diff --git a/provider_gemini.go b/provider_gemini.go
new file mode 100644
index 0000000..f00bfa9
--- /dev/null
+++ b/provider_gemini.go
@@ -0,0 +1,271 @@
+package traefikllmgateway
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"io"
+	"net/http"
+	"strings"
+	"time"
+)
+
+// geminiAdapter is the providerAdapter for provider type "gemini".
+type geminiAdapter struct {
+	client      *http.Client
+	adapterName string
+	baseURL     string
+	apiKey      string
+}
+
+// newGeminiAdapter returns a geminiAdapter for provider name, with base as
+// its already-defaulted, trailing-slash-trimmed base URL and apiKey as
+// already resolved by resolveSecret. Like anthropic and unlike openai-type
+// providers, Gemini's API requires an API key on every request — a
+// resolved-empty apiKey is a constructor error here, mirroring ruling (c).
+func newGeminiAdapter(name, base, apiKey string) (*geminiAdapter, error) {
+	if apiKey == "" {
+		return nil, fmt.Errorf("llmgateway: provider %q: gemini requires an API key", name)
+	}
+	return &geminiAdapter{
+		adapterName: name,
+		baseURL:     base,
+		apiKey:      apiKey,
+		client:      newAdapterHTTPClient(),
+	}, nil
+}
+
+// name implements providerAdapter.
+func (a *geminiAdapter) name() string { return a.adapterName }
+
+// typeName implements providerAdapter.
+func (a *geminiAdapter) typeName() string { return providerTypeGemini }
+
+// base implements providerAdapter.
+func (a *geminiAdapter) base() string { return a.baseURL }
+
+// injectAuth implements providerAdapter: Gemini's x-goog-api-key header
+// scheme. Never a no-op — the constructor already rejected an empty apiKey.
+func (a *geminiAdapter) injectAuth(r *http.Request) {
+	r.Header.Set("x-goog-api-key", a.apiKey)
+}
+
+// requestHeaders returns the headers for one upstream request: a
+// Content-Type when the request carries a JSON body, plus whatever
+// injectAuth adds.
+func (a *geminiAdapter) requestHeaders(hasBody bool) http.Header {
+	h := http.Header{}
+	if hasBody {
+		h.Set("Content-Type", "application/json")
+	}
+	a.injectAuth(&http.Request{Header: h})
+	return h
+}
+
+// geminiModelFromRequest reads req["model"] — already rewritten to the
+// provider's own model id by the caller, per the providerAdapter interface
+// doc — and strips a leading "models/" prefix if present, so a client (or
+// an upstream config) that already qualified the id doesn't produce a
+// doubled "models/models/..." URL.
+func geminiModelFromRequest(req map[string]any) string {
+	raw, _ := req["model"].(string)
+	return strings.TrimPrefix(raw, "models/")
+}
+
+// chatCompletion implements providerAdapter: translates req to a Gemini
+// generateContent/streamGenerateContent request via geminiRequestFromOpenAI,
+// then forwards the (possibly streamed) response back translated to
+// OpenAI's shape. Unlike anthropic and openai-type, the model id is never
+// part of the JSON body — Gemini puts it in the URL path — so it is read
+// once here and threaded through to both URL construction and the response
+// translation's "model" field. A translation failure — a *translateError
+// from geminiRequestFromOpenAI — is returned as-is, before any upstream
+// request is made.
+func (a *geminiAdapter) chatCompletion(ctx context.Context, w http.ResponseWriter, req map[string]any) (usage, error) {
+	model := geminiModelFromRequest(req)
+	streaming, _ := req["stream"].(bool)
+
+	body, err := geminiRequestFromOpenAI(req)
+	if err != nil {
+		return usage{}, err
+	}
+
+	hdr := a.requestHeaders(true)
+	url := a.baseURL + geminiAPIPrefix + model + ":generateContent"
+	if streaming {
+		// Ask the upstream for an SSE response. The Content-Type check
+		// below still decides which forwarding path runs.
+		hdr.Set("Accept", "text/event-stream")
+		url = a.baseURL + geminiAPIPrefix + model + ":streamGenerateContent?alt=sse"
+	}
+
+	resp, err := upstreamJSON(ctx, a.client, http.MethodPost, url, hdr, body)
+	if err != nil {
+		return usage{}, err
+	}
+	defer resp.Body.Close() //nolint:errcheck // read-side close; nothing actionable on failure
+
+	if resp.StatusCode < 200 || resp.StatusCode >= 300 {
+		return usage{}, newProviderHTTPError(resp)
+	}
+
+	if streaming && strings.Contains(strings.ToLower(resp.Header.Get("Content-Type")), "text/event-stream") {
+		return a.forwardStream(w, resp.Body, model)
+	}
+	// Either a non-streaming request, or a streaming one whose upstream
+	// ignored stream=true and answered with a normal JSON body.
+	return a.forwardJSON(w, resp, model)
+}
+
+// forwardJSON reads resp's full body (capped at maxResponseBytes),
+// translates it to an OpenAI chat.completion response via
+// openAIResponseFromGemini, and writes the translated JSON to w.
+// gatewayModel becomes the translated response's "model" field, and the
+// response's "created" timestamp is generated here — Gemini's response
+// carries neither.
+func (a *geminiAdapter) forwardJSON(w http.ResponseWriter, resp *http.Response, gatewayModel string) (usage, error) {
+	raw, err := io.ReadAll(io.LimitReader(resp.Body, maxResponseBytes))
+	if err != nil {
+		return usage{}, fmt.Errorf("%w: read response body: %w", errUpstream, err)
+	}
+
+	out, u, err := openAIResponseFromGemini(raw, gatewayModel, time.Now().Unix())
+	if err != nil {
+		return usage{}, err
+	}
+
+	b, err := json.Marshal(out)
+	if err != nil {
+		return usage{}, fmt.Errorf("%w: marshal translated response: %w", errUpstream, err)
+	}
+
+	w.Header().Set("Content-Type", "application/json")
+	w.WriteHeader(http.StatusOK)
+	_, _ = w.Write(b)
+	return u, nil
+}
+
+// forwardStream reads body as an upstream Gemini text/event-stream,
+// translating each event via a geminiStreamState and forwarding every
+// resulting OpenAI-shaped chunk to w through a fresh sseWriter, flushing
+// per event. gatewayModel and the stream's start time become every chunk's
+// "model"/"created" fields. The terminal "[DONE]" is written once readSSE
+// returns without error, mirroring ruling (b) — Gemini's stream has no
+// distinct terminal event of its own to key off of; a clean EOF is what
+// signals the stream ended. On a translate error (a malformed chunk, or a
+// genuine stream read error), forwardStream returns without writing
+// "[DONE]".
+func (a *geminiAdapter) forwardStream(w http.ResponseWriter, body io.Reader, gatewayModel string) (usage, error) {
+	st := newGeminiStreamState(gatewayModel, time.Now().Unix())
+	sw := newSSEWriter(w)
+
+	err := readSSE(body, func(ev sseEvent) error {
+		chunks, terr := st.translate(ev)
+		for _, c := range chunks {
+			if werr := sw.writeData(c); werr != nil {
+				return werr
+			}
+		}
+		return terr
+	})
+	if err != nil {
+		return st.usage(), fmt.Errorf("%w: stream: %w", errUpstream, err)
+	}
+	sw.writeDone()
+	return st.usage(), nil
+}
+
+// embeddings implements providerAdapter: translates req's OpenAI "input"
+// (a string or an array of strings) to a Gemini :embedContent or
+// :batchEmbedContents request, per the embeddings mapping table, and
+// translates the response back to an OpenAI embeddings list response.
+// Embeddings never stream.
+func (a *geminiAdapter) embeddings(ctx context.Context, w http.ResponseWriter, req map[string]any) (usage, error) {
+	model := geminiModelFromRequest(req)
+
+	texts, single, err := geminiEmbeddingInputTexts(req["input"])
+	if err != nil {
+		return usage{}, err
+	}
+
+	var body map[string]any
+	var url string
+	if single {
+		body = geminiEmbedContentRequest(texts[0])
+		url = a.baseURL + geminiAPIPrefix + model + ":embedContent"
+	} else {
+		body = geminiBatchEmbedContentsRequest(model, texts)
+		url = a.baseURL + geminiAPIPrefix + model + ":batchEmbedContents"
+	}
+
+	resp, err := upstreamJSON(ctx, a.client, http.MethodPost, url, a.requestHeaders(true), body)
+	if err != nil {
+		return usage{}, err
+	}
+	defer resp.Body.Close() //nolint:errcheck
+
+	if resp.StatusCode < 200 || resp.StatusCode >= 300 {
+		return usage{}, newProviderHTTPError(resp)
+	}
+
+	raw, err := io.ReadAll(io.LimitReader(resp.Body, maxResponseBytes))
+	if err != nil {
+		return usage{}, fmt.Errorf("%w: read response body: %w", errUpstream, err)
+	}
+
+	out, u, err := openAIEmbeddingResponseFromGemini(raw, single, model, estimatedEmbeddingTokens(texts))
+	if err != nil {
+		return usage{}, err
+	}
+
+	b, err := json.Marshal(out)
+	if err != nil {
+		return usage{}, fmt.Errorf("%w: marshal translated embeddings response: %w", errUpstream, err)
+	}
+
+	w.Header().Set("Content-Type", "application/json")
+	w.WriteHeader(http.StatusOK)
+	_, _ = w.Write(b)
+	return u, nil
+}
+
+// geminiModelsListPayload is the "models": [{"name": "models/..."}, ...]
+// shape Gemini's models.list endpoint returns (its "name" field is the
+// full "models/{model}" resource name, unlike OpenAI's bare "id").
+type geminiModelsListPayload struct {
+	Models []struct {
+		Name string `json:"name"`
+	} `json:"models"`
+}
+
+// listModels implements providerAdapter: GET {base}/v1beta/models,
+// stripping the "models/" prefix off each entry's "name" to return bare
+// model ids, matching the shape openai-type and anthropic-type adapters
+// both return.
+func (a *geminiAdapter) listModels(ctx context.Context) ([]string, error) {
+	resp, err := upstreamJSON(ctx, a.client, http.MethodGet, a.baseURL+"/v1beta/models", a.requestHeaders(false), nil)
+	if err != nil {
+		return nil, err
+	}
+	defer resp.Body.Close() //nolint:errcheck
+
+	if resp.StatusCode < 200 || resp.StatusCode >= 300 {
+		return nil, newProviderHTTPError(resp)
+	}
+
+	body, err := io.ReadAll(io.LimitReader(resp.Body, maxResponseBytes))
+	if err != nil {
+		return nil, fmt.Errorf("%w: read response body: %w", errUpstream, err)
+	}
+
+	var payload geminiModelsListPayload
+	if err := json.Unmarshal(body, &payload); err != nil {
+		return nil, fmt.Errorf("%w: decode models response: %w", errUpstream, err)
+	}
+
+	ids := make([]string, len(payload.Models))
+	for i, m := range payload.Models {
+		ids[i] = strings.TrimPrefix(m.Name, "models/")
+	}
+	return ids, nil
+}
diff --git a/provider_gemini_test.go b/provider_gemini_test.go
new file mode 100644
index 0000000..599ba6f
--- /dev/null
+++ b/provider_gemini_test.go
@@ -0,0 +1,495 @@
+package traefikllmgateway
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+)
+
+// TestNewGeminiAdapter_KeylessIsAConstructorError proves an empty resolved
+// API key is a constructor error for gemini-type providers, mirroring
+// ruling (c) and anthropic's equivalent test.
+func TestNewGeminiAdapter_KeylessIsAConstructorError(t *testing.T) {
+	if _, err := newGeminiAdapter("p1", "https://generativelanguage.googleapis.com", ""); err == nil {
+		t.Fatalf("newGeminiAdapter: want error for empty API key, got nil")
+	}
+}
+
+// TestGeminiAdapter_InjectAuth proves injectAuth sets the x-goog-api-key
+// header unconditionally.
+func TestGeminiAdapter_InjectAuth(t *testing.T) {
+	a, err := newGeminiAdapter("p1", "https://generativelanguage.googleapis.com", "AIza-test")
+	if err != nil {
+		t.Fatalf("newGeminiAdapter: %v", err)
+	}
+	req := httptest.NewRequest(http.MethodPost, "/", nil)
+	a.injectAuth(req)
+
+	if got := req.Header.Get("x-goog-api-key"); got != "AIza-test" { //nolint:gosec // test-only literal, not a real credential
+		t.Errorf("x-goog-api-key = %q, want %q", got, "AIza-test")
+	}
+}
+
+// TestGeminiAdapter_ChatCompletion_NonStreaming proves the adapter builds
+// the {base}/v1beta/models/{model}:generateContent URL from req["model"],
+// sends the x-goog-api-key header, translates the OpenAI request to
+// Gemini's wire format, and translates the Gemini response back to an
+// OpenAI-shaped chat.completion body on the wire, with usage extracted
+// from usageMetadata.
