]> git.giacomolli.uk - onechat/blob - responses_test.go
Lay out turn requests for the prompt cache
[onechat] / responses_test.go
1 package main
2
3 import (
4 "bufio"
5 "encoding/json"
6 "fmt"
7 "io"
8 "net/http"
9 "net/http/httptest"
10 "os"
11 "path/filepath"
12 "sync"
13 "testing"
14 "time"
15
16 "github.com/stretchr/testify/require"
17 )
18
19 // testKey is a fake API key.
20 const testKey = "sk-test-not-a-real-key"
21
22 func TestClientSendsTheFixedRequestFields(t *testing.T) {
23 server := newFakeResponses(t)
24 client := newTestClient(t, server)
25
26 _, err := client.Create(t.Context(), Request{
27 Model: "gpt-test",
28 Effort: "high",
29 Instructions: "You are a test.",
30 Tools: []Tool{{
31 Type: "function",
32 Name: "date",
33 Parameters: map[string]any{"type": "object"},
34 }},
35 Input: []any{userMessage("<chat>\n</chat>\n", "hello")},
36 })
37 require.NoError(t, err)
38
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"))
43
44 body := call.body
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{
55 "role": "user",
56 "content": []any{
57 map[string]any{"type": "input_text", "text": "<chat>\n</chat>\n"},
58 map[string]any{"type": "input_text", "text": "hello"},
59 },
60 }}, body["input"])
61 require.Len(t, body["tools"], 1)
62 }
63
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")
68
69 testCases := []struct {
70 name string
71 auth authorizer
72 want string
73 }{
74 {
75 name: "ChatGPT backend",
76 auth: auth,
77 want: "agent-key",
78 },
79 {
80 name: "OpenAI API",
81 auth: apiKey(testKey),
82 want: "",
83 },
84 }
85
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)
92
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"])
96 })
97 }
98 }
99
100 func TestClientReadsTheStream(t *testing.T) {
101 server := newFakeResponses(t)
102 server.reply = func(w http.ResponseWriter, call fakeCall) {
103 writeEvents(w,
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),
110 )
111 }
112
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{
119 Input: 1200,
120 Cached: 1024,
121 Output: 30,
122 }, response.Usage)
123 }
124
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))
129 }
130
131 response, err := newTestClient(t, server).Create(t.Context(), testRequest())
132 require.NoError(t, err)
133 require.Equal(t, "only here", response.Text)
134 }
135
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"}`)
146
147 return
148 }
149
150 writeEvents(w, completedEvent([]any{messageItem("ok")}, 500, 400, 7))
151 }
152
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]")
160
161 _, err = client.Create(t.Context(), testRequest())
162 require.ErrorContains(t, err, "HTTP 429")
163 require.ErrorContains(t, err, "slow down")
164
165 records := readUsage(t, client.usage.path)
166 require.Len(t, records, 2)
167
168 quota := map[string]string{
169 "x-codex-primary-used-percent": "12",
170 "x-codex-plan-type": "plus",
171 }
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)
183
184 require.Nil(t, records[1].Node)
185 require.Zero(t, records[1].TextBytes)
186
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")
190
191 data, err := os.ReadFile(client.usage.path)
192 require.NoError(t, err)
193 require.NotContains(t, string(data), testKey)
194
195 info, err := os.Stat(client.usage.path)
196 require.NoError(t, err)
197 require.Equal(t, os.FileMode(0o600), info.Mode().Perm())
198 }
199
200 func TestClientStopsAtTheFirstOutputItem(t *testing.T) {
201 server := newFakeResponses(t)
202 server.reply = func(w http.ResponseWriter, call fakeCall) {
203 writeEvents(w,
204 map[string]any{"type": "response.created"},
205 map[string]any{"type": "response.output_item.added", "item": messageItem("")},
206 )
207 }
208
209 client := newTestClient(t, server)
210 request := testRequest()
211 request.StopAtOutput = true
212 request.Turn = "warmup"
213 response, err := client.Create(t.Context(), request)
214 require.NoError(t, err)
215 require.Empty(t, response.Output)
216
217 request.StopAtOutput = false
218 _, err = client.Create(t.Context(), request)
219 require.ErrorContains(t, err, "event stream ended before the response completed")
220 }
221
222 func TestClientRefusesCallsOverTheLimit(t *testing.T) {
223 server := newFakeResponses(t)
224 client := newTestClient(t, server)
225 client.LimitCalls(2)
226
227 for range 2 {
228 require.False(t, client.LimitReached())
229 _, err := client.Create(t.Context(), testRequest())
230 require.NoError(t, err)
231 }
232
233 require.True(t, client.LimitReached())
234 _, err := client.Create(t.Context(), testRequest())
235 require.ErrorIs(t, err, errCallLimit)
236 require.Equal(t, 2, server.Calls())
237 require.Len(t, readUsage(t, client.usage.path), 2)
238 }
239
240 func TestClientFailsWhenTheResponseFails(t *testing.T) {
241 testCases := []struct {
242 name string
243 events []any
244 want string
245 }{
246 {
247 name: "failed response",
248 events: []any{map[string]any{"type": "response.failed", "response": map[string]any{"error": map[string]any{"code": "server_error", "message": "busy"}}}},
249 want: "server_error: busy",
250 },
251 {
252 name: "incomplete response",
253 events: []any{map[string]any{"type": "response.incomplete", "response": map[string]any{"incomplete_details": map[string]any{"reason": "content_filter"}}}},
254 want: "content_filter",
255 },
256 {
257 name: "error event",
258 events: []any{map[string]any{"type": "error", "code": "rate_limit", "message": "too many"}},
259 want: "rate_limit: too many",
260 },
261 {
262 name: "stream ends early",
263 events: []any{map[string]any{"type": "response.output_item.done", "item": messageItem("cut")}},
264 want: "ended before the response completed",
265 },
266 }
267
268 for _, tc := range testCases {
269 t.Run(tc.name, func(t *testing.T) {
270 server := newFakeResponses(t)
271 server.reply = func(w http.ResponseWriter, call fakeCall) {
272 writeEvents(w, tc.events...)
