claude/sigit-cloud-agent-expansion-reox0a
claude/github-issue-40-0au5ua
claude/linked-issue-acp-tui-u4vvpa
claude/minor-release-cut-vkugow
claude/sigit-acp-local-chat-cx6380
claude/sigit-cloud-agent-expansion-reox0a
claude/tool-permission-system
claude/zen-feynman-0u78dk
development
feature/acp-cancel-request
feature/acp-session-list-cursor
feature/acp-v1-conformance
feature/agent-tools-multiedit-glob-todos-remember
feature/cargo-workspace
feature/hooks
feature/load-local-model-explicitly
feature/sigit-agent-profile
feature/sigit-code-cloud
feature/xcode-mcp-server
fix/acp-replay-loaded-session-history
fix/acp-replay-reasoning-and-attachments
fix/dsml-doubled-bar-markers
main
release/v1.5.10
release/v1.5.6
release/v1.6.1
| 1 | //! End-to-end ACP permission round-trip against the real binary. |
| 2 | //! |
| 3 | //! Spawns `sigit` in ACP mode (stdin piped, so not a TTY) wired to a scripted |
| 4 | //! OpenAI-compatible SSE endpoint via the `OPENAI_BASE_URL` override, then |
| 5 | //! drives newline-delimited JSON-RPC over stdio. The scripted model calls |
| 6 | //! `run_command` — a mutating tool — so the agent must send |
| 7 | //! `session/request_permission` mid-turn (the exact path the spawned-handler / |
| 8 | //! `turn_lock` design exists for). The test answers it twice: |
| 9 | //! |
| 10 | //! 1. `cancelled` — the prompt must stop with `stopReason: "cancelled"`, and |
| 11 | //! the *next* request to the endpoint must show the abandoned round closed |
| 12 | //! out with `role: "tool"` results, or a strict OpenAI-compatible endpoint |
| 13 | //! would reject the whole session. |
| 14 | //! 2. `selected: allow_once` — the tool must actually execute and its output |
| 15 | //! travel back to the endpoint as a tool result. |
| 16 | //! |
| 17 | //! A second test covers live progress: a slow command's `in_progress` tool call |
| 18 | //! must reach the client while the command is still running, which is what the |
| 19 | //! editor's spinner is driven from. A third test covers what the client |
| 20 | //! renders: text from two tool rounds must reach it as separate paragraphs, |
| 21 | //! since ACP clients concatenate consecutive agent-message chunks into one |
| 22 | //! block. |
| 23 | |
| 24 | use std::collections::VecDeque; |
| 25 | use std::io::{BufRead, BufReader, Read, Write}; |
| 26 | use std::net::TcpListener; |
| 27 | use std::process::{Child, ChildStdin, Command, Stdio}; |
| 28 | use std::sync::mpsc::{Receiver, channel}; |
| 29 | use std::sync::{Arc, Mutex}; |
| 30 | use std::time::{Duration, Instant}; |
| 31 | |
| 32 | use serde_json::{Value, json}; |
| 33 | |
| 34 | const TIMEOUT: Duration = Duration::from_secs(60); |
| 35 | |
| 36 | // ── Scripted OpenAI-compatible endpoint ───────────────────────────────────── |
| 37 | |
| 38 | fn sse_body(events: &[Value]) -> String { |
| 39 | let mut body = String::new(); |
| 40 | for event in events { |
| 41 | body.push_str("data: "); |
| 42 | body.push_str(&event.to_string()); |
| 43 | body.push_str("\n\n"); |
| 44 | } |
| 45 | body.push_str("data: [DONE]\n\n"); |
| 46 | body |
| 47 | } |
| 48 | |
| 49 | fn sse_tool_call(id: &str, name: &str, arguments: &str) -> String { |
| 50 | sse_body(&[json!({ |
| 51 | "choices": [{"delta": {"tool_calls": [{ |
| 52 | "index": 0, |
| 53 | "id": id, |
| 54 | "function": {"name": name, "arguments": arguments}, |
| 55 | }]}}] |
| 56 | })]) |
| 57 | } |
| 58 | |
| 59 | fn sse_text(text: &str) -> String { |
| 60 | sse_body(&[json!({"choices": [{"delta": {"content": text}}]})]) |
| 61 | } |
| 62 | |