+func TestGeminiAdapter_ChatCompletion_NonStreaming(t *testing.T) {
+	const geminiResp = `{"candidates":[{"content":{"role":"model","parts":[{"text":"hi there"}]},"finishReason":"STOP"}],"usageMetadata":{"promptTokenCount":10,"candidatesTokenCount":4},"responseId":"resp_01ABC"}`
+
+	var gotBody []byte
+	var gotAPIKey, gotPath, gotMethod string
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		gotPath = r.URL.Path
+		gotMethod = r.Method
+		gotBody, _ = io.ReadAll(r.Body)
+		gotAPIKey = r.Header.Get("x-goog-api-key")
+		w.Header().Set("Content-Type", "application/json")
+		w.WriteHeader(http.StatusOK)
+		_, _ = w.Write([]byte(geminiResp))
+	}))
+	defer srv.Close()
+
+	a, err := newGeminiAdapter("p1", srv.URL, "AIza-test")
+	if err != nil {
+		t.Fatalf("newGeminiAdapter: %v", err)
+	}
+	rec := httptest.NewRecorder()
+
+	req := map[string]any{
+		"model":    "gemini-2.5-pro",
+		"messages": []any{map[string]any{"role": "user", "content": "hello"}},
+	}
+	u, err := a.chatCompletion(context.Background(), rec, req)
+	if err != nil {
+		t.Fatalf("chatCompletion: %v", err)
+	}
+
+	if gotPath != "/v1beta/models/gemini-2.5-pro:generateContent" {
+		t.Errorf("path = %q, want /v1beta/models/gemini-2.5-pro:generateContent", gotPath)
+	}
+	if gotMethod != http.MethodPost {
+		t.Errorf("method = %q, want POST", gotMethod)
+	}
+	if u.prompt != 10 || u.completion != 4 {
+		t.Errorf("usage = %+v, want {prompt:10 completion:4}", u)
+	}
+	if gotAPIKey != "AIza-test" { //nolint:gosec // test-only literal, not a real credential
+		t.Errorf("x-goog-api-key = %q, want %q", gotAPIKey, "AIza-test")
+	}
+
+	sent := decodeJSONBody(t, gotBody)
+	if _, hasModel := sent["model"]; hasModel {
+		t.Errorf("upstream body carries a \"model\" field = %v, want none (Gemini's model is URL-only)", sent["model"])
+	}
+	contents, ok := sent["contents"].([]any)
+	if !ok || len(contents) != 1 {
+		t.Fatalf("upstream body contents = %v, want a single-element array", sent["contents"])
+	}
+
+	if got := rec.Header().Get("Content-Type"); got != "application/json" {
+		t.Errorf("Content-Type = %q, want application/json", got)
+	}
+	var out map[string]any
+	if err := json.Unmarshal(rec.Body.Bytes(), &out); err != nil {
+		t.Fatalf("client body is not valid JSON: %v (%s)", err, rec.Body.String())
+	}
+	if out["object"] != "chat.completion" {
+		t.Errorf("object = %v, want %q", out["object"], "chat.completion")
+	}
+	if out["id"] != "chatcmpl-resp_01ABC" {
+		t.Errorf("id = %v, want %q", out["id"], "chatcmpl-resp_01ABC")
+	}
+	choices, ok := out["choices"].([]any)
+	if !ok || len(choices) != 1 {
+		t.Fatalf("choices = %v, want a single-element array", out["choices"])
+	}
+	choice := choices[0].(map[string]any)
+	message := choice["message"].(map[string]any)
+	if message["content"] != "hi there" {
+		t.Errorf("message.content = %v, want %q", message["content"], "hi there")
+	}
+	if choice["finish_reason"] != "stop" {
+		t.Errorf("finish_reason = %v, want %q", choice["finish_reason"], "stop")
+	}
+}
+
+// TestGeminiAdapter_ChatCompletion_ModelsPrefixStripped proves a
+// req["model"] that already carries a "models/" prefix is not doubled into
+// the URL.
+func TestGeminiAdapter_ChatCompletion_ModelsPrefixStripped(t *testing.T) {
+	var gotPath string
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		gotPath = r.URL.Path
+		w.Header().Set("Content-Type", "application/json")
+		w.WriteHeader(http.StatusOK)
+		_, _ = w.Write([]byte(`{"candidates":[{"content":{"role":"model","parts":[{"text":"ok"}]},"finishReason":"STOP"}]}`))
+	}))
+	defer srv.Close()
+
+	a, err := newGeminiAdapter("p1", srv.URL, "AIza-test")
+	if err != nil {
+		t.Fatalf("newGeminiAdapter: %v", err)
+	}
+	rec := httptest.NewRecorder()
+
+	_, err = a.chatCompletion(context.Background(), rec, map[string]any{
+		"model":    "models/gemini-2.5-pro",
+		"messages": []any{map[string]any{"role": "user", "content": "hi"}},
+	})
+	if err != nil {
+		t.Fatalf("chatCompletion: %v", err)
+	}
+	if gotPath != "/v1beta/models/gemini-2.5-pro:generateContent" {
+		t.Errorf("path = %q, want /v1beta/models/gemini-2.5-pro:generateContent (no doubled models/ prefix)", gotPath)
+	}
+}
+
+// geminiStreamFrames are the raw SSE "data:" lines a fake streaming
+// upstream emits for the tests below: a role/text chunk, a text-only
+// continuation, then a finish chunk carrying usageMetadata. Gemini's SSE
+// stream has no "event:" field to set, unlike Anthropic's.
+var geminiStreamFrames = []string{
+	`{"candidates":[{"content":{"role":"model","parts":[{"text":"Hi!"}]}}],"responseId":"resp_STREAM"}`,
+	`{"candidates":[{"content":{"role":"model","parts":[]},"finishReason":"STOP"}],"usageMetadata":{"promptTokenCount":6,"candidatesTokenCount":3}}`,
+}
+
+// TestGeminiAdapter_ChatCompletion_Streaming proves the adapter requests
+// the :streamGenerateContent?alt=sse URL, forwards a streaming Gemini
+// response as OpenAI-shaped SSE chunks on the wire, flushing per event, and
+// terminates with "[DONE]" once the upstream stream ends — mirroring
+// ruling (b): there is no distinct terminal event of Gemini's own to key
+// off of.
+func TestGeminiAdapter_ChatCompletion_Streaming(t *testing.T) {
+	var gotPath, gotQuery, gotAccept string
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		gotPath = r.URL.Path
+		gotQuery = r.URL.RawQuery
+		gotAccept = r.Header.Get("Accept")
+		w.Header().Set("Content-Type", "text/event-stream")
+		w.WriteHeader(http.StatusOK)
+		f, ok := w.(http.Flusher)
+		if !ok {
+			t.Errorf("httptest ResponseWriter does not implement http.Flusher")
+			return
+		}
+		for _, data := range geminiStreamFrames {
+			fmt.Fprintf(w, "data: %s\n\n", data)
+			f.Flush()
+		}
+	}))
+	defer srv.Close()
+
+	a, err := newGeminiAdapter("p1", srv.URL, "AIza-test")
+	if err != nil {
+		t.Fatalf("newGeminiAdapter: %v", err)
+	}
+	rec := newRecordingResponseWriter()
+
+	req := map[string]any{
+		"model":    "gemini-2.5-pro",
+		"stream":   true,
+		"messages": []any{map[string]any{"role": "user", "content": "hello"}},
+	}
+	u, err := a.chatCompletion(context.Background(), rec, req)
+	if err != nil {
+		t.Fatalf("chatCompletion: %v", err)
+	}
+
+	if gotPath != "/v1beta/models/gemini-2.5-pro:streamGenerateContent" {
+		t.Errorf("path = %q, want /v1beta/models/gemini-2.5-pro:streamGenerateContent", gotPath)
+	}
+	if gotQuery != "alt=sse" {
+		t.Errorf("query = %q, want %q", gotQuery, "alt=sse")
+	}
+	if gotAccept != "text/event-stream" {
+		t.Errorf("Accept = %q, want text/event-stream", gotAccept)
+	}
+	if u.prompt != 6 || u.completion != 3 {
+		t.Errorf("usage = %+v, want {prompt:6 completion:3}", u)
+	}
+
+	body := rec.body.String()
+	if got := rec.header.Get("Content-Type"); got != "text/event-stream" {
+		t.Errorf("Content-Type = %q, want text/event-stream", got)
+	}
+
+	wantSubstrings := []string{
+		`"delta":{"role":"assistant"}`,
+		`"delta":{"content":"Hi!"}`,
+		`"finish_reason":"stop"`,
+		"data: [DONE]\n\n",
+	}
+	for _, want := range wantSubstrings {
+		if !strings.Contains(body, want) {
+			t.Errorf("streamed body missing %q; got:\n%s", want, body)
+		}
+	}
+	// 3 chunks (role, content delta, finish) plus one more flush for the
+	// terminal [DONE]: 4 flushes, proving the adapter flushes per event
+	// rather than buffering the whole stream.
+	const wantFlushes = 4
+	if rec.flushed < wantFlushes {
+		t.Errorf("flushed = %d, want >= %d", rec.flushed, wantFlushes)
+	}
+}
+
+// TestGeminiAdapter_ChatCompletion_Streaming_CaseInsensitiveContentType
+// proves the streaming-vs-JSON-fallback gate matches an upstream's
+// Content-Type case-insensitively, per lesson (g) — tested with
+// "Text/Event-Stream; charset=utf-8".
+func TestGeminiAdapter_ChatCompletion_Streaming_CaseInsensitiveContentType(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "Text/Event-Stream; charset=utf-8")
+		w.WriteHeader(http.StatusOK)
+		f, ok := w.(http.Flusher)
+		if !ok {
+			t.Errorf("httptest ResponseWriter does not implement http.Flusher")
+			return
+		}
+		for _, data := range geminiStreamFrames {
+			fmt.Fprintf(w, "data: %s\n\n", data)
+			f.Flush()
+		}
+	}))
+	defer srv.Close()
+
+	a, err := newGeminiAdapter("p1", srv.URL, "AIza-test")
+	if err != nil {
+		t.Fatalf("newGeminiAdapter: %v", err)
+	}
+	rec := newRecordingResponseWriter()
+
+	req := map[string]any{
+		"model":    "gemini-2.5-pro",
+		"stream":   true,
+		"messages": []any{map[string]any{"role": "user", "content": "hello"}},
+	}
+	u, err := a.chatCompletion(context.Background(), rec, req)
+	if err != nil {
+		t.Fatalf("chatCompletion: %v", err)
+	}
+	if u.prompt != 6 || u.completion != 3 {
+		t.Errorf("usage = %+v, want {prompt:6 completion:3} (proves the streaming path, not JSON fallback, was taken)", u)
+	}
+	if got := rec.header.Get("Content-Type"); got != "text/event-stream" {
+		t.Errorf("Content-Type on client response = %q, want %q (sseWriter's own header; JSON fallback would copy the upstream's mixed-case value verbatim)", got, "text/event-stream")
+	}
+}
+
+// TestGeminiAdapter_ChatCompletion_Upstream429 proves a non-2xx upstream
+// response returns a *providerHTTPError with nothing written to the
+// client.
+func TestGeminiAdapter_ChatCompletion_Upstream429(t *testing.T) {
+	const errBody = `{"error":{"code":429,"message":"rate limited","status":"RESOURCE_EXHAUSTED"}}`
+
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		w.WriteHeader(http.StatusTooManyRequests)
+		_, _ = w.Write([]byte(errBody))
+	}))
+	defer srv.Close()
+
+	a, err := newGeminiAdapter("p1", srv.URL, "AIza-test")
+	if err != nil {
+		t.Fatalf("newGeminiAdapter: %v", err)
+	}
+	rec := httptest.NewRecorder()
+
+	u, err := a.chatCompletion(context.Background(), rec, map[string]any{
+		"model":    "gemini-2.5-pro",
+		"messages": []any{map[string]any{"role": "user", "content": "hi"}},
+	})
+	if err == nil {
+		t.Fatalf("chatCompletion: want error, got nil (usage=%+v)", u)
+	}
+	var httpErr *providerHTTPError
+	if !errors.As(err, &httpErr) {
+		t.Fatalf("err = %v (%T), want *providerHTTPError", err, err)
+	}
+	if httpErr.status != http.StatusTooManyRequests {
+		t.Errorf("status = %d, want %d", httpErr.status, http.StatusTooManyRequests)
+	}
+	if rec.Body.Len() != 0 {
+		t.Errorf("client response body = %q, want empty", rec.Body.String())
+	}
+}
+
+// TestGeminiAdapter_ChatCompletion_TranslateErrorSkipsUpstream proves a
+// request field geminiRequestFromOpenAI rejects (n>1, here) never reaches
+// the upstream at all.
+func TestGeminiAdapter_ChatCompletion_TranslateErrorSkipsUpstream(t *testing.T) {
+	called := false
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		called = true
+		w.WriteHeader(http.StatusOK)
+	}))
+	defer srv.Close()
+
+	a, err := newGeminiAdapter("p1", srv.URL, "AIza-test")
+	if err != nil {
+		t.Fatalf("newGeminiAdapter: %v", err)
+	}
+	rec := httptest.NewRecorder()
+
+	_, err = a.chatCompletion(context.Background(), rec, map[string]any{
+		"model":    "gemini-2.5-pro",
+		"messages": []any{map[string]any{"role": "user", "content": "hi"}},
+		"n":        float64(2),
+	})
+	var terr *translateError
+	if !errors.As(err, &terr) {
+		t.Fatalf("err = %v (%T), want *translateError", err, err)
+	}
+	if called {
+		t.Errorf("upstream was called, want the translate error to short-circuit before any request")
+	}
+}
+
+// TestGeminiAdapter_Embeddings_Single proves embeddings builds the
+// :embedContent URL and translates the response to an OpenAI embeddings
+// list body for a single-string "input".
+func TestGeminiAdapter_Embeddings_Single(t *testing.T) {
+	var gotPath string
+	var gotBody []byte
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		gotPath = r.URL.Path
+		gotBody, _ = io.ReadAll(r.Body)
+		w.Header().Set("Content-Type", "application/json")
+		w.WriteHeader(http.StatusOK)
+		_, _ = w.Write([]byte(`{"embedding":{"values":[0.1,0.2,0.3]}}`))
+	}))
+	defer srv.Close()
+
+	a, err := newGeminiAdapter("p1", srv.URL, "AIza-test")
+	if err != nil {
+		t.Fatalf("newGeminiAdapter: %v", err)
+	}
+	rec := httptest.NewRecorder()
+
+	u, err := a.embeddings(context.Background(), rec, map[string]any{
+		"model": "text-embedding-004",
+		"input": "hello world",
+	})
+	if err != nil {
+		t.Fatalf("embeddings: %v", err)
+	}
+	if gotPath != "/v1beta/models/text-embedding-004:embedContent" {
+		t.Errorf("path = %q, want /v1beta/models/text-embedding-004:embedContent", gotPath)
+	}
+	sent := decodeJSONBody(t, gotBody)
+	if _, hasRequests := sent["requests"]; hasRequests {
+		t.Errorf("single-input body carries \"requests\", want the :embedContent shape")
+	}
+	if !u.estimated {
+		t.Errorf("usage.estimated = false, want true (Gemini reports no embeddings usage)")
+	}
+
+	var out map[string]any
+	if err := json.Unmarshal(rec.Body.Bytes(), &out); err != nil {
+		t.Fatalf("client body is not valid JSON: %v (%s)", err, rec.Body.String())
+	}
+	if out["object"] != "list" {
+		t.Errorf("object = %v, want %q", out["object"], "list")
+	}
+	data, ok := out["data"].([]any)
+	if !ok || len(data) != 1 {
+		t.Fatalf("data = %v, want a single-element array", out["data"])
+	}
+}
+
+// TestGeminiAdapter_Embeddings_Batch proves an array "input" selects
+// Gemini's :batchEmbedContents endpoint and translates its response to an
+// OpenAI embeddings list body with one entry per input string.
