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

   1 package main
   2
   3 import (
   4         "context"
   5         "io"
   6         "log/slog"
   7         "net/http"
   8         "os"
   9         "path/filepath"
  10         "strconv"
  11         "strings"
  12         "sync/atomic"
  13         "testing"
  14         "time"
  15
  16         "github.com/stretchr/testify/require"
  17 )
  18
  19 // subagentTestFullContext is the error body of a call whose input does not fit
  20 // in the context of the model.
  21 const subagentTestFullContext = `{"error": {"message": "Your input exceeds the context window of this model.", "type": "invalid_request_error", "code": "context_length_exceeded"}}`
  22
  23 func TestSpawnRunsASubagentInTheBackground(t *testing.T) {
  24         server := newModelFake(t)
  25
  26         var test *TurnTestLoop
  27
  28         server.reply = func(w http.ResponseWriter, call ModelFakeCall) {
  29                 last := subagentTestLastItem(call)
  30                 switch {
  31                 case subagentTestCall(call) && last["type"] == "function_call_output":
  32                         // The report comes after the reply of the turn of the
  33                         // spawn.
  34                         test.events.WaitFor(t, "reply", 1)
  35                         modelWriteEvents(w, modelCompletedEvent([]any{modelMessageItem("There are 3 files.")}, 100, 0, 5))
  36                 case subagentTestCall(call):
  37                         modelWriteEvents(w, modelCompletedEvent([]any{modelFunctionCallItem("fc_1", "call_1", "shell", `{"command":"pwd"}`)}, 100, 0, 5))
  38                 case last["type"] == "function_call_output":
  39                         modelWriteEvents(w, modelCompletedEvent([]any{modelMessageItem("Started.")}, 100, 0, 5))
  40                 case strings.HasPrefix(subagentTestText(last), "["):
  41                         modelWriteEvents(w, modelCompletedEvent([]any{modelMessageItem("Noted.")}, 100, 0, 5))
  42                 default:
  43                         modelWriteEvents(w, modelCompletedEvent([]any{modelFunctionCallItem("fc_1", "call_1", "spawn", `{"tasks":["Count the files."],"effort":null}`)}, 100, 0, 5))
  44                 }
  45         }
  46
  47         test = newTestTurnLoop(t, server)
  48         _, err := test.loop.Submit(terminalSource, "Count the files in a subagent.")
  49         require.NoError(t, err)
  50         test.events.WaitFor(t, "idle", 2)
  51
  52         name := test.events.Last("subagent")
  53         require.True(t, subagentNameValid(name), name)
  54
  55         work := filepath.Join(test.loop.runs.subagents.dir, name, "work")
  56         require.DirExists(t, work)
  57
  58         report := "[" + name + "]\nThere are 3 files.\n\nFull chat: zoom(\"" + name + "\")"
  59         require.Equal(t, []string{
  60                 "user: Count the files in a subagent.",
  61                 `tool: spawn {"tasks":["Count the files."],"effort":null}`,
  62                 "echo: Started " + name + ". Each one's report will reach you as a message when it finishes.",
  63                 "reply: Started.",
  64                 "work: " + report,
  65                 "reply: Noted.",
  66         }, test.Log(t))
  67
  68         require.Equal(t, Event{
  69                 Kind:   "work",
  70                 ID:     4,
  71                 Text:   report,
  72                 Source: name,
  73                 Status: subagentFinished,
  74         }, test.events.LastEvent("work"))
  75
  76         require.Equal(t, "<6>subagent started name="+name+"\n<6>subagent done name="+name+" after=0s\n", test.logs.String())
  77
  78         var first, second *ModelFakeCall
  79
  80         for n := range server.Calls() {
  81                 call := server.Call(n)
  82                 if !subagentTestCall(call) {
  83                         continue
  84                 }
  85
  86                 if first == nil {
  87                         first = &call
  88                 } else {
  89                         second = &call
  90                 }
  91         }
  92
  93         require.NotNil(t, second)
  94
  95         // The first request of the subagent holds the view, the state, and its
  96         // whole log, with its own prompt and tools.
