MCPcopy Create free account
hub / github.com/colbymchenry/codegraph / getTools

Method getTools

src/mcp/tools.ts:2361–2430  ·  view source on GitHub ↗

* Get tool definitions with dynamic descriptions based on project size. * The codegraph_explore tool description includes a budget recommendation * scaled to the number of indexed files. Honors the CODEGRAPH_MCP_TOOLS * allowlist so a trimmed surface is reflected in ListTools.

()

Source from the content-addressed store, hash-verified

2359 * allowlist so a trimmed surface is reflected in ListTools.
2360 */
2361 getTools(): ToolDefinition[] {
2362 const allow = this.toolAllowlist();
2363 // No explicit allowlist → the default 4-tool surface (see
2364 // DEFAULT_MCP_TOOLS for the evidence). An allowlist replaces the
2365 // default entirely, so any defined tool can be re-enabled.
2366 let visible = allow
2367 ? tools.filter(t => allow.has(t.name.replace(/^codegraph_/, '')))
2368 : tools.filter(t => DEFAULT_MCP_TOOLS.has(t.name.replace(/^codegraph_/, '')));
2369 // No default project loaded → no-root-index case (#993): a gateway server
2370 // started outside any repo, or a monorepo root whose indexes live in
2371 // sub-projects. With nothing to fall back to, EVERY call needs an explicit
2372 // projectPath, so mark it required in the schema — a high-salience nudge the
2373 // agent acts on, where SERVER_INSTRUCTIONS_NO_ROOT_INDEX's prose alone
2374 // wasn't enough (the reporter had to add an AGENTS.md note). `this.cg` is
2375 // settled by `retryInitIfNeeded()` before `handleToolsList` calls us, so a
2376 // null here means "genuinely no default", not a startup race. When a default
2377 // IS open we leave projectPath optional (below): a bare call falls back to
2378 // it, exactly as in the common single-project launch.
2379 if (!this.cg) return withRequiredProjectPath(visible);
2380
2381 try {
2382 const stats = this.cg.getStats();
2383 const budget = getExploreBudget(stats.fileCount);
2384
2385 // Tiny-repo tool gating: on projects under TINY_REPO_FILE_THRESHOLD
2386 // files, only expose the core trio (search, node, explore) — one
2387 // below even the 4-tool default: at this scale callers, too, reduces
2388 // to one grep. (Historical note: the audit below ran when context and
2389 // trace still existed; its "5 core tools" are today's trio.)
2390 //
2391 // n=2 audits ruled out cutting below 5 tools:
2392 // - 3-tool gate (search + context + trace): cost regressed on
2393 // cobra/ky/sinatra. The agent fell back to raw Reads to cover
2394 // what codegraph_node + codegraph_explore would have answered.
2395 // - 1-tool gate (search only): catastrophic regression — express
2396 // went from -43% WIN to +107% LOSS. With only search, the agent
2397 // can't navigate the call graph structurally and reads everything.
2398 //
2399 // 5 is the empirical lower bound. Tools beyond search/context/
2400 // node/explore/trace pay overhead that the agent doesn't recoup
2401 // on tiny-repo flow questions.
2402 // ITER4: raise threshold 150 → 500 so single-file frameworks
2403 // (sinatra at 159, slim_framework around 200) also get the
2404 // 5-tool surface. The empirical 5-tool floor was set on <150
2405 // probes; iter3 measurement showed sinatra is structurally the
2406 // SAME problem as cobra (single-file WITHOUT-arm Read wins),
2407 // so it deserves the same gating.
2408 const TINY_REPO_FILE_THRESHOLD = 500;
2409 const TINY_REPO_CORE_TOOLS = new Set([
2410 'codegraph_explore',
2411 'codegraph_search',
2412 'codegraph_node',
2413 ]);
2414 if (stats.fileCount < TINY_REPO_FILE_THRESHOLD) {
2415 visible = visible.filter(t => TINY_REPO_CORE_TOOLS.has(t.name));
2416 }
2417
2418 return visible.map(tool => {

Callers 4

handleToolsListMethod · 0.80
listedFunction · 0.80

Calls 6

toolAllowlistMethod · 0.95
withRequiredProjectPathFunction · 0.85
getExploreBudgetFunction · 0.85
hasMethod · 0.80
mapMethod · 0.80
getStatsMethod · 0.45

Tested by 1

listedFunction · 0.64