+func TestGeminiAdapter_Embeddings_Batch(t *testing.T) {
+	var gotPath string
+	var gotBody []byte
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		gotPath = r.URL.Path
+		gotBody, _ = io.ReadAll(r.Body)
+		w.Header().Set("Content-Type", "application/json")
+		w.WriteHeader(http.StatusOK)
+		_, _ = w.Write([]byte(`{"embeddings":[{"values":[0.1,0.2]},{"values":[0.3,0.4]}]}`))
+	}))
+	defer srv.Close()
+
+	a, err := newGeminiAdapter("p1", srv.URL, "AIza-test")
+	if err != nil {
+		t.Fatalf("newGeminiAdapter: %v", err)
+	}
+	rec := httptest.NewRecorder()
+
+	_, err = a.embeddings(context.Background(), rec, map[string]any{
+		"model": "text-embedding-004",
+		"input": []any{"hello", "world!!"},
+	})
+	if err != nil {
+		t.Fatalf("embeddings: %v", err)
+	}
+	if gotPath != "/v1beta/models/text-embedding-004:batchEmbedContents" {
+		t.Errorf("path = %q, want /v1beta/models/text-embedding-004:batchEmbedContents", gotPath)
+	}
+	sent := decodeJSONBody(t, gotBody)
+	requests, ok := sent["requests"].([]any)
+	if !ok || len(requests) != 2 {
+		t.Fatalf("requests = %v, want a 2-element array", sent["requests"])
+	}
+	first := requests[0].(map[string]any)
+	if first["model"] != "models/text-embedding-004" {
+		t.Errorf("requests[0].model = %v, want %q", first["model"], "models/text-embedding-004")
+	}
+
+	var out map[string]any
+	if err := json.Unmarshal(rec.Body.Bytes(), &out); err != nil {
+		t.Fatalf("client body is not valid JSON: %v (%s)", err, rec.Body.String())
+	}
+	data, ok := out["data"].([]any)
+	if !ok || len(data) != 2 {
+		t.Fatalf("data = %v, want a 2-element array", out["data"])
+	}
+}
+
+// TestGeminiAdapter_ListModels proves listModels parses Gemini's
+// "models": [{"name": "models/..."}] Models API response into a flat
+// slice of bare ids, stripping the "models/" prefix.
+func TestGeminiAdapter_ListModels(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.URL.Path != "/v1beta/models" {
+			t.Errorf("path = %q, want /v1beta/models", r.URL.Path)
+			return
+		}
+		if r.Method != http.MethodGet {
+			t.Errorf("method = %q, want GET", r.Method)
+			return
+		}
+		if got := r.Header.Get("x-goog-api-key"); got != "AIza-test" {
+			t.Errorf("x-goog-api-key = %q, want %q", got, "AIza-test")
+			return
+		}
+		w.Header().Set("Content-Type", "application/json")
+		_, _ = w.Write([]byte(`{"models":[{"name":"models/gemini-2.5-pro"},{"name":"models/gemini-2.5-flash"}]}`))
+	}))
+	defer srv.Close()
+
+	a, err := newGeminiAdapter("p1", srv.URL, "AIza-test")
+	if err != nil {
+		t.Fatalf("newGeminiAdapter: %v", err)
+	}
+
+	got, err := a.listModels(context.Background())
+	if err != nil {
+		t.Fatalf("listModels: %v", err)
+	}
+	want := []string{"gemini-2.5-pro", "gemini-2.5-flash"}
+	if len(got) != len(want) || got[0] != want[0] || got[1] != want[1] {
+		t.Errorf("got %#v, want %#v", got, want)
+	}
+}
diff --git a/provider_openai_test.go b/provider_openai_test.go
index a99b2a0..ec13f8b 100644
--- a/provider_openai_test.go
+++ b/provider_openai_test.go
@@ -629,13 +629,32 @@ func TestBuildAdapters(t *testing.T) {
 		}
 	})
 
-	t.Run("edge: gemini type not yet implemented", func(t *testing.T) {
+	t.Run("good: gemini with an API key builds successfully", func(t *testing.T) {
 		cfg := &Config{Providers: map[string]*ProviderConfig{
 			"p1": {Type: "gemini", APIKey: "sk-test"},
 		}}
-		_, err := buildAdapters(cfg)
-		if err == nil {
-			t.Fatalf("buildAdapters: want placeholder error for gemini, got nil")
+		adapters, err := buildAdapters(cfg)
+		if err != nil {
+			t.Fatalf("buildAdapters: %v", err)
+		}
+		a, ok := adapters["p1"]
+		if !ok {
+			t.Fatalf("adapters[%q] missing", "p1")
+		}
+		if a.typeName() != "gemini" {
+			t.Errorf("typeName() = %q, want %q", a.typeName(), "gemini")
+		}
+		if a.base() != defaultBaseGemini {
+			t.Errorf("base() = %q, want %q", a.base(), defaultBaseGemini)
+		}
+	})
+
+	t.Run("bad: gemini without an API key is a constructor error", func(t *testing.T) {
+		cfg := &Config{Providers: map[string]*ProviderConfig{
+			"p1": {Type: "gemini"},
+		}}
+		if _, err := buildAdapters(cfg); err == nil {
+			t.Fatalf("buildAdapters: want error for keyless gemini provider, got nil")
 		}
 	})
 }
diff --git a/providers.go b/providers.go
index e6b24d6..5612b9b 100644
--- a/providers.go
+++ b/providers.go
@@ -182,8 +182,8 @@ func upstreamJSON(ctx context.Context, client *http.Client, method, url string,
 // keyless upstream for openai-type providers, but a constructor error for
 // anthropic-type ones — newAnthropicAdapter enforces that, per ruling (c);
 // this function just propagates whatever error it returns. A configured
-// gemini provider currently errors too — Task 10 fills in that switch
-// case.
+// gemini provider follows the same constructor-error ruling —
+// newGeminiAdapter enforces it too.
 func buildAdapters(cfg *Config) (map[string]providerAdapter, error) {
 	adapters := make(map[string]providerAdapter, len(cfg.Providers))
 	for name, pc := range cfg.Providers {
@@ -207,7 +207,11 @@ func buildAdapters(cfg *Config) (map[string]providerAdapter, error) {
 			}
 			adapters[name] = a
 		case providerTypeGemini:
-			return nil, fmt.Errorf("llmgateway: provider %q: type %q not yet implemented", name, pc.Type)
+			g, err := newGeminiAdapter(name, base, apiKey)
+			if err != nil {
+				return nil, err
+			}
+			adapters[name] = g
 		default:
 			return nil, fmt.Errorf("llmgateway: provider %q: unknown type %q", name, pc.Type)
 		}
diff --git a/testdata/gemini/chat_basic_in.json b/testdata/gemini/chat_basic_in.json
new file mode 100644
index 0000000..9123c6d
--- /dev/null
+++ b/testdata/gemini/chat_basic_in.json
@@ -0,0 +1,8 @@
+{
+  "model": "gemini-2.5-pro",
+  "messages": [
+    {"role": "user", "content": "Hello there"}
+  ],
+  "max_tokens": 256,
+  "temperature": 0.7
+}
diff --git a/testdata/gemini/chat_basic_want.json b/testdata/gemini/chat_basic_want.json
new file mode 100644
index 0000000..6e1063d
--- /dev/null
+++ b/testdata/gemini/chat_basic_want.json
@@ -0,0 +1,9 @@
+{
+  "contents": [
+    {"role": "user", "parts": [{"text": "Hello there"}]}
+  ],
+  "generationConfig": {
+    "maxOutputTokens": 256,
+    "temperature": 0.7
+  }
+}
diff --git a/testdata/gemini/chat_images_http_error_in.json b/testdata/gemini/chat_images_http_error_in.json
new file mode 100644
index 0000000..a1707ac
--- /dev/null
+++ b/testdata/gemini/chat_images_http_error_in.json
@@ -0,0 +1,11 @@
+{
+  "model": "gemini-2.5-pro",
+  "messages": [
+    {
+      "role": "user",
+      "content": [
+        {"type": "image_url", "image_url": {"url": "https://example.com/cat.png"}}
+      ]
+    }
+  ]
+}
diff --git a/testdata/gemini/chat_images_http_error_wanterr.json b/testdata/gemini/chat_images_http_error_wanterr.json
new file mode 100644
index 0000000..0578360
--- /dev/null
+++ b/testdata/gemini/chat_images_http_error_wanterr.json
@@ -0,0 +1 @@
+{"errContains": "must be a data: URI", "notSupported": false}
diff --git a/testdata/gemini/chat_images_in.json b/testdata/gemini/chat_images_in.json
new file mode 100644
index 0000000..3ee1202
--- /dev/null
+++ b/testdata/gemini/chat_images_in.json
@@ -0,0 +1,12 @@
+{
+  "model": "gemini-2.5-pro",
+  "messages": [
+    {
+      "role": "user",
+      "content": [
+        {"type": "text", "text": "What is in this image?"},
+        {"type": "image_url", "image_url": {"url": "data:image/jpeg;charset=utf-8;base64,/9j/4AAQSkZJRg=="}}
+      ]
+    }
+  ]
+}
diff --git a/testdata/gemini/chat_images_want.json b/testdata/gemini/chat_images_want.json
new file mode 100644
index 0000000..32f422f
--- /dev/null
+++ b/testdata/gemini/chat_images_want.json
@@ -0,0 +1,11 @@
+{
+  "contents": [
+    {
+      "role": "user",
+      "parts": [
+        {"text": "What is in this image?"},
+        {"inlineData": {"mimeType": "image/jpeg", "data": "/9j/4AAQSkZJRg=="}}
+      ]
+    }
+  ]
+}
diff --git a/testdata/gemini/chat_parallel_tool_results_in.json b/testdata/gemini/chat_parallel_tool_results_in.json
new file mode 100644
index 0000000..84e87a6
--- /dev/null
+++ b/testdata/gemini/chat_parallel_tool_results_in.json
@@ -0,0 +1,16 @@
+{
+  "model": "gemini-2.5-pro",
+  "messages": [
+    {"role": "user", "content": "What's the weather in Paris and Tokyo?"},
+    {
+      "role": "assistant",
+      "content": null,
+      "tool_calls": [
+        {"id": "call_1", "type": "function", "function": {"name": "get_weather", "arguments": "{\"location\":\"Paris\"}"}},
+        {"id": "call_2", "type": "function", "function": {"name": "get_weather", "arguments": "{\"location\":\"Tokyo\"}"}}
+      ]
+    },
+    {"role": "tool", "tool_call_id": "call_1", "content": "{\"temp_c\":18}"},
+    {"role": "tool", "tool_call_id": "call_2", "content": "{\"temp_c\":27}"}
+  ]
+}
diff --git a/testdata/gemini/chat_parallel_tool_results_want.json b/testdata/gemini/chat_parallel_tool_results_want.json
new file mode 100644
index 0000000..90abdbd
--- /dev/null
+++ b/testdata/gemini/chat_parallel_tool_results_want.json
@@ -0,0 +1,19 @@
+{
+  "contents": [
+    {"role": "user", "parts": [{"text": "What's the weather in Paris and Tokyo?"}]},
+    {
+      "role": "model",
+      "parts": [
+        {"functionCall": {"name": "get_weather", "args": {"location": "Paris"}}},
+        {"functionCall": {"name": "get_weather", "args": {"location": "Tokyo"}}}
+      ]
+    },
+    {
+      "role": "user",
+      "parts": [
+        {"functionResponse": {"name": "get_weather", "response": {"content": "{\"temp_c\":18}"}}},
+        {"functionResponse": {"name": "get_weather", "response": {"content": "{\"temp_c\":27}"}}}
+      ]
+    }
+  ]
+}
diff --git a/testdata/gemini/chat_system_array_error_in.json b/testdata/gemini/chat_system_array_error_in.json
new file mode 100644
index 0000000..d6e848f
--- /dev/null
+++ b/testdata/gemini/chat_system_array_error_in.json
@@ -0,0 +1,7 @@
+{
+  "model": "gemini-2.5-pro",
+  "messages": [
+    {"role": "system", "content": [{"type": "image_url", "image_url": {"url": "https://x"}}]},
+    {"role": "user", "content": "Hi"}
+  ]
+}
diff --git a/testdata/gemini/chat_system_array_error_wanterr.json b/testdata/gemini/chat_system_array_error_wanterr.json
new file mode 100644
index 0000000..9f97b90
--- /dev/null
+++ b/testdata/gemini/chat_system_array_error_wanterr.json
@@ -0,0 +1 @@
+{"errContains": "not supported for gemini models", "notSupported": false}
diff --git a/testdata/gemini/chat_system_array_in.json b/testdata/gemini/chat_system_array_in.json
new file mode 100644
index 0000000..c36a3ab
--- /dev/null
+++ b/testdata/gemini/chat_system_array_in.json
@@ -0,0 +1,7 @@
+{
+  "model": "gemini-2.5-pro",
+  "messages": [
+    {"role": "system", "content": [{"type": "text", "text": "Line one."}, {"type": "text", "text": "Line two."}]},
+    {"role": "user", "content": "Hi"}
+  ]
+}
diff --git a/testdata/gemini/chat_system_array_want.json b/testdata/gemini/chat_system_array_want.json
new file mode 100644
index 0000000..9ea40ed
--- /dev/null
+++ b/testdata/gemini/chat_system_array_want.json
@@ -0,0 +1,6 @@
+{
+  "systemInstruction": {"parts": [{"text": "Line one.\nLine two."}]},
+  "contents": [
+    {"role": "user", "parts": [{"text": "Hi"}]}
+  ]
+}
diff --git a/testdata/gemini/chat_tool_choice_none_in.json b/testdata/gemini/chat_tool_choice_none_in.json
new file mode 100644
index 0000000..5408483
--- /dev/null
+++ b/testdata/gemini/chat_tool_choice_none_in.json
@@ -0,0 +1,11 @@
+{
+  "model": "gemini-2.5-pro",
+  "messages": [{"role": "user", "content": "hi"}],
+  "tools": [
+    {
+      "type": "function",
+      "function": {"name": "noop", "description": "does nothing", "parameters": {"type": "object", "properties": {}}}
+    }
+  ],
+  "tool_choice": "none"
+}
diff --git a/testdata/gemini/chat_tool_choice_none_want.json b/testdata/gemini/chat_tool_choice_none_want.json
new file mode 100644
index 0000000..85836ed
--- /dev/null
+++ b/testdata/gemini/chat_tool_choice_none_want.json
@@ -0,0 +1,15 @@
+{
+  "contents": [
+    {"role": "user", "parts": [{"text": "hi"}]}
+  ],
+  "tools": [
+    {
+      "functionDeclarations": [
+        {"name": "noop", "description": "does nothing", "parameters": {"type": "object", "properties": {}}}
+      ]
+    }
+  ],
+  "toolConfig": {
+    "functionCallingConfig": {"mode": "NONE"}
+  }
+}
diff --git a/testdata/gemini/chat_tools_in.json b/testdata/gemini/chat_tools_in.json
new file mode 100644
index 0000000..10bee49
--- /dev/null
+++ b/testdata/gemini/chat_tools_in.json
@@ -0,0 +1,22 @@
+{
+  "model": "gemini-2.5-pro",
+  "messages": [
+    {"role": "system", "content": "You are a helpful weather bot."},
+    {"role": "user", "content": "What's the weather in Paris?"}
+  ],
+  "tools": [
+    {
+      "type": "function",
+      "function": {
+        "name": "get_weather",
+        "description": "Get current weather for a location",
+        "parameters": {
+          "type": "object",
+          "properties": {"location": {"type": "string"}},
+          "required": ["location"]
+        }
+      }
+    }
+  ],
+  "tool_choice": {"type": "function", "function": {"name": "get_weather"}}
+}
diff --git a/testdata/gemini/chat_tools_want.json b/testdata/gemini/chat_tools_want.json
new file mode 100644
index 0000000..693c32e
--- /dev/null
+++ b/testdata/gemini/chat_tools_want.json
@@ -0,0 +1,27 @@
+{
+  "systemInstruction": {"parts": [{"text": "You are a helpful weather bot."}]},
+  "contents": [
+    {"role": "user", "parts": [{"text": "What's the weather in Paris?"}]}
+  ],
+  "tools": [
+    {
+      "functionDeclarations": [
+        {
+          "name": "get_weather",
+          "description": "Get current weather for a location",
+          "parameters": {
+            "type": "object",
+            "properties": {"location": {"type": "string"}},
+            "required": ["location"]
+          }
+        }
+      ]
+    }
+  ],
+  "toolConfig": {
+    "functionCallingConfig": {
+      "mode": "ANY",
+      "allowedFunctionNames": ["get_weather"]
+    }
+  }
+}
diff --git a/testdata/gemini/chat_unsupported_logit_bias_in.json b/testdata/gemini/chat_unsupported_logit_bias_in.json
new file mode 100644
index 0000000..48bccd9
--- /dev/null
+++ b/testdata/gemini/chat_unsupported_logit_bias_in.json
@@ -0,0 +1,5 @@
+{
+  "model": "gemini-2.5-pro",
+  "messages": [{"role": "user", "content": "hi"}],
+  "logit_bias": {"1234": 10}
+}
diff --git a/testdata/gemini/chat_unsupported_logit_bias_wanterr.json b/testdata/gemini/chat_unsupported_logit_bias_wanterr.json
new file mode 100644
index 0000000..5d3dfef
--- /dev/null
+++ b/testdata/gemini/chat_unsupported_logit_bias_wanterr.json
@@ -0,0 +1 @@
+{"errContains": "logit_bias is not supported for gemini models", "notSupported": false}
diff --git a/testdata/gemini/chat_unsupported_n_in.json b/testdata/gemini/chat_unsupported_n_in.json
new file mode 100644
index 0000000..1617f9f
--- /dev/null
+++ b/testdata/gemini/chat_unsupported_n_in.json
@@ -0,0 +1,5 @@
+{
+  "model": "gemini-2.5-pro",
+  "messages": [{"role": "user", "content": "hi"}],
+  "n": 2
+}
diff --git a/testdata/gemini/chat_unsupported_n_wanterr.json b/testdata/gemini/chat_unsupported_n_wanterr.json
new file mode 100644
index 0000000..2a14758
--- /dev/null
+++ b/testdata/gemini/chat_unsupported_n_wanterr.json
@@ -0,0 +1 @@
+{"errContains": "n is not supported for gemini models", "notSupported": false}
diff --git a/testdata/gemini/embed_batch_in.json b/testdata/gemini/embed_batch_in.json
new file mode 100644
index 0000000..0e63ccf
--- /dev/null
+++ b/testdata/gemini/embed_batch_in.json
@@ -0,0 +1,4 @@
+{
+  "model": "text-embedding-004",
+  "input": ["hello", "world!!"]
