]> git.giacomolli.uk - onechat/blob - compactor_test.go
Build the summary nodes with a model
[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|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(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>\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         }, parts)
51 }
52
53 func TestCompactorWritesTheMergeTask(t *testing.T) {
54         text, err := compactionTask(Task{
55                 Level: 2,
56                 Index: 3,
57                 Input: []string{"left line", "right line"},
58         })
59         require.NoError(t, err)
60         require.Equal(t, "Compaction: merge lines 12+2 and 14+2, adjacent, into one line of at most\n"+
61                 "512 bytes (about 70 words), the length of this ruler:\n"+
62                 strings.Repeat("-", 512)+"\n"+
63                 "<chat> may hold their messages, 12 to 15, in more detail: take details\n"+
64                 "of them from there too.\n"+
65                 "<input>\nleft line\nright line\n</input>", text)
66 }
67
68 func TestCompactorRetriesALongLineWithTheNewestNote(t *testing.T) {
69         long := []string{strings.Repeat("a", 600), strings.Repeat("b", 550), strings.Repeat("c", 400)}
70         server := newFakeResponses(t)
71         server.reply = func(w http.ResponseWriter, call fakeCall) {
72                 writeEvents(w, completedEvent([]any{messageItem(long[call.number])}, 100, 0, 5))
73         }
74
75         line, err := newTestCompactor(t, server).Summarize(t.Context(), leafTask())
76         require.NoError(t, err)
77         require.Equal(t, long[2], line)
78         require.Equal(t, 3, server.Calls())
79
80         first := inputParts(t, server.Call(0))
81         require.Len(t, first, 2)
82         for n, want := range []string{long[0], long[1]} {
83                 call := server.Call(n + 1)
84                 parts := inputParts(t, call)
85                 require.Equal(t, first, parts[:2])
86                 require.Equal(t, []string{tooLong(want)}, parts[2:])
87                 require.Contains(t, parts[2], "was "+strconv.Itoa(len(want))+" bytes,")
88                 require.Contains(t, parts[2], "\n"+want[:512]+"| ← LIMIT")
89
90                 note, err := json.Marshal(map[string]string{
91                         "type": "input_text",
92                         "text": parts[2],
93                 })
94                 require.NoError(t, err)
95                 require.LessOrEqual(t, len(call.raw), len(server.Call(0).raw)+len(",")+len(note))
96         }
97 }
98
99 func TestCompactorKeepsTheShortestOfFiveLines(t *testing.T) {
100         sizes := []int{700, 530, 600, 520, 560, 513}
101         server := newFakeResponses(t)
102         server.reply = func(w http.ResponseWriter, call fakeCall) {
103                 writeEvents(w, completedEvent([]any{messageItem(strings.Repeat("z", sizes[call.number]))}, 100, 0, 5))
104         }
105
106         line, err := newTestCompactor(t, server).Summarize(t.Context(), leafTask())
107         require.NoError(t, err)
108         require.Equal(t, compactionTries, server.Calls())
109         require.Len(t, line, 520)
110 }
111
112 func TestCompactorFailsWithoutALine(t *testing.T) {
113         testCases := []struct {
114                 name  string
115                 reply func(http.ResponseWriter, fakeCall)
116                 want  string
117         }{
118                 {
119                         name: "failed call",
120                         reply: func(w http.ResponseWriter, call fakeCall) {
121                                 w.WriteHeader(http.StatusInternalServerError)
122                                 _, _ = io.WriteString(w, "down")
123                         },
124                         want: "HTTP 500",
125                 },
126                 {
127                         name: "tool call and no text",
128                         reply: func(w http.ResponseWriter, call fakeCall) {
129                                 writeEvents(w, completedEvent([]any{map[string]any{"type": "function_call", "name": "zoom", "arguments": "{}"}}, 100, 0, 5))
130                         },
131                         want: "no line",
132                 },
133         }
134
135         for _, tc := range testCases {
136                 t.Run(tc.name, func(t *testing.T) {
137                         server := newFakeResponses(t)
138                         server.reply = tc.reply
139
140                         _, err := newTestCompactor(t, server).Summarize(t.Context(), leafTask())
141                         require.ErrorContains(t, err, tc.want)
142                         require.Equal(t, 1, server.Calls())
143                 })
144         }
145 }
146
147 func TestCutLineKeepsWholeCharacters(t *testing.T) {
148         line := strings.Repeat("a", 511) + "é" + "tail"
149         cut := cutLine(line)
150         require.Equal(t, strings.Repeat("a", 511), cut)
151         require.True(t, utf8.ValidString(cut))
152         require.Equal(t, "short", cutLine("short"))
153 }
154
155 func TestCompactorBuildsNodesInTheCore(t *testing.T) {
156         server := newFakeResponses(t)
157         server.reply = func(w http.ResponseWriter, call fakeCall) {
158                 if call.number == 0 {
159                         w.WriteHeader(http.StatusServiceUnavailable)
160                         return
161                 }
162
163                 writeEvents(w, completedEvent([]any{messageItem("summary of the long output")}, 100, 0, 5))
164         }
165
166         memory, err := openMemory(t.Context(), t.TempDir(), newTestClock().Now, newTestCompactor(t, server))
167         require.NoError(t, err)
168
169         defer func() {
170                 require.NoError(t, memory.Close())
171         }()
172
173         _, err = memory.Append("echo", strings.Repeat("output ", 200))
174         require.NoError(t, err)
175         require.NoError(t, memory.Wait(t.Context()))
176         require.False(t, isBuilt(t, memory.tree, 0, 0))
177
178         _, err = memory.Append("user", "next")
179         require.NoError(t, err)
180         require.NoError(t, memory.Wait(t.Context()))
181
182         node := readNode(t, memory.tree, 0, 0)
183         require.Equal(t, "summary of the long output", node.Text)
184         require.Equal(t, "<chat>\n0+1|summary of the long output\n1+1|user: next\n</chat>", memory.View())
185         require.Equal(t, 2, server.Calls())
186 }
187
188 func newTestCompactor(t *testing.T, server *fakeResponses) *Compactor {
189         t.Helper()
190
191         client := newClient(server.URL, apiKey(testKey), "agent-key", filepath.Join(t.TempDir(), "usage.jsonl"))
192
193         return newCompactor(client, "Ada", ModelConfig{
194                 Model:  "gpt-test",
195                 Effort: "high",
196         })
197 }
198
199 func leafTask() Task {
200         return Task{
201                 Level:   0,
202                 Index:   0,
203                 Context: "<chat>\n</chat>",
204                 Input:   []string{"user: " + strings.Repeat("y", 600)},
205         }
206 }
207
208 // inputParts returns the texts of the one user message of call.
209 func inputParts(t *testing.T, call fakeCall) []string {
210         t.Helper()
211
212         var body struct {
213                 Input []struct {
214                         Role    string `json:"role"`
215                         Content []struct {
216                                 Type string `json:"type"`
217                                 Text string `json:"text"`
218                         } `json:"content"`
219                 } `json:"input"`
220         }
221
222         require.NoError(t, json.Unmarshal(call.raw, &body))
223         require.Len(t, body.Input, 1)
224         require.Equal(t, "user", body.Input[0].Role)
225
226         texts := make([]string, len(body.Input[0].Content))
227         for n, part := range body.Input[0].Content {
228                 require.Equal(t, "input_text", part.Type)
229                 texts[n] = part.Text
230         }
231
232         return texts
233 }