| 63 | /// One round that says something and then asks for a tool — the shape that used |
| 64 | /// to leave the next round's text glued to this sentence. |
| 65 | fn sse_text_then_tool_call(text: &str, id: &str, name: &str, arguments: &str) -> String { |
| 66 | sse_body(&[ |
| 67 | json!({"choices": [{"delta": {"content": text}}]}), |
| 68 | json!({ |
| 69 | "choices": [{"delta": {"tool_calls": [{ |
| 70 | "index": 0, |
| 71 | "id": id, |
| 72 | "function": {"name": name, "arguments": arguments}, |
| 73 | }]}}] |
| 74 | }), |
| 75 | ]) |
| 76 | } |
| 77 | |
| 78 | /// Serves one scripted SSE response per request and records each request body. |
| 79 | struct FakeEndpoint { |
| 80 | port: u16, |
| 81 | requests: Arc<Mutex<Vec<Value>>>, |
| 82 | } |
| 83 | |
| 84 | fn start_fake_endpoint(responses: Vec<String>) -> FakeEndpoint { |
| 85 | let listener = TcpListener::bind("127.0.0.1:0").expect("bind fake endpoint"); |
| 86 | let port = listener.local_addr().unwrap().port(); |
| 87 | let requests: Arc<Mutex<Vec<Value>>> = Arc::default(); |
| 88 | let recorded = Arc::clone(&requests); |
| 89 | let queue = Mutex::new(VecDeque::from(responses)); |
| 90 | |
| 91 | std::thread::spawn(move || { |
| 92 | // `connection: close` below means one request per connection, so the |
| 93 | // serial accept loop matches the agent's serial completion requests. |
| 94 | for stream in listener.incoming() { |
| 95 | let Ok(mut stream) = stream else { continue }; |
| 96 | let mut reader = BufReader::new(match stream.try_clone() { |
| 97 | Ok(clone) => clone, |
| 98 | Err(_) => continue, |
| 99 | }); |
| 100 | let mut content_length = 0usize; |
| 101 | loop { |
| 102 | let mut line = String::new(); |
| 103 | if reader.read_line(&mut line).unwrap_or(0) == 0 { |
| 104 | break; |
| 105 | } |
| 106 | let line = line.trim(); |
| 107 | if line.is_empty() { |
| 108 | break; |
| 109 | } |
| 110 | if let Some(length) = line.to_ascii_lowercase().strip_prefix("content-length:") { |
| 111 | content_length = length.trim().parse().unwrap_or(0); |
| 112 | } |
| 113 | } |
| 114 | let mut body = vec![0u8; content_length]; |
| 115 | if reader.read_exact(&mut body).is_err() { |
| 116 | continue; |
| 117 | } |
| 118 | if let Ok(request) = serde_json::from_slice::<Value>(&body) { |
| 119 | recorded.lock().unwrap().push(request); |
| 120 | } |
| 121 | let payload = queue |
| 122 | .lock() |
| 123 | .unwrap() |
| 124 | .pop_front() |
| 125 | .unwrap_or_else(|| sse_text("out of scripted responses")); |
| 126 | let response = format!( |
| 127 | "HTTP/1.1 200 OK\r\ncontent-type: text/event-stream\r\n\ |
| 128 | content-length: {}\r\nconnection: close\r\n\r\n{}", |
| 129 | payload.len(), |
| 130 | payload |
| 131 | ); |
| 132 | let _ = stream.write_all(response.as_bytes()); |
| 133 | } |
| 134 | }); |
| 135 | |
| 136 | FakeEndpoint { port, requests } |
| 137 | } |
| 138 | |
| 139 | // ── ACP client over the binary's stdio ────────────────────────────────────── |
| 140 | |
| 141 | struct AgentUnderTest { |
| 142 | child: Child, |
| 143 | stdin: ChildStdin, |
| 144 | incoming: Receiver<Value>, |
| 145 | next_id: u64, |
| 146 | } |
| 147 | |
| 148 | fn spawn_agent(port: u16, config_dir: &std::path::Path) -> AgentUnderTest { |
| 149 | spawn_agent_with_permissions(port, config_dir, None) |
| 150 | } |
| 151 | |
| 152 | /// `permissions` sets `SIGIT_PERMISSIONS`; `None` leaves the default `ask` mode |
| 153 | /// in place, which is what the permission round-trip needs. |
| 154 | fn spawn_agent_with_permissions( |
| 155 | port: u16, |
| 156 | config_dir: &std::path::Path, |
