diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index d8112c1..75a16b8 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -86,3 +86,110 @@ jobs:
 
     - name: Run Tests
       run: pnpm --filter web test
+
+  indexer-go:
+    runs-on: ubuntu-latest
+    defaults:
+      run:
+        working-directory: ./services/indexer
+
+    steps:
+    - uses: actions/checkout@v4
+
+    - name: Set up Go
+      uses: actions/setup-go@v5
+      with:
+        go-version: '1.23'
+        cache-dependency-path: services/indexer/go.sum
+
+    # Every tree-sitter grammar is a cgo package with its own generated parser.c, so this
+    # job cannot use CGO_ENABLED=0 and needs a C toolchain. ubuntu-latest ships gcc, but
+    # pinning the expectation here makes the failure obvious if that ever changes.
+    - name: Verify C toolchain is present (grammars are cgo)
+      run: cc --version
+
+    - name: Formatting
+      run: |
+        # gen/ is protoc output and is checked by the codegen-drift step instead.
+        unformatted=$(gofmt -l . | grep -v '^gen/' || true)
+        if [ -n "$unformatted" ]; then
+          echo "::error::gofmt needed for: $unformatted"
+          gofmt -d $unformatted
+          exit 1
+        fi
+
+    - name: Vet
+      run: CGO_ENABLED=1 go vet ./...
+
+    - name: Test
+      run: CGO_ENABLED=1 go test -race ./...
+
+    - name: Build
+      run: CGO_ENABLED=1 go build -o /tmp/indexer .
+
+  # The contract is shared between two languages that generate stubs independently. Without
+  # this, editing indexer.proto and regenerating only one side produces a mismatch that
+  # nothing catches until runtime -- the Go server and Python client would simply disagree.
+  proto-codegen-drift:
+    runs-on: ubuntu-latest
+
+    steps:
+    - uses: actions/checkout@v4
+
+    - name: Set up Go
+      uses: actions/setup-go@v5
+      with:
+        go-version: '1.23'
+        cache-dependency-path: services/indexer/go.sum
+
+    - name: Set up Python
+      uses: actions/setup-python@v5
+      with:
+        python-version: "3.11"
+
+    - name: Install protoc and plugins
+      run: |
+        PROTOC_VERSION=29.3
+        curl -fsSL -o /tmp/protoc.zip \
+          "https://github.com/protocolbuffers/protobuf/releases/download/v${PROTOC_VERSION}/protoc-${PROTOC_VERSION}-linux-x86_64.zip"
+        unzip -q /tmp/protoc.zip -d /tmp/protoc
+        echo "/tmp/protoc/bin" >> "$GITHUB_PATH"
+        # Pinned, not @latest. Generated files embed the generator version
+        # ("protoc-gen-go v1.36.12", "Protobuf Python Version: 7.35.1"), so an unpinned
+        # plugin would eventually change the output and fail this check for a reason that
+        # has nothing to do with the proto. These are the exact versions that produced the
+        # committed stubs; bump them and the stubs together.
+        go install google.golang.org/protobuf/cmd/protoc-gen-go@v1.36.12
+        go install google.golang.org/grpc/cmd/protoc-gen-go-grpc@v1.6.2
+        echo "$(go env GOPATH)/bin" >> "$GITHUB_PATH"
+        python -m pip install --quiet "grpcio-tools==1.83.0"
+
+    - name: Regenerate Go stubs
+      working-directory: ./services/indexer
+      run: |
+        protoc --proto_path=proto --go_out=gen --go_opt=paths=source_relative \
+               --go-grpc_out=gen --go-grpc_opt=paths=source_relative proto/indexer.proto
+
+    - name: Regenerate Python stubs
+      working-directory: ./apps/api
+      run: |
+        python -m grpc_tools.protoc \
+          --proto_path=../../services/indexer/proto \
+          --python_out=src/core/indexer --grpc_python_out=src/core/indexer \
+          ../../services/indexer/proto/indexer.proto
+        # The generated grpc stub uses a flat import; the committed copy is package-relative.
