16 "github.com/stretchr/testify/require"
19 // testKey is a fake API key.
20 const testKey = "sk-test-not-a-real-key"
22 func TestClientSendsTheFixedRequestFields(t *testing.T) {
23 server := newFakeResponses(t)
24 client := newTestClient(t, server)
26 _, err := client.Create(t.Context(), Request{
29 Instructions: "You are a test.",
33 Parameters: map[string]any{"type": "object"},
35 Input: []any{userMessage("<chat>\n</chat>\n", "hello")},
37 require.NoError(t, err)
39 call := server.Call(0)
40 require.Equal(t, "Bearer "+testKey, call.header.Get("Authorization"))
41 require.Equal(t, "text/event-stream", call.header.Get("Accept"))
42 require.Equal(t, "application/json", call.header.Get("Content-Type"))
45 require.Equal(t, "gpt-test", body["model"])
46 require.Equal(t, "You are a test.", body["instructions"])
47 require.Equal(t, false, body["store"])
48 require.Equal(t, true, body["stream"])
49 require.Equal(t, []any{"reasoning.encrypted_content"}, body["include"])
50 require.Equal(t, "agent-key", body["prompt_cache_key"])
51 require.Equal(t, map[string]any{"effort": "high"}, body["reasoning"])
52 require.NotContains(t, body, "max_output_tokens")
53 require.NotContains(t, body, "metadata")
54 require.Equal(t, []any{map[string]any{
57 map[string]any{"type": "input_text", "text": "<chat>\n</chat>\n"},
58 map[string]any{"type": "input_text", "text": "hello"},
61 require.Len(t, body["tools"], 1)
64 func TestClientSendsTheSessionHeaderToChatGPT(t *testing.T) {
65 login := newFakeAuth(t)
66 auth := newTestAuth(t, login)
67 writeTestTokens(t, auth, login.token(time.Hour), "refresh-0")
69 testCases := []struct {
75 name: "ChatGPT backend",
81 auth: apiKey(testKey),
86 for _, tc := range testCases {
87 t.Run(tc.name, func(t *testing.T) {
88 server := newFakeResponses(t)
89 client := newClient(server.URL, tc.auth, "agent-key", filepath.Join(t.TempDir(), "usage.jsonl"))
90 _, err := client.Create(t.Context(), testRequest())
91 require.NoError(t, err)
93 call := server.Call(0)
94 require.Equal(t, tc.want, call.header.Get("session_id"))
95 require.Equal(t, "agent-key", call.body["prompt_cache_key"])
100 func TestClientReadsTheStream(t *testing.T) {
101 server := newFakeResponses(t)
102 server.reply = func(w http.ResponseWriter, call fakeCall) {
104 map[string]any{"type": "response.created", "response": map[string]any{}},
105 map[string]any{"type": "response.output_item.done", "item": map[string]any{"type": "reasoning", "encrypted_content": "opaque"}},
106 map[string]any{"type": "response.output_text.delta", "delta": "Hel"},
107 map[string]any{"type": "response.output_item.done", "item": messageItem("Hello")},
108 map[string]any{"type": "response.output_item.done", "item": messageItem(" world")},
109 completedEvent(nil, 1200, 1024, 30),
113 response, err := newTestClient(t, server).Create(t.Context(), testRequest())
114 require.NoError(t, err)
115 require.Equal(t, "Hello world", response.Text)
116 require.Len(t, response.Output, 3)
117 require.JSONEq(t, `{"type": "reasoning", "encrypted_content": "opaque"}`, string(response.Output[0]))
118 require.Equal(t, Usage{
125 func TestClientTakesTheOutputOfTheCompletedResponse(t *testing.T) {
126 server := newFakeResponses(t)
127 server.reply = func(w http.ResponseWriter, call fakeCall) {
128 writeEvents(w, completedEvent([]any{messageItem("only here")}, 10, 0, 2))
131 response, err := newTestClient(t, server).Create(t.Context(), testRequest())
132 require.NoError(t, err)
133 require.Equal(t, "only here", response.Text)
136 func TestClientRecordsEachCall(t *testing.T) {
137 server := newFakeResponses(t)
138 server.reply = func(w http.ResponseWriter, call fakeCall) {
139 w.Header().Set("X-Codex-Primary-Used-Percent", "12")
140 w.Header().Set("X-Codex-Plan-Type", "plus")
141 w.Header().Set("X-Other", "not quota")
142 w.Header().Set("X-Codex-Turn-State", "opaque state")
143 if call.number == 1 {
144 w.WriteHeader(http.StatusTooManyRequests)
145 _, _ = io.WriteString(w, `{"detail": "slow down"}`)
150 writeEvents(w, completedEvent([]any{messageItem("ok")}, 500, 400, 7))
153 client := newTestClient(t, server)
154 request := testRequest()
155 request.Node = &nodeKey{2, 5}
156 _, err := client.Create(t.Context(), request)
157 require.NoError(t, err)
158 require.NotContains(t, server.Call(0).body, "node")
159 require.NotContains(t, string(server.Call(0).raw), "[2,5]")
161 _, err = client.Create(t.Context(), testRequest())
162 require.ErrorContains(t, err, "HTTP 429")
163 require.ErrorContains(t, err, "slow down")
165 records := readUsage(t, client.usage.path)
166 require.Len(t, records, 2)
168 quota := map[string]string{
169 "x-codex-primary-used-percent": "12",
170 "x-codex-plan-type": "plus",
172 require.Equal(t, "gpt-test", records[0].Model)
173 require.Equal(t, http.StatusOK, records[0].Status)
174 require.Equal(t, 500, records[0].InputTokens)
175 require.Equal(t, 400, records[0].CachedTokens)
176 require.Equal(t, 7, records[0].OutputTokens)
177 require.Equal(t, quota, records[0].Quota)
178 require.Equal(t, &[2]int{2, 5}, records[0].Node)