  97         texts := turnInputTexts(t, *first)
  98         view := strings.Join(texts[:len(texts)-2], "")
  99         require.True(t, strings.HasPrefix(view, "<chat>\n0+1|user: Count the files in a subagent.\n1+1|tool: spawn"), view)
 100         require.Equal(t, "</chat>\n\nNow: Thursday 2026-01-01 12:00 +00:00", texts[len(texts)-2])
 101         require.Equal(t, "0|note: subagent\n\n1|user: Count the files.", texts[len(texts)-1])
 102
 103         instructions, err := test.agent.SubagentPrompt()
 104         require.NoError(t, err)
 105         require.Equal(t, instructions, first.body["instructions"])
 106
 107         var tools []string
 108
 109         for _, tool := range first.body["tools"].([]any) {
 110                 tools = append(tools, tool.(map[string]any)["name"].(string))
 111         }
 112
 113         require.Equal(t, []string{"zoom", "date", "shell", "read", "write", "edit"}, tools)
 114         require.Equal(t, work+"\n", subagentTestLastItem(*second)["output"])
 115
 116         subagent, err := test.loop.runs.subagents.Get(t.Context(), name)
 117         require.NoError(t, err)
 118
 119         log, err := subagent.Show()
 120         require.NoError(t, err)
 121         require.Equal(t, "0|note: subagent\n\n1|user: Count the files.\n\n2|tool: shell {\"command\":\"pwd\"}\n\n3|echo: "+work+"\n\n\n4|reply: There are 3 files.", log)
 122
 123         var agents []string
 124
 125         for _, record := range modelReadUsage(t, test.usagePath) {
 126                 agents = append(agents, record.Agent)
 127         }
 128
 129         require.ElementsMatch(t, []string{"", "", "", name, name}, agents)
 130 }
 131
 132 func TestReportJoinsTheNextRequestOfARunningTurn(t *testing.T) {
 133         server := newModelFake(t)
 134         arrived := make(chan struct{})
 135         release := make(chan struct{})
 136         server.reply = func(w http.ResponseWriter, call ModelFakeCall) {
 137                 if call.number == 0 {
 138                         close(arrived)
 139                         <-release
 140                         modelWriteEvents(w, modelCompletedEvent([]any{modelFunctionCallItem("fc_1", "call_1", "date", `{"id":0}`)}, 100, 0, 5))
 141
 142                         return
 143                 }
 144
 145                 modelWriteEvents(w, modelCompletedEvent([]any{modelMessageItem("done")}, 100, 0, 5))
 146         }
 147
 148         test := newTestTurnLoop(t, server)
 149         _, err := test.loop.Submit(terminalSource, "first")
 150         require.NoError(t, err)
 151
 152         <-arrived
 153
 154         report := subagentReport("GreenOtter", []string{"It is done."})
 155         test.loop.Report("GreenOtter", report, subagentFinished)
 156         close(release)
 157
 158         test.events.WaitFor(t, "idle", 1)
 159         require.Equal(t, report, subagentTestText(subagentTestLastItem(server.Call(1))))
 160
 161         time.Sleep(50 * time.Millisecond)
 162         require.Equal(t, 2, server.Calls())
 163         require.Equal(t, 1, test.events.Count("busy"))
 164 }
 165
 166 func TestReportOfAStoppedSubagentStartsNoTurn(t *testing.T) {
 167         server := newModelFake(t)
 168         test := newTestTurnLoop(t, server)
 169         report := subagentReport("GreenOtter", []string{"failed: context canceled"})
 170         test.loop.Report("GreenOtter", report, subagentStopped)
 171
 172         test.events.WaitFor(t, "work", 1)
 173         time.Sleep(50 * time.Millisecond)
 174         require.Equal(t, []string{"work: " + report}, test.Log(t))
 175         require.Zero(t, server.Calls())
 176         require.Zero(t, test.events.Count("busy"))
 177 }
 178
 179 func TestFullContextGivesAFailedReport(t *testing.T) {
 180         server := newModelFake(t)
 181         server.reply = func(w http.ResponseWriter, call ModelFakeCall) {
 182                 if subagentTestCall(call) {
 183                         w.WriteHeader(http.StatusBadRequest)
 184
 185                         _, _ = io.WriteString(w, subagentTestFullContext)
 186
 187                         return
 188                 }
 189
 190                 modelWriteEvents(w, modelCompletedEvent([]any{modelMessageItem("It failed.")}, 100, 0, 5))
 191         }
 192
 193         test := newTestTurnLoop(t, server)