+        sed -i 's/^import indexer_pb2 as indexer__pb2$/from src.core.indexer import indexer_pb2 as indexer__pb2/' \
+          src/core/indexer/indexer_pb2_grpc.py
+
+    - name: Fail if committed stubs differ from the proto
+      run: |
+        # A strict whole-file diff is safe because protoc, both Go plugins and grpcio-tools
+        # are all pinned above; the round-trip was verified byte-identical.
+        if ! git diff --quiet -- services/indexer/gen apps/api/src/core/indexer; then
+          echo "::error::Generated stubs are out of date with proto/indexer.proto."
+          echo "Regenerate both sides and commit; see services/indexer/README.md."
+          git diff --stat -- services/indexer/gen apps/api/src/core/indexer
+          git diff -- services/indexer/gen apps/api/src/core/indexer | head -60
+          exit 1
+        fi
+        echo "Generated stubs match the proto on both sides."
diff --git a/services/indexer/README.md b/services/indexer/README.md
index 7b21a25..0639128 100644
--- a/services/indexer/README.md
+++ b/services/indexer/README.md
@@ -61,6 +61,32 @@ CGO_ENABLED=1 go build -o bin/indexer .
 `CGO_ENABLED=1` is mandatory: every grammar is cgo. That means a C toolchain in any builder
 image, no `scratch` base, and a slower uncached build than the all-wheels Python image.
 
+## Regenerating the stubs
+
+Both sides are generated from `proto/indexer.proto` and both are committed. CI regenerates
+them and fails if they differ, so an edit to the proto that updates only one language
+cannot merge.
+
+```bash
+# Go
+cd services/indexer
+protoc --proto_path=proto --go_out=gen --go_opt=paths=source_relative \
+       --go-grpc_out=gen --go-grpc_opt=paths=source_relative proto/indexer.proto
+
+# Python
+cd apps/api
+python -m grpc_tools.protoc --proto_path=../../services/indexer/proto \
+  --python_out=src/core/indexer --grpc_python_out=src/core/indexer \
+  ../../services/indexer/proto/indexer.proto
+sed -i 's/^import indexer_pb2 as indexer__pb2$/from src.core.indexer import indexer_pb2 as indexer__pb2/' \
+  src/core/indexer/indexer_pb2_grpc.py
+```
+
+Generator versions are **pinned in CI** (protoc 29.3, protoc-gen-go v1.36.12,
+protoc-gen-go-grpc v1.6.2, grpcio-tools 1.83.0) because generated files embed the
+generator version — an unpinned plugin would change the output and fail the drift check
+for a reason unrelated to the contract. Bump the pins and the committed stubs together.
+
 ## Grammar scope, stated rather than implied
 
 This build links **5** grammars (python, javascript, typescript, tsx, go) against the
@@ -109,5 +135,6 @@ The service is **not wired into the indexing pipeline** — `IndexingService` st
 Python parser, and `IndexerClient` is opt-in. Doing that swap needs a decision this PR does
 not make: the service must be co-located with the API on the shared volume, because clones
 live under `./data/repos/<owner>/<name>` and a Fly/EBS-style volume attaches to exactly one
-machine. There is also no CI job for Go yet (`ci.yml` has Python 3.11 and Node 20 only, no
-cgo toolchain, no grammar-compile caching, no cross-language contract test).
+machine. CI now covers this service: `indexer-go` runs gofmt, `go vet`, `go test -race` and a cgo
+build, and `proto-codegen-drift` regenerates both languages' stubs and fails if they differ
+from what is committed.
diff --git a/services/indexer/internal/parse/parse.go b/services/indexer/internal/parse/parse.go
index 60ae5a4..887cf99 100644
--- a/services/indexer/internal/parse/parse.go
+++ b/services/indexer/internal/parse/parse.go
@@ -97,9 +97,9 @@ var configs = map[string]langConfig{
 var extToLang = map[string]string{
 	".py": "python",
 	".js": "javascript", ".jsx": "javascript",
-	".ts": "typescript",
+	".ts":  "typescript",
 	".tsx": "tsx", // never "typescript"
-	".go": "go",
+	".go":  "go",
 }
 
 // LanguageFor reports the grammar for an extension, and whether one exists.
