diff --git a/src/review/rag-index.ts b/src/review/rag-index.ts
index be702b39db..a1bfb98f82 100644
--- a/src/review/rag-index.ts
+++ b/src/review/rag-index.ts
@@ -15,8 +15,11 @@
 //      Vectorize/AI binding, a GitHub error, an oversized repo, or a partial batch degrades to "indexed less /
 //      nothing" rather than failing the job. `upsertChunks` itself already no-ops to 0 when infra is absent.
 //   2. FREE-TIER — `isIndexablePath` filters the tree to CODE (not the content/data corpus), source is
-//      prioritized, and a hard MAX_CHUNKS_PER_REPO cap bounds stored vectors per repo (the same cap retrieval
-//      assumes). We stop fetching once the cap is reached.
+//      prioritized ahead of docs, manifest/config files (package.json, tsconfig*.json, wrangler.*,
+//      pnpm-workspace.yaml, go.mod, Cargo.toml, pyproject.toml, ...) are prioritized ahead of THAT
+//      (manifestPriority — on a repo over the cap they'd otherwise tie every other source file and lose
+//      the alphabetical tiebreaker), and a hard MAX_CHUNKS_PER_REPO cap bounds stored vectors per repo
+//      (the same cap retrieval assumes). We stop fetching once the cap is reached.
 //
 // GATING — the caller (processors.ts) only DISPATCHES indexing when `isRagEnabled(env)` is true, and the cron
 // only ENQUEUES the fan-out under the same flag; flag-OFF (the default) this module is never invoked, makes no
@@ -24,6 +27,7 @@
 
 import { createInstallationToken } from "../github/app";
 import { githubRateLimitAdmissionKeyForInstallation, timeoutFetch, type GitHubRateLimitAdmissionKey } from "../github/client";
+import { isConfigFile, isDependencyManifestFile } from "../signals/path-matchers";
 import { repoParts } from "../utils/json";
 import { createReviewAdapters } from "./adapters";
 import {
@@ -42,6 +46,23 @@ import {
 /** A single indexable entry from the repo git tree (path + size, used by isIndexablePath's size guard). */
 type TreeEntry = { path: string; size?: number | undefined };
 
+/**
+ * Sort key that puts small, high-value manifest/config files (package.json, tsconfig*.json,
+ * wrangler.jsonc, pnpm-workspace.yaml, go.mod, Cargo.toml, pyproject.toml, requirements*.txt, ...)
+ * AHEAD of filePriority's code/doc split. On a repo whose file count exceeds MAX_CHUNKS_PER_REPO,
+ * `indexRepo`'s per-file loop stops once the cap is hit — with only `filePriority` (code=0, doc=1)
+ * as the sort key, a manifest file ties every other source file at priority 0 and then loses on the
+ * alphabetical tiebreaker, so it can be starved out entirely by volume (verified in prod: gittensory's
+ * own package.json never got indexed). These files are already indexable code (JSON/TOML/YAML all
+ * match CODE_EXT_RE in `./rag`) — this only reorders them, it does not change what's included.
+ * Reuses the same "manifest-like filename" classifiers signals/path-matchers.ts already exports for
+ * slop classification (isDependencyManifestFile / isConfigFile) rather than inventing a second
+ * filename vocabulary.
+ */
+function manifestPriority(path: string): number {
+  return isDependencyManifestFile(path) || isConfigFile(path) ? -1 : filePriority(path);
+}
+
 /** Cap on how many chunks we upsert per Vectorize/D1 write batch (bounds the bound-param + neuron cost per call;
  *  embedTexts itself batches the AI calls at EMBED_BATCH internally). */
 const UPSERT_BATCH = 50;
@@ -230,8 +251,9 @@ export type IndexRepoResult = { indexed: number; files: number; capped: boolean
 
