fix(explore): report extension-less paths the index doesn't hold (#1830)

CHANGELOG.md

@@ -153,6 +153,8 @@ and adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
 - The prompt hook no longer injects unrelated projects when run from your home directory or a broader directory containing a stray workspace manifest. (#1454)
 
+- `codegraph_explore` now says so when a query names an extension-less file the index doesn't hold. A path like `scripts/deploy` has no extension on its last segment, so it failed the shape test that decides a span is a path beyond doubt — the name was left in the query, shredded into `scripts` and `deploy`, and the answer came back as a pile of unrelated source with no hint that the file you named was never consulted. Such a span is now checked against the project directory: if it is a real file, the answer carries the same `No indexed file uniquely matches ...` note a misspelled `src/foo.ts` already got. Slashed prose — `and/or`, `input/output`, `gen_server:call/2` — has no file behind it and is still left in the query untouched. (#1830)
+
 - Indexing now succeeds when Node.js's SQLite lacks FTS5, with search falling back to name and fuzzy matching; thanks @aniruddhaadak80. (#1532)
 
 - `codegraph_explore` now makes clear that suggested call counts are advisory, so agents keep exploring when an answer is incomplete; thanks @rongbc. (#1504, #1570)

__tests__/query-paths.test.ts

@@ -236,3 +236,84 @@ describe('extractQueryPaths — extension-less kebab basenames', () => {
     expect(out.strippedQuery).toBe('background-image-table then');
   });
 });
+
+/**
+ * Dotless slashed spans (#1830). `scripts/deploy` names a real file that the
+ * index does not hold (no recognized extension), but the shape test demands a
+ * dot-extension on the last segment — so the span was neither pinned NOR
+ * reported, and its fragments (`scripts`, `deploy`) went on to feed FTS. The
+ * agent got a pile of unrelated source with no hint that the file it named was
+ * never consulted. Shape alone cannot decide this (`and/or` is the same shape),
+ * so the caller injects an `existsOnDisk` predicate and the file's existence
+ * decides. These tests also pin that the predicate is genuinely CONSULTED —
+ * a predicate that is never called would make the whole arm vacuous.
+ */
+describe('extractQueryPaths — dotless slashed spans, decided on disk', () => {
+  /** Records every span the predicate is asked about. */
+  const probe = (onDisk: readonly string[]) => {
+    const asked: string[] = [];
+    return {
+      asked,
+      existsOnDisk: (rel: string) => { asked.push(rel); return onDisk.includes(rel); },
+    };
+  };
+
+  it('reports a dotless path that exists on disk but is not indexed', () => {
+    const p = probe(['scripts/deploy']);
+    const out = extractQueryPaths(
+      'why does scripts/deploy fail on release', INDEX, { existsOnDisk: p.existsOnDisk },
+    );
+    expect(out.pinnedFiles).toEqual([]);
+    expect(out.unresolvedPathSpans).toEqual(['scripts/deploy']);
+    expect(out.strippedQuery).toBe('why does fail on release');
+    // Vacuity guard: the verdict came from the predicate, not from some other arm.
+    expect(p.asked).toContain('scripts/deploy');
+  });
+
+  it('leaves the same span alone when no predicate is injected', () => {
+    const q = 'why does scripts/deploy fail on release';
+    const out = extractQueryPaths(q, INDEX);
+    expect(out.unresolvedPathSpans).toEqual([]);
+    expect(out.strippedQuery).toBe(q);
+  });
+
+  it('leaves `and/or` prose alone even though a predicate is injected', () => {
+    const p = probe(['scripts/deploy']);
+    const q = 'does gen_server:call/2 block and/or timeout';
+    const out = extractQueryPaths(q, INDEX, { existsOnDisk: p.existsOnDisk });
+    expect(out.pinnedFiles).toEqual([]);
+    expect(out.unresolvedPathSpans).toEqual([]);
+    expect(out.strippedQuery).toBe(q);
+    // Consulted and refused — not skipped by shape.
+    expect(p.asked).toContain('and/or');
+  });
+
+  it('leaves a slashed word pair that is not a file on disk alone', () => {
+    const p = probe(['scripts/deploy']);
+    const q = 'trace the input/output buffering path';
+    const out = extractQueryPaths(q, INDEX, { existsOnDisk: p.existsOnDisk });
+    expect(out.unresolvedPathSpans).toEqual([]);
+    expect(out.strippedQuery).toBe(q);
+    expect(p.asked).toContain('input/output');
+  });
+
+  it('still reports a dotted span that matches nothing and is not on disk', () => {
+    const p = probe([]);
+    const out = extractQueryPaths(
+      'crash in src/routes/gone/missing-page.svelte on load', INDEX,
+      { existsOnDisk: p.existsOnDisk },
+    );
+    expect(out.unresolvedPathSpans).toEqual(['src/routes/gone/missing-page.svelte']);
+    expect(out.strippedQuery).toBe('crash in on load');
+  });
+
+  it('pins an indexed dotless path by resolution, never asking about it', () => {
+    const p = probe(['scripts/pre-commit']);
+    const out = extractQueryPaths(
+      'what does scripts/pre-commit run', INDEX, { existsOnDisk: p.existsOnDisk },
+    );
+    expect(out.pinnedFiles).toEqual(['scripts/pre-commit']);
+    expect(out.unresolvedPathSpans).toEqual([]);
+    expect(p.asked).not.toContain('scripts/pre-commit');
+  });
+});