+}
diff --git a/testdata/gemini/embed_batch_resp_in.json b/testdata/gemini/embed_batch_resp_in.json
new file mode 100644
index 0000000..bebebb1
--- /dev/null
+++ b/testdata/gemini/embed_batch_resp_in.json
@@ -0,0 +1,6 @@
+{
+  "embeddings": [
+    {"values": [0.1, 0.2]},
+    {"values": [0.3, 0.4]}
+  ]
+}
diff --git a/testdata/gemini/embed_batch_resp_want.json b/testdata/gemini/embed_batch_resp_want.json
new file mode 100644
index 0000000..3db08cb
--- /dev/null
+++ b/testdata/gemini/embed_batch_resp_want.json
@@ -0,0 +1,9 @@
+{
+  "object": "list",
+  "data": [
+    {"object": "embedding", "index": 0, "embedding": [0.1, 0.2]},
+    {"object": "embedding", "index": 1, "embedding": [0.3, 0.4]}
+  ],
+  "model": "gw-embed-model",
+  "usage": {"prompt_tokens": 3, "total_tokens": 3}
+}
diff --git a/testdata/gemini/embed_batch_want.json b/testdata/gemini/embed_batch_want.json
new file mode 100644
index 0000000..18116b7
--- /dev/null
+++ b/testdata/gemini/embed_batch_want.json
@@ -0,0 +1,6 @@
+{
+  "requests": [
+    {"model": "models/text-embedding-004", "content": {"parts": [{"text": "hello"}]}},
+    {"model": "models/text-embedding-004", "content": {"parts": [{"text": "world!!"}]}}
+  ]
+}
diff --git a/testdata/gemini/embed_single_in.json b/testdata/gemini/embed_single_in.json
new file mode 100644
index 0000000..7e5fe7c
--- /dev/null
+++ b/testdata/gemini/embed_single_in.json
@@ -0,0 +1,4 @@
+{
+  "model": "text-embedding-004",
+  "input": "hello world"
+}
diff --git a/testdata/gemini/embed_single_resp_in.json b/testdata/gemini/embed_single_resp_in.json
new file mode 100644
index 0000000..71b5e0d
--- /dev/null
+++ b/testdata/gemini/embed_single_resp_in.json
@@ -0,0 +1,3 @@
+{
+  "embedding": {"values": [0.1, 0.2, 0.3]}
+}
diff --git a/testdata/gemini/embed_single_resp_want.json b/testdata/gemini/embed_single_resp_want.json
new file mode 100644
index 0000000..2d8da6a
--- /dev/null
+++ b/testdata/gemini/embed_single_resp_want.json
@@ -0,0 +1,8 @@
+{
+  "object": "list",
+  "data": [
+    {"object": "embedding", "index": 0, "embedding": [0.1, 0.2, 0.3]}
+  ],
+  "model": "gw-embed-model",
+  "usage": {"prompt_tokens": 3, "total_tokens": 3}
+}
diff --git a/testdata/gemini/embed_single_want.json b/testdata/gemini/embed_single_want.json
new file mode 100644
index 0000000..b2cb5a8
--- /dev/null
+++ b/testdata/gemini/embed_single_want.json
@@ -0,0 +1,3 @@
+{
+  "content": {"parts": [{"text": "hello world"}]}
+}
diff --git a/testdata/gemini/resp_funcall_in.json b/testdata/gemini/resp_funcall_in.json
new file mode 100644
index 0000000..8275639
--- /dev/null
+++ b/testdata/gemini/resp_funcall_in.json
@@ -0,0 +1,18 @@
+{
+  "candidates": [
+    {
+      "content": {
+        "role": "model",
+        "parts": [
+          {"functionCall": {"id": "fc_1", "name": "get_weather", "args": {"location": "Paris"}}}
+        ]
+      },
+      "finishReason": "STOP"
+    }
+  ],
+  "usageMetadata": {
+    "promptTokenCount": 12,
+    "candidatesTokenCount": 6
+  },
+  "responseId": "resp_FUNC1"
+}
diff --git a/testdata/gemini/resp_funcall_want.json b/testdata/gemini/resp_funcall_want.json
new file mode 100644
index 0000000..9390d4a
--- /dev/null
+++ b/testdata/gemini/resp_funcall_want.json
@@ -0,0 +1,24 @@
+{
+  "id": "chatcmpl-resp_FUNC1",
+  "object": "chat.completion",
+  "created": 1734000000,
+  "model": "gw-gemini-model",
+  "choices": [
+    {
+      "index": 0,
+      "message": {
+        "role": "assistant",
+        "content": null,
+        "tool_calls": [
+          {
+            "id": "fc_1",
+            "type": "function",
+            "function": {"name": "get_weather", "arguments": "{\"location\":\"Paris\"}"}
+          }
+        ]
+      },
+      "finish_reason": "tool_calls"
+    }
+  ],
+  "usage": {"prompt_tokens": 12, "completion_tokens": 6, "total_tokens": 18}
+}
diff --git a/testdata/gemini/resp_maxtokens_in.json b/testdata/gemini/resp_maxtokens_in.json
new file mode 100644
index 0000000..6327a5d
--- /dev/null
+++ b/testdata/gemini/resp_maxtokens_in.json
@@ -0,0 +1,13 @@
+{
+  "candidates": [
+    {
+      "content": {
+        "role": "model",
+        "parts": [{"text": "This response got cut off because it"}]
+      },
+      "finishReason": "MAX_TOKENS"
+    }
+  ],
+  "usageMetadata": {"promptTokenCount": 5, "candidatesTokenCount": 50},
+  "responseId": "resp_MAXTOK"
+}
diff --git a/testdata/gemini/resp_maxtokens_want.json b/testdata/gemini/resp_maxtokens_want.json
new file mode 100644
index 0000000..36e29d5
--- /dev/null
+++ b/testdata/gemini/resp_maxtokens_want.json
@@ -0,0 +1,14 @@
+{
+  "id": "chatcmpl-resp_MAXTOK",
+  "object": "chat.completion",
+  "created": 1734000000,
+  "model": "gw-gemini-model",
+  "choices": [
+    {
+      "index": 0,
+      "message": {"role": "assistant", "content": "This response got cut off because it"},
+      "finish_reason": "length"
+    }
+  ],
+  "usage": {"prompt_tokens": 5, "completion_tokens": 50, "total_tokens": 55}
+}
diff --git a/testdata/gemini/resp_safety_in.json b/testdata/gemini/resp_safety_in.json
new file mode 100644
index 0000000..19fa5a6
--- /dev/null
+++ b/testdata/gemini/resp_safety_in.json
@@ -0,0 +1,10 @@
+{
+  "candidates": [
+    {
+      "content": {"role": "model", "parts": []},
+      "finishReason": "SAFETY"
+    }
+  ],
+  "usageMetadata": {"promptTokenCount": 4, "candidatesTokenCount": 0},
+  "responseId": "resp_SAFETY1"
+}
diff --git a/testdata/gemini/resp_safety_want.json b/testdata/gemini/resp_safety_want.json
new file mode 100644
index 0000000..2d4c7b0
--- /dev/null
+++ b/testdata/gemini/resp_safety_want.json
@@ -0,0 +1,14 @@
+{
+  "id": "chatcmpl-resp_SAFETY1",
+  "object": "chat.completion",
+  "created": 1734000000,
+  "model": "gw-gemini-model",
+  "choices": [
+    {
+      "index": 0,
+      "message": {"role": "assistant", "content": null},
+      "finish_reason": "content_filter"
+    }
+  ],
+  "usage": {"prompt_tokens": 4, "completion_tokens": 0, "total_tokens": 4}
+}
diff --git a/testdata/gemini/resp_text_in.json b/testdata/gemini/resp_text_in.json
new file mode 100644
index 0000000..439fa6d
--- /dev/null
+++ b/testdata/gemini/resp_text_in.json
@@ -0,0 +1,18 @@
+{
+  "candidates": [
+    {
+      "content": {
+        "role": "model",
+        "parts": [{"text": "Hello! How can I help you today?"}]
+      },
+      "finishReason": "STOP"
+    }
+  ],
+  "usageMetadata": {
+    "promptTokenCount": 8,
+    "candidatesTokenCount": 9,
+    "totalTokenCount": 17
+  },
+  "modelVersion": "gemini-2.5-pro-001",
+  "responseId": "resp_ABC123"
+}
diff --git a/testdata/gemini/resp_text_want.json b/testdata/gemini/resp_text_want.json
new file mode 100644
index 0000000..4ea17be
--- /dev/null
+++ b/testdata/gemini/resp_text_want.json
@@ -0,0 +1,14 @@
+{
+  "id": "chatcmpl-resp_ABC123",
+  "object": "chat.completion",
+  "created": 1734000000,
+  "model": "gw-gemini-model",
+  "choices": [
+    {
+      "index": 0,
+      "message": {"role": "assistant", "content": "Hello! How can I help you today?"},
+      "finish_reason": "stop"
+    }
+  ],
+  "usage": {"prompt_tokens": 8, "completion_tokens": 9, "total_tokens": 17}
+}
diff --git a/testdata/gemini/stream_text.sse b/testdata/gemini/stream_text.sse
new file mode 100644
index 0000000..006fa9e
--- /dev/null
+++ b/testdata/gemini/stream_text.sse
@@ -0,0 +1,6 @@
+data: {"candidates":[{"content":{"role":"model","parts":[{"text":"Hello"}]}}],"responseId":"resp_STREAM1"}
+
+data: {"candidates":[{"content":{"role":"model","parts":[{"text":", world!"}]}}]}
+
+data: {"candidates":[{"content":{"role":"model","parts":[]},"finishReason":"STOP"}],"usageMetadata":{"promptTokenCount":5,"candidatesTokenCount":3}}
+
diff --git a/testdata/gemini/stream_text_then_funcall.sse b/testdata/gemini/stream_text_then_funcall.sse
new file mode 100644
index 0000000..15de92d
--- /dev/null
+++ b/testdata/gemini/stream_text_then_funcall.sse
@@ -0,0 +1,6 @@
+data: {"candidates":[{"content":{"role":"model","parts":[{"text":"Let me check that."}]}}],"responseId":"resp_STREAM2"}
+
+data: {"candidates":[{"content":{"role":"model","parts":[{"functionCall":{"name":"get_weather","args":{"location":"Berlin"}}}]}}]}
+
+data: {"candidates":[{"content":{"role":"model","parts":[]},"finishReason":"STOP"}],"usageMetadata":{"promptTokenCount":10,"candidatesTokenCount":4}}
+
diff --git a/testdata/gemini/stream_text_then_funcall_usage.json b/testdata/gemini/stream_text_then_funcall_usage.json
new file mode 100644
index 0000000..ad04e9b
--- /dev/null
+++ b/testdata/gemini/stream_text_then_funcall_usage.json
@@ -0,0 +1 @@
+{"prompt": 10, "completion": 4}
diff --git a/testdata/gemini/stream_text_then_funcall_want.json b/testdata/gemini/stream_text_then_funcall_want.json
new file mode 100644
index 0000000..a219082
--- /dev/null
+++ b/testdata/gemini/stream_text_then_funcall_want.json
@@ -0,0 +1,6 @@
+[
+  {"id":"chatcmpl-resp_STREAM2","object":"chat.completion.chunk","created":1734000000,"model":"gw-gemini-model","choices":[{"index":0,"delta":{"role":"assistant"},"finish_reason":null}]},
+  {"id":"chatcmpl-resp_STREAM2","object":"chat.completion.chunk","created":1734000000,"model":"gw-gemini-model","choices":[{"index":0,"delta":{"content":"Let me check that."},"finish_reason":null}]},
+  {"id":"chatcmpl-resp_STREAM2","object":"chat.completion.chunk","created":1734000000,"model":"gw-gemini-model","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"id":"call_0","type":"function","function":{"name":"get_weather","arguments":"{\"location\":\"Berlin\"}"}}]},"finish_reason":null}]},
+  {"id":"chatcmpl-resp_STREAM2","object":"chat.completion.chunk","created":1734000000,"model":"gw-gemini-model","choices":[{"index":0,"delta":{},"finish_reason":"tool_calls"}]}
+]
diff --git a/testdata/gemini/stream_text_usage.json b/testdata/gemini/stream_text_usage.json
new file mode 100644
index 0000000..ee07b50
--- /dev/null
+++ b/testdata/gemini/stream_text_usage.json
@@ -0,0 +1 @@
+{"prompt": 5, "completion": 3}
diff --git a/testdata/gemini/stream_text_want.json b/testdata/gemini/stream_text_want.json
new file mode 100644
index 0000000..22d8c84
--- /dev/null
+++ b/testdata/gemini/stream_text_want.json
@@ -0,0 +1,6 @@
+[
+  {"id":"chatcmpl-resp_STREAM1","object":"chat.completion.chunk","created":1734000000,"model":"gw-gemini-model","choices":[{"index":0,"delta":{"role":"assistant"},"finish_reason":null}]},
+  {"id":"chatcmpl-resp_STREAM1","object":"chat.completion.chunk","created":1734000000,"model":"gw-gemini-model","choices":[{"index":0,"delta":{"content":"Hello"},"finish_reason":null}]},
+  {"id":"chatcmpl-resp_STREAM1","object":"chat.completion.chunk","created":1734000000,"model":"gw-gemini-model","choices":[{"index":0,"delta":{"content":", world!"},"finish_reason":null}]},
+  {"id":"chatcmpl-resp_STREAM1","object":"chat.completion.chunk","created":1734000000,"model":"gw-gemini-model","choices":[{"index":0,"delta":{},"finish_reason":"stop"}]}
+]
diff --git a/translate_gemini.go b/translate_gemini.go
new file mode 100644
index 0000000..fde58eb
--- /dev/null
+++ b/translate_gemini.go
@@ -0,0 +1,836 @@
+package traefikllmgateway
+
+import (
+	"encoding/json"
+	"fmt"
+	"strconv"
+	"strings"
+)
+
+// geminiAPIPrefix is the path segment every Gemini generateContent-family
+// and embedding-family URL is built under, followed by "{model}:{method}".