diff --git a/services/indexer/internal/parse/parse_test.go b/services/indexer/internal/parse/parse_test.go
new file mode 100644
index 0000000..3918ff6
--- /dev/null
+++ b/services/indexer/internal/parse/parse_test.go
@@ -0,0 +1,174 @@
+package parse
+
+import (
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+)
+
+func write(t *testing.T, dir, name, content string) string {
+	t.Helper()
+	p := filepath.Join(dir, name)
+	if err := os.MkdirAll(filepath.Dir(p), 0o755); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(p, []byte(content), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	return p
+}
+
+// The plain TypeScript grammar cannot parse JSX. Mapping .tsx to it produces a tree full
+// of ERROR nodes for every .tsx file -- the exact bug that shipped in the Python parser.
+func TestTSXUsesTheTSXGrammar(t *testing.T) {
+	dir := t.TempDir()
+	src := "export function Widget({ label }: { label: string }) {\n" +
+		"  return <div className=\"w\">{label}</div>;\n}\n"
+	abs := write(t, dir, "Widget.tsx", src)
+
+	chunks, err := File(abs, "Widget.tsx")
+	if err != nil {
+		t.Fatalf("parse: %v", err)
+	}
+	if len(chunks) == 0 {
+		t.Fatal("expected at least one chunk")
+	}
+	for _, c := range chunks {
+		if c.Language != "tsx" {
+			t.Errorf("language = %q, want tsx (the plain TS grammar cannot parse JSX)", c.Language)
+		}
+		if c.HadParseError {
+			t.Errorf("chunk %q has a parse error; JSX should parse cleanly under tsx", c.Name)
+		}
+	}
+}
+
+// tree-sitter offsets are BYTE offsets. Slicing a decoded string with them corrupts every
+// chunk after the first multi-byte character.
+func TestChunksAfterNonASCIIAreNotMisaligned(t *testing.T) {
+	dir := t.TempDir()
+	src := "# héllo — wörld 日本語 🎉\n" +
+		"def after_unicode(x):\n    return x * 2\n" +
+		"class Café:\n    def método(self):\n        return \"ok\"\n"
+	abs := write(t, dir, "unicode.py", src)
+
+	chunks, err := File(abs, "unicode.py")
+	if err != nil {
+		t.Fatalf("parse: %v", err)
+	}
+
+	byName := map[string]Chunk{}
+	for _, c := range chunks {
+		if c.Name != "" {
+			byName[c.Name] = c
+		}
+	}
+
+	for _, want := range []struct{ name, prefix string }{
+		{"after_unicode", "def after_unicode"},
+		{"Café", "class Café:"},
+		{"método", "def método(self):"},
+	} {
+		got, ok := byName[want.name]
+		if !ok {
+			t.Errorf("missing chunk %q (got %v)", want.name, keys(byName))
+			continue
+		}
+		if !strings.HasPrefix(got.Content, want.prefix) {
+			t.Errorf("chunk %q content = %q, want prefix %q -- byte/char offset mismatch",
+				want.name, truncateForMsg(got.Content), want.prefix)
+		}
+	}
+}
+
+// tree-sitter returns a partial tree rather than failing on a syntax error, so the flag is
+// the only way a caller can decide to fall back to raw indexing.