 194         _, err := test.loop.runs.Spawn(t.Context(), []string{"Read the big file."}, nil)
 195         require.NoError(t, err)
 196         test.events.WaitFor(t, "idle", 1)
 197
 198         name := test.events.Last("subagent")
 199         require.Equal(t, []string{
 200                 "work: [" + name + "]\nfailed: create model response: model call: HTTP 400: " + subagentTestFullContext + "\n\nFull chat: zoom(\"" + name + "\")",
 201                 "reply: It failed.",
 202         }, test.Log(t))
 203
 204         require.Equal(t, subagentFailed, test.events.LastEvent("work").Status)
 205         require.Equal(t, "<6>subagent started name="+name+"\n<3>subagent failed name="+name+" after=0s: create model response: model call: HTTP 400: "+subagentTestFullContext+"\n", test.logs.String())
 206 }
 207
 208 func TestSubagentHasItsOwnEffortAndNoSpawn(t *testing.T) {
 209         server := newModelFake(t)
 210         server.reply = func(w http.ResponseWriter, call ModelFakeCall) {
 211                 if subagentTestCall(call) && subagentTestLastItem(call)["type"] != "function_call_output" {
 212                         modelWriteEvents(w, modelCompletedEvent([]any{modelFunctionCallItem("fc_1", "call_1", "spawn", `{"tasks":["More."],"effort":null}`)}, 100, 0, 5))
 213
 214                         return
 215                 }
 216
 217                 modelWriteEvents(w, modelCompletedEvent([]any{modelMessageItem("ok")}, 100, 0, 5))
 218         }
 219
 220         test := newTestTurnLoop(t, server)
 221         runs := test.loop.runs
 222         _, err := runs.Spawn(t.Context(), nil, nil)
 223         require.EqualError(t, err, "no tasks")
 224
 225         effort := "max"
 226         _, err = runs.Spawn(t.Context(), []string{"Think."}, &effort)
 227         require.EqualError(t, err, `no effort "max": one of low, medium, high, xhigh`)
 228
 229         effort = "high"
 230         _, err = runs.Spawn(t.Context(), []string{"Think hard."}, &effort)
 231         require.NoError(t, err)
 232         test.events.WaitFor(t, "idle", 1)
 233
 234         for n := range server.Calls() {
 235                 call := server.Call(n)
 236                 if subagentTestCall(call) {
 237                         require.Equal(t, "high", call.body["reasoning"].(map[string]any)["effort"])
 238                 }
 239         }
 240
 241         subagent, err := runs.subagents.Get(t.Context(), test.events.Last("subagent"))
 242         require.NoError(t, err)
 243
 244         log, err := subagent.Show()
 245         require.NoError(t, err)
 246         require.Equal(t, "0|note: subagent at effort high\n\n1|user: Think hard.\n\n2|tool: spawn {\"tasks\":[\"More.\"],\"effort\":null}\n\n3|echo: error: unknown tool \"spawn\"\n\n4|reply: ok", log)
 247
 248         entries, err := os.ReadDir(runs.subagents.dir)
 249         require.NoError(t, err)
 250         require.Len(t, entries, 1)
 251 }
 252
 253 func TestTellReachesARunningSubagentBetweenItsToolCalls(t *testing.T) {
 254         server := newModelFake(t)
 255         arrived := make(chan struct{})
 256         told := make(chan struct{})
 257         server.reply = func(w http.ResponseWriter, call ModelFakeCall) {
 258                 switch {
 259                 case !subagentTestCall(call):
 260                         modelWriteEvents(w, modelCompletedEvent([]any{modelMessageItem("Noted.")}, 100, 0, 5))
 261                 case call.number == 0:
 262                         close(arrived)
 263                         <-told
 264                         modelWriteEvents(w, modelCompletedEvent([]any{modelFunctionCallItem("fc_1", "call_1", "shell", `{"command":"true"}`)}, 100, 0, 5))
 265                 default:
 266                         modelWriteEvents(w, modelCompletedEvent([]any{modelMessageItem("Used the second disk.")}, 100, 0, 5))
 267                 }
 268         }
 269
 270         test := newTestTurnLoop(t, server)
 271         runs := test.loop.runs
 272         _, err := runs.Spawn(t.Context(), []string{"Copy the files."}, nil)
 273         require.NoError(t, err)
 274
 275         <-arrived
 276
 277         name := test.events.Last("subagent")
 278         output, err := runs.Tell(t.Context(), name, "Use the second disk.")
