12 "github.com/stretchr/testify/require"
15 // subagentTestFullContext is the error body of a call whose input does not fit
16 // in the context of the model.
17 const subagentTestFullContext = `{"error": {"message": "Your input exceeds the context window of this model.", "type": "invalid_request_error", "code": "context_length_exceeded"}}`
19 func TestSpawnRunsASubagentInTheBackground(t *testing.T) {
20 server := newModelFake(t)
21 var test *TurnTestLoop
22 server.reply = func(w http.ResponseWriter, call ModelFakeCall) {
23 last := subagentTestLastItem(call)
25 case subagentTestCall(call) && last["type"] == "function_call_output":
26 // The report comes after the reply of the turn of the spawn.
27 test.events.WaitFor(t, "reply", 1)
28 modelWriteEvents(w, modelCompletedEvent([]any{modelMessageItem("There are 3 files.")}, 100, 0, 5))
29 case subagentTestCall(call):
30 modelWriteEvents(w, modelCompletedEvent([]any{modelFunctionCallItem("fc_1", "call_1", "shell", `{"command":"pwd"}`)}, 100, 0, 5))
31 case last["type"] == "function_call_output":
32 modelWriteEvents(w, modelCompletedEvent([]any{modelMessageItem("Started.")}, 100, 0, 5))
33 case strings.HasPrefix(subagentTestText(last), "["):
34 modelWriteEvents(w, modelCompletedEvent([]any{modelMessageItem("Noted.")}, 100, 0, 5))
36 modelWriteEvents(w, modelCompletedEvent([]any{modelFunctionCallItem("fc_1", "call_1", "spawn", `{"tasks":["Count the files."],"effort":null}`)}, 100, 0, 5))
40 test = newTestTurnLoop(t, server)
41 _, err := test.loop.Submit(terminalSource, "Count the files in a subagent.")
42 require.NoError(t, err)
43 test.events.WaitFor(t, "idle", 2)
45 name := test.events.Last("subagent")
46 require.True(t, subagentNameValid(name), name)
47 work := filepath.Join(test.loop.runs.subagents.dir, name, "work")
48 require.DirExists(t, work)
50 report := "[" + name + "]\nThere are 3 files.\n\nFull chat: zoom(\"" + name + "\")"
51 require.Equal(t, []string{
52 "user: Count the files in a subagent.",
53 `tool: spawn {"tasks":["Count the files."],"effort":null}`,
54 "echo: Started " + name + ". Each one's report will reach you as a message when it finishes.",
60 var first, second *ModelFakeCall
61 for n := range server.Calls() {
62 call := server.Call(n)
63 if !subagentTestCall(call) {
74 require.NotNil(t, second)
76 // The first request of the subagent holds the view, the state, and its
77 // whole log, with its own prompt and tools.
78 texts := turnInputTexts(t, *first)
79 view := strings.Join(texts[:len(texts)-2], "")
80 require.True(t, strings.HasPrefix(view, "<chat>\n0+1|user: Count the files in a subagent.\n1+1|tool: spawn"), view)
81 require.Equal(t, "</chat>\n\nNow: Thursday 2026-01-01 12:00 +00:00", texts[len(texts)-2])
82 require.Equal(t, "0|note: subagent\n\n1|user: Count the files.", texts[len(texts)-1])
84 instructions, err := test.agent.SubagentPrompt()
85 require.NoError(t, err)
86 require.Equal(t, instructions, first.body["instructions"])
89 for _, tool := range first.body["tools"].([]any) {
90 tools = append(tools, tool.(map[string]any)["name"].(string))
93 require.Equal(t, []string{"zoom", "date", "shell", "read", "write", "edit"}, tools)
94 require.Equal(t, work+"\n", subagentTestLastItem(*second)["output"])
96 subagent, err := test.loop.runs.subagents.Get(t.Context(), name)
97 require.NoError(t, err)
99 log, err := subagent.Show()
100 require.NoError(t, err)
101 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)
104 for _, record := range modelReadUsage(t, test.usagePath) {
105 agents = append(agents, record.Agent)
108 require.ElementsMatch(t, []string{"", "", "", name, name}, agents)
111 func TestReportJoinsTheNextRequestOfARunningTurn(t *testing.T) {
112 server := newModelFake(t)
113 arrived := make(chan struct{})
114 release := make(chan struct{})
115 server.reply = func(w http.ResponseWriter, call ModelFakeCall) {
116 if call.number == 0 {
119 modelWriteEvents(w, modelCompletedEvent([]any{modelFunctionCallItem("fc_1", "call_1", "date", `{"id":0}`)}, 100, 0, 5))
124 modelWriteEvents(w, modelCompletedEvent([]any{modelMessageItem("done")}, 100, 0, 5))
127 test := newTestTurnLoop(t, server)
128 _, err := test.loop.Submit(terminalSource, "first")
129 require.NoError(t, err)
132 report := subagentReport("GreenOtter", []string{"It is done."})
133 test.loop.Report(report, false)
136 test.events.WaitFor(t, "idle", 1)