+func TestSyntaxErrorIsReportedNotSwallowed(t *testing.T) {
+	dir := t.TempDir()
+	abs := write(t, dir, "broken.py", "def ok():\n    return 1\n\nclass ((( :\n")
+
+	chunks, err := File(abs, "broken.py")
+	if err != nil {
+		t.Fatalf("parse should not fail on a syntax error: %v", err)
+	}
+	if len(chunks) == 0 {
+		t.Fatal("expected chunks even from a file with errors")
+	}
+	found := false
+	for _, c := range chunks {
+		if c.HadParseError {
+			found = true
+		}
+	}
+	if !found {
+		t.Error("HadParseError was never set on a file with a syntax error")
+	}
+}
+
+func TestExtensionMapping(t *testing.T) {
+	cases := map[string]string{
+		".py": "python", ".ts": "typescript", ".tsx": "tsx",
+		".js": "javascript", ".jsx": "javascript", ".go": "go",
+	}
+	for ext, want := range cases {
+		got, ok := LanguageFor(ext)
+		if !ok || got != want {
+			t.Errorf("LanguageFor(%q) = %q,%v; want %q", ext, got, ok, want)
+		}
+	}
+	if _, ok := LanguageFor(".rs"); ok {
+		t.Error("LanguageFor(.rs) should report unlinked: this build has 5 grammars, not 9")
+	}
+}
+
+func TestUnknownExtensionYieldsNoChunks(t *testing.T) {
+	dir := t.TempDir()
+	abs := write(t, dir, "notes.txt", "plain text")
+	chunks, err := File(abs, "notes.txt")
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if len(chunks) != 0 {
+		t.Errorf("expected no chunks for an unlinked extension, got %d", len(chunks))
+	}
+}
+
+// A file with no extractable declarations still carries meaning; dropping it would lose
+// content the Python side keeps.
+func TestFileWithNoDeclarationsFallsBackToModuleChunk(t *testing.T) {
+	dir := t.TempDir()
+	abs := write(t, dir, "consts.py", "A = 1\nB = 2\n")
+	chunks, err := File(abs, "consts.py")
+	if err != nil {
+		t.Fatalf("parse: %v", err)
+	}
+	if len(chunks) != 1 || chunks[0].ChunkType != "module" {
+		t.Fatalf("want one module chunk, got %+v", chunks)
+	}
+}
+
+func TestLinkedReportsEveryMappedExtension(t *testing.T) {
+	linked := Linked()
+	if len(linked) == 0 {
+		t.Fatal("Linked() must report grammars so callers never assume parity with Python")
+	}
+	if exts, ok := linked["tsx"]; !ok || len(exts) == 0 {
+		t.Error("tsx must be reported separately from typescript")
+	}
+}
+
+func keys(m map[string]Chunk) []string {
+	out := make([]string, 0, len(m))
+	for k := range m {
+		out = append(out, k)
+	}
+	return out
+}
+
+func truncateForMsg(s string) string {
+	if len(s) > 60 {
+		return s[:60] + "..."
+	}
+	return s
+}
diff --git a/services/indexer/internal/walk/walk.go b/services/indexer/internal/walk/walk.go
index 6a179b0..372a664 100644
--- a/services/indexer/internal/walk/walk.go
+++ b/services/indexer/internal/walk/walk.go
@@ -30,7 +30,7 @@ var indexedExts = map[string]struct{}{
 	".cs": {}, ".csx": {},
 	".rb": {}, ".rake": {}, ".gemspec": {}, ".php": {}, ".swift": {}, ".kt": {},
 	".erb": {},
-	".md": {}, ".json": {},
+	".md":  {}, ".json": {},
 }
 
 // Mirrors INDEXED_FILENAMES: extensionless files worth indexing.