+const geminiAPIPrefix = "/v1beta/models/"
+
+// isTruthyMap reports whether v decodes to a non-empty JSON object. Per
+// lesson (i), an OpenAI field this translator rejects when present (
+// logit_bias) is only rejected when it actually carries content — a
+// present-but-empty {} or a JSON null must not 400 a request that would
+// otherwise translate cleanly.
+func isTruthyMap(v any) bool {
+	m, ok := v.(map[string]any)
+	return ok && len(m) > 0
+}
+
+// geminiUnsupportedFieldError returns a *translateError for an OpenAI
+// request field Gemini's API has no equivalent for. Deliberately not
+// translate_anthropic.go's unsupportedFieldError: that helper's message
+// hardcodes "not supported for anthropic models", which would misreport the
+// provider on a Gemini request, and this file may only reuse translateError
+// itself, not touch translate_anthropic.go to parameterize it.
+func geminiUnsupportedFieldError(field string) *translateError {
+	return &translateError{msg: fmt.Sprintf("%s is not supported for gemini models", field)}
+}
+
+// geminiSystemTextFromContentParts joins an OpenAI system message's
+// array-form content ([{"type":"text","text":...}, ...]) into plain text,
+// joined with "\n", per lesson (e). Any part that is not {"type":"text"}
+// has no equivalent in Gemini's plain-string systemInstruction text, so it
+// is a *translateError rather than being silently dropped. Deliberately
+// not translate_anthropic.go's systemTextFromContentParts, for the same
+// hardcoded-provider-name reason as geminiUnsupportedFieldError above.
+func geminiSystemTextFromContentParts(parts []any) (string, error) {
+	texts := make([]string, 0, len(parts))
+	for _, part := range parts {
+		pm, ok := part.(map[string]any)
+		if !ok {
+			return "", &translateError{msg: "system message content[] entries must be objects"}
+		}
+		ptype, _ := pm["type"].(string)
+		if ptype != "text" {
+			return "", &translateError{msg: fmt.Sprintf("system message content part type %q is not supported for gemini models", ptype)}
+		}
+		text, _ := pm["text"].(string)
+		texts = append(texts, text)
+	}
+	return strings.Join(texts, "\n"), nil
+}
+
+// geminiInlineDataFromDataURI parses an OpenAI image_url.url as a
+// "data:<media-type-and-params>,<data>" URI and returns Gemini's inlineData
+// object. Per lesson (d), the media type is everything before the first
+// ";" after "data:" — any parameters after it (";charset=utf-8;base64", for
+// instance) are stripped rather than rejected or included. Any URL that is
+// not a data: URI (an http(s) image URL, which Gemini's generateContent API
+// does not fetch) returns a *translateError, per the request mapping
+// table's image row.
+func geminiInlineDataFromDataURI(uri string) (map[string]any, error) {
+	rest, ok := strings.CutPrefix(uri, "data:")
+	if !ok {
+		return nil, &translateError{msg: "image_url must be a data: URI (http(s) image URLs are not supported for gemini models)"}
+	}
+	header, data, ok := strings.Cut(rest, ",")
+	if !ok {
+		return nil, &translateError{msg: "malformed data: URI in image_url"}
+	}
+	mimeType, _, _ := strings.Cut(header, ";")
+	return map[string]any{"mimeType": mimeType, "data": data}, nil
+}
+
+// geminiPartFromOpenAI maps one OpenAI content-array entry ({"type":"text",
+// ...} or {"type":"image_url",...}) to a Gemini content part. An entry of
+// any other type is dropped (nil, nil) rather than erroring — an
+// unrecognized part type is forwarded content this translator does not
+// understand, not necessarily an invalid request.
+func geminiPartFromOpenAI(pm map[string]any) (map[string]any, error) {
+	switch ptype, _ := pm["type"].(string); ptype {
+	case "text":
+		text, _ := pm["text"].(string)
+		return map[string]any{"text": text}, nil
+	case "image_url":
+		iu, _ := pm["image_url"].(map[string]any)
+		url, _ := iu["url"].(string)
+		inline, err := geminiInlineDataFromDataURI(url)
+		if err != nil {
+			return nil, err
+		}
+		return map[string]any{"inlineData": inline}, nil
+	default:
+		return nil, nil
+	}
+}
+
+// geminiFunctionCallPartFromOpenAI maps one assistant tool_calls[] entry
+// ({"id","type":"function","function":{"name","arguments"}}) to a Gemini
+// {"functionCall":{"name","args"}} content part, parsing the OpenAI
+// arguments JSON string into Gemini's structured "args" object. It also
+// returns the entry's id and name, so the caller can record them in the
+// pending tool_call id map a later role:"tool" message's functionResponse
+// needs (Gemini's functionResponse has no id field to correlate by; this
+// translator resolves "name" instead, per the request mapping table's
+// role:tool row).
+func geminiFunctionCallPartFromOpenAI(tc any) (part map[string]any, id, name string, err error) {
+	tm, ok := tc.(map[string]any)
+	if !ok {
+		return nil, "", "", &translateError{msg: "assistant tool_calls[] entry is malformed"}
+	}
+	id, _ = tm["id"].(string)
+	fn, _ := tm["function"].(map[string]any)
+	name, _ = fn["name"].(string)
+	argsStr, _ := fn["arguments"].(string)
+
+	args := any(map[string]any{})
+	if argsStr != "" {
+		if err := json.Unmarshal([]byte(argsStr), &args); err != nil {
+			return nil, "", "", &translateError{msg: "assistant tool_calls[].function.arguments is not valid JSON"}
+		}
+	}
+	return map[string]any{"functionCall": map[string]any{"name": name, "args": args}}, id, name, nil
+}
+
+// geminiMessageFromOpenAI maps one OpenAI user/assistant message to a
+// Gemini {"role","parts"} contents entry. geminiRole is "user" or "model"
+// (assistant's Gemini-side role name, per the request mapping table).
+// toolNameByID accumulates id -> name for every tool_calls[] entry seen on
+// an assistant message, so a later role:"tool" message can resolve its
+// functionResponse's "name".
+func geminiMessageFromOpenAI(msg map[string]any, geminiRole string, toolNameByID map[string]string) (map[string]any, error) {
+	var parts []any
+	switch c := msg["content"].(type) {
+	case string:
+		if c != "" {
+			parts = append(parts, map[string]any{"text": c})
+		}
+	case []any:
+		for _, p := range c {
+			pm, ok := p.(map[string]any)
+			if !ok {
+				continue
+			}
+			part, err := geminiPartFromOpenAI(pm)
+			if err != nil {
+				return nil, err
+			}
+			if part != nil {
+				parts = append(parts, part)
+			}
+		}
+	}
+
+	if geminiRole == "model" {
+		toolCalls, _ := msg["tool_calls"].([]any)
+		for _, tc := range toolCalls {
+			part, id, name, err := geminiFunctionCallPartFromOpenAI(tc)
+			if err != nil {
+				return nil, err
+			}
+			if id != "" {
+				toolNameByID[id] = name
+			}
+			parts = append(parts, part)
+		}
+	}
+
+	if parts == nil {
+		parts = []any{}
+	}
+	return map[string]any{"role": geminiRole, "parts": parts}, nil
+}
+
+// geminiFunctionResponsePart maps one OpenAI role:"tool" message to a
+// single Gemini {"functionResponse":{"name","response":{"content"}}}
+// content part, per the request mapping table's role:tool row. name is
+// looked up from toolNameByID — the id -> name map built while converting
+// the assistant message that produced this tool_call_id — and an unknown
+// id is a *translateError rather than a silently empty name.
+func geminiFunctionResponsePart(msg map[string]any, toolNameByID map[string]string) (map[string]any, error) {
+	toolCallID, _ := msg["tool_call_id"].(string)
+	name, ok := toolNameByID[toolCallID]
+	if !ok {
+		return nil, &translateError{msg: fmt.Sprintf("role:tool message references unknown tool_call_id %q", toolCallID)}
+	}
+	return map[string]any{
+		"functionResponse": map[string]any{
+			"name":     name,
+			"response": map[string]any{"content": msg["content"]},
+		},
+	}, nil
+}
+
+// geminiToolsFromOpenAI maps OpenAI's tools[].function{name, description,
+// parameters} to Gemini's tools:[{functionDeclarations:[{name,description,
+// parameters}]}], per the request mapping table's tools row.
+func geminiToolsFromOpenAI(toolsRaw []any) []any {
+	decls := make([]any, 0, len(toolsRaw))
+	for _, t := range toolsRaw {
+		tm, ok := t.(map[string]any)
+		if !ok {
+			continue
+		}
+		fn, _ := tm["function"].(map[string]any)
+		entry := map[string]any{"name": fn["name"]}
+		if desc, ok := fn["description"]; ok {
+			entry["description"] = desc
+		}
+		if params, ok := fn["parameters"]; ok {
+			entry["parameters"] = params
+		}
+		decls = append(decls, entry)
+	}
+	return []any{map[string]any{"functionDeclarations": decls}}
+}
+
+// geminiToolConfigFromOpenAI maps OpenAI's "tool_choice" to Gemini's
+// toolConfig.functionCallingConfig object, per the request mapping table's
+// tool_choice rows. Unlike Anthropic, Gemini has a real NONE mode, so
+// "none" needs no special "drop tools" handling. An unrecognized shape
+// returns nil — the caller omits "toolConfig" entirely rather than sending
+// a malformed one.
+func geminiToolConfigFromOpenAI(tc any) map[string]any {
+	switch t := tc.(type) {
+	case string:
+		switch t {
+		case "none":
+			return map[string]any{"functionCallingConfig": map[string]any{"mode": "NONE"}}
+		case "auto":
+			return map[string]any{"functionCallingConfig": map[string]any{"mode": "AUTO"}}
+		case "required":
+			return map[string]any{"functionCallingConfig": map[string]any{"mode": "ANY"}}
+		}
+	case map[string]any:
+		if fn, ok := t["function"].(map[string]any); ok {
+			if name, ok := fn["name"].(string); ok {
+				return map[string]any{"functionCallingConfig": map[string]any{
+					"mode":                 "ANY",
+					"allowedFunctionNames": []any{name},
+				}}
+			}
+		}
+	}
+	return nil
+}
+
+// geminiRequestFromOpenAI maps an OpenAI chat-completion request body to a
+// Gemini generateContent/streamGenerateContent request body, per this
+// file's request mapping table. It returns a *translateError — never
+// wrapped — for a field Gemini has no equivalent for (n>1, a truthy
+// logit_bias) or a content part this translator cannot map (a non-data-URI
+// image URL, a malformed tool_calls argument string, a role:"tool" message
+// whose tool_call_id was never seen on a preceding assistant message).
+// Consecutive role:"tool" messages are merged into one Gemini {"role":
+// "user"} contents entry carrying all of their functionResponse parts, per
+// lesson (f). The request's "model" field is deliberately never read here —
+// Gemini's body carries no model field at all; the adapter reads it
+// separately to build the URL.