| 157 | permissions: Option<&str>, |
| 158 | ) -> AgentUnderTest { |
| 159 | let mut command = Command::new(env!("CARGO_BIN_EXE_sigit")); |
| 160 | command |
| 161 | .env("OPENAI_BASE_URL", format!("http://127.0.0.1:{port}")) |
| 162 | .env("OPENAI_API_KEY", "test-key") |
| 163 | .env("SIGIT_MODEL", "scripted-model") |
| 164 | .env("SIGIT_CONFIG_DIR", config_dir) |
| 165 | .env("SIGIT_MCP", "off") |
| 166 | .env_remove("SIGIT_LOCAL_INFERENCE") |
| 167 | .stdin(Stdio::piped()) |
| 168 | .stdout(Stdio::piped()) |
| 169 | .stderr(Stdio::null()); |
| 170 | match permissions { |
| 171 | // A fresh config dir means the default permission mode, `ask` — make |
| 172 | // sure the environment can't turn the gate off underneath the test. |
| 173 | None => command.env_remove("SIGIT_PERMISSIONS"), |
| 174 | Some(mode) => command.env("SIGIT_PERMISSIONS", mode), |
| 175 | }; |
| 176 | let mut child = command.spawn().expect("spawn sigit in ACP mode"); |
| 177 | |
| 178 | let stdout = child.stdout.take().unwrap(); |
| 179 | let (message_tx, incoming) = channel(); |
| 180 | std::thread::spawn(move || { |
| 181 | for line in BufReader::new(stdout).lines() { |
| 182 | let Ok(line) = line else { break }; |
| 183 | if let Ok(message) = serde_json::from_str::<Value>(&line) |
| 184 | && message_tx.send(message).is_err() |
| 185 | { |
| 186 | break; |
| 187 | } |
| 188 | } |
| 189 | }); |
| 190 | |
| 191 | let stdin = child.stdin.take().unwrap(); |
| 192 | AgentUnderTest { |
| 193 | child, |
| 194 | stdin, |
| 195 | incoming, |
| 196 | next_id: 0, |
| 197 | } |
| 198 | } |
| 199 | |
| 200 | impl AgentUnderTest { |
| 201 | fn send(&mut self, message: Value) { |
| 202 | let mut line = message.to_string(); |
| 203 | line.push('\n'); |
| 204 | self.stdin |
| 205 | .write_all(line.as_bytes()) |
| 206 | .expect("write to agent stdin"); |
| 207 | self.stdin.flush().expect("flush agent stdin"); |
| 208 | } |
| 209 | |
| 210 | fn request(&mut self, method: &str, params: Value) -> u64 { |
| 211 | self.next_id += 1; |
| 212 | let id = self.next_id; |
| 213 | self.send(json!({"jsonrpc": "2.0", "id": id, "method": method, "params": params})); |
| 214 | id |
| 215 | } |
| 216 | |
| 217 | fn respond(&mut self, id: Value, result: Value) { |
| 218 | self.send(json!({"jsonrpc": "2.0", "id": id, "result": result})); |
| 219 | } |
| 220 | |
| 221 | /// Skip notifications and unrelated traffic until `matches` is satisfied. |
| 222 | fn wait_for(&mut self, what: &str, matches: impl Fn(&Value) -> bool) -> Value { |
| 223 | let deadline = Instant::now() + TIMEOUT; |
| 224 | loop { |
| 225 | let remaining = deadline.saturating_duration_since(Instant::now()); |
| 226 | match self.incoming.recv_timeout(remaining) { |
| 227 | Ok(message) if matches(&message) => return message, |
| 228 | Ok(_) => continue, |
| 229 | Err(_) => panic!("timed out waiting for {what}"), |
| 230 | } |
| 231 | } |
| 232 | } |
| 233 | |
| 234 | /// The response to one of *our* requests (has our id, no `method`). |
| 235 | fn wait_for_response(&mut self, id: u64) -> Value { |
| 236 | let response = self.wait_for(&format!("response to request {id}"), |message| { |
| 237 | message["id"] == id && message.get("method").is_none() |
| 238 | }); |
| 239 | assert!( |
| 240 | response.get("error").is_none(), |
| 241 | "request {id} failed: {response}" |
| 242 | ); |
| 243 | response |
| 244 | } |
| 245 | |
| 246 | /// The first `session/update` whose payload satisfies `matches`. |
| 247 | /// |
| 248 | /// Only used by tests that run with the permission gate off, so a |
