| 🛟 Updated. 28d5985 k33g 18h ago | 1 | package acp_test |
| 2 | |
| 3 | import ( |
| 4 | "errors" |
| 5 | "strings" |
| 6 | "testing" |
| 7 | |
| 8 | "codeberg.org/turbo-editors/turbo-core/acp" |
| 9 | ) |
| 10 | |
| 11 | // start returns a session talking to a fake agent over an in-memory pipe. |
| 12 | func start(t *testing.T, options acp.Options) (*acp.Session, *fakeAgent) { |
| 13 | t.Helper() |
| 14 | |
| 15 | agent, stream := newFakeAgent(t) |
| 16 | session := acp.NewSession(stream, acp.Agent{Name: "Bob"}, t.TempDir(), options) |
| 17 | t.Cleanup(func() { _ = session.Close() }) |
| 18 | return session, agent |
| 19 | } |
| 20 | |
| 21 | func TestTheHandshakeMakesASessionReady(t *testing.T) { |
| 22 | session, agent := start(t, acp.Options{}) |
| 23 | |
| 24 | agent.handshake() |
| 25 | waitFor(t, "the session to be ready", session.Ready) |
| 26 | |
| 27 | if err := session.Err(); err != nil { |
| 28 | t.Fatalf("Err() = %v", err) |
| 29 | } |
| 30 | // The configured name wins: docker agent reports "docker agent", which is |
| 31 | // the runtime rather than the assistant, and the window is titled with what |
| 32 | // the user chose. Two names for one thing is worse than one. |
| 33 | if got := session.AgentName(); got != "Bob" { |
| 34 | t.Errorf("AgentName() = %q, want the configured name", got) |
| 35 | } |
| 36 | if got := session.AgentInfo().Version; got != "v1.139.0" { |
| 37 | t.Errorf("AgentInfo().Version = %q; the handshake's own details are still kept", got) |
| 38 | } |
| 39 | } |
| 40 | |
| 41 | func TestTheClientSendsTheProtocolVersionAndItsCapabilities(t *testing.T) { |
| 42 | _, agent := start(t, acp.Options{Client: acp.Implementation{Name: "turbo-test"}}) |
| 43 | |
| 44 | initialize := agent.read() |
| 45 | params, _ := initialize["params"].(map[string]any) |
| 46 | if params["protocolVersion"] != float64(acp.ProtocolVersion) { |
| 47 | t.Errorf("the client asked for version %v, want %d", params["protocolVersion"], acp.ProtocolVersion) |
| 48 | } |
| 49 | |
| 50 | capabilities, _ := params["clientCapabilities"].(map[string]any) |
| 51 | files, _ := capabilities["fs"].(map[string]any) |
| 52 | if files["readTextFile"] != true || files["writeTextFile"] != true { |
| 53 | t.Errorf("the client offered %v, want both file operations", capabilities) |
| 54 | } |
| 55 | // Terminal is deliberately not offered, so a conforming agent never asks. |
| 56 | if _, offered := capabilities["terminal"]; offered { |
| 57 | t.Errorf("the client offered a terminal capability it does not implement: %v", capabilities) |
| 58 | } |
| 59 | if info, _ := params["clientInfo"].(map[string]any); info["name"] != "turbo-test" { |
| 60 | t.Errorf("the client called itself %v", info) |
| 61 | } |
| 62 | } |
| 63 | |
| 64 | func TestAnAgentSpeakingAnotherVersionIsRefusedWithAReadableReason(t *testing.T) { |
| 65 | session, agent := start(t, acp.Options{}) |
| 66 | |
| 67 | initialize := agent.read() |
| 68 | agent.reply(initialize["id"], `{"protocolVersion":99,"agentInfo":{"name":"future"}}`) |
| 69 | |
| 70 | waitFor(t, "the session to fail", func() bool { return session.Err() != nil }) |
| 71 | if got := session.Err().Error(); !strings.Contains(got, "99") { |
| 72 | t.Errorf("Err() = %q, want both versions named", got) |
| 73 | } |
| 74 | if notice := textOf(session.Entries(), acp.EntryNotice); !strings.Contains(notice, "99") { |
| 75 | t.Errorf("the window says %q; the reason must be visible in it", notice) |
| 76 | } |
| 77 | } |
| 78 | |
| 79 | func TestAnAgentWantingALoginSaysSoRatherThanHanging(t *testing.T) { |
| 80 | session, agent := start(t, acp.Options{}) |