+func geminiRequestFromOpenAI(req map[string]any) (map[string]any, error) {
+	if v, ok := req["n"]; ok {
+		if n, ok2 := toFloat64(v); ok2 && n > 1 {
+			return nil, geminiUnsupportedFieldError("n")
+		}
+	}
+	if v, ok := req["logit_bias"]; ok && isTruthyMap(v) {
+		return nil, geminiUnsupportedFieldError("logit_bias")
+	}
+
+	out := map[string]any{}
+
+	msgsRaw, _ := req["messages"].([]any)
+	var systemParts []string
+	contents := make([]any, 0, len(msgsRaw))
+	toolNameByID := map[string]string{}
+
+	// pendingFuncResponses accumulates consecutive role:"tool" messages'
+	// functionResponse parts. Gemini, like Anthropic, expects the N
+	// tool-result messages following a parallel-tool-call assistant turn
+	// to collapse into one contents entry, not N separate ones — see
+	// geminiFunctionResponsePart's doc comment for the id resolution this
+	// depends on.
+	var pendingFuncResponses []any
+	flushPending := func() {
+		if len(pendingFuncResponses) == 0 {
+			return
+		}
+		contents = append(contents, map[string]any{"role": "user", "parts": pendingFuncResponses})
+		pendingFuncResponses = nil
+	}
+
+	for _, m := range msgsRaw {
+		msg, ok := m.(map[string]any)
+		if !ok {
+			continue
+		}
+		switch role, _ := msg["role"].(string); role {
+		case "system":
+			switch c := msg["content"].(type) {
+			case string:
+				systemParts = append(systemParts, c)
+			case []any:
+				text, err := geminiSystemTextFromContentParts(c)
+				if err != nil {
+					return nil, err
+				}
+				systemParts = append(systemParts, text)
+			}
+		case "user":
+			flushPending()
+			gm, err := geminiMessageFromOpenAI(msg, "user", toolNameByID)
+			if err != nil {
+				return nil, err
+			}
+			contents = append(contents, gm)
+		case "assistant":
+			flushPending()
+			gm, err := geminiMessageFromOpenAI(msg, "model", toolNameByID)
+			if err != nil {
+				return nil, err
+			}
+			contents = append(contents, gm)
+		case "tool":
+			part, err := geminiFunctionResponsePart(msg, toolNameByID)
+			if err != nil {
+				return nil, err
+			}
+			pendingFuncResponses = append(pendingFuncResponses, part)
+		}
+	}
+	flushPending()
+
+	if len(systemParts) > 0 {
+		out["systemInstruction"] = map[string]any{"parts": []any{map[string]any{"text": strings.Join(systemParts, "\n\n")}}}
+	}
+	out["contents"] = contents
+
+	genConfig := map[string]any{}
+	var maxTokens int64
+	var hasMaxTokens bool
+	if v, ok := req["max_completion_tokens"]; ok {
+		if n, ok2 := toInt64(v); ok2 {
+			maxTokens, hasMaxTokens = n, true
+		}
+	} else if v, ok := req["max_tokens"]; ok {
+		if n, ok2 := toInt64(v); ok2 {
+			maxTokens, hasMaxTokens = n, true
+		}
+	}
+	if hasMaxTokens {
+		genConfig["maxOutputTokens"] = maxTokens
+	}
+	if v, ok := req["temperature"]; ok {
+		genConfig["temperature"] = v
+	}
+	if v, ok := req["top_p"]; ok {
+		genConfig["topP"] = v
+	}
+	if v, ok := req["stop"]; ok {
+		seqs, err := stopSequencesFrom(v)
+		if err != nil {
+			return nil, err
+		}
+		if len(seqs) > 0 {
+			genConfig["stopSequences"] = seqs
+		}
+	}
+	if len(genConfig) > 0 {
+		out["generationConfig"] = genConfig
+	}
+
+	if toolsRaw, ok := req["tools"].([]any); ok && len(toolsRaw) > 0 {
+		out["tools"] = geminiToolsFromOpenAI(toolsRaw)
+	}
+	if tc, ok := req["tool_choice"]; ok {
+		if toolConfig := geminiToolConfigFromOpenAI(tc); toolConfig != nil {
+			out["toolConfig"] = toolConfig
+		}
+	}
+
+	return out, nil
+}
+
+// geminiFunctionCall is the subset of a Gemini "functionCall" content
+// part's JSON this translator reads. ID is optional on Gemini's side (the
+// "Unique identifier of the function call" field, populated only when the
+// model chooses to set it); openAIResponseFromGemini and geminiStreamState
+// both fall back to a synthesized id when it is empty, since OpenAI's
+// tool_calls[].id is required.
+type geminiFunctionCall struct {
+	ID   string          `json:"id,omitempty"`
+	Name string          `json:"name"`
+	Args json.RawMessage `json:"args"`
+}
+
+// geminiPart is the subset of one Gemini response content part's JSON this
+// translator reads: a text part carries Text, a function-call part carries
+// FunctionCall.
+type geminiPart struct {
+	FunctionCall *geminiFunctionCall `json:"functionCall,omitempty"`
+	Text         string              `json:"text,omitempty"`
+}
+
+// geminiContent is a Gemini response candidate's "content" object.
+type geminiContent struct {
+	Role  string       `json:"role"`
+	Parts []geminiPart `json:"parts"`
+}
+
+// geminiCandidate is one entry of a Gemini response's "candidates" array.
+type geminiCandidate struct {
+	FinishReason string        `json:"finishReason"`
+	Content      geminiContent `json:"content"`
+}
+
+// geminiUsageMetadata is the "usageMetadata" object on a Gemini
+// GenerateContentResponse, non-streaming or streaming alike.
+type geminiUsageMetadata struct {
+	PromptTokenCount     int64 `json:"promptTokenCount"`
+	CandidatesTokenCount int64 `json:"candidatesTokenCount"`
+}
+
+// geminiGenerateContentResponse is the subset of a Gemini
+// GenerateContentResponse this translator reads — the shape of both a
+// non-streaming response body and every streaming chunk (each chunk is a
+// full GenerateContentResponse increment, per the stream mapping table).
+// UsageMetadata is a pointer so the translator can tell "this chunk carries
+// no usage" (nil) apart from "this chunk reports zero tokens" (non-nil,
+// zero fields) — the stream mapping table's "last usageMetadata wins" rule
+// depends on that distinction.
+type geminiGenerateContentResponse struct {
+	ResponseID    string               `json:"responseId"`
+	UsageMetadata *geminiUsageMetadata `json:"usageMetadata"`
+	Candidates    []geminiCandidate    `json:"candidates"`
+}
+
+// geminiFinishReason maps a Gemini finishReason to an OpenAI finish_reason,
+// per the response mapping table: MAX_TOKENS -> "length"; SAFETY,
+// RECITATION, BLOCKLIST, PROHIBITED_CONTENT, and SPII -> "content_filter";
+// STOP and any unrecognized value -> "stop", unless hasFunctionCall is set,
+// in which case that default branch reports "tool_calls" instead — Gemini
+// has no distinct finishReason value for a function-calling turn the way
+// Anthropic's "tool_use" stop_reason does, so the caller's own knowledge of
+// whether the candidate carried a functionCall part is what this
+// translator uses instead.
+func geminiFinishReason(reason string, hasFunctionCall bool) string {
+	switch reason {
+	case "MAX_TOKENS":
+		return "length"
+	case "SAFETY", "RECITATION", "BLOCKLIST", "PROHIBITED_CONTENT", "SPII":
+		return "content_filter"
+	default:
+		if hasFunctionCall {
+			return "tool_calls"
+		}
+		return "stop"
+	}
+}
+
+// geminiToolCallID returns a synthesized OpenAI tool_calls[].id for a
+// Gemini functionCall part that did not carry its own optional "id" field:
+// "call_" followed by ordinal, the part's 0-based position among tool
+// calls seen so far in this response or stream (per lesson (a), a
+// contiguous per-response/per-stream counter, never a raw content-part
+// index).
+func geminiToolCallID(ordinal int) string {
+	return "call_" + strconv.Itoa(ordinal)
+}
+
+// openAIResponseFromGemini maps a non-streaming Gemini
+// GenerateContentResponse body to an OpenAI chat.completion response, per
+// this file's response mapping table. Only candidates[0] is read, matching
+// the table. model and created are supplied by the caller — the gateway-
+// facing model id and a response timestamp, since Gemini's own response
+// carries neither.
+func openAIResponseFromGemini(body []byte, model string, created int64) (map[string]any, usage, error) {
+	var resp geminiGenerateContentResponse
+	if err := json.Unmarshal(body, &resp); err != nil {
+		return nil, usage{}, fmt.Errorf("%w: decode gemini response: %w", errUpstream, err)
+	}
+
+	var textParts []string
+	var toolCalls []any
+	var finishReasonRaw string
+	if len(resp.Candidates) > 0 {
+		c := resp.Candidates[0]
+		finishReasonRaw = c.FinishReason
+		ordinal := 0
+		for _, p := range c.Content.Parts {
+			switch {
+			case p.FunctionCall != nil:
+				args, err := reMarshalToolInput(p.FunctionCall.Args)
+				if err != nil {
+					return nil, usage{}, err
+				}
+				id := p.FunctionCall.ID
+				if id == "" {
+					id = geminiToolCallID(ordinal)
+				}
+				toolCalls = append(toolCalls, map[string]any{
+					"id":   id,
+					"type": "function",
+					"function": map[string]any{
+						"name":      p.FunctionCall.Name,
+						"arguments": args,
+					},
+				})
+				ordinal++
+			case p.Text != "":
+				textParts = append(textParts, p.Text)
+			}
+		}
+	}
+
+	message := map[string]any{"role": "assistant"}
+	if len(textParts) > 0 {
+		message["content"] = strings.Join(textParts, "")
+	} else {
+		message["content"] = nil
+	}
+	if len(toolCalls) > 0 {
+		message["tool_calls"] = toolCalls
+	}
+
+	u := usage{prompt: resp.UsageMetadata.safePrompt(), completion: resp.UsageMetadata.safeCompletion()}
+
+	out := map[string]any{
+		"id":      chatCompletionIDPrefix + resp.ResponseID,
+		"object":  "chat.completion",
+		"created": created,
+		"model":   model,
+		"choices": []any{map[string]any{
+			"index":         0,
+			"message":       message,
+			"finish_reason": geminiFinishReason(finishReasonRaw, len(toolCalls) > 0),
+		}},
+		"usage": map[string]any{
+			"prompt_tokens":     u.prompt,
+			"completion_tokens": u.completion,
+			"total_tokens":      u.total(),
+		},
+	}
+	return out, u, nil
+}
+
+// safePrompt returns m.PromptTokenCount, or 0 when m is nil (no
+// usageMetadata was present on the response at all).
+func (m *geminiUsageMetadata) safePrompt() int64 {
+	if m == nil {
+		return 0
+	}
+	return m.PromptTokenCount
+}
+
+// safeCompletion returns m.CandidatesTokenCount, or 0 when m is nil.
+func (m *geminiUsageMetadata) safeCompletion() int64 {
+	if m == nil {
+		return 0
+	}
+	return m.CandidatesTokenCount
+}
+
+// geminiStreamState is one streaming chat completion's translation state:
+// the OpenAI-shaped chunk envelope fields (id, captured from the first
+// chunk that carries a non-empty "responseId"; model and created, fixed at
+// construction), the running usage total, and the per-stream tool-call
+// ordinal counter lesson (a) requires. A caller constructs one per stream
+// and calls translate once per upstream sseEvent, in order.
+type geminiStreamState struct {
+	id      string
+	model   string
+	u       usage
+	created int64
+	// nextToolOrdinal is the per-stream contiguous tool-call counter lesson
+	// (a) requires.
+	nextToolOrdinal int
+	startedRole     bool
+	// sawFunctionCall latches true the first time any chunk in this stream
+	// carries a functionCall part, and stays true for the rest of the
+	// stream. Gemini's finishReason is "STOP" whether the turn ended in
+	// text or a function call — unlike Anthropic's distinct "tool_use"
+	// stop_reason — and the finishReason usually arrives on a later,
+	// parts-less chunk of its own. A per-call local flag would forget a
+	// functionCall part seen on an earlier chunk by the time the finish
+	// chunk arrives, so this must be state on the struct, not a translate
+	// local.
+	sawFunctionCall bool
+}
+
+// newGeminiStreamState returns a geminiStreamState for one streaming chat
+// completion. model is the gateway-facing model id and created the
+// response timestamp every emitted chunk carries.
+func newGeminiStreamState(model string, created int64) *geminiStreamState {
+	return &geminiStreamState{model: model, created: created}
+}
+
+// usage returns the prompt/completion token counts captured so far, from
+// whichever chunk's usageMetadata was translated most recently.
+func (st *geminiStreamState) usage() usage {
+	return st.u
+}
+
+// chunk builds one OpenAI-shaped "chat.completion.chunk" payload carrying
+// delta, and finishReason as its choices[0].finish_reason — nil for the
+// JSON null every non-terminal chunk carries.
+func (st *geminiStreamState) chunk(delta map[string]any, finishReason *string) []byte {
+	var fr any
+	if finishReason != nil {
+		fr = *finishReason
+	}
+	b, err := json.Marshal(map[string]any{
+		"id":      chatCompletionIDPrefix + st.id,
+		"object":  "chat.completion.chunk",
+		"created": st.created,
+		"model":   st.model,
+		"choices": []any{map[string]any{
+			"index":         0,
+			"delta":         delta,
+			"finish_reason": fr,
+		}},
+	})
+	if err != nil {
+		// delta is always built from this file's own map[string]any/string/
+		// int64 literals — never a value that can fail to marshal.
+		panic(fmt.Sprintf("llmgateway: gemini stream chunk failed to marshal: %v", err))
+	}
+	return b
+}
+
+// translate maps one upstream Gemini SSE event to zero or more
+// OpenAI-shaped "chat.completion.chunk" payloads, per this file's stream
+// mapping table. Every event's data is a full GenerateContentResponse
+// increment — there is no "event:" type field to switch on the way
+// Anthropic's stream has, so translate always attempts to decode ev.data
+// as one. A malformed data payload returns an error wrapping errUpstream
+// rather than being silently skipped, per lesson (c):
+//
+//   - The first chunk translate ever emits (across the whole stream) is
+//     always a {role:"assistant"} delta.
+//   - A text part becomes a {content: text} delta.