| 249 | /// `session/request_permission` means the run is misconfigured — say that |
| 250 | /// rather than waiting out the timeout on a request nobody will answer. |
| 251 | fn wait_for_update(&mut self, what: &str, matches: impl Fn(&Value) -> bool) -> Value { |
| 252 | let message = self.wait_for(what, |message| { |
| 253 | assert!( |
| 254 | message["method"] != "session/request_permission", |
| 255 | "unexpected permission request while waiting for {what}: {message}" |
| 256 | ); |
| 257 | message["method"] == "session/update" && matches(&message["params"]["update"]) |
| 258 | }); |
| 259 | message["params"]["update"].clone() |
| 260 | } |
| 261 | |
| 262 | /// The response to `session/prompt`, paired with every `agent_message_chunk` |
| 263 | /// seen on the way, concatenated exactly as a client renders them. |
| 264 | fn wait_for_prompt(&mut self, id: u64) -> (Value, String) { |
| 265 | let mut rendered = String::new(); |
| 266 | let deadline = Instant::now() + TIMEOUT; |
| 267 | loop { |
| 268 | let remaining = deadline.saturating_duration_since(Instant::now()); |
| 269 | let Ok(message) = self.incoming.recv_timeout(remaining) else { |
| 270 | panic!("timed out waiting for the prompt response"); |
| 271 | }; |
| 272 | let update = &message["params"]["update"]; |
| 273 | if message["method"] == "session/update" |
| 274 | && update["sessionUpdate"] == "agent_message_chunk" |
| 275 | && let Some(chunk) = update["content"]["text"].as_str() |
| 276 | { |
| 277 | rendered.push_str(chunk); |
| 278 | } |
| 279 | if message["id"] == id && message.get("method").is_none() { |
| 280 | assert!(message.get("error").is_none(), "prompt failed: {message}"); |
| 281 | return (message, rendered); |
| 282 | } |
| 283 | } |
| 284 | } |
| 285 | |
| 286 | /// A request *from* the agent (has a `method` and its own id). |
| 287 | fn wait_for_agent_request(&mut self, method: &str) -> Value { |
| 288 | self.wait_for(&format!("agent request {method}"), |message| { |
| 289 | message["method"] == method && message.get("id").is_some() |
| 290 | }) |
| 291 | } |
| 292 | } |
| 293 | |
| 294 | impl Drop for AgentUnderTest { |
| 295 | fn drop(&mut self) { |
| 296 | let _ = self.child.kill(); |
| 297 | let _ = self.child.wait(); |
| 298 | } |
| 299 | } |
| 300 | |
| 301 | // ── The round-trip ────────────────────────────────────────────────────────── |
| 302 | |
| 303 | #[test] |
| 304 | fn permission_round_trip_cancel_then_allow() { |
| 305 | let endpoint = start_fake_endpoint(vec![ |
| 306 | sse_tool_call("call_1", "run_command", r#"{"command":"echo sigit-first"}"#), |
| 307 | sse_tool_call( |
| 308 | "call_2", |
| 309 | "run_command", |
| 310 | r#"{"command":"echo sigit-approved"}"#, |
| 311 | ), |
| 312 | sse_text("done"), |
| 313 | ]); |
| 314 | |
| 315 | let scratch = std::env::temp_dir().join(format!("sigit_acp_perm_{}", std::process::id())); |
| 316 | let config_dir = scratch.join("config"); |
| 317 | let cwd = scratch.join("cwd"); |
| 318 | std::fs::create_dir_all(&config_dir).unwrap(); |
| 319 | std::fs::create_dir_all(&cwd).unwrap(); |
| 320 | |
| 321 | let mut agent = spawn_agent(endpoint.port, &config_dir); |
| 322 | |
| 323 | let id = agent.request( |
| 324 | "initialize", |
| 325 | json!({"protocolVersion": 1, "clientCapabilities": {}}), |
| 326 | ); |
| 327 | agent.wait_for_response(id); |
| 328 | |
| 329 | let id = agent.request("session/new", json!({"cwd": cwd, "mcpServers": []})); |
| 330 | let session_id = agent.wait_for_response(id)["result"]["sessionId"] |
| 331 | .as_str() |
| 332 | .expect("session id") |