| 81 | |
| 82 | initialize := agent.read() |
| 83 | agent.reply(initialize["id"], `{"protocolVersion":1,"agentInfo":{"name":"paid"},"authMethods":[{"id":"oauth","name":"Log in with OAuth"}]}`) |
| 84 | |
| 85 | waitFor(t, "the session to fail", func() bool { return session.Err() != nil }) |
| 86 | if got := session.Err().Error(); !strings.Contains(got, "Log in with OAuth") { |
| 87 | t.Errorf("Err() = %q, want the method named so a user knows what to run", got) |
| 88 | } |
| 89 | } |
| 90 | |
| 91 | func TestWhatIsTypedBeforeTheAgentIsReadyIsSentWhenItIs(t *testing.T) { |
| 92 | // The handshake may mean reaching a model, which takes seconds. Losing |
| 93 | // what somebody typed in that window would be the most annoying possible |
| 94 | // version of this. |
| 95 | session, agent := start(t, acp.Options{}) |
| 96 | |
| 97 | session.Prompt("early") |
| 98 | waitFor(t, "the prompt to appear in the window", func() bool { |
| 99 | return strings.Contains(textOf(session.Entries(), acp.EntryUser), "early") |
| 100 | }) |
| 101 | |
| 102 | id := agent.handshake() |
| 103 | prompt := agent.read() |
| 104 | if prompt["method"] != acp.MethodPrompt { |
| 105 | t.Fatalf("the client sent %v, want the held prompt", prompt["method"]) |
| 106 | } |
| 107 | |
| 108 | params, _ := prompt["params"].(map[string]any) |
| 109 | if params["sessionId"] != id { |
| 110 | t.Errorf("the prompt quotes session %v, want %q", params["sessionId"], id) |
| 111 | } |
| 112 | blocks, _ := params["prompt"].([]any) |
| 113 | first, _ := blocks[0].(map[string]any) |
| 114 | if first["text"] != "early" || first["type"] != "text" { |
| 115 | t.Errorf("the prompt carried %v", blocks) |
| 116 | } |
| 117 | } |
| 118 | |
| 119 | func TestAReplyArrivesChunkByChunkAndReadsAsOneMessage(t *testing.T) { |
| 120 | // This is exactly what a real agent does: one update per token. An entry |
| 121 | // per chunk could be neither wrapped nor told apart from a code fence. |
| 122 | session, agent := start(t, acp.Options{}) |
| 123 | id := agent.handshake() |
| 124 | waitFor(t, "ready", session.Ready) |
| 125 | |
| 126 | session.Prompt("hello") |
| 127 | prompt := agent.read() |
| 128 | |
| 129 | for _, chunk := range []string{"I", " found", " three", " files", "."} { |
| 130 | agent.update(id, `{"sessionUpdate":"agent_message_chunk","content":{"type":"text","text":"`+chunk+`"}}`) |
| 131 | } |
| 132 | agent.reply(prompt["id"], `{"stopReason":"end_turn"}`) |
| 133 | |
| 134 | waitFor(t, "the whole reply", func() bool { |
| 135 | return textOf(session.Entries(), acp.EntryAgent) == "I found three files." |
| 136 | }) |
| 137 | |
| 138 | agentEntries := 0 |
| 139 | for _, entry := range session.Entries() { |
| 140 | if entry.Kind == acp.EntryAgent { |
| 141 | agentEntries++ |
| 142 | } |
| 143 | } |
| 144 | if agentEntries != 1 { |
| 145 | t.Errorf("five chunks became %d entries, want 1", agentEntries) |
| 146 | } |
| 147 | } |
| 148 | |
| 149 | func TestAThoughtIsNotFoldedIntoTheReplyBesideIt(t *testing.T) { |
| 150 | session, agent := start(t, acp.Options{}) |
| 151 | id := agent.handshake() |
| 152 | waitFor(t, "ready", session.Ready) |
| 153 | |
| 154 | agent.update(id, `{"sessionUpdate":"agent_thought_chunk","content":{"type":"text","text":"let me look"}}`) |
| 155 | agent.update(id, `{"sessionUpdate":"agent_message_chunk","content":{"type":"text","text":"here it is"}}`) |
| 156 | |
| 157 | waitFor(t, "both", func() bool { |
| 158 | return textOf(session.Entries(), acp.EntryAgent) == "here it is" |
| 159 | }) |
| 160 | if got := textOf(session.Entries(), acp.EntryThought); got != "let me look" { |