 279         require.NoError(t, err)
 280         require.Equal(t, "Sent. "+name+" reads it between its tool calls, and its report answers it.", output)
 281         close(told)
 282
 283         test.events.WaitFor(t, "idle", 1)
 284         require.Equal(t, "Use the second disk.", subagentTestText(subagentTestLastItem(server.Call(1))))
 285         require.Equal(t, "work: "+subagentReport(name, []string{"Used the second disk."}), test.Log(t)[0])
 286 }
 287
 288 func TestTellAfterTheFinalAnswerStartsAnotherChain(t *testing.T) {
 289         server := newModelFake(t)
 290         arrived := make(chan struct{})
 291         told := make(chan struct{})
 292         server.reply = func(w http.ResponseWriter, call ModelFakeCall) {
 293                 switch {
 294                 case !subagentTestCall(call):
 295                         modelWriteEvents(w, modelCompletedEvent([]any{modelMessageItem("Noted.")}, 100, 0, 5))
 296                 case call.number == 0:
 297                         close(arrived)
 298                         <-told
 299                         modelWriteEvents(w, modelCompletedEvent([]any{modelMessageItem("First.")}, 100, 0, 5))
 300                 default:
 301                         modelWriteEvents(w, modelCompletedEvent([]any{modelMessageItem("Second.")}, 100, 0, 5))
 302                 }
 303         }
 304
 305         test := newTestTurnLoop(t, server)
 306         runs := test.loop.runs
 307         _, err := runs.Spawn(t.Context(), []string{"Count."}, nil)
 308         require.NoError(t, err)
 309
 310         <-arrived
 311
 312         name := test.events.Last("subagent")
 313         _, err = runs.Tell(t.Context(), name, "Count again.")
 314         require.NoError(t, err)
 315         close(told)
 316
 317         test.events.WaitFor(t, "idle", 1)
 318         require.Equal(t, "0|note: subagent\n\n1|user: Count.\n\n2|user: Count again.\n\n3|reply: First.", subagentTestText(subagentTestLastItem(server.Call(1))))
 319         require.Equal(t, "work: "+subagentReport(name, []string{"First.", "Second."}), test.Log(t)[0])
 320 }
 321
 322 func TestTellToAFinishedSubagentGoesOnFromItsLog(t *testing.T) {
 323         server := newModelFake(t)
 324         server.reply = func(w http.ResponseWriter, call ModelFakeCall) {
 325                 text := "Noted."
 326                 if subagentTestCall(call) {
 327                         text = "Reply " + strconv.Itoa(call.number) + "."
 328                 }
 329
 330                 modelWriteEvents(w, modelCompletedEvent([]any{modelMessageItem(text)}, 100, 0, 5))
 331         }
 332
 333         test := newTestTurnLoop(t, server)
 334         runs := test.loop.runs
 335         effort := "high"
 336         _, err := runs.Spawn(t.Context(), []string{"Count."}, &effort)
 337         require.NoError(t, err)
 338         test.events.WaitFor(t, "idle", 1)
 339
 340         name := test.events.Last("subagent")
 341         output, err := runs.Tell(t.Context(), name, "And the hidden files?")
