git.giacomolli.uk - onechat/blob - 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", "/AGENT/instructions.md", ""), 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 TestCompactorWritesTheMergeTask(t *testing.T) {
  54         text, err := compactorTask(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 := newModelFake(t)
  71         server.reply = func(w http.ResponseWriter, call ModelFakeCall) {
  72                 modelWriteEvents(w, modelCompletedEvent([]any{modelMessageItem(long[call.number])}, 100, 0, 5))
  73         }
  74
  75         task := compactorLeafTask()
  76         task.Context = "<chat>\n0+1|a\n1+1|b\n2+1|c\n3+1|d\n4+1|e\n</chat>"
  77
  78         compactor := newTestCompactor(t, server)
  79         line, err := compactor.Summarize(t.Context(), task)
  80         require.NoError(t, err)
  81         require.Equal(t, long[2], line)
  82         require.Equal(t, 3, server.Calls())
  83
  84         records := modelReadUsage(t, compactor.client.usage.path)
  85         require.Len(t, records, 3)
  86
  87         for n, record := range records {
  88                 require.Equal(t, &[2]int{0, 0}, record.Node)
  89                 require.Equal(t, len(long[n]), record.TextBytes)
  90         }
  91
  92         text, err := compactorTask(task)
  93         require.NoError(t, err)
  94
  95         first := compactorInputTexts(t, server.Call(0))
  96         require.Equal(t, []string{task.Context + "\n" + text}, first)
  97
  98         for n, want := range []string{long[0], long[1]} {
  99                 call := server.Call(n + 1)
 100                 texts := compactorInputTexts(t, call)
 101                 require.Equal(t, first, texts[:1])
 102                 require.Equal(t, []string{compactorTooLong(want)}, texts[1:])
 103                 require.Contains(t, texts[1], "was "+strconv.Itoa(len(want))+" bytes,")
 104                 require.Contains(t, texts[1], "\n"+want[:512]+"| <- LIMIT")
 105
 106                 note, err := json.Marshal(modelUserMessage(texts[1]))
 107                 require.NoError(t, err)
 108                 require.LessOrEqual(t, len(call.raw), len(server.Call(0).raw)+len(",")+len(note))
 109         }
 110 }
 111
 112 func TestCompactorKeepsTheShortestOfFiveLines(t *testing.T) {
 113         sizes := []int{700, 530, 600, 520, 560, 513}
 114         server := newModelFake(t)
 115         server.reply = func(w http.ResponseWriter, call ModelFakeCall) {
 116                 modelWriteEvents(w, modelCompletedEvent([]any{modelMessageItem(strings.Repeat("z", sizes[call.number]))}, 100, 0, 5))
 117         }
 118
 119         line, err := newTestCompactor(t, server).Summarize(t.Context(), compactorLeafTask())
 120         require.NoError(t, err)
 121         require.Equal(t, compactorTries, server.Calls())
 122         require.Len(t, line, 520)
 123 }
 124
 125 func TestCompactorFailsWithoutALine(t *testing.T) {
 126         testCases := []struct {
 127                 name  string
 128                 reply func(http.ResponseWriter, ModelFakeCall)
 129                 want  string
 130         }{
 131                 {
 132                         name: "failed call",
 133                         reply: func(w http.ResponseWriter, _ ModelFakeCall) {
 134                                 w.WriteHeader(http.StatusInternalServerError)
 135
 136                                 _, _ = io.WriteString(w, "down")
 137                         },
 138                         want: "HTTP 500",
 139                 },
 140                 {
 141                         name: "tool call and no text",
 142                         reply: func(w http.ResponseWriter, _ ModelFakeCall) {
 143                                 modelWriteEvents(w, modelCompletedEvent([]any{map[string]any{"type": "function_call", "name": "zoom", "arguments": "{}"}}, 100, 0, 5))
 144                         },
 145                         want: "no line",
 146                 },
 147         }
 148
 149         for _, tc := range testCases {
 150                 t.Run(tc.name, func(t *testing.T) {
 151                         server := newModelFake(t)
 152                         server.reply = tc.reply
