]> git.giacomolli.uk - onechat/blob - responses_test.go
Leave the turn state out of the usage records
[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
15 "github.com/stretchr/testify/require"
16 )
17
18 // testKey is a fake API key.
19 const testKey = "sk-test-not-a-real-key"
20
21 func TestClientSendsTheFixedRequestFields(t *testing.T) {
22 server := newFakeResponses(t)
23 client := newTestClient(t, server)
24
25 _, err := client.Create(t.Context(), Request{
26 Model: "gpt-test",
27 Effort: "high",
28 Instructions: "You are a test.",
29 Tools: []Tool{{
30 Type: "function",
31 Name: "date",
32 Parameters: map[string]any{"type": "object"},
33 }},
34 Input: []any{userMessage("<chat>\n</chat>\n", "hello")},
35 })
36 require.NoError(t, err)
37
38 call := server.Call(0)
39 require.Equal(t, "Bearer "+testKey, call.header.Get("Authorization"))
40 require.Equal(t, "text/event-stream", call.header.Get("Accept"))
41 require.Equal(t, "application/json", call.header.Get("Content-Type"))
42
43 body := call.body
44 require.Equal(t, "gpt-test", body["model"])
45 require.Equal(t, "You are a test.", body["instructions"])
46 require.Equal(t, false, body["store"])
47 require.Equal(t, true, body["stream"])
48 require.Equal(t, []any{"reasoning.encrypted_content"}, body["include"])
49 require.Equal(t, "agent-key", body["prompt_cache_key"])
50 require.Equal(t, map[string]any{"effort": "high"}, body["reasoning"])
51 require.NotContains(t, body, "max_output_tokens")
52 require.NotContains(t, body, "metadata")
53 require.Equal(t, []any{map[string]any{
54 "role": "user",
55 "content": []any{
56 map[string]any{"type": "input_text", "text": "<chat>\n</chat>\n"},
57 map[string]any{"type": "input_text", "text": "hello"},
58 },
59 }}, body["input"])
60 require.Len(t, body["tools"], 1)
61 }
62
63 func TestClientReadsTheStream(t *testing.T) {
64 server := newFakeResponses(t)
65 server.reply = func(w http.ResponseWriter, call fakeCall) {
66 writeEvents(w,
67 map[string]any{"type": "response.created", "response": map[string]any{}},
68 map[string]any{"type": "response.output_item.done", "item": map[string]any{"type": "reasoning", "encrypted_content": "opaque"}},
69 map[string]any{"type": "response.output_text.delta", "delta": "Hel"},
70 map[string]any{"type": "response.output_item.done", "item": messageItem("Hello")},
71 map[string]any{"type": "response.output_item.done", "item": messageItem(" world")},
72 completedEvent(nil, 1200, 1024, 30),
73 )
74 }
75
76 response, err := newTestClient(t, server).Create(t.Context(), testRequest())
77 require.NoError(t, err)
78 require.Equal(t, "Hello world", response.Text)
79 require.Len(t, response.Output, 3)
80 require.JSONEq(t, `{"type": "reasoning", "encrypted_content": "opaque"}`, string(response.Output[0]))
81 require.Equal(t, Usage{
82 Input: 1200,
83 Cached: 1024,
84 Output: 30,
85 }, response.Usage)
86 }
87
88 func TestClientTakesTheOutputOfTheCompletedResponse(t *testing.T) {
89 server := newFakeResponses(t)
90 server.reply = func(w http.ResponseWriter, call fakeCall) {
91 writeEvents(w, completedEvent([]any{messageItem("only here")}, 10, 0, 2))
92 }
93
94 response, err := newTestClient(t, server).Create(t.Context(), testRequest())
95 require.NoError(t, err)
96 require.Equal(t, "only here", response.Text)
97 }
98
99 func TestClientRecordsEachCall(t *testing.T) {
100 server := newFakeResponses(t)
101 server.reply = func(w http.ResponseWriter, call fakeCall) {
102 w.Header().Set("X-Codex-Primary-Used-Percent", "12")
103 w.Header().Set("X-Codex-Plan-Type", "plus")