 342         require.NoError(t, err)
 343         require.Equal(t, "Sent. "+name+"'s reply will reach you as a message.", output)
 344         test.events.WaitFor(t, "idle", 2)
 345
 346         call := server.Call(2)
 347         require.True(t, subagentTestCall(call))
 348         require.Equal(t, "high", call.body["reasoning"].(map[string]any)["effort"])
 349         require.Equal(t, "0|note: subagent at effort high\n\n1|user: Count.\n\n2|reply: Reply 0.\n\n3|user: And the hidden files?", subagentTestText(subagentTestLastItem(call)))
 350
 351         // The view of the new run holds the chat of now, with the first report.
 352         texts := turnInputTexts(t, call)
 353         require.Contains(t, strings.Join(texts[:len(texts)-2], ""), "work: ["+name+"]")
 354         require.Equal(t, "work: "+subagentReport(name, []string{"Reply 2."}), test.Log(t)[2])
 355
 356         _, err = runs.Tell(t.Context(), "BlueFox", "Hello.")
 357         require.EqualError(t, err, "no agent BlueFox")
 358 }
 359
 360 func TestStopEndsASubagentAndItsCommand(t *testing.T) {
 361         server := newModelFake(t)
 362         server.reply = func(w http.ResponseWriter, _ ModelFakeCall) {
 363                 modelWriteEvents(w, modelCompletedEvent([]any{modelFunctionCallItem("fc_1", "call_1", "shell", `{"command":"touch started && sleep 30"}`)}, 100, 0, 5))
 364         }
 365
 366         test := newTestTurnLoop(t, server)
 367         runs := test.loop.runs
 368         _, err := runs.Spawn(t.Context(), []string{"Wait long."}, nil)
 369         require.NoError(t, err)
 370
 371         name := test.events.Last("subagent")
 372         subagent, err := runs.subagents.Get(t.Context(), name)
 373         require.NoError(t, err)
 374
 375         require.Eventually(t, func() bool {
 376                 _, err := os.Stat(filepath.Join(subagent.WorkDir(), "started"))
 377                 return err == nil
 378         }, 10*time.Second, time.Millisecond)
 379
 380         start := time.Now()
 381         found, err := test.loop.StopSubagent(name)
 382         require.NoError(t, err)
 383         require.True(t, found)
 384
 385         test.events.WaitFor(t, "work", 1)
 386         require.Less(t, time.Since(start), 10*time.Second)
 387         require.Equal(t, []string{
 388                 "note: Stopped by the user: " + name + ".",
 389                 "work: " + subagentReport(name, []string{"stopped"}),
 390         }, test.Log(t))
 391
 392         require.Equal(t, subagentStopped, test.events.LastEvent("work").Status)
 393         require.Equal(t, "<6>subagent started name="+name+"\n<6>subagent stopped name="+name+" after=0s\n", test.logs.String())
 394
 395         log, err := subagent.Show()
 396         require.NoError(t, err)
 397         require.Contains(t, log, "echo: \n[interrupted]")
 398
 399         found, err = test.loop.StopSubagent(name)
 400         require.NoError(t, err)
 401         require.False(t, found)
 402
 403         time.Sleep(50 * time.Millisecond)
 404         require.Zero(t, test.events.Count("busy"))
 405 }
 406
 407 func TestCloseEndsTheSubagentsWithNoReport(t *testing.T) {
 408         server := newModelFake(t)
 409         arrived := make(chan struct{})
 410         release := make(chan struct{})
 411         t.Cleanup(func() {
 412                 close(release)
 413         })
 414
 415         server.reply = func(_ http.ResponseWriter, _ ModelFakeCall) {
 416                 close(arrived)
 417                 <-release
 418         }
 419
 420         test := newTestTurnLoop(t, server)
 421         runs := test.loop.runs
 422         _, err := runs.Spawn(t.Context(), []string{"Wait."}, nil)
 423         require.NoError(t, err)
 424
 425         <-arrived
 426         runs.Close()
 427         require.Empty(t, runs.running)
 428         require.Empty(t, test.Log(t))
 429
 430         _, err = runs.Spawn(t.Context(), []string{"Again."}, nil)
 431         require.ErrorIs(t, err, errSubagentClosed)
 432
 433         entries, err := os.ReadDir(runs.subagents.dir)
 434         require.NoError(t, err)
 435         require.Len(t, entries, 1)
 436 }
 437
 438 func TestRestartResumesTheCutSubagents(t *testing.T) {