137 require.Equal(t, report, subagentTestText(subagentTestLastItem(server.Call(1))))
139 time.Sleep(50 * time.Millisecond)
140 require.Equal(t, 2, server.Calls())
141 require.Equal(t, 1, test.events.Count("busy"))
144 func TestReportOfAStoppedSubagentStartsNoTurn(t *testing.T) {
145 server := newModelFake(t)
146 test := newTestTurnLoop(t, server)
147 report := subagentReport("GreenOtter", []string{"failed: context canceled"})
148 test.loop.Report(report, true)
150 test.events.WaitFor(t, "work", 1)
151 time.Sleep(50 * time.Millisecond)
152 require.Equal(t, []string{"work: " + report}, test.Log(t))
153 require.Zero(t, server.Calls())
154 require.Zero(t, test.events.Count("busy"))
157 func TestFullContextGivesAFailedReport(t *testing.T) {
158 server := newModelFake(t)
159 server.reply = func(w http.ResponseWriter, call ModelFakeCall) {
160 if subagentTestCall(call) {
161 w.WriteHeader(http.StatusBadRequest)
162 _, _ = io.WriteString(w, subagentTestFullContext)
167 modelWriteEvents(w, modelCompletedEvent([]any{modelMessageItem("It failed.")}, 100, 0, 5))
170 test := newTestTurnLoop(t, server)
171 _, err := test.loop.runs.Spawn(t.Context(), []string{"Read the big file."}, nil)
172 require.NoError(t, err)
173 test.events.WaitFor(t, "idle", 1)
175 name := test.events.Last("subagent")
176 require.Equal(t, []string{
177 "work: [" + name + "]\nfailed: model call: HTTP 400: " + subagentTestFullContext + "\n\nFull chat: zoom(\"" + name + "\")",
182 func TestSubagentHasItsOwnEffortAndNoSpawn(t *testing.T) {
183 server := newModelFake(t)
184 server.reply = func(w http.ResponseWriter, call ModelFakeCall) {
185 if subagentTestCall(call) && subagentTestLastItem(call)["type"] != "function_call_output" {
186 modelWriteEvents(w, modelCompletedEvent([]any{modelFunctionCallItem("fc_1", "call_1", "spawn", `{"tasks":["More."],"effort":null}`)}, 100, 0, 5))
191 modelWriteEvents(w, modelCompletedEvent([]any{modelMessageItem("ok")}, 100, 0, 5))
194 test := newTestTurnLoop(t, server)
195 runs := test.loop.runs
196 _, err := runs.Spawn(t.Context(), nil, nil)
197 require.EqualError(t, err, "no tasks")
200 _, err = runs.Spawn(t.Context(), []string{"Think."}, &effort)
201 require.EqualError(t, err, `no effort "max": one of low, medium, high, xhigh`)
204 _, err = runs.Spawn(t.Context(), []string{"Think hard."}, &effort)
205 require.NoError(t, err)
206 test.events.WaitFor(t, "idle", 1)
208 for n := range server.Calls() {
209 call := server.Call(n)
210 if subagentTestCall(call) {
211 require.Equal(t, "high", call.body["reasoning"].(map[string]any)["effort"])
215 subagent, err := runs.subagents.Get(t.Context(), test.events.Last("subagent"))
216 require.NoError(t, err)
218 log, err := subagent.Show()
219 require.NoError(t, err)
220 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)
222 entries, err := os.ReadDir(runs.subagents.dir)
223 require.NoError(t, err)
224 require.Len(t, entries, 1)
227 func TestCloseEndsTheSubagentsWithNoReport(t *testing.T) {
228 server := newModelFake(t)
229 arrived := make(chan struct{})
230 release := make(chan struct{})
235 server.reply = func(w http.ResponseWriter, call ModelFakeCall) {
240 test := newTestTurnLoop(t, server)
241 runs := test.loop.runs
242 _, err := runs.Spawn(t.Context(), []string{"Wait."}, nil)
243 require.NoError(t, err)
247 require.Empty(t, runs.running)
248 require.Empty(t, test.Log(t))
250 _, err = runs.Spawn(t.Context(), []string{"Again."}, nil)
251 require.ErrorIs(t, err, errSubagentClosed)
253 entries, err := os.ReadDir(runs.subagents.dir)
254 require.NoError(t, err)
255 require.Len(t, entries, 1)
258 // subagentTestCall tells if call is a call of a subagent.
259 func subagentTestCall(call ModelFakeCall) bool {
260 instructions, _ := call.body["instructions"].(string)
262 return strings.HasPrefix(instructions, "You are a subagent")
265 // subagentTestLastItem returns the last input item of call.
266 func subagentTestLastItem(call ModelFakeCall) map[string]any {
267 input := call.body["input"].([]any)
269 return input[len(input)-1].(map[string]any)
272 // subagentTestText returns the text of the user message item, or "" for
274 func subagentTestText(item map[string]any) string {
275 content, _ := item["content"].([]any)
276 var text strings.Builder
277 for _, part := range content {
278 value, _ := part.(map[string]any)["text"].(string)
279 text.WriteString(value)