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