diff --git a/services/indexer/internal/walk/walk_test.go b/services/indexer/internal/walk/walk_test.go
new file mode 100644
index 0000000..d0ec1d5
--- /dev/null
+++ b/services/indexer/internal/walk/walk_test.go
@@ -0,0 +1,134 @@
+package walk
+
+import (
+	"fmt"
+	"os"
+	"path/filepath"
+	"testing"
+)
+
+func seed(t *testing.T, dir string, files map[string]string) {
+	t.Helper()
+	for name, body := range files {
+		p := filepath.Join(dir, name)
+		if err := os.MkdirAll(filepath.Dir(p), 0o755); err != nil {
+			t.Fatal(err)
+		}
+		if err := os.WriteFile(p, []byte(body), 0o644); err != nil {
+			t.Fatal(err)
+		}
+	}
+}
+
+func TestSkipsExcludedDirectories(t *testing.T) {
+	dir := t.TempDir()
+	seed(t, dir, map[string]string{
+		"src/app.ts":                "x",
+		"node_modules/pkg/index.js": "x",
+		".git/config":               "x",
+		"dist/bundle.js":            "x",
+		".venv/lib/mod.py":          "x",
+	})
+
+	got, err := Find(dir, Options{})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(got.Paths) != 1 || got.Paths[0] != "src/app.ts" {
+		t.Errorf("paths = %v, want only src/app.ts", got.Paths)
+	}
+}
+
+// The Python original's cap did not work: its `break` left only the inner filename loop,
+// so os.walk continued into the next directory and kept appending. This is the regression
+// test for that class of bug.
+func TestMaxFilesActuallyStopsTheWalk(t *testing.T) {
+	dir := t.TempDir()
+	files := map[string]string{}
+	for d := 0; d < 5; d++ {
+		for f := 0; f < 10; f++ {
+			files[fmt.Sprintf("d%d/f%d.py", d, f)] = "x"
+		}
+	}
+	seed(t, dir, files)
+
+	got, err := Find(dir, Options{MaxFiles: 12})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(got.Paths) > 12 {
+		t.Errorf("found %d files with MaxFiles=12: the cap leaked", len(got.Paths))
+	}
+	if !got.Truncated {
+		t.Error("Truncated must be set, or a partial index looks like full coverage")
+	}
+}
+
+func TestTruncatedIsFalseWhenEverythingFits(t *testing.T) {
+	dir := t.TempDir()
+	seed(t, dir, map[string]string{"a.py": "x", "b.py": "x"})
+	got, err := Find(dir, Options{MaxFiles: 100})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if got.Truncated {
+		t.Error("Truncated set despite the walk completing")
+	}
+}
+
+func TestSkipsOversizeFiles(t *testing.T) {
+	dir := t.TempDir()
+	big := make([]byte, 3*1024)
+	for i := range big {
+		big[i] = 'x'
+	}
+	seed(t, dir, map[string]string{"small.py": "x"})
+	if err := os.WriteFile(filepath.Join(dir, "big.py"), big, 0o644); err != nil {
+		t.Fatal(err)
+	}
+
+	got, err := Find(dir, Options{MaxFileSizeKB: 2})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(got.Paths) != 1 || got.Paths[0] != "small.py" {
+		t.Errorf("paths = %v, want only small.py", got.Paths)
+	}
+	if got.Skipped != 1 {
+		t.Errorf("Skipped = %d, want 1", got.Skipped)
+	}
+}
+
+func TestIndexesExtensionlessKnownFilenames(t *testing.T) {
+	dir := t.TempDir()
+	seed(t, dir, map[string]string{
+		"Gemfile": "x", "Rakefile": "x", "config.ru": "x", "LICENSE": "x",
+	})
+	got, err := Find(dir, Options{})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(got.Paths) != 3 {
+		t.Errorf("paths = %v, want the three known Ruby filenames and not LICENSE", got.Paths)
+	}
+}
+
+func TestPathsAreRelativeAndSlashSeparated(t *testing.T) {
+	dir := t.TempDir()
+	seed(t, dir, map[string]string{"a/b/c.py": "x"})
+	got, err := Find(dir, Options{})
+	if err != nil {
+		t.Fatal(err)
+	}
+	// Must match CodeFile.path on the Python side, which is repo-relative with forward
+	// slashes regardless of platform.
+	if len(got.Paths) != 1 || got.Paths[0] != "a/b/c.py" {
+		t.Errorf("paths = %v, want [a/b/c.py]", got.Paths)
+	}
+}
+
+func TestMissingRootIsNotAPanic(t *testing.T) {
+	if _, err := Find(filepath.Join(t.TempDir(), "nope"), Options{}); err != nil {
+		t.Errorf("a missing root should return empty, not error: %v", err)
+	}
+}