179 require.Equal(t, len("ok"), records[0].TextBytes)
180 require.Greater(t, records[0].Seconds, 0.0)
181 require.Empty(t, records[0].Error)
182 require.NotEmpty(t, records[0].Date)
184 require.Nil(t, records[1].Node)
185 require.Zero(t, records[1].TextBytes)
187 require.Equal(t, http.StatusTooManyRequests, records[1].Status)
188 require.Equal(t, quota, records[1].Quota)
189 require.Contains(t, records[1].Error, "HTTP 429")
191 data, err := os.ReadFile(client.usage.path)
192 require.NoError(t, err)
193 require.NotContains(t, string(data), testKey)
195 info, err := os.Stat(client.usage.path)
196 require.NoError(t, err)
197 require.Equal(t, os.FileMode(0o600), info.Mode().Perm())
200 func TestClientRefusesCallsOverTheLimit(t *testing.T) {
201 server := newFakeResponses(t)
202 client := newTestClient(t, server)
206 require.False(t, client.LimitReached())
207 _, err := client.Create(t.Context(), testRequest())
208 require.NoError(t, err)
211 require.True(t, client.LimitReached())
212 _, err := client.Create(t.Context(), testRequest())
213 require.ErrorIs(t, err, errCallLimit)
214 require.Equal(t, 2, server.Calls())
215 require.Len(t, readUsage(t, client.usage.path), 2)
218 func TestClientFailsWhenTheResponseFails(t *testing.T) {
219 testCases := []struct {
225 name: "failed response",
226 events: []any{map[string]any{"type": "response.failed", "response": map[string]any{"error": map[string]any{"code": "server_error", "message": "busy"}}}},
227 want: "server_error: busy",
230 name: "incomplete response",
231 events: []any{map[string]any{"type": "response.incomplete", "response": map[string]any{"incomplete_details": map[string]any{"reason": "content_filter"}}}},
232 want: "content_filter",
236 events: []any{map[string]any{"type": "error", "code": "rate_limit", "message": "too many"}},
237 want: "rate_limit: too many",
240 name: "stream ends early",
241 events: []any{map[string]any{"type": "response.output_item.done", "item": messageItem("cut")}},
242 want: "ended before the response completed",
246 for _, tc := range testCases {
247 t.Run(tc.name, func(t *testing.T) {
248 server := newFakeResponses(t)
249 server.reply = func(w http.ResponseWriter, call fakeCall) {
250 writeEvents(w, tc.events...)
253 client := newTestClient(t, server)
254 _, err := client.Create(t.Context(), testRequest())
255 require.ErrorContains(t, err, tc.want)
257 records := readUsage(t, client.usage.path)
258 require.Len(t, records, 1)
259 require.Contains(t, records[0].Error, tc.want)
264 // fakeResponses is an httptest Responses endpoint. It keeps each call, and
265 // reply writes the response. The default reply is one short message.
266 type fakeResponses struct {
270 reply func(http.ResponseWriter, fakeCall)
273 // fakeCall is one call to a fake endpoint, numbered from 0.
274 type fakeCall struct {
281 func newFakeResponses(t *testing.T) *fakeResponses {
285 reply: func(w http.ResponseWriter, call fakeCall) {
286 writeEvents(w, completedEvent([]any{messageItem("a line")}, 100, 0, 5))
290 f.Server = httptest.NewServer(http.HandlerFunc(f.serve))
296 func (f *fakeResponses) serve(w http.ResponseWriter, r *http.Request) {
297 raw, err := io.ReadAll(r.Body)
299 http.Error(w, err.Error(), http.StatusBadRequest)
304 header: r.Header.Clone(),
308 err = json.Unmarshal(raw, &call.body)
310 http.Error(w, err.Error(), http.StatusBadRequest)
315 call.number = len(f.calls)
316 f.calls = append(f.calls, call)
323 // Call returns call number n.
324 func (f *fakeResponses) Call(n int) fakeCall {
326 defer f.mutex.Unlock()
331 // Calls returns the number of calls.
332 func (f *fakeResponses) Calls() int {
334 defer f.mutex.Unlock()
339 func newTestClient(t *testing.T, server *fakeResponses) *Client {
342 return newClient(server.URL, apiKey(testKey), "agent-key", filepath.Join(t.TempDir(), "usage.jsonl"))
345 func testRequest() Request {
349 Input: []any{userMessage("hello")},
353 // writeEvents writes events as a server-sent event stream.
354 func writeEvents(w http.ResponseWriter, events ...any) {
355 w.Header().Set("Content-Type", "text/event-stream")
356 for _, event := range events {
357 data, err := json.Marshal(event)
362 _, err = fmt.Fprintf(w, "event: %s\ndata: %s\n\n", event.(map[string]any)["type"], data)
369 func messageItem(text string) map[string]any {
370 return map[string]any{
374 map[string]any{"type": "output_text", "text": text},
379 func completedEvent(output []any, input, cached, out int) map[string]any {
384 return map[string]any{
385 "type": "response.completed",
386 "response": map[string]any{
388 "usage": map[string]any{
389 "input_tokens": input,
390 "input_tokens_details": map[string]any{"cached_tokens": cached},
391 "output_tokens": out,
397 func readUsage(t *testing.T, path string) []usageRecord {
400 file, err := os.Open(path)
401 require.NoError(t, err)
404 require.NoError(t, file.Close())
407 var records []usageRecord
408 scanner := bufio.NewScanner(file)
410 var record usageRecord
411 require.NoError(t, json.Unmarshal(scanner.Bytes(), &record))
412 records = append(records, record)
415 require.NoError(t, scanner.Err())