104 w.Header().Set("X-Other", "not quota")
105 w.Header().Set("X-Codex-Turn-State", "opaque state")
106 if call.number == 1 {
107 w.WriteHeader(http.StatusTooManyRequests)
108 _, _ = io.WriteString(w, `{"detail": "slow down"}`)
109
110 return
111 }
112
113 writeEvents(w, completedEvent([]any{messageItem("ok")}, 500, 400, 7))
114 }
115
116 client := newTestClient(t, server)
117 _, err := client.Create(t.Context(), testRequest())
118 require.NoError(t, err)
119
120 _, err = client.Create(t.Context(), testRequest())
121 require.ErrorContains(t, err, "HTTP 429")
122 require.ErrorContains(t, err, "slow down")
123
124 records := readUsage(t, client.usage.path)
125 require.Len(t, records, 2)
126
127 quota := map[string]string{
128 "x-codex-primary-used-percent": "12",
129 "x-codex-plan-type": "plus",
130 }
131 require.Equal(t, "gpt-test", records[0].Model)
132 require.Equal(t, http.StatusOK, records[0].Status)
133 require.Equal(t, 500, records[0].InputTokens)
134 require.Equal(t, 400, records[0].CachedTokens)
135 require.Equal(t, 7, records[0].OutputTokens)
136 require.Equal(t, quota, records[0].Quota)
137 require.Empty(t, records[0].Error)
138 require.NotEmpty(t, records[0].Date)
139
140 require.Equal(t, http.StatusTooManyRequests, records[1].Status)
141 require.Equal(t, quota, records[1].Quota)
142 require.Contains(t, records[1].Error, "HTTP 429")
143
144 data, err := os.ReadFile(client.usage.path)
145 require.NoError(t, err)
146 require.NotContains(t, string(data), testKey)
147
148 info, err := os.Stat(client.usage.path)
149 require.NoError(t, err)
150 require.Equal(t, os.FileMode(0o600), info.Mode().Perm())
151 }
152
153 func TestClientFailsWhenTheResponseFails(t *testing.T) {
154 testCases := []struct {
155 name string
156 events []any
157 want string
158 }{
159 {
160 name: "failed response",
161 events: []any{map[string]any{"type": "response.failed", "response": map[string]any{"error": map[string]any{"code": "server_error", "message": "busy"}}}},
162 want: "server_error: busy",
163 },
164 {
165 name: "incomplete response",
166 events: []any{map[string]any{"type": "response.incomplete", "response": map[string]any{"incomplete_details": map[string]any{"reason": "content_filter"}}}},
167 want: "content_filter",
168 },
169 {
170 name: "error event",
171 events: []any{map[string]any{"type": "error", "code": "rate_limit", "message": "too many"}},
172 want: "rate_limit: too many",
173 },
174 {
175 name: "stream ends early",
176 events: []any{map[string]any{"type": "response.output_item.done", "item": messageItem("cut")}},
177 want: "ended before the response completed",
178 },
179 }
180
181 for _, tc := range testCases {
182 t.Run(tc.name, func(t *testing.T) {
183 server := newFakeResponses(t)
184 server.reply = func(w http.ResponseWriter, call fakeCall) {
185 writeEvents(w, tc.events...)
186 }
187
188 client := newTestClient(t, server)
189 _, err := client.Create(t.Context(), testRequest())
190 require.ErrorContains(t, err, tc.want)
191
192 records := readUsage(t, client.usage.path)
193 require.Len(t, records, 1)
194 require.Contains(t, records[0].Error, tc.want)
195 })
196 }
197 }
198
199 // fakeResponses is an httptest Responses endpoint. It keeps each call, and
200 // reply writes the response. The default reply is one short message.
201 type fakeResponses struct {
202 *httptest.Server
203 mutex sync.Mutex
204 calls []fakeCall
205 reply func(http.ResponseWriter, fakeCall)
206 }
207
208 // fakeCall is one call to a fake endpoint, numbered from 0.