 153
 154                         _, err := newTestCompactor(t, server).Summarize(t.Context(), compactorLeafTask())
 155                         require.ErrorContains(t, err, tc.want)
 156                         require.Equal(t, 1, server.Calls())
 157                 })
 158         }
 159 }
 160
 161 func TestCutLineKeepsWholeCharacters(t *testing.T) {
 162         line := strings.Repeat("a", 511) + "é" + "tail"
 163         cut := compactorCutLine(line)
 164         require.Equal(t, strings.Repeat("a", 511), cut)
 165         require.True(t, utf8.ValidString(cut))
 166         require.Equal(t, "short", compactorCutLine("short"))
 167 }
 168
 169 func TestCompactorBuildsNodesInTheCore(t *testing.T) {
 170         server := newModelFake(t)
 171         server.reply = func(w http.ResponseWriter, call ModelFakeCall) {
 172                 if call.number == 0 {
 173                         w.WriteHeader(http.StatusServiceUnavailable)
 174
 175                         return
 176                 }
 177
 178                 modelWriteEvents(w, modelCompletedEvent([]any{modelMessageItem("summary of the long output")}, 100, 0, 5))
 179         }
 180
 181         memory, err := openMemory(t.Context(), t.TempDir(), newTestClock().Now, newTestCompactor(t, server), discardLog())
 182         require.NoError(t, err)
 183
 184         defer func() {
 185                 require.NoError(t, memory.Close())
 186         }()
 187
 188         _, err = memory.Append("echo", strings.Repeat("output ", 200))
 189         require.NoError(t, err)
 190         require.NoError(t, memory.Wait(t.Context()))
 191         require.False(t, treeIsBuilt(t, memory.tree, 0, 0))
 192
 193         _, err = memory.Append("user", "next")
 194         require.NoError(t, err)
 195         require.NoError(t, memory.Wait(t.Context()))
 196
 197         node := treeReadNode(t, memory.tree, 0, 0)
 198         require.Equal(t, "summary of the long output", node.Text)
 199         require.Equal(t, "<chat>\n0+1|summary of the long output\n1+1|user: next\n</chat>", memory.View())
 200         require.Equal(t, 2, server.Calls())
 201 }
 202
 203 func newTestCompactor(t *testing.T, server *ModelFake) *Compactor {
 204         t.Helper()
 205
 206         client := newModelClient(server.URL, ModelAPIKey(modelTestKey), "agent-key", filepath.Join(t.TempDir(), "usage.jsonl"))
 207
 208         return newCompactor(client, compactorAdaPrompt, ModelConfig{
 209                 Model:  "gpt-test",
 210                 Effort: "high",
 211         })
 212 }
 213
 214 // compactorInputTexts returns the text of each user message of call. Each
 215 // message has one text part.
 216 func compactorInputTexts(t *testing.T, call ModelFakeCall) []string {
 217         t.Helper()
 218
 219         var body struct {
 220                 Input []struct {
 221                         Role    string `json:"role"`
 222                         Content []struct {
 223                                 Type string `json:"type"`
 224                                 Text string `json:"text"`
 225                         } `json:"content"`
 226                 } `json:"input"`
 227         }
 228
 229         require.NoError(t, json.Unmarshal(call.raw, &body))
 230
 231         texts := make([]string, len(body.Input))
 232         for n, message := range body.Input {
 233                 require.Equal(t, "user", message.Role)
 234                 require.Len(t, message.Content, 1)
 235                 require.Equal(t, "input_text", message.Content[0].Type)
 236
 237                 texts[n] = message.Content[0].Text
 238         }
 239
 240         return texts
 241 }
 242
 243 func compactorLeafTask() Task {
 244         return Task{
 245                 Level:   0,
 246                 Index:   0,
 247                 Context: "<chat>\n</chat>",
 248                 Input:   []string{"user: " + strings.Repeat("y", 600)},
 249         }
 250 }
 251
 252 // compactorAdaPrompt returns the prompt of the agent Ada, without instructions.
 253 func compactorAdaPrompt() (string, error) {
 254         return prompt("Ada", "/AGENT/instructions.md", ""), nil
 255 }