 /**
  * FULL (re)index of a repo's CODE into the RAG index. Fetches the git tree at the default branch, filters to
- * indexable code/docs (isIndexablePath), prioritizes source over docs, fetches each file's content, chunks it
- * (chunkFile), and upserts (embed + Vectorize + repo_chunks via upsertChunks) up to MAX_CHUNKS_PER_REPO.
+ * indexable code/docs (isIndexablePath), prioritizes manifest/config files first, then source over docs
+ * (manifestPriority), fetches each file's content, chunks it (chunkFile), and upserts (embed + Vectorize +
+ * repo_chunks via upsertChunks) up to MAX_CHUNKS_PER_REPO.
  *
  * Idempotent: chunk ids are stable (namespace|path::idx) so re-running upserts (ON CONFLICT updates) the same
  * rows rather than duplicating. Fully FAIL-SAFE — any error (no infra, GitHub down, bad file) degrades to
@@ -263,7 +285,7 @@ export async function indexRepo(
     if (rawTree === null) return empty;
     const tree = rawTree
       .filter((entry) => isIndexablePath(entry.path, entry.size))
-      .sort((a, b) => filePriority(a.path) - filePriority(b.path) || a.path.localeCompare(b.path));
+      .sort((a, b) => manifestPriority(a.path) - manifestPriority(b.path) || a.path.localeCompare(b.path));
     await pruneMissingPaths(infra, project, repoName, new Set(tree.map((entry) => entry.path)));
     if (tree.length === 0) return empty;
 
diff --git a/src/review/rag.ts b/src/review/rag.ts
index 1db1e9b8c7..e65653a6d6 100644
--- a/src/review/rag.ts
+++ b/src/review/rag.ts
@@ -133,7 +133,10 @@ const CODE_EXT_RE =
 // spellings were missing here, so e.g. NOTES.markdown / guide.asciidoc were
 // misclassified as skip instead of doc.
 const DOC_EXT_RE = /\.(md|mdx|markdown|rst|adoc|asciidoc|txt)$/i;
-const ALLOW_EXTLESS_RE = /(^|\/)(Dockerfile|Makefile|Justfile|Procfile)$/i;
+// `go.mod`/`go.work` (Go's extensionless dependency manifests) belong here for the same reason as
+// Dockerfile/Makefile: no recognized extension, but a real, high-value source file that must not
+// fall through to "skip" (go.sum/go.work.sum are resolved-tree lockfiles, already excluded above).
+const ALLOW_EXTLESS_RE = /(^|\/)(Dockerfile|Makefile|Justfile|Procfile|go\.mod|go\.work)$/i;
 
 /** code | doc | skip. Skips dependency/build/content/data/binary paths — RAG indexes code for code
  *  review, not the (potentially huge) submission/content corpus. */
diff --git a/test/unit/rag-index.test.ts b/test/unit/rag-index.test.ts
index 1cf0ddd03f..763ebe185d 100644
--- a/test/unit/rag-index.test.ts
+++ b/test/unit/rag-index.test.ts
@@ -302,6 +302,85 @@ describe("indexRepo: MAX_CHUNKS_PER_REPO cap holds", () => {
     expect(await countChunks(env, PROJECT, "gittensory")).toBeLessThanOrEqual(MAX_CHUNKS_PER_REPO);
     expect(result.indexed).toBe(MAX_CHUNKS_PER_REPO);
   });
+
+  it("still indexes package.json (and other root manifest/config files) on a repo whose file count exceeds the cap (regression: manifestPriority)", async () => {
+    const { env } = indexEnv();
+    // Alphabetically, "package.json" sorts AFTER most of "src/f0.ts".."src/f<N>.ts" — so without
+    // manifest-first prioritization it would be starved out once the cap is reached, exactly as it
+    // was in prod for gittensory's own repo (#confirmed via repo_chunks query).
+    const overCap = MAX_CHUNKS_PER_REPO + 25;
+    const tree: Array<{ path: string; size: number }> = [
+      { path: "package.json", size: 20 },
+      { path: "tsconfig.json", size: 20 },
+      { path: "pnpm-workspace.yaml", size: 20 },
+      { path: "go.mod", size: 20 },
+      { path: "Cargo.toml", size: 20 },
+      { path: "pyproject.toml", size: 20 },
+      { path: "requirements.txt", size: 20 },
+      { path: "wrangler.jsonc", size: 20 },
+      ...Array.from({ length: overCap }, (_, i) => ({ path: `src/f${i}.ts`, size: 20 })),
+    ];
+    const files: Record<string, string> = {
+      "package.json": '{"name":"gittensory"}\n',
+      "tsconfig.json": "{}\n",
+      "pnpm-workspace.yaml": "packages:\n",
+      "go.mod": "module example.com/foo\n",
+      "Cargo.toml": "[package]\n",
+      "pyproject.toml": "[project]\n",
+      "requirements.txt": "flask\n",
+      "wrangler.jsonc": "{}\n",
+    };
+    for (let i = 0; i < overCap; i++) files[`src/f${i}.ts`] = `export const f${i} = ${i};\n`;
+    stubGithub({ tree, files });
+
+    const result = await indexRepo(env, PROJECT, REPO);
+
+    expect(result.capped).toBe(true);
+    expect(result.indexed).toBe(MAX_CHUNKS_PER_REPO);
+    const indexedPaths = await pathsFor(env, PROJECT, "gittensory");
+    for (const manifest of [
+      "package.json",
+      "tsconfig.json",
+      "pnpm-workspace.yaml",
+      "go.mod",
+      "Cargo.toml",
+      "pyproject.toml",
+      "requirements.txt",
+      "wrangler.jsonc",
+    ]) {
+      expect(indexedPaths).toContain(manifest);
+    }
+  });
+});
+
+describe("manifestPriority ordering (via indexRepo, well under the cap)", () => {
+  afterEach(() => vi.unstubAllGlobals());
+
+  it("orders manifest/config files first, then other code, then docs — byte-identical file SET to before when the cap never matters", async () => {
+    const { env } = indexEnv();
+    stubGithub({
+      tree: [
+        { path: "README.md", size: 10 },
+        { path: "src/z.ts", size: 10 },
+        { path: "package.json", size: 10 },
+        { path: "src/a.ts", size: 10 },
+      ],
+      files: {
+        "README.md": "# Title\n",
+        "src/z.ts": "export const z = 1;\n",
+        "package.json": '{"name":"x"}\n',
+        "src/a.ts": "export const a = 1;\n",
+      },
+    });
+
+    const result = await indexRepo(env, PROJECT, REPO);
+
+    // Nowhere near the cap: every indexable file is still indexed (same SET as filePriority alone
+    // would have produced) — only the ORDER of indexing/upsert changed, not the outcome.
+    expect(result.capped).toBe(false);
+    expect(result.files).toBe(4);
+    expect(await pathsFor(env, PROJECT, "gittensory")).toEqual(["README.md", "package.json", "src/a.ts", "src/z.ts"]);
+  });
 });
 
