]> git.giacomolli.uk - onechat/blob - compactor_test.go
Open the chat when no command is given
[onechat] / compactor_test.go
1 package main
2
3 import (
4 "encoding/json"
5 "io"
6 "net/http"
7 "path/filepath"
8 "strconv"
9 "strings"
10 "testing"
11 "unicode/utf8"
12
13 "github.com/stretchr/testify/require"
14 )
15
16 func TestCompactorSendsThePromptToolsAndTask(t *testing.T) {
17 server := newModelFake(t)
18 compactor := newTestCompactor(t, server)
19
20 line, err := compactor.Summarize(t.Context(), Task{
21 Level: 0,
22 Index: 7,
23 Context: "<chat>\n0+4|user: earlier\n</chat>",
24 Input: []string{"user: " + strings.Repeat("x", 600)},
25 })
26 require.NoError(t, err)
27 require.Equal(t, "a line", line)
28 require.Equal(t, 1, server.Calls())
29
30 body := server.Call(0).body
31 require.Equal(t, prompt("Ada", ""), body["instructions"])
32 require.Equal(t, "gpt-test", body["model"])
33 require.Equal(t, map[string]any{"effort": "high"}, body["reasoning"])
34 require.Equal(t, "agent-key", body["prompt_cache_key"])
35
36 tools, err := json.Marshal(promptTools())
37 require.NoError(t, err)
38
39 sent, err := json.Marshal(body["tools"])
40 require.NoError(t, err)
41 require.JSONEq(t, string(tools), string(sent))
42
43 texts := compactorInputTexts(t, server.Call(0))
44 require.Equal(t, []string{
45 "<chat>\n0+4|user: earlier\n</chat>\n" +
46 "Compaction: compress message 7 into one line of at most 512 bytes\n" +
47 "(about 70 words), the length of this ruler:\n" +
48 strings.Repeat("-", 512) + "\n" +
49 "<input>\nuser: " + strings.Repeat("x", 600) + "\n</input>",
50 }, texts)
51 }
52
53 func TestCompactorSendsEachWholeBlockOfTheViewAsAMessage(t *testing.T) {
54 testCases := []struct {
55 name string
56 context string
57 want []string
58 }{
59 {
60 name: "empty view goes with the task",
61 context: "<chat>\n</chat>",
62 want: []string{"<chat>\n</chat>\nTASK"},
63 },
64 {
65 name: "lines short of a block go with the task",
66 context: "<chat>\n0+1|a\n1+1|b\n2+1|c\n</chat>",
67 want: []string{"<chat>\n0+1|a\n1+1|b\n2+1|c\n</chat>\nTASK"},
68 },
69 {
70 name: "one whole block and the task",
71 context: "<chat>\n0+1|a\n1+1|b\n2+1|c\n3+1|d\n</chat>",
72 want: []string{"<chat>\n0+1|a\n1+1|b\n2+1|c\n3+1|d\n", "</chat>\nTASK"},
73 },
74 {
75 name: "two whole blocks and the rest with the task",
76 context: "<chat>\n0+2|a\n2+1|b\n3+1|c\n4+1|d\n5+1|e\n6+1|f\n7+1|g\n8+1|h\n9+1|i\n</chat>",
77 want: []string{
78 "<chat>\n0+2|a\n2+1|b\n3+1|c\n4+1|d\n",
79 "5+1|e\n6+1|f\n7+1|g\n8+1|h\n",
80 "9+1|i\n</chat>\nTASK",
81 },
82 },
83 }
84
85 for _, tc := range testCases {
86 t.Run(tc.name, func(t *testing.T) {
87 server := newModelFake(t)
88 task := compactorLeafTask()
89 task.Context = tc.context
90
91 _, err := newTestCompactor(t, server).Summarize(t.Context(), task)
92 require.NoError(t, err)
93
94 text, err := compactorTask(task)
95 require.NoError(t, err)
96
97 want := make([]string, len(tc.want))
98 for n, message := range tc.want {
99 want[n] = strings.Replace(message, "TASK", text, 1)
100 }
101
102 texts := compactorInputTexts(t, server.Call(0))
103 require.Equal(t, want, texts)
104 require.Equal(t, tc.context+"\n"+text, strings.Join(texts, ""))
105 })
106 }
107 }
108
109 func TestCompactorSendsTheSameBlocksForViewsWithTheSameFirstLines(t *testing.T) {
110 shared := "<chat>\n0+4|a\n4+1|b\n5+1|c\n6+1|d\n7+1|e\n8+1|f\n9+1|g\n10+1|h\n"
111 server := newModelFake(t)
112 compactor := newTestCompactor(t, server)
113 _, err := compactor.Summarize(t.Context(), Task{
114 Level: 0,
115 Index: 11,
116 Context: shared + "11+1|i\n</chat>",
117 Input: []string{"user: eleven"},
