git.giacomolli.uk - onechat/blob - daemon_children_test.go

   1 package main
   2
   3 import (
   4         "context"
   5         "encoding/json"
   6         "net/http"
   7         "os"
   8         "path/filepath"
   9         "strings"
  10         "testing"
  11         "time"
  12
  13         "github.com/stretchr/testify/require"
  14 )
  15
  16 func TestDaemonToolsCreateAndTellAChildDuringATurn(t *testing.T) {
  17         dir := t.TempDir()
  18         server := newModelFake(t)
  19         agents, err := openDaemonAgents(t.Context(), dir, server.URL, ModelAPIKey(modelTestKey), defaultConfig().Compactor, newTestClock().Now, discardLog())
  20         require.NoError(t, err)
  21
  22         t.Cleanup(func() {
  23                 require.NoError(t, agents.Close())
  24         })
  25
  26         root := agents[0]
  27         next := 0
  28         server.reply = func(w http.ResponseWriter, call ModelFakeCall) {
  29                 if call.body["prompt_cache_key"] != root.agent.CacheKey() {
  30                         modelWriteEvents(w, modelCompletedEvent([]any{modelMessageItem("Approve the gate size.")}, 100, 0, 5))
  31
  32                         return
  33                 }
  34
  35                 next++
  36
  37                 var items []any
  38
  39                 switch next {
  40                 case 1:
  41                         items = []any{modelFunctionCallItem("fc_1", "call_1", "agent_create", `{"name":"House","responsibility":"Keep garden decisions."}`)}
  42                 case 2:
  43                         items = []any{modelFunctionCallItem("fc_2", "call_2", "tell", `{"name":"House","message":"Check the gate size."}`)}
  44                 case 3:
  45                         items = []any{modelFunctionCallItem("fc_3", "call_3", "zoom", `{"id":0,"n":1,"name":"House"}`)}
  46                 default:
  47                         items = []any{modelMessageItem("I will ask the user.")}
  48                 }
  49
  50                 modelWriteEvents(w, modelCompletedEvent(items, 100, 0, 5))
  51         }
  52         whatsapp, _, _ := whatsappStartTest(t, dir, newTestClock())
  53         listener, err := daemonListenSocket(filepath.Join(dir, socketFileName))
  54         require.NoError(t, err)
  55
  56         ctx, cancel := context.WithCancel(t.Context())
  57         defer cancel()
  58
  59         done := make(chan error, 1)
  60         go func() {
  61                 done <- daemonServe(ctx, listener, nil, whatsapp, root.schedules, func(notify func(Event)) (*TurnLoop, error) {
  62                         return root.Loop(whatsapp, notify)
  63                 })
  64         }()
  65
  66         t.Cleanup(func() {
  67                 cancel()
  68                 require.NoError(t, <-done)
  69         })
  70
  71         client := terminalStartTestClient(t, listener.Addr().String())
  72         client.WaitFor(t, "[connected to OneChat]")
  73         client.Type("Create a child and ask it about the gate.")
  74         client.WaitFor(t, "[House] Approve the gate size.")
  75         client.WaitFor(t, "I will ask the user.")
  76         require.Contains(t, client.Output(), "Started House.")
  77         require.Contains(t, client.Output(), "user: Check the gate size.")
  78
  79         child, err := openAgent(filepath.Join(root.agent.dir, "children", "House"))
  80         require.NoError(t, err)
  81         require.Equal(t, root.agent.config.Model, child.config.Model)
  82         require.Equal(t, root.agent.config.Effort, child.config.Effort)
  83         require.Equal(t, root.agent.config.Workdir, child.config.Workdir)
  84         require.Equal(t, root.agent.config.Name, child.config.Parent)
  85         require.NotEqual(t, root.agent.CacheKey(), child.CacheKey())
  86
  87         client.Type("Show the garden responsibility.")
  88         require.Eventually(t, func() bool {
  89                 for i := range server.Calls() {
  90                         call := server.Call(i)
  91                         texts := turnInputTexts(t, call)
  92                         if len(texts) > 0 && texts[len(texts)-1] == "Show the garden responsibility." {
  93                                 return strings.Contains(strings.Join(texts, "\n"), "Children:\nHouse: Keep garden decisions.")