 439         server := newModelFake(t)
 440         arrived := make(chan struct{}, 1)
 441         release := make(chan struct{})
 442         t.Cleanup(func() {
 443                 close(release)
 444         })
 445
 446         var restarted atomic.Bool
 447
 448         server.reply = func(w http.ResponseWriter, call ModelFakeCall) {
 449                 switch {
 450                 case subagentTestCall(call) && !restarted.Load():
 451                         arrived <- struct{}{}
 452
 453                         <-release
 454                 case subagentTestCall(call):
 455                         modelWriteEvents(w, modelCompletedEvent([]any{modelMessageItem("Built.")}, 100, 0, 5))
 456                 default:
 457                         modelWriteEvents(w, modelCompletedEvent([]any{modelMessageItem("Noted.")}, 100, 0, 5))
 458                 }
 459         }
 460
 461         agent, err := openAgent(t.TempDir())
 462         require.NoError(t, err)
 463
 464         first := &TurnEventRecorder{}
 465         loop, _, stop := subagentTestStart(t, agent, server, first, discardLog())
 466         _, err = loop.runs.Spawn(t.Context(), []string{"Build the project."}, nil)
 467         require.NoError(t, err)
 468
 469         <-arrived
 470
 471         name := first.Last("subagent")
 472
 473         stop()
 474         require.Zero(t, first.Count("work"))
 475
 476         data, err := os.ReadFile(filepath.Join(agent.dir, subagentWorkFileName))
 477         require.NoError(t, err)
 478         require.JSONEq(t, `{"running": ["`+name+`"]}`, string(data))
 479
 480         restarted.Store(true)
 481
 482         second := &TurnEventRecorder{}
 483         loop, memory, stop := subagentTestStart(t, agent, server, second, discardLog())
 484         defer stop()
 485
 486         second.WaitFor(t, "idle", 1)
 487
 488         note := "The daemon restarted, cutting " + name + "; nothing was lost: go on."
 489         log := subagentTestChat(t, memory)
 490         require.Equal(t, []string{
 491                 "note: " + note,
 492                 "work: " + subagentReport(name, []string{"Built."}),
 493                 "reply: Noted.",
 494         }, log[len(log)-3:])
 495
 496         subagent, err := loop.runs.subagents.Get(t.Context(), name)
 497         require.NoError(t, err)
 498
 499         text, err := subagent.Show()
 500         require.NoError(t, err)
 501         require.Equal(t, "0|note: subagent\n\n1|user: Build the project.\n\n2|user: "+note+"\n\n3|reply: Built.", text)
 502
 503         data, err = os.ReadFile(filepath.Join(agent.dir, subagentWorkFileName))
 504         require.NoError(t, err)
 505         require.JSONEq(t, `{"running": []}`, string(data))
 506 }
 507
 508 func TestRestartReportsASubagentThatCannotGoOn(t *testing.T) {
 509         server := newModelFake(t)
 510         agent, err := openAgent(t.TempDir())
 511         require.NoError(t, err)
 512         require.NoError(t, os.WriteFile(filepath.Join(agent.dir, subagentWorkFileName), []byte(`{"running": ["BlueFox"]}`), 0o600))
 513
 514         events := &TurnEventRecorder{}
 515         logs := &SyncBuffer{}
 516         _, memory, stop := subagentTestStart(t, agent, server, events, daemonTestLogger(logs))
 517         defer stop()
 518
 519         events.WaitFor(t, "idle", 1)
 520         require.Equal(t, []string{
 521                 "note: The daemon restarted, cutting BlueFox; nothing was lost: go on.",
 522                 "work: " + subagentReport("BlueFox", []string{"failed: no agent BlueFox"}),
 523                 "reply: a line",
 524         }, subagentTestChat(t, memory))
 525
 526         require.Equal(t, subagentFailed, events.LastEvent("work").Status)
 527         require.Equal(t, "<3>subagent lost at restart name=BlueFox: no agent BlueFox\n", logs.String())
 528
 529         data, err := os.ReadFile(filepath.Join(agent.dir, subagentWorkFileName))
 530         require.NoError(t, err)
 531         require.JSONEq(t, `{"running": []}`, string(data))
 532 }
 533
 534 // subagentTestStart runs a loop of agent with its memory and subagents in the
 535 // agent directory, as the daemon does. The returned function stops the loop
 536 // and closes the files.