| 333 | .to_string(); |
| 334 | |
| 335 | // ── Turn 1: cancel at the permission gate ─────────────────────────── |
| 336 | let prompt_id = agent.request( |
| 337 | "session/prompt", |
| 338 | json!({ |
| 339 | "sessionId": session_id, |
| 340 | "prompt": [{"type": "text", "text": "run the first command"}], |
| 341 | }), |
| 342 | ); |
| 343 | |
| 344 | let permission = agent.wait_for_agent_request("session/request_permission"); |
| 345 | let params = &permission["params"]; |
| 346 | assert_eq!(params["sessionId"], session_id.as_str()); |
| 347 | let title = params["toolCall"]["title"].as_str().expect("title"); |
| 348 | assert!( |
| 349 | title.contains("run_command") && title.contains("echo sigit-first"), |
| 350 | "the dialog must show the tool and its arguments, got: {title}" |
| 351 | ); |
| 352 | assert_eq!( |
| 353 | params["toolCall"]["rawInput"]["command"], "echo sigit-first", |
| 354 | "full arguments must travel as rawInput" |
| 355 | ); |
| 356 | let option_ids: Vec<&str> = params["options"] |
| 357 | .as_array() |
| 358 | .expect("options") |
| 359 | .iter() |
| 360 | .map(|option| option["optionId"].as_str().unwrap_or_default()) |
| 361 | .collect(); |
| 362 | assert_eq!(option_ids, ["allow_once", "allow_session", "reject_once"]); |
| 363 | |
| 364 | agent.respond( |
| 365 | permission["id"].clone(), |
| 366 | json!({"outcome": {"outcome": "cancelled"}}), |
| 367 | ); |
| 368 | |
| 369 | let response = agent.wait_for_response(prompt_id); |
| 370 | assert_eq!(response["result"]["stopReason"], "cancelled"); |
| 371 | |
| 372 | // ── Turn 2: history must be repaired; then approve once ───────────── |
| 373 | let prompt_id = agent.request( |
| 374 | "session/prompt", |
| 375 | json!({ |
| 376 | "sessionId": session_id, |
| 377 | "prompt": [{"type": "text", "text": "run the second command"}], |
| 378 | }), |
| 379 | ); |
| 380 | |
| 381 | let permission = agent.wait_for_agent_request("session/request_permission"); |
| 382 | agent.respond( |
| 383 | permission["id"].clone(), |
| 384 | json!({"outcome": {"outcome": "selected", "optionId": "allow_once"}}), |
| 385 | ); |
| 386 | |
| 387 | let response = agent.wait_for_response(prompt_id); |
| 388 | assert_eq!(response["result"]["stopReason"], "end_turn"); |
| 389 | |
| 390 | // ── What the endpoint saw ──────────────────────────────────────────── |
| 391 | let requests = endpoint.requests.lock().unwrap(); |
| 392 | assert_eq!(requests.len(), 3, "expected exactly three completions"); |
| 393 | |
| 394 | // The client explicitly requests automatic tool selection. Some |
| 395 | // OpenAI-compatible gateways do not honour OpenAI's implicit `auto` |
| 396 | // default, which otherwise lets the model describe a tool action in prose |
| 397 | // and end the turn without a tool call. |
| 398 | assert_eq!(requests[0]["tool_choice"], "auto"); |
| 399 | |
| 400 | // Request 2 replays the full history: the cancelled round's tool call |
| 401 | // must be answered by a `role: "tool"` message, not left dangling. |
| 402 | let messages = requests[1]["messages"].as_array().expect("messages"); |
| 403 | let call_position = messages |
| 404 | .iter() |
| 405 | .position(|message| message["tool_calls"][0]["id"] == "call_1") |
| 406 | .expect("cancelled turn's assistant tool call in replayed history"); |
| 407 | let repair = &messages[call_position + 1]; |
| 408 | assert_eq!(repair["role"], "tool", "dangling tool call not closed out"); |
| 409 | assert_eq!(repair["tool_call_id"], "call_1"); |
| 410 | assert!( |
| 411 | repair["content"] |
| 412 | .as_str() |
