]> git.giacomolli.uk - onechat/blob - compactor_test.go
Send the compaction view in blocks of four lines
[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 := newFakeResponses(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|a\n4+1|b\n5+1|c\n6+1|d\n7+1|e\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(agentTools())
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         parts := inputParts(t, server.Call(0))
44         require.Equal(t, []string{
45                 "<chat>\n",
46                 "0+4|a\n4+1|b\n5+1|c\n6+1|d\n",
47                 "7+1|e\n</chat>\n",
48                 "Compaction: compress message 7 into one line of at most 512 bytes\n" +
49                         "(about 70 words), the length of this ruler:\n" +
50                         strings.Repeat("-", 512) + "\n" +
51                         "<input>\nuser: " + strings.Repeat("x", 600) + "\n</input>",
52         }, parts)
53 }
54
55 func TestViewPartsSendWholeBlocksOfLines(t *testing.T) {
56         testCases := []struct {
57                 name    string
58                 context string
59                 want    []string
60         }{
61                 {
62                         name:    "empty view has the tags only",
63                         context: "<chat>\n</chat>",
64                         want:    []string{"<chat>\n", "</chat>\n"},
65                 },
66                 {
67                         name:    "lines short of a block go with the closing tag",
68                         context: "<chat>\n0+1|a\n1+1|b\n2+1|c\n</chat>",
69                         want:    []string{"<chat>\n", "0+1|a\n1+1|b\n2+1|c\n</chat>\n"},
70                 },
71                 {
72                         name:    "one whole block",
73                         context: "<chat>\n0+1|a\n1+1|b\n2+1|c\n3+1|d\n</chat>",
74                         want:    []string{"<chat>\n", "0+1|a\n1+1|b\n2+1|c\n3+1|d\n", "</chat>\n"},
75                 },
76                 {
77                         name:    "two whole blocks and a rest",
78                         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>",
79                         want: []string{
80                                 "<chat>\n",
81                                 "0+2|a\n2+1|b\n3+1|c\n4+1|d\n",
82                                 "5+1|e\n6+1|f\n7+1|g\n8+1|h\n",
83                                 "9+1|i\n</chat>\n",
84                         },
85                 },
86         }
87
88         for _, tc := range testCases {
89                 t.Run(tc.name, func(t *testing.T) {
90                         parts := viewParts(tc.context)
91                         require.Equal(t, tc.want, parts)
92                         require.Equal(t, tc.context+"\n", strings.Join(parts, ""))
93                 })
94         }
95 }
96
97 func TestViewPartsKeepTheBlocksOfAShorterView(t *testing.T) {
98         shorter := viewParts("<chat>\n0+1|a\n1+1|b\n2+1|c\n3+1|d\n4+1|e\n</chat>")
99         longer := viewParts("<chat>\n0+1|a\n1+1|b\n2+1|c\n3+1|d\n4+1|e\n5+1|f\n6+1|g\n7+1|h\n</chat>")
100         require.Equal(t, shorter[:2], longer[:2])
101         require.NotEqual(t, shorter[2], longer[2])
102 }
103
104 func TestCompactorWritesTheMergeTask(t *testing.T) {
105         text, err := compactionTask(Task{
106                 Level: 2,
107                 Index: 3,
108                 Input: []string{"left line", "right line"},
109         })
110         require.NoError(t, err)
111         require.Equal(t, "Compaction: merge lines 12+2 and 14+2, adjacent, into one line of at most\n"+
112                 "512 bytes (about 70 words), the length of this ruler:\n"+
113                 strings.Repeat("-", 512)+"\n"+
114                 "<chat> may hold their messages, 12 to 15, in more detail: take details\n"+
115                 "of them from there too.\n"+
116                 "<input>\nleft line\nright line\n</input>", text)
117 }
118
119 func TestCompactorRetriesALongLineWithTheNewestNote(t *testing.T) {
120         long := []string{strings.Repeat("a", 600), strings.Repeat("b", 550), strings.Repeat("c", 400)}
121         server := newFakeResponses(t)
122         server.reply = func(w http.ResponseWriter, call fakeCall) {
123                 writeEvents(w, completedEvent([]any{messageItem(long[call.number])}, 100, 0, 5))
124         }
125
126         compactor := newTestCompactor(t, server)
127         line, err := compactor.Summarize(t.Context(), leafTask())
128         require.NoError(t, err)
129         require.Equal(t, long[2], line)
130         require.Equal(t, 3, server.Calls())
131
132         records := readUsage(t, compactor.client.usage.path)
133         require.Len(t, records, 3)
134         for n, record := range records {
135                 require.Equal(t, &[2]int{0, 0}, record.Node)
136                 require.Equal(t, len(long[n]), record.TextBytes)
137         }
138
139         first := inputParts(t, server.Call(0))
140         require.Len(t, first, 3)
141         for n, want := range []string{long[0], long[1]} {
142                 call := server.Call(n + 1)
143                 parts := inputParts(t, call)
144                 require.Equal(t, first, parts[:3])
145                 require.Equal(t, []string{tooLong(want)}, parts[3:])
146                 require.Contains(t, parts[3], "was "+strconv.Itoa(len(want))+" bytes,")
147                 require.Contains(t, parts[3], "\n"+want[:512]+"| <- LIMIT")
148
149                 note, err := json.Marshal(map[string]string{
150                         "type": "input_text",
151                         "text": parts[3],
152                 })
153                 require.NoError(t, err)