118 })
119 require.NoError(t, err)
120
121 _, err = compactor.Summarize(t.Context(), Task{
122 Level: 1,
123 Index: 5,
124 Context: shared + "11+1|j\n12+1|k\n</chat>",
125 Input: []string{"left line", "right line"},
126 })
127 require.NoError(t, err)
128
129 first := compactorInputTexts(t, server.Call(0))
130 second := compactorInputTexts(t, server.Call(1))
131 require.Len(t, first, 3)
132 require.Len(t, second, 3)
133 require.Equal(t, first[:2], second[:2])
134 require.NotEqual(t, first[2], second[2])
135 }
136
137 func TestCompactorWritesTheMergeTask(t *testing.T) {
138 text, err := compactorTask(Task{
139 Level: 2,
140 Index: 3,
141 Input: []string{"left line", "right line"},
142 })
143 require.NoError(t, err)
144 require.Equal(t, "Compaction: merge lines 12+2 and 14+2, adjacent, into one line of at most\n"+
145 "512 bytes (about 70 words), the length of this ruler:\n"+
146 strings.Repeat("-", 512)+"\n"+
147 "<chat> may hold their messages, 12 to 15, in more detail: take details\n"+
148 "of them from there too.\n"+
149 "<input>\nleft line\nright line\n</input>", text)
150 }
151
152 func TestCompactorRetriesALongLineWithTheNewestNote(t *testing.T) {
153 long := []string{strings.Repeat("a", 600), strings.Repeat("b", 550), strings.Repeat("c", 400)}
154 server := newModelFake(t)
155 server.reply = func(w http.ResponseWriter, call ModelFakeCall) {
156 modelWriteEvents(w, modelCompletedEvent([]any{modelMessageItem(long[call.number])}, 100, 0, 5))
157 }
158
159 task := compactorLeafTask()
160 task.Context = "<chat>\n0+1|a\n1+1|b\n2+1|c\n3+1|d\n4+1|e\n</chat>"
161
162 compactor := newTestCompactor(t, server)
163 line, err := compactor.Summarize(t.Context(), task)
164 require.NoError(t, err)
165 require.Equal(t, long[2], line)
166 require.Equal(t, 3, server.Calls())
167
168 records := modelReadUsage(t, compactor.client.usage.path)
169 require.Len(t, records, 3)
170 for n, record := range records {
171 require.Equal(t, &[2]int{0, 0}, record.Node)
172 require.Equal(t, len(long[n]), record.TextBytes)
173 }
174
175 first := compactorInputTexts(t, server.Call(0))
176 require.Len(t, first, 2)
177 for n, want := range []string{long[0], long[1]} {
178 call := server.Call(n + 1)
179 texts := compactorInputTexts(t, call)
180 require.Equal(t, first, texts[:2])
181 require.Equal(t, []string{compactorTooLong(want)}, texts[2:])
182 require.Contains(t, texts[2], "was "+strconv.Itoa(len(want))+" bytes,")
183 require.Contains(t, texts[2], "\n"+want[:512]+"| <- LIMIT")
184
185 note, err := json.Marshal(modelUserMessage(texts[2]))
186 require.NoError(t, err)
187 require.LessOrEqual(t, len(call.raw), len(server.Call(0).raw)+len(",")+len(note))
188 }
189 }
190
191 func TestCompactorKeepsTheShortestOfFiveLines(t *testing.T) {
192 sizes := []int{700, 530, 600, 520, 560, 513}
193 server := newModelFake(t)
194 server.reply = func(w http.ResponseWriter, call ModelFakeCall) {
195 modelWriteEvents(w, modelCompletedEvent([]any{modelMessageItem(strings.Repeat("z", sizes[call.number]))}, 100, 0, 5))
196 }
197
198 line, err := newTestCompactor(t, server).Summarize(t.Context(), compactorLeafTask())
199 require.NoError(t, err)
200 require.Equal(t, compactorTries, server.Calls())
201 require.Len(t, line, 520)
202 }
203
204 func TestCompactorFailsWithoutALine(t *testing.T) {
205 testCases := []struct {
206 name string
207 reply func(http.ResponseWriter, ModelFakeCall)
208 want string
209 }{
210 {
211 name: "failed call",
212 reply: func(w http.ResponseWriter, call ModelFakeCall) {
213 w.WriteHeader(http.StatusInternalServerError)
214 _, _ = io.WriteString(w, "down")
215 },