 537 func subagentTestStart(t *testing.T, agent *Agent, server *ModelFake, events *TurnEventRecorder, logger *slog.Logger) (*TurnLoop, *Memory, func()) {
 538         t.Helper()
 539
 540         agent.config.Workdir = t.TempDir()
 541         clock := newTestClock()
 542         memory, err := openMemory(t.Context(), agent.dir, clock.Now, newFakeSummarizer(), discardLog())
 543         require.NoError(t, err)
 544
 545         subagents := newSubagents(filepath.Join(agent.dir, subagentDirName), clock.Now, subagentNameRandom)
 546         client := newModelClient(server.URL, ModelAPIKey(modelTestKey), agent.CacheKey(), filepath.Join(agent.dir, "usage.jsonl"))
 547         whatsapp, _, _ := whatsappStartTest(t, t.TempDir(), clock)
 548         loop, err := newTurnLoop(agent, memory, subagents, whatsapp, newTestSchedules(t, clock.Now), client, clock.Now, logger, events.Add)
 549         require.NoError(t, err)
 550
 551         ctx, cancel := context.WithCancel(t.Context())
 552         done := make(chan error)
 553         go func() {
 554                 done <- loop.Run(ctx)
 555         }()
 556
 557         return loop, memory, func() {
 558                 cancel()
 559                 require.ErrorIs(t, <-done, context.Canceled)
 560                 require.NoError(t, subagents.Close())
 561                 require.NoError(t, memory.Close())
 562         }
 563 }
 564
 565 // subagentTestChat returns each message of the chat in memory as
 566 // "kind: text".
 567 func subagentTestChat(t *testing.T, memory *Memory) []string {
 568         t.Helper()
 569
 570         var lines []string
 571
 572         for id := range memory.Len() {
 573                 message, err := memory.Message(id)
 574                 require.NoError(t, err)
 575
 576                 lines = append(lines, message.Kind+": "+message.Text)
 577         }
 578
 579         return lines
 580 }
 581
 582 // subagentTestCall tells if call is a call of a subagent.
 583 func subagentTestCall(call ModelFakeCall) bool {
 584         instructions, _ := call.body["instructions"].(string)
 585
 586         return strings.HasPrefix(instructions, "You are a subagent")
 587 }
 588
 589 // subagentTestLastItem returns the last input item of call.
 590 func subagentTestLastItem(call ModelFakeCall) map[string]any {
 591         input := call.body["input"].([]any)
 592
 593         return input[len(input)-1].(map[string]any)
 594 }
 595
 596 // subagentTestText returns the text of the user message item, or "" for
 597 // another item.
 598 func subagentTestText(item map[string]any) string {
 599         content, _ := item["content"].([]any)
 600
 601         var text strings.Builder
 602
 603         for _, part := range content {
 604                 value, _ := part.(map[string]any)["text"].(string)
 605                 text.WriteString(value)
 606         }
 607
 608         return text.String()
 609 }