src/mcp/tools.ts

@@ -157,6 +157,25 @@ export function normalizeQuerySpelling(query: string): string {
     );
 }
 
+/**
+ * Does this query-named span point at a real FILE inside the project?
+ *
+ * The `existsOnDisk` predicate `extractQueryPaths` takes (that module is pure —
+ * no DB, no fs — so the fs access lives here, where the project root is known).
+ * Only a REGULAR FILE counts: a directory span (`src/search`) is not a file
+ * reference and must keep flowing to the normal matching pipeline. Containment
+ * is enforced by `validatePathWithinRoot`, so a `../` span in a query cannot
+ * probe outside the project, and every fs error answers `false`.
+ */
+function pathIsProjectFile(projectRoot: string, relPath: string): boolean {
+  try {
+    const abs = validatePathWithinRoot(projectRoot, relPath);
+    return abs !== null && statSync(abs).isFile();
+  } catch {
+    return false;
+  }
+}
+
 /**
  * Calculate the recommended number of codegraph_explore calls based on project size.
  * Larger codebases need more exploration calls to cover their surface area,
@@ -3318,7 +3337,7 @@ export class ToolHandler {
         const extraction = extractQueryPaths(
           rawQuery,
           cg.getFiles().map((f) => f.path),
-          { maxPins: maxFiles },
+          { maxPins: maxFiles, existsOnDisk: (rel) => pathIsProjectFile(projectRoot, rel) },
         );
         if (extraction.pinnedFiles.length > 0 || extraction.unresolvedPathSpans.length > 0) {
           pinnedFiles = extraction.pinnedFiles;

src/search/query-paths.ts

@@ -24,7 +24,11 @@
  * spans removed.
  * Callers treat pinned files as first-class: guaranteed admission, top rank,
  * funded first. Pure string work — no DB, no fs — so it is trivially testable
- * and safe inside the query-pool workers.
+ * and safe inside the query-pool workers. The one question string shape cannot
+ * answer (is this dotless slashed span, `scripts/deploy`, a real file that is
+ * merely unindexed, or is it prose like `and/or`?) is delegated to an OPTIONAL
+ * `existsOnDisk` predicate the caller injects — the fs access stays at the call
+ * site, which owns the project root, and this module stays pure.
  */
 
 export interface QueryPathExtraction {
@@ -189,7 +193,18 @@ function resolveSpan(
 export function extractQueryPaths(
   query: string,
   indexedPaths: readonly string[],
-  opts: { maxPins?: number; maxMatchesPerSpan?: number } = {},
+  opts: {
+    maxPins?: number;
+    maxMatchesPerSpan?: number;
+    /**
+     * Does this repo-relative span name a real FILE in the project? Optional,
+     * injected by the caller (see the module docstring): it is the only way to
+     * tell a dotless path the index simply doesn't hold (`scripts/deploy`)
+     * from slashed prose (`and/or`), and it must stay out of this module.
+     * Must not throw — the caller absorbs fs errors and returns false.
+     */
+    existsOnDisk?: (relPath: string) => boolean;
+  } = {},
 ): QueryPathExtraction {
   const maxPins = Math.max(1, opts.maxPins ?? 8);
   const maxMatchesPerSpan = Math.max(1, opts.maxMatchesPerSpan ?? 3);
@@ -235,10 +250,19 @@ export function extractQueryPaths(
         pinnedSeen.add(m);
         pinned.push(m);
       }
-    } else if (ambiguous || isClearlyPathShaped(normalized)) {
+    } else if (
+      ambiguous
+      || isClearlyPathShaped(normalized)
+      || (normalized.includes('/') && opts.existsOnDisk?.(normalized) === true)
+    ) {
       // A real path that didn't resolve to a usable set. Keeping it in the
       // query is strictly worse — its fragments are what minted the junk
       // matches this module exists to stop — so strip it and say so.
+      // The third arm covers the DOTLESS slashed span (`scripts/deploy`,
+      // `bin/build`): shape alone cannot tell it from `and/or` or
+      // `input/output`, so the file's existence on disk decides. Loosening the
+      // SHAPE test instead would strip that prose out of every query and mint a
+      // false "no indexed file matches `and/or`" caveat.
       consumed.add(i);
       if (unresolved.length < 4) unresolved.push(normalized);
     }