+//   - A functionCall part becomes a delta.tool_calls[0] chunk carrying the
+//     0-based per-stream tool ordinal as its index (lesson (a)), the
+//     part's own "id" when Gemini set one or else a synthesized one, and
+//     the re-marshaled "args" as function.arguments.
+//   - A non-empty finishReason becomes a finish_reason chunk, mapped per
+//     geminiFinishReason — using st.sawFunctionCall, latched true the first
+//     time any chunk in this stream carried a functionCall part, since the
+//     finish chunk is typically its own later, parts-less chunk with no
+//     functionCall of its own to inspect.
+//   - Any usageMetadata present on the chunk overwrites the running usage
+//     total — "last usageMetadata wins", per the stream mapping table.
+func (st *geminiStreamState) translate(ev sseEvent) ([][]byte, error) {
+	var resp geminiGenerateContentResponse
+	if err := json.Unmarshal(ev.data, &resp); err != nil {
+		return nil, fmt.Errorf("%w: decode gemini stream chunk: %w", errUpstream, err)
+	}
+	if resp.ResponseID != "" {
+		st.id = resp.ResponseID
+	}
+	if resp.UsageMetadata != nil {
+		st.u = usage{prompt: resp.UsageMetadata.PromptTokenCount, completion: resp.UsageMetadata.CandidatesTokenCount}
+	}
+
+	var chunks [][]byte
+	if !st.startedRole {
+		st.startedRole = true
+		chunks = append(chunks, st.chunk(map[string]any{"role": "assistant"}, nil))
+	}
+
+	if len(resp.Candidates) == 0 {
+		return chunks, nil
+	}
+	c := resp.Candidates[0]
+
+	for _, p := range c.Content.Parts {
+		switch {
+		case p.FunctionCall != nil:
+			st.sawFunctionCall = true
+			args, err := reMarshalToolInput(p.FunctionCall.Args)
+			if err != nil {
+				return chunks, err
+			}
+			id := p.FunctionCall.ID
+			if id == "" {
+				id = geminiToolCallID(st.nextToolOrdinal)
+			}
+			delta := map[string]any{"tool_calls": []any{map[string]any{
+				"index": st.nextToolOrdinal,
+				"id":    id,
+				"type":  "function",
+				"function": map[string]any{
+					"name":      p.FunctionCall.Name,
+					"arguments": args,
+				},
+			}}}
+			st.nextToolOrdinal++
+			chunks = append(chunks, st.chunk(delta, nil))
+		case p.Text != "":
+			chunks = append(chunks, st.chunk(map[string]any{"content": p.Text}, nil))
+		}
+	}
+
+	if c.FinishReason != "" {
+		reason := geminiFinishReason(c.FinishReason, st.sawFunctionCall)
+		chunks = append(chunks, st.chunk(map[string]any{}, &reason))
+	}
+
+	return chunks, nil
+}
+
+// geminiEmbeddingValues is the "embedding" or one "embeddings[]" entry's
+// JSON shape on a Gemini embedContent/batchEmbedContents response.
+type geminiEmbeddingValues struct {
+	Values []float64 `json:"values"`
+}
+
+// geminiEmbedContentResponse is the response body shape for Gemini's
+// :embedContent endpoint (the OpenAI single-string-input case).
+type geminiEmbedContentResponse struct {
+	Embedding geminiEmbeddingValues `json:"embedding"`
+}
+
+// geminiBatchEmbedContentsResponse is the response body shape for Gemini's
+// :batchEmbedContents endpoint (the OpenAI array-input case).
+type geminiBatchEmbedContentsResponse struct {
+	Embeddings []geminiEmbeddingValues `json:"embeddings"`
+}
+
+// geminiEmbedContentRequest builds the :embedContent request body for a
+// single OpenAI embeddings input string, per the embeddings mapping table.
+func geminiEmbedContentRequest(text string) map[string]any {
+	return map[string]any{
+		"content": map[string]any{"parts": []any{map[string]any{"text": text}}},
+	}
+}
+
+// geminiBatchEmbedContentsRequest builds the :batchEmbedContents request
+// body for an array of OpenAI embeddings input strings, per the embeddings
+// mapping table. model is the bare (no "models/" prefix) Gemini model id;
+// each request entry's own "model" field needs the prefix restored, since
+// Gemini's per-request model reference is always the full resource name.
+func geminiBatchEmbedContentsRequest(model string, texts []string) map[string]any {
+	requests := make([]any, 0, len(texts))
+	for _, text := range texts {
+		requests = append(requests, map[string]any{
+			"model":   "models/" + model,
+			"content": map[string]any{"parts": []any{map[string]any{"text": text}}},
+		})
+	}
+	return map[string]any{"requests": requests}
+}
+
+// geminiEmbeddingInputTexts extracts the OpenAI "input" field (a string or
+// an array of strings) as a slice of one-or-more strings, and reports
+// whether the original input was a single string — selecting the
+// :embedContent endpoint — rather than an array, which selects
+// :batchEmbedContents.
+func geminiEmbeddingInputTexts(input any) (texts []string, single bool, err error) {
+	switch v := input.(type) {
+	case string:
+		return []string{v}, true, nil
+	case []any:
+		texts = make([]string, 0, len(v))
+		for _, e := range v {
+			s, ok := e.(string)
+			if !ok {
+				return nil, false, &translateError{msg: "input[] entries must be strings"}
+			}
+			texts = append(texts, s)
+		}
+		return texts, false, nil
+	default:
+		return nil, false, &translateError{msg: "input must be a string or an array of strings"}
+	}
+}
+
+// estimatedEmbeddingTokens estimates a token count for texts the way
+// Gemini's embedding endpoints do not: neither :embedContent nor
+// :batchEmbedContents reports any usage at all, so this translator
+// estimates ceil(total character count / 4) summed across every input
+// string, per the embeddings mapping table.
+func estimatedEmbeddingTokens(texts []string) int64 {
+	var chars int64
+	for _, t := range texts {
+		chars += int64(len(t))
+	}
+	return (chars + 3) / 4
+}
+
+// openAIEmbeddingResponseFromGemini maps a Gemini embedContent or
+// batchEmbedContents response body to an OpenAI embeddings list response,
+// per the embeddings mapping table. single selects which of Gemini's two
+// response shapes body decodes as. model is the gateway-facing model id and
+// estimatedTokens the caller's already-computed estimatedEmbeddingTokens
+// result — both usage.prompt_tokens and usage.total_tokens report it, and
+// the returned usage is marked estimated, since Gemini reported none.
+func openAIEmbeddingResponseFromGemini(body []byte, single bool, model string, estimatedTokens int64) (map[string]any, usage, error) {
+	var vectors [][]float64
+	if single {
+		var resp geminiEmbedContentResponse
+		if err := json.Unmarshal(body, &resp); err != nil {
+			return nil, usage{}, fmt.Errorf("%w: decode gemini embedContent response: %w", errUpstream, err)
+		}
+		vectors = [][]float64{resp.Embedding.Values}
+	} else {
+		var resp geminiBatchEmbedContentsResponse
+		if err := json.Unmarshal(body, &resp); err != nil {
+			return nil, usage{}, fmt.Errorf("%w: decode gemini batchEmbedContents response: %w", errUpstream, err)
+		}
+		vectors = make([][]float64, len(resp.Embeddings))
+		for i, e := range resp.Embeddings {
+			vectors[i] = e.Values
+		}
+	}
+
+	data := make([]any, len(vectors))
+	for i, v := range vectors {
+		data[i] = map[string]any{"object": "embedding", "index": i, "embedding": v}
+	}
+
+	u := usage{prompt: estimatedTokens, estimated: true}
+	out := map[string]any{
+		"object": "list",
+		"data":   data,
+		"model":  model,
+		"usage": map[string]any{
+			"prompt_tokens": u.prompt,
+			"total_tokens":  u.prompt,
+		},
+	}
+	return out, u, nil
+}
diff --git a/translate_gemini_test.go b/translate_gemini_test.go
new file mode 100644
index 0000000..4df654b
--- /dev/null
+++ b/translate_gemini_test.go
@@ -0,0 +1,442 @@
+package traefikllmgateway
+
+import (
+	"encoding/json"
+	"errors"
+	"os"
+	"path/filepath"
+	"reflect"
+	"sort"
+	"strings"
+	"testing"
+)
+
+// geminiGoldenTestdataDir is where every fixture this file walks lives.
+const geminiGoldenTestdataDir = "testdata/gemini"
+
+// geminiGoldenGatewayModel and geminiGoldenCreated are the fixed model/
+// created values every response- and stream-translation golden fixture's
+// "_want.json" bakes in — openAIResponseFromGemini and
+// newGeminiStreamState both take these as caller-supplied parameters
+// (Gemini's own wire formats carry neither), so the golden tests must
+// supply the same fixed values the "_want.json" files were written
+// against.
+const (
+	geminiGoldenGatewayModel = "gw-gemini-model"
+	geminiGoldenCreated      = int64(1734000000)
+)
+
+// geminiReadGoldenJSON reads and json.Unmarshals name (relative to
+// testdata/gemini) into a fresh map[string]any.
+func geminiReadGoldenJSON(t *testing.T, name string) map[string]any {
+	t.Helper()
+	b, err := os.ReadFile(filepath.Join(geminiGoldenTestdataDir, name))
+	if err != nil {
+		t.Fatalf("read %s: %v", name, err)
+	}
+	var m map[string]any
+	if err := json.Unmarshal(b, &m); err != nil {
+		t.Fatalf("decode %s: %v", name, err)
+	}
+	return m
+}
+
+// geminiAssertTranslateError asserts err is a *translateError matching want.
+func geminiAssertTranslateError(t *testing.T, err error, want wantErrFixture) {
+	t.Helper()
+	if err == nil {
+		t.Fatalf("want a *translateError containing %q, got nil", want.ErrContains)
+	}
+	var terr *translateError
+	if !errors.As(err, &terr) {
+		t.Fatalf("err = %v (%T), want *translateError", err, err)
+	}
+	if !strings.Contains(terr.Error(), want.ErrContains) {
+		t.Errorf("error message = %q, want it to contain %q", terr.Error(), want.ErrContains)
+	}
+	if terr.notSupported != want.NotSupported {
+		t.Errorf("notSupported = %v, want %v", terr.notSupported, want.NotSupported)
+	}
+}
+
+// TestGeminiRequestGolden walks every "chat_*_in.json" fixture under
+// testdata/gemini, translating it via geminiRequestFromOpenAI, and asserts
+// the result against the sibling "_want.json" fixture (a successful
+// translation) or "_wanterr.json" fixture (a *translateError), whichever is
+// present — exactly one of the two exists per fixture.
+func TestGeminiRequestGolden(t *testing.T) {
+	entries, err := os.ReadDir(geminiGoldenTestdataDir)
+	if err != nil {
+		t.Fatalf("read testdata dir: %v", err)
+	}
+
+	var names []string
+	for _, e := range entries {
+		if n, ok := strings.CutSuffix(e.Name(), "_in.json"); ok && strings.HasPrefix(n, "chat_") {
+			names = append(names, n)
+		}
+	}
+	sort.Strings(names)
+	if len(names) == 0 {
+		t.Fatalf("no chat_*_in.json fixtures found under %s", geminiGoldenTestdataDir)
+	}
+
+	for _, name := range names {
+		t.Run(name, func(t *testing.T) {
+			in := geminiReadGoldenJSON(t, name+"_in.json")
+
+			got, err := geminiRequestFromOpenAI(in)
+
+			wantErrPath := filepath.Join(geminiGoldenTestdataDir, name+"_wanterr.json")
+			if _, statErr := os.Stat(wantErrPath); statErr == nil {
+				var want wantErrFixture
+				b, rerr := os.ReadFile(wantErrPath)
+				if rerr != nil {
+					t.Fatalf("read %s: %v", wantErrPath, rerr)
+				}
+				if uerr := json.Unmarshal(b, &want); uerr != nil {
+					t.Fatalf("decode %s: %v", wantErrPath, uerr)
+				}
+				geminiAssertTranslateError(t, err, want)
+				return
+			}
+
+			if err != nil {
+				t.Fatalf("geminiRequestFromOpenAI: %v", err)
+			}
+			want := geminiReadGoldenJSON(t, name+"_want.json")
+			gotN := normalizeJSON(t, got)
+			wantN := normalizeJSON(t, want)
+			if !reflect.DeepEqual(gotN, wantN) {
+				gotB, _ := json.MarshalIndent(gotN, "", "  ")
+				wantB, _ := json.MarshalIndent(wantN, "", "  ")
+				t.Errorf("geminiRequestFromOpenAI(%s) mismatch:\ngot:\n%s\nwant:\n%s", name, gotB, wantB)
+			}
+		})
+	}
+}
+
+// TestGeminiResponseGolden walks every "resp_*_in.json" fixture,
+// translating it via openAIResponseFromGemini with the fixed
+// geminiGoldenGatewayModel/geminiGoldenCreated, and asserts the result
+// against the sibling "_want.json" fixture.
+func TestGeminiResponseGolden(t *testing.T) {
+	entries, err := os.ReadDir(geminiGoldenTestdataDir)
+	if err != nil {
+		t.Fatalf("read testdata dir: %v", err)
+	}
+
+	var names []string
+	for _, e := range entries {
+		if n, ok := strings.CutSuffix(e.Name(), "_in.json"); ok && strings.HasPrefix(n, "resp_") {
+			names = append(names, n)
+		}
+	}
+	sort.Strings(names)
+	if len(names) == 0 {
+		t.Fatalf("no resp_*_in.json fixtures found under %s", geminiGoldenTestdataDir)
+	}
+
+	for _, name := range names {
+		t.Run(name, func(t *testing.T) {
+			body, err := os.ReadFile(filepath.Join(geminiGoldenTestdataDir, name+"_in.json"))
+			if err != nil {
+				t.Fatalf("read %s_in.json: %v", name, err)
+			}
+
+			got, _, err := openAIResponseFromGemini(body, geminiGoldenGatewayModel, geminiGoldenCreated)
+			if err != nil {
+				t.Fatalf("openAIResponseFromGemini: %v", err)
+			}
+
+			want := geminiReadGoldenJSON(t, name+"_want.json")
+			gotN := normalizeJSON(t, got)
+			wantN := normalizeJSON(t, want)
+			if !reflect.DeepEqual(gotN, wantN) {
+				gotB, _ := json.MarshalIndent(gotN, "", "  ")
+				wantB, _ := json.MarshalIndent(wantN, "", "  ")
+				t.Errorf("openAIResponseFromGemini(%s) mismatch:\ngot:\n%s\nwant:\n%s", name, gotB, wantB)
+			}
+		})
+	}
+}
+
+// geminiRunStreamGolden translates every event in
+// testdata/gemini/{name}.sse through a fresh geminiStreamState, in order,
+// and asserts the flattened list of emitted chunks against
+// "{name}_want.json" and the final usage() against "{name}_usage.json".
