]> git.giacomolli.uk - onechat/blob - responses_test.go
Send the cache key in the session_id header
[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         _, err := client.Create(t.Context(), testRequest())
155         require.NoError(t, err)
156
157         _, err = client.Create(t.Context(), testRequest())
158         require.ErrorContains(t, err, "HTTP 429")
159         require.ErrorContains(t, err, "slow down")
160
161         records := readUsage(t, client.usage.path)
162         require.Len(t, records, 2)
163
164         quota := map[string]string{
165                 "x-codex-primary-used-percent": "12",
166                 "x-codex-plan-type":            "plus",
167         }
168         require.Equal(t, "gpt-test", records[0].Model)
169         require.Equal(t, http.StatusOK, records[0].Status)
170         require.Equal(t, 500, records[0].InputTokens)
171         require.Equal(t, 400, records[0].CachedTokens)
172         require.Equal(t, 7, records[0].OutputTokens)
173         require.Equal(t, quota, records[0].Quota)
174         require.Empty(t, records[0].Error)
175         require.NotEmpty(t, records[0].Date)
176
177         require.Equal(t, http.StatusTooManyRequests, records[1].Status)
178         require.Equal(t, quota, records[1].Quota)
179         require.Contains(t, records[1].Error, "HTTP 429")
180
181         data, err := os.ReadFile(client.usage.path)
182         require.NoError(t, err)
183         require.NotContains(t, string(data), testKey)
184
185         info, err := os.Stat(client.usage.path)
186         require.NoError(t, err)
187         require.Equal(t, os.FileMode(0o600), info.Mode().Perm())
188 }
189
190 func TestClientFailsWhenTheResponseFails(t *testing.T) {
191         testCases := []struct {
192                 name   string
193                 events []any
194                 want   string
195         }{
196                 {
197                         name:   "failed response",
198                         events: []any{map[string]any{"type": "response.failed", "response": map[string]any{"error": map[string]any{"code": "server_error", "message": "busy"}}}},
199                         want:   "server_error: busy",
200                 },
201                 {
202                         name:   "incomplete response",
203                         events: []any{map[string]any{"type": "response.incomplete", "response": map[string]any{"incomplete_details": map[string]any{"reason": "content_filter"}}}},
204                         want:   "content_filter",
205                 },
206                 {
207                         name:   "error event",
208                         events: []any{map[string]any{"type": "error", "code": "rate_limit", "message": "too many"}},
209                         want:   "rate_limit: too many",
210                 },
211                 {
212                         name:   "stream ends early",
213                         events: []any{map[string]any{"type": "response.output_item.done", "item": messageItem("cut")}},
214                         want:   "ended before the response completed",
215                 },
216         }
217
218         for _, tc := range testCases {
219                 t.Run(tc.name, func(t *testing.T) {
220                         server := newFakeResponses(t)
221                         server.reply = func(w http.ResponseWriter, call fakeCall) {
222                                 writeEvents(w, tc.events...)
223                         }
224
225                         client := newTestClient(t, server)
226                         _, err := client.Create(t.Context(), testRequest())
227                         require.ErrorContains(t, err, tc.want)
228
229                         records := readUsage(t, client.usage.path)
230                         require.Len(t, records, 1)
231                         require.Contains(t, records[0].Error, tc.want)
232                 })
233         }
234 }
235
236 // fakeResponses is an httptest Responses endpoint. It keeps each call, and
237 // reply writes the response. The default reply is one short message.
238 type fakeResponses struct {
239         *httptest.Server
240         mutex sync.Mutex
241         calls []fakeCall
242         reply func(http.ResponseWriter, fakeCall)
243 }
244
245 // fakeCall is one call to a fake endpoint, numbered from 0.