| 161 | t.Errorf("the thought reads %q", got) |
| 162 | } |
| 163 | } |
| 164 | |
| 165 | func TestAToolCallIsOneLineThatFillsInAsItRuns(t *testing.T) { |
| 166 | // The real order, recorded from docker agent: the tool_call arrives, then |
| 167 | // a tool_call_update carrying the output, both naming the same id. |
| 168 | session, agent := start(t, acp.Options{}) |
| 169 | id := agent.handshake() |
| 170 | waitFor(t, "ready", session.Ready) |
| 171 | |
| 172 | agent.update(id, `{"sessionUpdate":"tool_call","toolCallId":"call_1","title":"Shell","kind":"execute","status":"pending","rawInput":{"cmd":"ls -1","cwd":".","timeout":30}}`) |
| 173 | waitFor(t, "the tool call", func() bool { return toolEntry(session, "call_1") != nil }) |
| 174 | |
| 175 | tool := toolEntry(session, "call_1") |
| 176 | if tool.Speaker != "Shell" { |
| 177 | t.Errorf("the tool is called %q", tool.Speaker) |
| 178 | } |
| 179 | if tool.Detail != "ls -1" { |
| 180 | t.Errorf("the detail is %q, want the command rather than the whole input object", tool.Detail) |
| 181 | } |
| 182 | if tool.Done() { |
| 183 | t.Error("a pending tool call reports itself as finished") |
| 184 | } |
| 185 | |
| 186 | agent.update(id, `{"sessionUpdate":"tool_call_update","toolCallId":"call_1","status":"completed","content":[{"type":"content","content":{"type":"text","text":"agent.yaml\nprobe.py"}}]}`) |
| 187 | waitFor(t, "the output", func() bool { |
| 188 | tool := toolEntry(session, "call_1") |
| 189 | return tool != nil && tool.Done() |
| 190 | }) |
| 191 | |
| 192 | tool = toolEntry(session, "call_1") |
| 193 | if !strings.Contains(tool.Text, "agent.yaml") { |
| 194 | t.Errorf("the output is %q", tool.Text) |
| 195 | } |
| 196 | if tool.Failed() { |
| 197 | t.Error("a completed tool call reports itself as failed") |
| 198 | } |
| 199 | if count := toolEntries(session); count != 1 { |
| 200 | t.Errorf("the update made a second entry: %d tool entries, want 1", count) |
| 201 | } |
| 202 | } |
| 203 | |
| 204 | func TestAPlanIsReplacedRatherThanRepeated(t *testing.T) { |
| 205 | // An agent republishes the whole plan every time a step changes, so |
| 206 | // appending would leave five copies of a four-line plan in the window. |
| 207 | session, agent := start(t, acp.Options{}) |
| 208 | id := agent.handshake() |
| 209 | waitFor(t, "ready", session.Ready) |
| 210 | |
| 211 | agent.update(id, `{"sessionUpdate":"plan","entries":[{"content":"look","status":"pending"}]}`) |
| 212 | agent.update(id, `{"sessionUpdate":"plan","entries":[{"content":"look","status":"completed"},{"content":"report","status":"pending"}]}`) |
| 213 | |
| 214 | waitFor(t, "the second plan", func() bool { |
| 215 | for _, entry := range session.Entries() { |
| 216 | if entry.Kind == acp.EntryPlan && len(entry.Plan) == 2 { |
| 217 | return true |
| 218 | } |
| 219 | } |
| 220 | return false |
| 221 | }) |
| 222 | |
| 223 | plans := 0 |
| 224 | for _, entry := range session.Entries() { |
| 225 | if entry.Kind == acp.EntryPlan { |
| 226 | plans++ |
| 227 | } |
| 228 | } |
| 229 | if plans != 1 { |
| 230 | t.Errorf("there are %d plans in the window, want 1", plans) |
| 231 | } |
| 232 | } |
| 233 | |
| 234 | func TestAnUpdateThisClientDoesNotKnowIsCountedAndIgnored(t *testing.T) { |
| 235 | // The protocol grows. An editor that stopped talking to an agent because |
| 236 | // it learnt a new message would be wrong more often than it was right. |
| 237 | session, agent := start(t, acp.Options{}) |
| 238 | id := agent.handshake() |
| 239 | waitFor(t, "ready", session.Ready) |
| 240 | |