  94                         }
  95                 }
  96
  97                 return false
  98         }, 5*time.Second, time.Millisecond)
  99 }
 100
 101 func TestChildCreationRefusesUsedNames(t *testing.T) {
 102         test := newDaemonChildrenTest(t)
 103         box := test.root.tools
 104         require.Contains(t, box.run(t.Context(), "agent_create", `{"name":"OneChat","responsibility":"Other plans."}`), "already used")
 105
 106         require.NoDirExists(t, filepath.Join(test.root.agent.dir, "children"))
 107         require.Contains(t, box.run(t.Context(), "agent_create", `{"name":"House","responsibility":" "}`), "responsibility")
 108         require.Contains(t, box.run(t.Context(), "agent_create", `{"name":"House","responsibility":"Keep garden plans."}`), "Started House.")
 109         require.Contains(t, box.run(t.Context(), "agent_create", `{"name":"House","responsibility":"Other plans."}`), "already used")
 110
 111         test.root.runs.subagents.draw = func() string {
 112                 return "BlueFox"
 113         }
 114         subagent, err := test.root.runs.subagents.Create(t.Context())
 115         require.NoError(t, err)
 116         require.Equal(t, "BlueFox", subagent.name)
 117         require.Contains(t, box.run(t.Context(), "agent_create", `{"name":"BlueFox","responsibility":"Other plans."}`), "already used by a subagent")
 118         require.NoError(t, test.root.runs.subagents.Reserve("Calendar"))
 119         require.Contains(t, box.run(t.Context(), "agent_create", `{"name":"Calendar","responsibility":"Keep dates."}`), "already used")
 120 }
 121
 122 func TestChildCreationAcceptsAreaNames(t *testing.T) {
 123         testCases := []struct {
 124                 name  string
 125                 child string
 126         }{
 127                 {
 128                         name:  "house area",
 129                         child: "House",
 130                 },
 131                 {
 132                         name:  "calendar area",
 133                         child: "Calendar",
 134                 },
 135                 {
 136                         name:  "two letters",
 137                         child: "Ab",
 138                 },
 139                 {
 140                         name:  "capital letters after the first",
 141                         child: "ABcD",
 142                 },
 143                 {
 144                         name:  "thirty-two letters",
 145                         child: "A" + strings.Repeat("b", 31),
 146                 },
 147         }
 148         for _, tc := range testCases {
 149                 t.Run(tc.name, func(t *testing.T) {
 150                         test := newDaemonChildrenTest(t)
 151                         arguments, err := json.Marshal(ToolAgentCreateArguments{
 152                                 Name:           tc.child,
 153                                 Responsibility: "Keep area decisions.",
 154                         })
 155                         require.NoError(t, err)
 156                         require.Contains(t, test.root.tools.run(t.Context(), "agent_create", string(arguments)), "Started "+tc.child+".")