| 413 | .unwrap_or_default() |
| 414 | .contains("cancelled"), |
| 415 | "repair message should say the turn was cancelled: {repair}" |
| 416 | ); |
| 417 | |
| 418 | // Request 3 carries the approved call's real output. |
| 419 | let messages = requests[2]["messages"].as_array().expect("messages"); |
| 420 | let result = messages |
| 421 | .iter() |
| 422 | .find(|message| message["role"] == "tool" && message["tool_call_id"] == "call_2") |
| 423 | .expect("tool result for the approved call"); |
| 424 | assert!( |
| 425 | result["content"] |
| 426 | .as_str() |
| 427 | .unwrap_or_default() |
| 428 | .contains("sigit-approved"), |
| 429 | "the approved command's output should reach the endpoint: {result}" |
| 430 | ); |
| 431 | |
| 432 | drop(agent); |
| 433 | let _ = std::fs::remove_dir_all(&scratch); |
| 434 | } |
| 435 | |
| 436 | #[test] |
| 437 | fn a_slow_commands_progress_reaches_the_client_while_it_runs() { |
| 438 | let scratch = std::env::temp_dir().join(format!("sigit_acp_progress_{}", std::process::id())); |
| 439 | let config_dir = scratch.join("config"); |
| 440 | let cwd = scratch.join("cwd"); |
| 441 | std::fs::create_dir_all(&config_dir).unwrap(); |
| 442 | std::fs::create_dir_all(&cwd).unwrap(); |
| 443 | |
| 444 | // The command has to run for a few seconds through `sh -c` / `cmd /C`. |
| 445 | // `sleep` is neither a cmd builtin nor on a Windows runner's PATH (Git's |
| 446 | // usr\bin, which carries the MSYS coreutils, is deliberately left off it), |
| 447 | // and the Windows job hung on it rather than failing fast. `ping` ships in |
| 448 | // System32, waits a second between echoes, and never reads stdin. |
| 449 | // |
| 450 | // Four seconds against the one-second assertion below is deliberate |
| 451 | // padding. A busy runner only ever widens the gap being measured — it |
| 452 | // cannot make the command finish sooner — so the one way a correct agent |
| 453 | // could fail here is this thread being descheduled for seconds between |
| 454 | // reading the two updates off the channel. Three seconds of slack covers |
| 455 | // that without weakening the assertion, which is separating four seconds |
| 456 | // from the 164µs the blocking version produced. |
| 457 | #[cfg(unix)] |
| 458 | let command = "sleep 4"; |
| 459 | #[cfg(windows)] |
| 460 | let command = "ping -n 5 127.0.0.1"; |
| 461 | |
| 462 | // `SIGIT_PERMISSIONS=allow` skips the gate, so the only thing between the |
| 463 | // two updates below is the command itself. |
| 464 | let endpoint = start_fake_endpoint(vec![ |
| 465 | sse_tool_call( |
| 466 | "call_1", |
| 467 | "run_command", |
| 468 | &json!({"command": command, "cwd": cwd}).to_string(), |
| 469 | ), |
| 470 | sse_text("done"), |
| 471 | ]); |
| 472 | |
| 473 | let mut agent = spawn_agent_with_permissions(endpoint.port, &config_dir, Some("allow")); |
| 474 | |
| 475 | let id = agent.request( |
| 476 | "initialize", |
| 477 | json!({"protocolVersion": 1, "clientCapabilities": {}}), |
| 478 | ); |
| 479 | agent.wait_for_response(id); |
| 480 | |
| 481 | let id = agent.request("session/new", json!({"cwd": cwd, "mcpServers": []})); |
| 482 | let session_id = agent.wait_for_response(id)["result"]["sessionId"] |
| 483 | .as_str() |
| 484 | .expect("session id") |
| 485 | .to_string(); |
| 486 | |
| 487 | let prompt_id = agent.request( |
| 488 | "session/prompt", |
| 489 | json!({ |
| 490 | "sessionId": session_id, |
| 491 | "prompt": [{"type": "text", "text": "run the slow command"}], |
| 492 | }), |
| 493 | ); |
| 494 | |