| 241 | agent.update(id, `{"sessionUpdate":"something_invented_next_year","whatever":1}`) |
| 242 | agent.update(id, `{"sessionUpdate":"agent_message_chunk","content":{"type":"text","text":"still here"}}`) |
| 243 | |
| 244 | waitFor(t, "the message after it", func() bool { |
| 245 | return textOf(session.Entries(), acp.EntryAgent) == "still here" |
| 246 | }) |
| 247 | if session.Unknown() != 1 { |
| 248 | t.Errorf("Unknown() = %d, want 1", session.Unknown()) |
| 249 | } |
| 250 | if err := session.Err(); err != nil { |
| 251 | t.Errorf("an unknown update ended the session: %v", err) |
| 252 | } |
| 253 | } |
| 254 | |
| 255 | func TestUsageAndCommandsAreKeptWithoutBeingDrawn(t *testing.T) { |
| 256 | session, agent := start(t, acp.Options{}) |
| 257 | id := agent.handshake() |
| 258 | waitFor(t, "ready", session.Ready) |
| 259 | |
| 260 | agent.update(id, `{"sessionUpdate":"usage_update","used":1617,"size":262144}`) |
| 261 | agent.update(id, `{"sessionUpdate":"available_commands_update","availableCommands":[{"name":"compact","description":"squash the history"},{"name":"web","description":"Search the web","input":{"hint":"query to search for"}}]}`) |
| 262 | |
| 263 | waitFor(t, "the usage", func() bool { |
| 264 | used, size := session.Usage() |
| 265 | return used == 1617 && size == 262144 |
| 266 | }) |
| 267 | waitFor(t, "the commands", func() bool { return len(session.Commands()) == 2 }) |
| 268 | |
| 269 | // The shape is the protocol's own: input is present exactly when the |
| 270 | // command wants something after its name, and carries the hint. |
| 271 | commands := session.Commands() |
| 272 | if commands[0].TakesInput() || commands[0].Hint() != "" { |
| 273 | t.Errorf("compact takes input: %+v", commands[0]) |
| 274 | } |
| 275 | if !commands[1].TakesInput() || commands[1].Hint() != "query to search for" { |
| 276 | t.Errorf("web lost its hint: %+v", commands[1]) |
| 277 | } |
| 278 | if !session.EmbedsContext() { |
| 279 | t.Error("EmbedsContext() = false after a handshake that declared embeddedContext") |
| 280 | } |
| 281 | |
| 282 | if session.Unknown() != 0 { |
| 283 | t.Errorf("Unknown() = %d; these two are understood", session.Unknown()) |
| 284 | } |
| 285 | for _, entry := range session.Entries() { |
| 286 | if strings.Contains(entry.Text, "262144") { |
| 287 | t.Errorf("the usage was drawn into the conversation: %q", entry.Text) |
| 288 | } |
| 289 | } |
| 290 | } |
| 291 | |
| 292 | func TestCancellingATurnNotifiesTheAgent(t *testing.T) { |
| 293 | session, agent := start(t, acp.Options{}) |
| 294 | id := agent.handshake() |
| 295 | waitFor(t, "ready", session.Ready) |
| 296 | |
| 297 | session.Prompt("something long") |
| 298 | agent.read() // the prompt |
| 299 | waitFor(t, "the turn to start", session.Running) |
| 300 | |
| 301 | session.Cancel() |
| 302 | cancel := agent.read() |
| 303 | if cancel["method"] != acp.MethodCancel { |
| 304 | t.Fatalf("the client sent %v, want session/cancel", cancel["method"]) |
| 305 | } |
| 306 | if _, carries := cancel["id"]; carries { |
| 307 | t.Error("session/cancel was sent as a request; it is a notification") |
| 308 | } |
| 309 | params, _ := cancel["params"].(map[string]any) |
| 310 | if params["sessionId"] != id { |
| 311 | t.Errorf("the cancellation quotes session %v", params["sessionId"]) |
| 312 | } |
| 313 | } |
| 314 | |
| 315 | func TestCancellingWithNoTurnRunningSendsNothing(t *testing.T) { |
| 316 | session, agent := start(t, acp.Options{}) |
| 317 | agent.handshake() |
| 318 | waitFor(t, "ready", session.Ready) |
| 319 | |
| 320 | session.Cancel() |
| 321 | |