 157
 158                         child := test.runtime.messages.Child(test.root.agent.config.Name, tc.child)
 159                         require.NotNil(t, child)
 160                         require.Equal(t, tc.child, child.agent.config.Name)
 161                         require.FileExists(t, filepath.Join(child.agent.dir, "config.json"))
 162                 })
 163         }
 164 }
 165
 166 func TestChildCreationRefusesInvalidAreaNames(t *testing.T) {
 167         testCases := []struct {
 168                 name  string
 169                 child string
 170         }{
 171                 {
 172                         name:  "empty name",
 173                         child: "",
 174                 },
 175                 {
 176                         name:  "one letter",
 177                         child: "H",
 178                 },
 179                 {
 180                         name:  "thirty-three letters",
 181                         child: "A" + strings.Repeat("b", 32),
 182                 },
 183                 {
 184                         name:  "lowercase initial",
 185                         child: "house",
 186                 },
 187                 {
 188                         name:  "two words",
 189                         child: "My House",
 190                 },
 191                 {
 192                         name:  "digit",
 193                         child: "House2",
 194                 },
 195                 {
 196                         name:  "path",
 197                         child: "../House",
 198                 },
 199                 {
 200                         name:  "punctuation",
 201                         child: "House-",
 202                 },
 203                 {
 204                         name:  "non-ASCII initial",
 205                         child: "\u00c1rea",
 206                 },
 207                 {
 208                         name:  "non-ASCII remainder",
 209                         child: "Hous\u00e9",
 210                 },
 211         }
 212         for _, tc := range testCases {
 213                 t.Run(tc.name, func(t *testing.T) {
 214                         test := newDaemonChildrenTest(t)
 215                         arguments, err := json.Marshal(ToolAgentCreateArguments{
 216                                 Name:           tc.child,
 217                                 Responsibility: "Keep area decisions.",
 218                         })
 219                         require.NoError(t, err)
 220                         require.Contains(t, test.root.tools.run(t.Context(), "agent_create", string(arguments)), "name must have 2 to 32 ASCII letters and start with a capital letter")
 221                         require.NoDirExists(t, filepath.Join(test.root.agent.dir, "children"))
 222                 })
 223         }
 224 }
 225
 226 func TestFailedChildPublicationKeepsStartupAndRetryWorking(t *testing.T) {
 227         if os.Geteuid() == 0 {
 228                 t.Skip("requires directory write permissions")
 229         }
 230
 231         dir := t.TempDir()
 232         root, err := openAgent(filepath.Join(dir, daemonAgentDirName))
 233         require.NoError(t, err)
 234
 235         subagents := newSubagents(filepath.Join(root.dir, subagentDirName), time.Now, subagentNameRandom)
 236         defer func() {
 237                 require.NoError(t, subagents.Close())
 238         }()
 239
 240         children := filepath.Join(root.dir, "children")
 241         require.NoError(t, os.Mkdir(children, 0o500))
 242         t.Cleanup(func() {
 243                 require.NoError(t, os.Chmod(children, 0o700))
 244         })
 245
 246         _, err = root.CreateChild("GreenOtter", "Keep garden plans.", subagents)
 247         require.ErrorContains(t, err, "publish child directory")
 248         require.NoDirExists(t, filepath.Join(children, "GreenOtter"))
 249
 250         staged, err := filepath.Glob(filepath.Join(root.dir, ".child-*"))
 251         require.NoError(t, err)
 252         require.Empty(t, staged)
 253
 254         server := newModelFake(t)
 255         reopened, err := openDaemonAgents(t.Context(), dir, server.URL, ModelAPIKey(modelTestKey), defaultConfig().Compactor, time.Now, discardLog())
 256         require.NoError(t, err)
 257         require.Len(t, reopened, 1)
 258
 259         defer func() {
 260                 require.NoError(t, reopened.Close())
 261         }()
 262
 263         require.NoError(t, os.Chmod(children, 0o700))
 264
 265         child, err := root.CreateChild("GreenOtter", "Keep garden plans.", subagents)
 266         require.NoError(t, err)
 267         require.Equal(t, filepath.Join(children, "GreenOtter"), child.dir)
 268 }
 269
 270 func TestInterruptedChildPreparationDoesNotStopStartup(t *testing.T) {
 271         dir := t.TempDir()
 272         root, err := openAgent(filepath.Join(dir, daemonAgentDirName))
 273         require.NoError(t, err)
 274
 275         staged, err := os.MkdirTemp(root.dir, ".child-*")
 276         require.NoError(t, err)
 277         require.NoError(t, os.WriteFile(filepath.Join(staged, ".config.json.partial"), []byte("{"), 0o600))
 278
 279         server := newModelFake(t)
 280         agents, err := openDaemonAgents(t.Context(), dir, server.URL, ModelAPIKey(modelTestKey), defaultConfig().Compactor, time.Now, discardLog())
 281         require.NoError(t, err)
 282
 283         defer func() {
 284                 require.NoError(t, agents.Close())
 285         }()
 286
 287         require.Len(t, agents, 1)
 288
 289         child, err := root.CreateChild("GreenOtter", "Keep garden plans.", agents[0].subagents)
 290         require.NoError(t, err)
 291
 292         children, err := root.Children()
 293         require.NoError(t, err)
 294         require.Len(t, children, 1)
 295         require.Equal(t, child.config, children[0].config)
 296         require.DirExists(t, staged)
 297 }
 298
 299 func TestChildPublicationDoesNotReplaceAnExistingDirectory(t *testing.T) {
 300         test := newDaemonChildrenTest(t)
 301         path := filepath.Join(test.root.agent.dir, "children", "GreenOtter")
 302         require.NoError(t, os.MkdirAll(path, 0o700))
 303
 304         _, err := test.runtime.Create("GreenOtter", "Keep garden plans.")