273 }
274
275 client := newTestClient(t, server)
276 _, err := client.Create(t.Context(), testRequest())
277 require.ErrorContains(t, err, tc.want)
278
279 records := readUsage(t, client.usage.path)
280 require.Len(t, records, 1)
281 require.Contains(t, records[0].Error, tc.want)
282 })
283 }
284 }
285
286 // fakeResponses is an httptest Responses endpoint. It keeps each call, and
287 // reply writes the response. The default reply is one short message.
288 type fakeResponses struct {
289 *httptest.Server
290 mutex sync.Mutex
291 calls []fakeCall
292 reply func(http.ResponseWriter, fakeCall)
293 }
294
295 // fakeCall is one call to a fake endpoint, numbered from 0.
296 type fakeCall struct {
297 number int
298 header http.Header
299 raw []byte
300 body map[string]any
301 }
302
303 func newFakeResponses(t *testing.T) *fakeResponses {
304 t.Helper()
305
306 f := &fakeResponses{
307 reply: func(w http.ResponseWriter, call fakeCall) {
308 writeEvents(w, completedEvent([]any{messageItem("a line")}, 100, 0, 5))
309 },
310 }
311
312 f.Server = httptest.NewServer(http.HandlerFunc(f.serve))
313 t.Cleanup(f.Close)
314
315 return f
316 }
317
318 func (f *fakeResponses) serve(w http.ResponseWriter, r *http.Request) {
319 raw, err := io.ReadAll(r.Body)
320 if err != nil {
321 http.Error(w, err.Error(), http.StatusBadRequest)
322 return
323 }
324
325 call := fakeCall{
326 header: r.Header.Clone(),
327 raw: raw,
328 }
329
330 err = json.Unmarshal(raw, &call.body)
331 if err != nil {
332 http.Error(w, err.Error(), http.StatusBadRequest)
333 return
334 }
335
336 f.mutex.Lock()
337 call.number = len(f.calls)
338 f.calls = append(f.calls, call)
339 reply := f.reply
340 f.mutex.Unlock()
341
342 reply(w, call)
343 }
344
345 // Call returns call number n.
346 func (f *fakeResponses) Call(n int) fakeCall {
347 f.mutex.Lock()
348 defer f.mutex.Unlock()
349
350 return f.calls[n]
351 }
352
353 // Calls returns the number of calls.
354 func (f *fakeResponses) Calls() int {
355 f.mutex.Lock()
356 defer f.mutex.Unlock()
357
358 return len(f.calls)
359 }
360
361 func newTestClient(t *testing.T, server *fakeResponses) *Client {
362 t.Helper()
363
364 return newClient(server.URL, apiKey(testKey), "agent-key", filepath.Join(t.TempDir(), "usage.jsonl"))
365 }
366
367 func testRequest() Request {
368 return Request{
369 Model: "gpt-test",
370 Effort: "low",
371 Input: []any{userMessage("hello")},
372 }
373 }
374
375 // writeEvents writes events as a server-sent event stream.
376 func writeEvents(w http.ResponseWriter, events ...any) {
377 w.Header().Set("Content-Type", "text/event-stream")
378 for _, event := range events {
379 data, err := json.Marshal(event)
380 if err != nil {
381 panic(err)
382 }
383
384 _, err = fmt.Fprintf(w, "event: %s\ndata: %s\n\n", event.(map[string]any)["type"], data)
385 if err != nil {
386 panic(err)
387 }
388 }
389 }
390
391 func messageItem(text string) map[string]any {
392 return map[string]any{
393 "type": "message",
394 "role": "assistant",
395 "content": []any{
396 map[string]any{"type": "output_text", "text": text},
397 },
398 }
399 }
400
401 func completedEvent(output []any, input, cached, out int) map[string]any {
402 if output == nil {
403 output = []any{}
404 }
405
406 return map[string]any{
407 "type": "response.completed",
408 "response": map[string]any{
409 "output": output,
410 "usage": map[string]any{
411 "input_tokens": input,
412 "input_tokens_details": map[string]any{"cached_tokens": cached},
413 "output_tokens": out,
414 },
415 },
416 }
417 }
418
419 func readUsage(t *testing.T, path string) []usageRecord {
420 t.Helper()
421
422 file, err := os.Open(path)
423 require.NoError(t, err)
424
425 defer func() {
426 require.NoError(t, file.Close())
427 }()
428
429 var records []usageRecord
430 scanner := bufio.NewScanner(file)
431 for scanner.Scan() {
432 var record usageRecord
433 require.NoError(t, json.Unmarshal(scanner.Bytes(), &record))
434 records = append(records, record)
435 }
436
437 require.NoError(t, scanner.Err())
438
439 return records
440 }