246 type fakeCall struct {
247         number int
248         header http.Header
249         raw    []byte
250         body   map[string]any
251 }
252
253 func newFakeResponses(t *testing.T) *fakeResponses {
254         t.Helper()
255
256         f := &fakeResponses{
257                 reply: func(w http.ResponseWriter, call fakeCall) {
258                         writeEvents(w, completedEvent([]any{messageItem("a line")}, 100, 0, 5))
259                 },
260         }
261
262         f.Server = httptest.NewServer(http.HandlerFunc(f.serve))
263         t.Cleanup(f.Close)
264
265         return f
266 }
267
268 func (f *fakeResponses) serve(w http.ResponseWriter, r *http.Request) {
269         raw, err := io.ReadAll(r.Body)
270         if err != nil {
271                 http.Error(w, err.Error(), http.StatusBadRequest)
272                 return
273         }
274
275         call := fakeCall{
276                 header: r.Header.Clone(),
277                 raw:    raw,
278         }
279
280         err = json.Unmarshal(raw, &call.body)
281         if err != nil {
282                 http.Error(w, err.Error(), http.StatusBadRequest)
283                 return
284         }
285
286         f.mutex.Lock()
287         call.number = len(f.calls)
288         f.calls = append(f.calls, call)
289         reply := f.reply
290         f.mutex.Unlock()
291
292         reply(w, call)
293 }
294
295 // Call returns call number n.
296 func (f *fakeResponses) Call(n int) fakeCall {
297         f.mutex.Lock()
298         defer f.mutex.Unlock()
299
300         return f.calls[n]
301 }
302
303 // Calls returns the number of calls.
304 func (f *fakeResponses) Calls() int {
305         f.mutex.Lock()
306         defer f.mutex.Unlock()
307
308         return len(f.calls)
309 }
310
311 func newTestClient(t *testing.T, server *fakeResponses) *Client {
312         t.Helper()
313
314         return newClient(server.URL, apiKey(testKey), "agent-key", filepath.Join(t.TempDir(), "usage.jsonl"))
315 }
316
317 func testRequest() Request {
318         return Request{
319                 Model:  "gpt-test",
320                 Effort: "low",
321                 Input:  []any{userMessage("hello")},
322         }
323 }
324
325 // writeEvents writes events as a server-sent event stream.
326 func writeEvents(w http.ResponseWriter, events ...any) {
327         w.Header().Set("Content-Type", "text/event-stream")
328         for _, event := range events {
329                 data, err := json.Marshal(event)
330                 if err != nil {
331                         panic(err)
332                 }
333
334                 _, err = fmt.Fprintf(w, "event: %s\ndata: %s\n\n", event.(map[string]any)["type"], data)
335                 if err != nil {
336                         panic(err)
337                 }
338         }
339 }
340
341 func messageItem(text string) map[string]any {
342         return map[string]any{
343                 "type": "message",
344                 "role": "assistant",
345                 "content": []any{
346                         map[string]any{"type": "output_text", "text": text},
347                 },
348         }
349 }
350
351 func completedEvent(output []any, input, cached, out int) map[string]any {
352         if output == nil {
353                 output = []any{}
354         }
355
356         return map[string]any{
357                 "type": "response.completed",
358                 "response": map[string]any{
359                         "output": output,
360                         "usage": map[string]any{
361                                 "input_tokens":         input,
362                                 "input_tokens_details": map[string]any{"cached_tokens": cached},
363                                 "output_tokens":        out,
364                         },
365                 },
366         }
367 }
368
369 func readUsage(t *testing.T, path string) []usageRecord {
370         t.Helper()
371
372         file, err := os.Open(path)
373         require.NoError(t, err)
374
375         defer func() {
376                 require.NoError(t, file.Close())
377         }()
378
379         var records []usageRecord
380         scanner := bufio.NewScanner(file)
381         for scanner.Scan() {
382                 var record usageRecord
383                 require.NoError(t, json.Unmarshal(scanner.Bytes(), &record))
384                 records = append(records, record)
385         }
386
387         require.NoError(t, scanner.Err())
388
389         return records
390 }