| 495 | let started = agent.wait_for_update("the tool call starting", |update| { |
| 496 | update["sessionUpdate"] == "tool_call" && update["toolCallId"] == "call_1" |
| 497 | }); |
| 498 | let in_progress_at = Instant::now(); |
| 499 | assert_eq!( |
| 500 | started["status"], "in_progress", |
| 501 | "a running tool call must be announced as in_progress: {started}" |
| 502 | ); |
| 503 | |
| 504 | let finished = agent.wait_for_update("the tool call finishing", |update| { |
| 505 | update["sessionUpdate"] == "tool_call_update" |
| 506 | && update["toolCallId"] == "call_1" |
| 507 | && update["status"] == "completed" |
| 508 | }); |
| 509 | |
| 510 | // The command runs between the two updates. If the turn blocked the |
| 511 | // connection's actors, both would only reach the client once it had |
| 512 | // already finished, and the editor would never draw a spinner. |
| 513 | assert!( |
| 514 | in_progress_at.elapsed() >= Duration::from_millis(1_000), |
| 515 | "in_progress arrived only {:?} before completed — the client had no \ |
| 516 | window to show progress in. `{command}` said: {}", |
| 517 | in_progress_at.elapsed(), |
| 518 | finished["rawOutput"] |
| 519 | ); |
| 520 | |
| 521 | let response = agent.wait_for_response(prompt_id); |
| 522 | assert_eq!(response["result"]["stopReason"], "end_turn"); |
| 523 | |
| 524 | drop(agent); |
| 525 | let _ = std::fs::remove_dir_all(&scratch); |
| 526 | } |
| 527 | |
| 528 | #[test] |
| 529 | fn text_from_two_tool_rounds_reaches_the_client_as_separate_paragraphs() { |
| 530 | let endpoint = start_fake_endpoint(vec![ |
| 531 | // `list_directory` is read-only, so it runs without a permission gate |
| 532 | // and the two rounds follow each other straight away. |
| 533 | sse_text_then_tool_call( |
| 534 | "I'll search for this pattern.", |
| 535 | "call_1", |
| 536 | "list_directory", |
| 537 | r#"{"path":"."}"#, |
| 538 | ), |
| 539 | sse_text("Let me broaden the search:"), |
| 540 | ]); |
| 541 | |
| 542 | let scratch = std::env::temp_dir().join(format!("sigit_acp_spacing_{}", std::process::id())); |
| 543 | let config_dir = scratch.join("config"); |
| 544 | let cwd = scratch.join("cwd"); |
| 545 | std::fs::create_dir_all(&config_dir).unwrap(); |
| 546 | std::fs::create_dir_all(&cwd).unwrap(); |
| 547 | |
| 548 | let mut agent = spawn_agent(endpoint.port, &config_dir); |
| 549 | |
| 550 | let id = agent.request( |
| 551 | "initialize", |
| 552 | json!({"protocolVersion": 1, "clientCapabilities": {}}), |
| 553 | ); |
| 554 | agent.wait_for_response(id); |
| 555 | |
| 556 | let id = agent.request("session/new", json!({"cwd": cwd, "mcpServers": []})); |
| 557 | let session_id = agent.wait_for_response(id)["result"]["sessionId"] |
| 558 | .as_str() |
| 559 | .expect("session id") |
| 560 | .to_string(); |
| 561 | |
| 562 | let prompt_id = agent.request( |
| 563 | "session/prompt", |
| 564 | json!({ |
| 565 | "sessionId": session_id, |
| 566 | "prompt": [{"type": "text", "text": "find it"}], |
| 567 | }), |
| 568 | ); |
| 569 | let (response, rendered) = agent.wait_for_prompt(prompt_id); |
| 570 | assert_eq!(response["result"]["stopReason"], "end_turn"); |
| 571 | |
| 572 | assert!( |
| 573 | rendered.contains("pattern.\n\nLet me broaden"), |
| 574 | "the second round must open a new paragraph, got: {rendered:?}" |
| 575 | ); |
| 576 | assert!( |
| 577 | !rendered.contains("pattern.Let me"), |
| 578 | "rounds must not run together into one sentence, got: {rendered:?}" |
| 579 | ); |
| 580 | |
| 581 | drop(agent); |
| 582 | let _ = std::fs::remove_dir_all(&scratch); |
| 583 | } |