool/builtin-lsp-tool.ts	~270	LocalTool definition: hand-written JSON Schema (Kun convention), position conversion, result simplification (file:// → path, symbol kind numbers → names)
kun/src/adapters/tool/builtin-tools.ts	+2	Register in buildBuiltinLocalTools()

Key design decisions

1. No new kun/ dependency. JSON-RPC framing is hand-rolled (~100 lines) rather than pulling in vscode-languageserver-protocol. The framing logic is straightforward (Content-Length header + body) and has no edge cases for the query-only operations we support.
2. Reference-counted sessions. Like the bash tool's session map, LSP servers are keyed by workspace root and kept alive between tool calls. The last release schedules a 30s cleanup timer; a new acquire cancels it. This avoids respawning the server on every call.
3. Graceful degradation. If typescript-language-server isn't installed, the tool returns a helpful error (Install with: npm install -g typescript-language-server typescript) instead of crashing the agent loop.
4. Read-only. All operations are queries — policy: 'auto', toolKind: 'tool_call'. No approval needed.
5. Result simplification. Raw LSP responses contain file:// URIs, 0-based positions, and numeric symbol kinds. The tool converts these to plain paths, 1-based positions, and human-readable kind names so the output is compact for the agent's context window.

Prerequisites

The user (or their project) must have typescript-language-server installed:

npm install -g typescript-language-server typescript

Or locally: npm install -D typescript-language-server typescript (the tool checks node_modules/.bin/ first).

Verification

• kun/ typecheck passes (tsc --noEmit)
• Root project typecheck passes
• kun/ tool tests pass (24/24)
• No new test failures introduced (3 pre-existing failures in compaction tests are unrelated to this PR — verified they fail on clean develop too)

Future work (not in this PR)

• Support more languages (pyright, gopls, rust-analyzer) — the client is language-agnostic; just needs a config table + binary detection
• Editor integration (completion, go-to-definition in Monaco/CodeMirror)
• Automatic server installation prompt when missing

[XingYu-Zhong]

hope Future work can fast coming，lsp is very useful

[whitelonng]

Thanks for the merge! Glad the LSP tool is useful  I'm working on a few more features (git worktree pool, memory management UI) and would love to discuss the project's roadmap and technical direction. If you're open to it, we could connect on WeChat or Telegram — my wechat is “xiaobailong0216” Happy to keep contributing either way! Martini ***@***.***