+func geminiRunStreamGolden(t *testing.T, name string) {
+	t.Helper()
+
+	f, err := os.Open(filepath.Join(geminiGoldenTestdataDir, name+".sse"))
+	if err != nil {
+		t.Fatalf("open %s.sse: %v", name, err)
+	}
+	defer f.Close() //nolint:errcheck
+
+	st := newGeminiStreamState(geminiGoldenGatewayModel, geminiGoldenCreated)
+	var got []any
+	terminalErr := readSSE(f, func(ev sseEvent) error {
+		chunks, terr := st.translate(ev)
+		for _, c := range chunks {
+			var v any
+			if uerr := json.Unmarshal(c, &v); uerr != nil {
+				t.Fatalf("translate(%s) produced invalid JSON: %v (%s)", ev.event, uerr, c)
+			}
+			got = append(got, v)
+		}
+		return terr
+	})
+	if terminalErr != nil {
+		t.Fatalf("readSSE/translate(%s): %v", name, terminalErr)
+	}
+
+	wantBytes, err := os.ReadFile(filepath.Join(geminiGoldenTestdataDir, name+"_want.json"))
+	if err != nil {
+		t.Fatalf("read %s_want.json: %v", name, err)
+	}
+	var want []any
+	if uerr := json.Unmarshal(wantBytes, &want); uerr != nil {
+		t.Fatalf("decode %s_want.json: %v", name, uerr)
+	}
+
+	if !reflect.DeepEqual(got, want) {
+		gotB, _ := json.MarshalIndent(got, "", "  ")
+		wantB, _ := json.MarshalIndent(want, "", "  ")
+		t.Errorf("translate(%s) chunk list mismatch:\ngot:\n%s\nwant:\n%s", name, gotB, wantB)
+	}
+
+	usageBytes, err := os.ReadFile(filepath.Join(geminiGoldenTestdataDir, name+"_usage.json"))
+	if err != nil {
+		t.Fatalf("read %s_usage.json: %v", name, err)
+	}
+	var wantUsage streamUsageFixture
+	if err := json.Unmarshal(usageBytes, &wantUsage); err != nil {
+		t.Fatalf("decode %s_usage.json: %v", name, err)
+	}
+	if gotUsage := st.usage(); gotUsage.prompt != wantUsage.Prompt || gotUsage.completion != wantUsage.Completion {
+		t.Errorf("usage() = %+v, want {prompt:%d completion:%d}", gotUsage, wantUsage.Prompt, wantUsage.Completion)
+	}
+}
+
+// TestGeminiStreamGolden_Text covers a text-only stream: a role chunk, two
+// content deltas across two chunks, and a finish chunk on a third,
+// parts-less chunk that also carries the final usageMetadata.
+func TestGeminiStreamGolden_Text(t *testing.T) {
+	geminiRunStreamGolden(t, "stream_text")
+}
+
+// TestGeminiStreamGolden_TextThenFuncall covers a text delta, then a
+// functionCall part with no upstream "id" (proving the ordinal-synthesized
+// "call_0" id, per lesson (a)), then a finish chunk on its own later,
+// parts-less chunk — proving st.sawFunctionCall correctly latches across
+// chunks so the finish_reason still maps to "tool_calls" rather than
+// "stop" even though the finish chunk itself carries no functionCall part.
+func TestGeminiStreamGolden_TextThenFuncall(t *testing.T) {
+	geminiRunStreamGolden(t, "stream_text_then_funcall")
+}
+
+// TestGeminiStreamState_MalformedEventReturnsError proves a malformed
+// (invalid-JSON) SSE data payload returns an error wrapping errUpstream,
+// rather than being swallowed as (nil, nil) — lesson (c).
+func TestGeminiStreamState_MalformedEventReturnsError(t *testing.T) {
+	st := newGeminiStreamState(geminiGoldenGatewayModel, geminiGoldenCreated)
+	chunks, err := st.translate(sseEvent{data: []byte(`{"candidates":[`)})
+	if err == nil {
+		t.Fatalf("translate(corrupt data): want error, got nil (chunks=%v)", chunks)
+	}
+	if !errors.Is(err, errUpstream) {
+		t.Errorf("translate(corrupt data): err = %v, want errors.Is(err, errUpstream)", err)
+	}
+}
+
+// TestGeminiStreamState_UsageMetadataOnlyOverwritesWhenPresent proves "last
+// usageMetadata wins" only among chunks that actually carry the field — a
+// later chunk with no usageMetadata at all must not reset a usage total an
+// earlier chunk already captured.
+func TestGeminiStreamState_UsageMetadataOnlyOverwritesWhenPresent(t *testing.T) {
+	st := newGeminiStreamState(geminiGoldenGatewayModel, geminiGoldenCreated)
+
+	if _, err := st.translate(sseEvent{data: []byte(`{"candidates":[{"content":{"role":"model","parts":[{"text":"hi"}]}}],"usageMetadata":{"promptTokenCount":7,"candidatesTokenCount":2}}`)}); err != nil {
+		t.Fatalf("translate: %v", err)
+	}
+	if got := st.usage(); got.prompt != 7 || got.completion != 2 {
+		t.Fatalf("usage() after first chunk = %+v, want {prompt:7 completion:2}", got)
+	}
+
+	if _, err := st.translate(sseEvent{data: []byte(`{"candidates":[{"content":{"role":"model","parts":[{"text":" there"}]}}]}`)}); err != nil {
+		t.Fatalf("translate: %v", err)
+	}
+	if got := st.usage(); got.prompt != 7 || got.completion != 2 {
+		t.Errorf("usage() after a usageMetadata-less chunk = %+v, want unchanged {prompt:7 completion:2}", got)
+	}
+}
+
+// TestGeminiRequestFromOpenAI_MaxCompletionTokensPrecedence proves
+// max_completion_tokens takes precedence over max_tokens when an OpenAI
+// request sends both — the newer field wins, mirroring the anthropic
+// translator's rule.
+func TestGeminiRequestFromOpenAI_MaxCompletionTokensPrecedence(t *testing.T) {
+	req := map[string]any{
+		"model":                 "gemini-2.5-pro",
+		"messages":              []any{map[string]any{"role": "user", "content": "hi"}},
+		"max_tokens":            float64(100),
+		"max_completion_tokens": float64(200),
+	}
+	got, err := geminiRequestFromOpenAI(req)
+	if err != nil {
+		t.Fatalf("geminiRequestFromOpenAI: %v", err)
+	}
+	genConfig, ok := got["generationConfig"].(map[string]any)
+	if !ok {
+		t.Fatalf("generationConfig = %v, want a map", got["generationConfig"])
+	}
+	if genConfig["maxOutputTokens"] != int64(200) {
+		t.Errorf("maxOutputTokens = %v, want 200 (max_completion_tokens takes precedence)", genConfig["maxOutputTokens"])
+	}
+}
+
+// TestGeminiRequestFromOpenAI_LogitBiasEmptyMapAllowed proves a
+// present-but-empty logit_bias does not 400 — lesson (i): reject only when
+// truthy/non-empty.
+func TestGeminiRequestFromOpenAI_LogitBiasEmptyMapAllowed(t *testing.T) {
+	req := map[string]any{
+		"model":      "gemini-2.5-pro",
+		"messages":   []any{map[string]any{"role": "user", "content": "hi"}},
+		"logit_bias": map[string]any{},
+	}
+	if _, err := geminiRequestFromOpenAI(req); err != nil {
+		t.Errorf("geminiRequestFromOpenAI with empty logit_bias: %v, want no error", err)
+	}
+}
+
+// TestGeminiRequestFromOpenAI_UnknownToolCallIDErrors proves a role:"tool"
+// message whose tool_call_id was never seen on a preceding assistant
+// message's tool_calls[] is a *translateError, per the request mapping
+// table's role:tool row — Gemini's functionResponse needs the name that
+// map resolves, and an unresolvable id must not silently produce an empty
+// name.
+func TestGeminiRequestFromOpenAI_UnknownToolCallIDErrors(t *testing.T) {
+	req := map[string]any{
+		"model": "gemini-2.5-pro",
+		"messages": []any{
+			map[string]any{"role": "user", "content": "hi"},
+			map[string]any{"role": "tool", "tool_call_id": "call_never_seen", "content": "42"},
+		},
+	}
+	_, err := geminiRequestFromOpenAI(req)
+	var terr *translateError
+	if !errors.As(err, &terr) {
+		t.Fatalf("err = %v (%T), want *translateError", err, err)
+	}
+	if !strings.Contains(terr.Error(), "call_never_seen") {
+		t.Errorf("error message = %q, want it to mention the unknown id", terr.Error())
+	}
+}
+
+// TestGeminiEmbeddings_Single proves a single-string OpenAI embeddings
+// input builds Gemini's :embedContent request body and translates its
+// response back to an OpenAI embeddings list response, with usage
+// estimated (ceil(chars/4)) and marked estimated — per the embeddings
+// mapping table.
+func TestGeminiEmbeddings_Single(t *testing.T) {
+	in := geminiReadGoldenJSON(t, "embed_single_in.json")
+
+	texts, single, err := geminiEmbeddingInputTexts(in["input"])
+	if err != nil {
+		t.Fatalf("geminiEmbeddingInputTexts: %v", err)
+	}
+	if !single {
+		t.Fatalf("single = false, want true for a string input")
+	}
+
+	gotReq := geminiEmbedContentRequest(texts[0])
+	wantReq := geminiReadGoldenJSON(t, "embed_single_want.json")
+	if !reflect.DeepEqual(normalizeJSON(t, gotReq), normalizeJSON(t, wantReq)) {
+		t.Errorf("geminiEmbedContentRequest mismatch: got %#v, want %#v", gotReq, wantReq)
+	}
+
+	respBody, err := os.ReadFile(filepath.Join(geminiGoldenTestdataDir, "embed_single_resp_in.json"))
+	if err != nil {
+		t.Fatalf("read embed_single_resp_in.json: %v", err)
+	}
+	estTokens := estimatedEmbeddingTokens(texts)
+	if estTokens != 3 {
+		t.Fatalf("estimatedEmbeddingTokens(%q) = %d, want 3", texts, estTokens)
+	}
+
+	gotResp, u, err := openAIEmbeddingResponseFromGemini(respBody, single, "gw-embed-model", estTokens)
+	if err != nil {
+		t.Fatalf("openAIEmbeddingResponseFromGemini: %v", err)
+	}
+	if !u.estimated {
+		t.Errorf("usage.estimated = false, want true")
+	}
+	if u.prompt != 3 {
+		t.Errorf("usage.prompt = %d, want 3", u.prompt)
+	}
+	wantResp := geminiReadGoldenJSON(t, "embed_single_resp_want.json")
+	if !reflect.DeepEqual(normalizeJSON(t, gotResp), normalizeJSON(t, wantResp)) {
+		gotB, _ := json.MarshalIndent(gotResp, "", "  ")
+		wantB, _ := json.MarshalIndent(wantResp, "", "  ")
+		t.Errorf("openAIEmbeddingResponseFromGemini mismatch:\ngot:\n%s\nwant:\n%s", gotB, wantB)
+	}
+}
+
+// TestGeminiEmbeddings_Batch mirrors TestGeminiEmbeddings_Single for an
+// array-of-strings OpenAI embeddings input, which selects Gemini's
+// :batchEmbedContents endpoint and restores the "models/" prefix on each
+// request entry's own "model" field, per the embeddings mapping table.
+func TestGeminiEmbeddings_Batch(t *testing.T) {
+	in := geminiReadGoldenJSON(t, "embed_batch_in.json")
+	rawModel, _ := in["model"].(string)
+
+	texts, single, err := geminiEmbeddingInputTexts(in["input"])
+	if err != nil {
+		t.Fatalf("geminiEmbeddingInputTexts: %v", err)
+	}
+	if single {
+		t.Fatalf("single = true, want false for an array input")
+	}
+
+	gotReq := geminiBatchEmbedContentsRequest(rawModel, texts)
+	wantReq := geminiReadGoldenJSON(t, "embed_batch_want.json")
+	if !reflect.DeepEqual(normalizeJSON(t, gotReq), normalizeJSON(t, wantReq)) {
+		gotB, _ := json.MarshalIndent(gotReq, "", "  ")
+		wantB, _ := json.MarshalIndent(wantReq, "", "  ")
+		t.Errorf("geminiBatchEmbedContentsRequest mismatch:\ngot:\n%s\nwant:\n%s", gotB, wantB)
+	}
+
+	respBody, err := os.ReadFile(filepath.Join(geminiGoldenTestdataDir, "embed_batch_resp_in.json"))
+	if err != nil {
+		t.Fatalf("read embed_batch_resp_in.json: %v", err)
+	}
+	estTokens := estimatedEmbeddingTokens(texts)
+	if estTokens != 3 {
+		t.Fatalf("estimatedEmbeddingTokens(%q) = %d, want 3", texts, estTokens)
+	}
+
+	gotResp, u, err := openAIEmbeddingResponseFromGemini(respBody, single, "gw-embed-model", estTokens)
+	if err != nil {
+		t.Fatalf("openAIEmbeddingResponseFromGemini: %v", err)
+	}
+	if !u.estimated {
+		t.Errorf("usage.estimated = false, want true")
+	}
+	wantResp := geminiReadGoldenJSON(t, "embed_batch_resp_want.json")
+	if !reflect.DeepEqual(normalizeJSON(t, gotResp), normalizeJSON(t, wantResp)) {
+		gotB, _ := json.MarshalIndent(gotResp, "", "  ")
+		wantB, _ := json.MarshalIndent(wantResp, "", "  ")
+		t.Errorf("openAIEmbeddingResponseFromGemini mismatch:\ngot:\n%s\nwant:\n%s", gotB, wantB)
+	}
+}
+
+// TestGeminiEmbeddingInputTexts_RejectsNonStringArrayEntry proves an
+// input[] array with a non-string entry is a *translateError.
+func TestGeminiEmbeddingInputTexts_RejectsNonStringArrayEntry(t *testing.T) {
+	_, _, err := geminiEmbeddingInputTexts([]any{"ok", float64(5)})
+	var terr *translateError
+	if !errors.As(err, &terr) {
+		t.Fatalf("err = %v (%T), want *translateError", err, err)
+	}
+}