154                 require.LessOrEqual(t, len(call.raw), len(server.Call(0).raw)+len(",")+len(note))
155         }
156 }
157
158 func TestCompactorKeepsTheShortestOfFiveLines(t *testing.T) {
159         sizes := []int{700, 530, 600, 520, 560, 513}
160         server := newFakeResponses(t)
161         server.reply = func(w http.ResponseWriter, call fakeCall) {
162                 writeEvents(w, completedEvent([]any{messageItem(strings.Repeat("z", sizes[call.number]))}, 100, 0, 5))
163         }
164
165         line, err := newTestCompactor(t, server).Summarize(t.Context(), leafTask())
166         require.NoError(t, err)
167         require.Equal(t, compactionTries, server.Calls())
168         require.Len(t, line, 520)
169 }
170
171 func TestCompactorFailsWithoutALine(t *testing.T) {
172         testCases := []struct {
173                 name  string
174                 reply func(http.ResponseWriter, fakeCall)
175                 want  string
176         }{
177                 {
178                         name: "failed call",
179                         reply: func(w http.ResponseWriter, call fakeCall) {
180                                 w.WriteHeader(http.StatusInternalServerError)
181                                 _, _ = io.WriteString(w, "down")
182                         },
183                         want: "HTTP 500",
184                 },
185                 {
186                         name: "tool call and no text",
187                         reply: func(w http.ResponseWriter, call fakeCall) {
188                                 writeEvents(w, completedEvent([]any{map[string]any{"type": "function_call", "name": "zoom", "arguments": "{}"}}, 100, 0, 5))
189                         },
190                         want: "no line",
191                 },
192         }
193
194         for _, tc := range testCases {
195                 t.Run(tc.name, func(t *testing.T) {
196                         server := newFakeResponses(t)
197                         server.reply = tc.reply
198
199                         _, err := newTestCompactor(t, server).Summarize(t.Context(), leafTask())
200                         require.ErrorContains(t, err, tc.want)
201                         require.Equal(t, 1, server.Calls())
202                 })
203         }
204 }
205
206 func TestCutLineKeepsWholeCharacters(t *testing.T) {
207         line := strings.Repeat("a", 511) + "é" + "tail"
208         cut := cutLine(line)
209         require.Equal(t, strings.Repeat("a", 511), cut)
210         require.True(t, utf8.ValidString(cut))
211         require.Equal(t, "short", cutLine("short"))
212 }
213
214 func TestCompactorBuildsNodesInTheCore(t *testing.T) {
215         server := newFakeResponses(t)
216         server.reply = func(w http.ResponseWriter, call fakeCall) {
217                 if call.number == 0 {
218                         w.WriteHeader(http.StatusServiceUnavailable)
219                         return
220                 }
221
222                 writeEvents(w, completedEvent([]any{messageItem("summary of the long output")}, 100, 0, 5))
223         }
224
225         memory, err := openMemory(t.Context(), t.TempDir(), newTestClock().Now, newTestCompactor(t, server))
226         require.NoError(t, err)
227
228         defer func() {
229                 require.NoError(t, memory.Close())
230         }()
231
232         _, err = memory.Append("echo", strings.Repeat("output ", 200))
233         require.NoError(t, err)
234         require.NoError(t, memory.Wait(t.Context()))
235         require.False(t, isBuilt(t, memory.tree, 0, 0))
236
237         _, err = memory.Append("user", "next")
238         require.NoError(t, err)
239         require.NoError(t, memory.Wait(t.Context()))
240
241         node := readNode(t, memory.tree, 0, 0)
242         require.Equal(t, "summary of the long output", node.Text)
243         require.Equal(t, "<chat>\n0+1|summary of the long output\n1+1|user: next\n</chat>", memory.View())
244         require.Equal(t, 2, server.Calls())
245 }
246
247 func newTestCompactor(t *testing.T, server *fakeResponses) *Compactor {
248         t.Helper()
249
250         client := newClient(server.URL, apiKey(testKey), "agent-key", filepath.Join(t.TempDir(), "usage.jsonl"))
251
252         return newCompactor(client, "Ada", ModelConfig{
253                 Model:  "gpt-test",
254                 Effort: "high",
255         })
256 }
257
258 func leafTask() Task {
259         return Task{
260                 Level:   0,
261                 Index:   0,
262                 Context: "<chat>\n</chat>",
263                 Input:   []string{"user: " + strings.Repeat("y", 600)},
264         }
265 }
266
267 // inputParts returns the texts of the one user message of call.
268 func inputParts(t *testing.T, call fakeCall) []string {
269         t.Helper()
270
271         var body struct {
272                 Input []struct {
273                         Role    string `json:"role"`
274                         Content []struct {
275                                 Type string `json:"type"`
276                                 Text string `json:"text"`
277                         } `json:"content"`
278                 } `json:"input"`
279         }
280
281         require.NoError(t, json.Unmarshal(call.raw, &body))
282         require.Len(t, body.Input, 1)
283         require.Equal(t, "user", body.Input[0].Role)
284
285         texts := make([]string, len(body.Input[0].Content))
286         for n, part := range body.Input[0].Content {
287                 require.Equal(t, "input_text", part.Type)
288                 texts[n] = part.Text
289         }
290
291         return texts
292 }