216 want: "HTTP 500",
217 },
218 {
219 name: "tool call and no text",
220 reply: func(w http.ResponseWriter, call ModelFakeCall) {
221 modelWriteEvents(w, modelCompletedEvent([]any{map[string]any{"type": "function_call", "name": "zoom", "arguments": "{}"}}, 100, 0, 5))
222 },
223 want: "no line",
224 },
225 }
226
227 for _, tc := range testCases {
228 t.Run(tc.name, func(t *testing.T) {
229 server := newModelFake(t)
230 server.reply = tc.reply
231
232 _, err := newTestCompactor(t, server).Summarize(t.Context(), compactorLeafTask())
233 require.ErrorContains(t, err, tc.want)
234 require.Equal(t, 1, server.Calls())
235 })
236 }
237 }
238
239 func TestCutLineKeepsWholeCharacters(t *testing.T) {
240 line := strings.Repeat("a", 511) + "é" + "tail"
241 cut := compactorCutLine(line)
242 require.Equal(t, strings.Repeat("a", 511), cut)
243 require.True(t, utf8.ValidString(cut))
244 require.Equal(t, "short", compactorCutLine("short"))
245 }
246
247 func TestCompactorBuildsNodesInTheCore(t *testing.T) {
248 server := newModelFake(t)
249 server.reply = func(w http.ResponseWriter, call ModelFakeCall) {
250 if call.number == 0 {
251 w.WriteHeader(http.StatusServiceUnavailable)
252 return
253 }
254
255 modelWriteEvents(w, modelCompletedEvent([]any{modelMessageItem("summary of the long output")}, 100, 0, 5))
256 }
257
258 memory, err := openMemory(t.Context(), t.TempDir(), newTestClock().Now, newTestCompactor(t, server))
259 require.NoError(t, err)
260
261 defer func() {
262 require.NoError(t, memory.Close())
263 }()
264
265 _, err = memory.Append("echo", strings.Repeat("output ", 200))
266 require.NoError(t, err)
267 require.NoError(t, memory.Wait(t.Context()))
268 require.False(t, treeIsBuilt(t, memory.tree, 0, 0))
269
270 _, err = memory.Append("user", "next")
271 require.NoError(t, err)
272 require.NoError(t, memory.Wait(t.Context()))
273
274 node := treeReadNode(t, memory.tree, 0, 0)
275 require.Equal(t, "summary of the long output", node.Text)
276 require.Equal(t, "<chat>\n0+1|summary of the long output\n1+1|user: next\n</chat>", memory.View())
277 require.Equal(t, 2, server.Calls())
278 }
279
280 func newTestCompactor(t *testing.T, server *ModelFake) *Compactor {
281 t.Helper()
282
283 client := newModelClient(server.URL, ModelAPIKey(modelTestKey), "agent-key", filepath.Join(t.TempDir(), "usage.jsonl"))
284
285 return newCompactor(client, compactorAdaPrompt, ModelConfig{
286 Model: "gpt-test",
287 Effort: "high",
288 })
289 }
290
291 // compactorInputTexts returns the text of each user message of call. Each
292 // message has one text part.
293 func compactorInputTexts(t *testing.T, call ModelFakeCall) []string {
294 t.Helper()
295
296 var body struct {
297 Input []struct {
298 Role string `json:"role"`
299 Content []struct {
300 Type string `json:"type"`
301 Text string `json:"text"`
302 } `json:"content"`
303 } `json:"input"`
304 }
305
306 require.NoError(t, json.Unmarshal(call.raw, &body))
307
308 texts := make([]string, len(body.Input))
309 for n, message := range body.Input {
310 require.Equal(t, "user", message.Role)
311 require.Len(t, message.Content, 1)
312 require.Equal(t, "input_text", message.Content[0].Type)
313 texts[n] = message.Content[0].Text
314 }
315
316 return texts
317 }
318
319 func compactorLeafTask() Task {
320 return Task{
321 Level: 0,
322 Index: 0,
323 Context: "<chat>\n</chat>",
324 Input: []string{"user: " + strings.Repeat("y", 600)},
325 }
326 }
327
328 // compactorAdaPrompt returns the prompt of the agent Ada, without instructions.
329 func compactorAdaPrompt() (string, error) {
330 return prompt("Ada", ""), nil
331 }