| 322 | // Nothing to read means nothing was sent; a stray cancel would confuse an |
| 323 | // agent that is between turns. |
| 324 | if got := agent.methods(); len(got) != 2 { |
| 325 | t.Errorf("the client has sent %v, want only the handshake", got) |
| 326 | } |
| 327 | } |
| 328 | |
| 329 | func TestATurnThatEndsUnusuallySaysSo(t *testing.T) { |
| 330 | session, agent := start(t, acp.Options{}) |
| 331 | agent.handshake() |
| 332 | waitFor(t, "ready", session.Ready) |
| 333 | |
| 334 | session.Prompt("go on") |
| 335 | prompt := agent.read() |
| 336 | agent.reply(prompt["id"], `{"stopReason":"max_tokens"}`) |
| 337 | |
| 338 | waitFor(t, "the notice", func() bool { |
| 339 | return strings.Contains(textOf(session.Entries(), acp.EntryNotice), "max_tokens") |
| 340 | }) |
| 341 | } |
| 342 | |
| 343 | func TestAnOrdinaryEndOfTurnSaysNothing(t *testing.T) { |
| 344 | // A notice after every successful turn would be noise in every |
| 345 | // conversation. |
| 346 | session, agent := start(t, acp.Options{}) |
| 347 | agent.handshake() |
| 348 | waitFor(t, "ready", session.Ready) |
| 349 | |
| 350 | session.Prompt("go on") |
| 351 | prompt := agent.read() |
| 352 | agent.reply(prompt["id"], `{"stopReason":"end_turn"}`) |
| 353 | |
| 354 | waitFor(t, "the turn to finish", func() bool { return !session.Running() }) |
| 355 | if notice := textOf(session.Entries(), acp.EntryNotice); notice != "" { |
| 356 | t.Errorf("an ordinary turn left %q in the window", notice) |
| 357 | } |
| 358 | } |
| 359 | |
| 360 | func TestAPromptIsRecordedBeforeItIsSent(t *testing.T) { |
| 361 | // The window has to show what you typed the moment you press Enter, not |
| 362 | // when the agent gets round to answering. |
| 363 | session, agent := start(t, acp.Options{}) |
| 364 | agent.handshake() |
| 365 | waitFor(t, "ready", session.Ready) |
| 366 | |
| 367 | session.Prompt("what does this do?") |
| 368 | if got := textOf(session.Entries(), acp.EntryUser); !strings.Contains(got, "what does this do?") { |
| 369 | t.Errorf("the window says %q straight after Prompt", got) |
| 370 | } |
| 371 | } |
| 372 | |
| 373 | func TestAnEmptyPromptDoesNothing(t *testing.T) { |
| 374 | session, agent := start(t, acp.Options{}) |
| 375 | agent.handshake() |
| 376 | waitFor(t, "ready", session.Ready) |
| 377 | |
| 378 | session.Prompt("") |
| 379 | if len(session.Entries()) != 0 { |
| 380 | t.Errorf("an empty prompt left %v in the window", session.Entries()) |
| 381 | } |
| 382 | } |
| 383 | |
| 384 | func TestClosingASessionTwiceIsHarmless(t *testing.T) { |
| 385 | session, agent := start(t, acp.Options{}) |
| 386 | agent.handshake() |
| 387 | |
| 388 | if err := session.Close(); err != nil && !errors.Is(err, acp.ErrMessageTooLarge) { |
| 389 | t.Fatalf("Close() error = %v", err) |
| 390 | } |
| 391 | if err := session.Close(); err != nil { |
| 392 | t.Errorf("the second Close() error = %v", err) |
| 393 | } |
| 394 | } |
| 395 | |
| 396 | // toolEntry returns the tool entry with an id, or nil. |
| 397 | func toolEntry(session *acp.Session, id string) *acp.Entry { |
| 398 | for _, entry := range session.Entries() { |
| 399 | if entry.Kind == acp.EntryTool && entry.ToolCallID == id { |
| 400 | found := entry |
| 401 | return &found |
| 402 | } |
| 403 | } |
| 404 | return nil |
| 405 | } |
| 406 | |
| 407 | // toolEntries counts the tool calls in a conversation. |
| 408 | func toolEntries(session *acp.Session) int { |
| 409 | count := 0 |
| 410 | for _, entry := range session.Entries() { |
| 411 | if entry.Kind == acp.EntryTool { |
| 412 | count++ |
| 413 | } |
| 414 | } |
| 415 | return count |
| 416 | } |