 305         require.ErrorContains(t, err, "publish child directory")
 306         require.NoFileExists(t, filepath.Join(path, "config.json"))
 307         require.NoError(t, test.root.runs.subagents.Reserve("GreenOtter"))
 308         test.root.runs.subagents.Release("GreenOtter")
 309 }
 310
 311 func TestChildNamesStayOutOfSubagentDraws(t *testing.T) {
 312         test := newDaemonChildrenTest(t)
 313         _, err := test.runtime.Create("GreenOtter", "Keep garden plans.")
 314         require.NoError(t, err)
 315
 316         names := []string{"GreenOtter", "BlueFox"}
 317         test.root.runs.subagents.draw = func() string {
 318                 name := names[0]
 319                 names = names[1:]
 320
 321                 return name
 322         }
 323         subagent, err := test.root.runs.subagents.Create(t.Context())
 324         require.NoError(t, err)
 325         require.Equal(t, "BlueFox", subagent.name)
 326 }
 327
 328 func TestChildToolsExcludeCreationAndWhatsApp(t *testing.T) {
 329         test := newDaemonChildrenTest(t)
 330         _, err := test.runtime.Create("GreenOtter", "Keep garden plans.")
 331         require.NoError(t, err)
 332
 333         child := test.runtime.messages.Child(test.root.agent.config.Name, "GreenOtter")
 334         require.NotNil(t, child)
 335
 336         names := promptToolNames(child.agent.Tools())
 337         require.NotContains(t, names, "agent_create")
 338         require.NotContains(t, names, "whatsapp_send")
 339         require.NotContains(t, names, "whatsapp_find")
 340         require.Contains(t, names, "spawn")
 341         require.Contains(t, names, "schedule_set")
 342         require.Contains(t, names, "schedule_list")
 343         require.Contains(t, names, "schedule_remove")
 344         require.Equal(t, `error: unknown tool "agent_create"`, child.tools.run(t.Context(), "agent_create", `{"name":"BlueFox","responsibility":"Other plans."}`))
 345         require.Nil(t, child.tools.messages.Child("GreenOtter", "OneChat"))
 346 }
 347
 348 func TestZoomReadsTheChildViewAndKeepsSubagentLogs(t *testing.T) {
 349         test := newDaemonChildrenTest(t)
 350         _, err := test.runtime.Create("GreenOtter", "Keep garden plans.")
 351         require.NoError(t, err)
 352
 353         child := test.runtime.messages.Child(test.root.agent.config.Name, "GreenOtter")
 354         _, err = child.memory.Append("note", "The gate is narrow.")
 355         require.NoError(t, err)
 356         require.NoError(t, child.memory.Settle(t.Context()))
 357
 358         output := test.root.tools.run(t.Context(), "zoom", `{"id":0,"n":1,"name":"GreenOtter"}`)
 359         require.Equal(t, strings.Join(child.memory.ViewLines(nil), ""), output)
 360         require.Contains(t, output, "note: The gate is narrow.")
 361
 362         test.root.runs.subagents.draw = func() string {
 363                 return "BlueFox"
 364         }
 365         subagent, err := test.root.runs.subagents.Create(t.Context())
 366         require.NoError(t, err)
 367
 368         _, err = subagent.Append("note", "A short task.")