 describe("reindexChangedPaths: delete + re-upsert only the changed paths", () => {
diff --git a/test/unit/rag.test.ts b/test/unit/rag.test.ts
index 6bbe3434d9..8dee900309 100644
--- a/test/unit/rag.test.ts
+++ b/test/unit/rag.test.ts
@@ -66,6 +66,12 @@ describe("rag: code-not-content filtering (free-tier cost guard)", () => {
     ]) {
       expect(classifyRepoFile(p)).toBe("code");
     }
+    // Go's extensionless dependency manifests (go.mod/go.work) are real, high-value source — same
+    // extensionless-allowlist treatment as Dockerfile/Makefile. Their resolved-tree lockfile
+    // siblings (go.sum/go.work.sum) stay excluded via SKIP_FILE_RE below.
+    expect(classifyRepoFile("go.mod")).toBe("code");
+    expect(classifyRepoFile("go.work")).toBe("code");
+    expect(classifyRepoFile("nested/module/go.mod")).toBe("code");
     expect(classifyRepoFile("README.md")).toBe("doc");
     expect(classifyRepoFile("docs/architecture.mdx")).toBe("doc");
     // long-form doc spellings (parity with signals/path-matchers DOCS_EXTENSIONS)
@@ -78,6 +84,9 @@ describe("rag: code-not-content filtering (free-tier cost guard)", () => {
     expect(classifyRepoFile("dist/bundle.js")).toBe("skip");
     expect(classifyRepoFile("package-lock.json")).toBe("skip");
     expect(classifyRepoFile("pnpm-lock.yaml")).toBe("skip");
+    // go.sum stays skipped despite go.mod/go.work now being recognized — SKIP_FILE_RE's lockfile
+    // check runs before ALLOW_EXTLESS_RE, so the resolved-tree lockfile never becomes indexable.
+    expect(classifyRepoFile("go.sum")).toBe("skip");
     expect(classifyRepoFile("public/logo.png")).toBe("skip");
     expect(classifyRepoFile("app.min.js")).toBe("skip");
     // more binary blobs: media/archives/fonts/compiled artifacts and ML model weights