209 type fakeCall struct {
210 number int
211 header http.Header
212 raw []byte
213 body map[string]any
214 }
215
216 func newFakeResponses(t *testing.T) *fakeResponses {
217 t.Helper()
218
219 f := &fakeResponses{
220 reply: func(w http.ResponseWriter, call fakeCall) {
221 writeEvents(w, completedEvent([]any{messageItem("a line")}, 100, 0, 5))
222 },
223 }
224
225 f.Server = httptest.NewServer(http.HandlerFunc(f.serve))
226 t.Cleanup(f.Close)
227
228 return f
229 }
230
231 func (f *fakeResponses) serve(w http.ResponseWriter, r *http.Request) {
232 raw, err := io.ReadAll(r.Body)
233 if err != nil {
234 http.Error(w, err.Error(), http.StatusBadRequest)
235 return
236 }
237
238 call := fakeCall{
239 header: r.Header.Clone(),
240 raw: raw,
241 }
242
243 err = json.Unmarshal(raw, &call.body)
244 if err != nil {
245 http.Error(w, err.Error(), http.StatusBadRequest)
246 return
247 }
248
249 f.mutex.Lock()
250 call.number = len(f.calls)
251 f.calls = append(f.calls, call)
252 reply := f.reply
253 f.mutex.Unlock()
254
255 reply(w, call)
256 }
257
258 // Call returns call number n.
259 func (f *fakeResponses) Call(n int) fakeCall {
260 f.mutex.Lock()
261 defer f.mutex.Unlock()
262
263 return f.calls[n]
264 }
265
266 // Calls returns the number of calls.
267 func (f *fakeResponses) Calls() int {
268 f.mutex.Lock()
269 defer f.mutex.Unlock()
270
271 return len(f.calls)
272 }
273
274 func newTestClient(t *testing.T, server *fakeResponses) *Client {
275 t.Helper()
276
277 return newClient(server.URL, apiKey(testKey), "agent-key", filepath.Join(t.TempDir(), "usage.jsonl"))
278 }
279
280 func testRequest() Request {
281 return Request{
282 Model: "gpt-test",
283 Effort: "low",
284 Input: []any{userMessage("hello")},
285 }
286 }
287
288 // writeEvents writes events as a server-sent event stream.
289 func writeEvents(w http.ResponseWriter, events ...any) {
290 w.Header().Set("Content-Type", "text/event-stream")
291 for _, event := range events {
292 data, err := json.Marshal(event)
293 if err != nil {
294 panic(err)
295 }
296
297 _, err = fmt.Fprintf(w, "event: %s\ndata: %s\n\n", event.(map[string]any)["type"], data)
298 if err != nil {
299 panic(err)
300 }
301 }
302 }
303
304 func messageItem(text string) map[string]any {
305 return map[string]any{
306 "type": "message",
307 "role": "assistant",
308 "content": []any{
309 map[string]any{"type": "output_text", "text": text},
310 },
311 }
312 }
313
314 func completedEvent(output []any, input, cached, out int) map[string]any {
315 if output == nil {
316 output = []any{}
317 }
318
319 return map[string]any{
320 "type": "response.completed",
321 "response": map[string]any{
322 "output": output,
323 "usage": map[string]any{
324 "input_tokens": input,
325 "input_tokens_details": map[string]any{"cached_tokens": cached},
326 "output_tokens": out,
327 },
328 },
329 }
330 }
331
332 func readUsage(t *testing.T, path string) []usageRecord {
333 t.Helper()
334
335 file, err := os.Open(path)
336 require.NoError(t, err)
337
338 defer func() {
339 require.NoError(t, file.Close())
340 }()
341
342 var records []usageRecord
343 scanner := bufio.NewScanner(file)
344 for scanner.Scan() {
345 var record usageRecord
346 require.NoError(t, json.Unmarshal(scanner.Bytes(), &record))
347 records = append(records, record)
348 }
349
350 require.NoError(t, scanner.Err())
351
352 return records
353 }