[colbymchenry]

Summary

#2462 has two halves. This PR fixes the second and records where the first stands.

1. The default build skip. On main a root build/ folder is still skipped by default. #2681, #2685 and #2687 re-opened only build folders below src/, for hand-written TS/JS-family, Go, C# and Rust files. The documented opt-in works: !/build/ in the project's .gitignore indexes it, and codegraph sync picks the opt-in up. I verified that on main and on v1.6.2, with the negation in .gitignore and in .git/info/exclude. I could not reproduce the report's "Already up to date". The reporter's repo has since committed exactly that opt-in, with a comment naming CodeGraph. Widening the default to a root build/ is a separate decision about the skip list: a root build/ is many toolchains' output folder, e.g. a CRA production bundle. This PR leaves the skip list alone.

2. Explore's silence on a miss. When a query names symbols the index has never seen, the answer now says so first, success-shaped, and points at what decides it, never at Read:

**Exploration: ReleaseNotesBuilder Program Main ChangelogParser PersistentData Entry AddFeatures**

Found 4 symbols across 2 files.

> **Not in the index:** `ReleaseNotesBuilder`, `ChangelogParser`, `PersistentData`, `AddFeatures`. No indexed symbol, file, signature or comment uses these names, so the code below matches the query's other words, not them. If they are defined in this project, the files defining them are not indexed: `codegraph status` shows what is. If they live in a folder skipped by default, like `build/` or `dist/`, the user can include that folder by adding a re-include such as `!/build/` to `.gitignore`.

That is newrelic/newrelic-dotnet-agent at 418a0f2f, the reporter's repo, with its committed !/build/ removed and the index synced. On main, the same query returns src/Agent/NewRelic/Profiler/VersionFileGenerator/Program.cs's Main and the C++ Sicily/Parser.cpp, with nothing to say four of the names matched nothing. That is the report, line for line.

Root cause

Explore answers whatever the query's words match. A name with no match simply contributes nothing, and the answer is built from the remaining words. The #2026 miss diagnostics ran only when nothing at all matched.

Fix

• Which tokens are candidates. In explore's named-symbol seeding, a token is a candidate if it is written like a code name and resolves to no node at all, not even through the camel-infix field-name fallback. Code-name shapes are a camelCase hump, snake_case, or a qualified a.b / a::b.
• Which candidates get named. getUnseenNames (new, in queries.ts) keeps the names the index has never seen: no node called that in any case, and no name, qualified name (a file's path), signature or docstring word that starts with it. It's bounded like getExploreMissDiagnostics: two index probes per name, at most 16 names, never a source scan.
• Qualified names. A qualified name is judged by its member. Express's app.handle = function handle(…) is indexed as handle, so a receiver the index doesn't record never makes a name missing. This was a real false positive on express's prose probe until fixed.
• Single words are never named. Program, Main, How and Parser can be English, and a question's ordinary words must never be reported as missing code.
• Placement. The note sits under the summary line, ahead of the source. Re-including a folder skipped by default is worded as the user's choice, since indexing scope is the user's call (AGENTS.md). The total-miss answer ends with the same note, worded without "the code below".
• Shared tokenizer. The seeding's query tokenizer moved into exploreQueryTokens, unchanged, so both paths read a query the same way.

Validation

Tests. New __tests__/explore-unindexed-names.test.ts, 8 tests; 6 fail on main. It uses the report's layout (a root build/ with the tool projects, an unrelated Program.cs, a C++ Parser) and covers:

• the note's content and placement, success-shaped, with no Read wording;
• single words and names the index has are never named, and a prose question gets no note;
• a name seen only in a docstring or a folder name counts as seen;
• qualified names are judged by their member;
• the total-miss answer;
• the note goes away once !/build/ is added and the project synced.

All 38 explore suites pass, plus the tool-allowlist and dynamic-boundary suites (466 tests).

Ordinary queries. 996 on express, flask, excalidraw, django, airflow and go-control-plane: the 50 longest functions by name, their files by path, caller/callee pairs, call chains, random names and flow questions. All are byte-identical to main.

Miss probes. 120 queries, 20 per repo: real pair queries with an invented code-style name appended, invented names alone, and prose that mentions one.

• Main never notes anything.
• The fix names every invented name in all 120, and names no real name in any.
• On a full response the note costs about 5 source lines (623 of 44,213 across the 120); in 28 of them nothing else changes.

Full suite on bb72ca6 (rebased onto 0953084 after main moved), with the native kernel built and CODEGRAPH_KERNEL_EXPECT=1: 716 files (710 passed, 6 skipped), 8,572 tests: 8,490 passed, 0 failed, 82 skipped. The pushed commit differs from bb72ca6 only in the note's wording, so including a skipped folder reads as the user's choice, and the test lines that pin it (git diff --stat bb72ca63 HEAD: 2 files, 7 insertions, 5 deletions). The new test file, the 38 explore suites and the tool-allowlist and dynamic-boundary suites were re-run on it: 466 passed.

Not in this PR

• The root build/ default (part 1 above), left for a separate decision.
• The report's third suggestion, sync missing a newly un-ignored path, did not reproduce on main or on v1.6.2 for this layout.
• No agent A/B: the change is deterministic, and the probes measure exactly what the agent receives.

Fixes #2462

🤖 Generated with Claude Code