 369         require.NoError(t, err)
 370         require.Equal(t, "0|note: A short task.", test.root.tools.run(t.Context(), "zoom", `{"id":0,"n":1,"name":"BlueFox"}`))
 371 }
 372
 373 func TestRuntimeChildJoinsBackupLocksAndStopsWithDaemon(t *testing.T) {
 374         test := newDaemonChildrenTest(t)
 375         _, err := test.runtime.Create("GreenOtter", "Keep garden plans.")
 376         require.NoError(t, err)
 377
 378         child := test.runtime.messages.Child(test.root.agent.config.Name, "GreenOtter")
 379         _, err = child.memory.Append("note", "Keep the gate plan.")
 380         require.NoError(t, err)
 381
 382         appended := make(chan error, 1)
 383         require.NoError(t, test.runtime.messages.SnapshotAgents(t.Context(), func() error {
 384                 require.FileExists(t, filepath.Join(child.agent.dir, "view.json"))
 385
 386                 go func() {
 387                         _, err := child.memory.Append("note", "A later plan.")
 388                         appended <- err
 389                 }()
 390
 391                 select {
 392                 case err := <-appended:
 393                         require.FailNow(t, "child append bypasses snapshot lock", "%v", err)
 394                 case <-time.After(20 * time.Millisecond):
 395                 }
 396
 397                 return nil
 398         }))
 399         require.NoError(t, <-appended)
 400         test.cancel()
 401         require.NoError(t, test.runtime.Close())
 402
 403         test.runtime.opened = nil
 404         reopened, err := test.root.openChild(t.Context(), child.agent)
 405         require.NoError(t, err)
 406         require.Equal(t, 2, reopened.memory.Len())
 407         require.NoError(t, reopened.Close())
 408 }
 409
 410 func TestRootStateListsChildResponsibilitiesOnlyWhenPresent(t *testing.T) {
 411         test := newDaemonChildrenTest(t)
 412         before := test.root.state()
 413         require.NotContains(t, before, "Children:")
 414
 415         _, err := test.runtime.Create("GreenOtter", "Keep garden plans.")
 416         require.NoError(t, err)
 417
 418         _, err = test.runtime.Create("BlueFox", "Keep book notes.")
 419         require.NoError(t, err)
 420         require.Equal(t, before+"\n\nChildren:\nBlueFox: Keep book notes.\nGreenOtter: Keep garden plans.", test.root.state())
 421
 422         child := test.runtime.messages.Child(test.root.agent.config.Name, "GreenOtter")
 423         require.NotContains(t, child.state(), "Children:")
 424 }
 425
 426 // DaemonChildrenTest owns one root and runtime children without a root turn.
 427 type DaemonChildrenTest struct {
 428         root    *TurnLoop
 429         runtime *DaemonChildren
 430         server  *ModelFake
 431         cancel  context.CancelFunc
 432 }
 433
 434 func newDaemonChildrenTest(t *testing.T) *DaemonChildrenTest {
 435         t.Helper()
 436
 437         test := &DaemonChildrenTest{
 438                 server: newModelFake(t),
 439         }
 440         agents, err := openDaemonAgents(t.Context(), t.TempDir(), test.server.URL, ModelAPIKey(modelTestKey), defaultConfig().Compactor, time.Now, discardLog())
 441         require.NoError(t, err)
 442
 443         test.root, err = agents[0].Loop(nil, func(Event) {
 444         })
 445         require.NoError(t, err)
 446
 447         messages, err := openDaemonMessages([]*TurnLoop{test.root})
 448         require.NoError(t, err)
 449
 450         ctx, cancel := context.WithCancel(t.Context())
 451         test.cancel = cancel
 452         test.runtime = &DaemonChildren{
 453                 ctx:      ctx,
 454                 cancel:   cancel,
 455                 root:     test.root,
 456                 messages: messages,
 457         }
 458         test.root.tools.children = test.runtime
 459         t.Cleanup(func() {
 460                 cancel()
 461                 require.NoError(t, test.runtime.Close())
 462                 test.root.runs.Close()
 463                 require.NoError(t, agents.Close())
 464         })
 465
 466         return test
 467 }