]> git.giacomolli.uk - onechat/blob - responses_test.go
Record open questions and reasons in issues
[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 TestClientFailsWhenTheResponseFails(t *testing.T) {
201         testCases := []struct {
202                 name   string
203                 events []any
204                 want   string
205         }{
206                 {
207                         name:   "failed response",
208                         events: []any{map[string]any{"type": "response.failed", "response": map[string]any{"error": map[string]any{"code": "server_error", "message": "busy"}}}},
209                         want:   "server_error: busy",
210                 },
211                 {
212                         name:   "incomplete response",
213                         events: []any{map[string]any{"type": "response.incomplete", "response": map[string]any{"incomplete_details": map[string]any{"reason": "content_filter"}}}},
214                         want:   "content_filter",
215                 },
216                 {
217                         name:   "error event",
218                         events: []any{map[string]any{"type": "error", "code": "rate_limit", "message": "too many"}},
219                         want:   "rate_limit: too many",
220                 },
221                 {
222                         name:   "stream ends early",
223                         events: []any{map[string]any{"type": "response.output_item.done", "item": messageItem("cut")}},
224                         want:   "ended before the response completed",
225                 },
226         }
227
228         for _, tc := range testCases {
229                 t.Run(tc.name, func(t *testing.T) {
230                         server := newFakeResponses(t)
231                         server.reply = func(w http.ResponseWriter, call fakeCall) {
232                                 writeEvents(w, tc.events...)
233                         }
234
235                         client := newTestClient(t, server)
236                         _, err := client.Create(t.Context(), testRequest())
237                         require.ErrorContains(t, err, tc.want)
238
239                         records := readUsage(t, client.usage.path)
240                         require.Len(t, records, 1)
241                         require.Contains(t, records[0].Error, tc.want)
242                 })
243         }
244 }
245
246 // fakeResponses is an httptest Responses endpoint. It keeps each call, and
247 // reply writes the response. The default reply is one short message.
248 type fakeResponses struct {
249         *httptest.Server
250         mutex sync.Mutex
251         calls []fakeCall
252         reply func(http.ResponseWriter, fakeCall)
253 }
254
255 // fakeCall is one call to a fake endpoint, numbered from 0.
256 type fakeCall struct {
257         number int
258         header http.Header
259         raw    []byte
260         body   map[string]any
261 }
262
263 func newFakeResponses(t *testing.T) *fakeResponses {
264         t.Helper()
265
266         f := &fakeResponses{
267                 reply: func(w http.ResponseWriter, call fakeCall) {
268                         writeEvents(w, completedEvent([]any{messageItem("a line")}, 100, 0, 5))
269                 },
270         }
271
272         f.Server = httptest.NewServer(http.HandlerFunc(f.serve))
273         t.Cleanup(f.Close)
274
275         return f
276 }
277
278 func (f *fakeResponses) serve(w http.ResponseWriter, r *http.Request) {
279         raw, err := io.ReadAll(r.Body)
280         if err != nil {
281                 http.Error(w, err.Error(), http.StatusBadRequest)
282                 return
283         }
284
285         call := fakeCall{
286                 header: r.Header.Clone(),
287                 raw:    raw,
288         }
289
290         err = json.Unmarshal(raw, &call.body)
291         if err != nil {
292                 http.Error(w, err.Error(), http.StatusBadRequest)
293                 return
294         }
295
296         f.mutex.Lock()
297         call.number = len(f.calls)
298         f.calls = append(f.calls, call)
299         reply := f.reply
300         f.mutex.Unlock()
301
302         reply(w, call)
303 }
304
305 // Call returns call number n.
306 func (f *fakeResponses) Call(n int) fakeCall {
307         f.mutex.Lock()
308         defer f.mutex.Unlock()
309
310         return f.calls[n]
311 }
312
313 // Calls returns the number of calls.
314 func (f *fakeResponses) Calls() int {
315         f.mutex.Lock()
316         defer f.mutex.Unlock()
317
318         return len(f.calls)
319 }
320
321 func newTestClient(t *testing.T, server *fakeResponses) *Client {
322         t.Helper()
323
324         return newClient(server.URL, apiKey(testKey), "agent-key", filepath.Join(t.TempDir(), "usage.jsonl"))
325 }
326
327 func testRequest() Request {
328         return Request{
329                 Model:  "gpt-test",
330                 Effort: "low",
331                 Input:  []any{userMessage("hello")},
332         }
333 }
334
335 // writeEvents writes events as a server-sent event stream.
336 func writeEvents(w http.ResponseWriter, events ...any) {
337         w.Header().Set("Content-Type", "text/event-stream")
338         for _, event := range events {
339                 data, err := json.Marshal(event)
340                 if err != nil {
341                         panic(err)
342                 }
343
344                 _, err = fmt.Fprintf(w, "event: %s\ndata: %s\n\n", event.(map[string]any)["type"], data)
345                 if err != nil {
346                         panic(err)
347                 }
348         }
349 }
350
351 func messageItem(text string) map[string]any {
352         return map[string]any{
353                 "type": "message",
354                 "role": "assistant",
355                 "content": []any{
356                         map[string]any{"type": "output_text", "text": text},
357                 },
358         }
359 }
360
361 func completedEvent(output []any, input, cached, out int) map[string]any {
362         if output == nil {
363                 output = []any{}
364         }
365
366         return map[string]any{
367                 "type": "response.completed",
368                 "response": map[string]any{
369                         "output": output,
370                         "usage": map[string]any{
371                                 "input_tokens":         input,
372                                 "input_tokens_details": map[string]any{"cached_tokens": cached},
373                                 "output_tokens":        out,
374                         },
375                 },
376         }
377 }
378
379 func readUsage(t *testing.T, path string) []usageRecord {
380         t.Helper()
381
382         file, err := os.Open(path)
383         require.NoError(t, err)
384
385         defer func() {
386                 require.NoError(t, file.Close())
387         }()
388
389         var records []usageRecord
390         scanner := bufio.NewScanner(file)
391         for scanner.Scan() {
392                 var record usageRecord
393                 require.NoError(t, json.Unmarshal(scanner.Bytes(), &record))
394                 records = append(records, record)
395         }
396
397         require.NoError(t, scanner.Err())
398
399         return records
400 }