fix/acp-replay-loaded-session-history
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 | fn wait_for_response_with_updates(&mut self, id: u64) -> (Value, Vec<Value>) { |
| 287 | let mut updates = Vec::new(); |
| 288 | let deadline = Instant::now() + TIMEOUT; |
| 289 | loop { |
| 290 | let remaining = deadline.saturating_duration_since(Instant::now()); |
| 291 | let Ok(message) = self.incoming.recv_timeout(remaining) else { |
| 292 | panic!("timed out waiting for response to request {id}"); |
| 293 | }; |
| 294 | if message["method"] == "session/update" { |
| 295 | updates.push(message["params"]["update"].clone()); |
| 296 | } |
| 297 | if message["id"] == id && message.get("method").is_none() { |
| 298 | assert!( |
| 299 | message.get("error").is_none(), |
| 300 | "request {id} failed: {message}" |
| 301 | ); |
| 302 | return (message, updates); |
| 303 | } |
| 304 | } |
| 305 | } |
| 306 | |
| 307 | /// A request *from* the agent (has a `method` and its own id). |
| 308 | fn wait_for_agent_request(&mut self, method: &str) -> Value { |
| 309 | self.wait_for(&format!("agent request {method}"), |message| { |
| 310 | message["method"] == method && message.get("id").is_some() |
| 311 | }) |
| 312 | } |
| 313 | } |
| 314 | |
| 315 | impl Drop for AgentUnderTest { |
| 316 | fn drop(&mut self) { |
| 317 | let _ = self.child.kill(); |
| 318 | let _ = self.child.wait(); |
| 319 | } |
| 320 | } |
| 321 | |
| 322 | // ── The round-trip ────────────────────────────────────────────────────────── |
| 323 | |
| 324 | #[test] |
| 325 | fn permission_round_trip_cancel_then_allow() { |
| 326 | let endpoint = start_fake_endpoint(vec![ |
| 327 | sse_tool_call("call_1", "run_command", r#"{"command":"echo sigit-first"}"#), |
| 328 | sse_tool_call( |
| 329 | "call_2", |
| 330 | "run_command", |
| 331 | r#"{"command":"echo sigit-approved"}"#, |
| 332 | ), |
| 333 | sse_text("done"), |
| 334 | ]); |
| 335 | |
| 336 | let scratch = std::env::temp_dir().join(format!("sigit_acp_perm_{}", std::process::id())); |
| 337 | let config_dir = scratch.join("config"); |
| 338 | let cwd = scratch.join("cwd"); |
| 339 | std::fs::create_dir_all(&config_dir).unwrap(); |
| 340 | std::fs::create_dir_all(&cwd).unwrap(); |
| 341 | |
| 342 | let mut agent = spawn_agent(endpoint.port, &config_dir); |
| 343 | |
| 344 | let id = agent.request( |
| 345 | "initialize", |
| 346 | json!({"protocolVersion": 1, "clientCapabilities": {}}), |
| 347 | ); |
| 348 | agent.wait_for_response(id); |
| 349 | |
| 350 | let id = agent.request("session/new", json!({"cwd": cwd, "mcpServers": []})); |
| 351 | let session_id = agent.wait_for_response(id)["result"]["sessionId"] |
| 352 | .as_str() |
| 353 | .expect("session id") |
| 354 | .to_string(); |
| 355 | |
| 356 | // ── Turn 1: cancel at the permission gate ─────────────────────────── |
| 357 | let prompt_id = agent.request( |
| 358 | "session/prompt", |
| 359 | json!({ |
| 360 | "sessionId": session_id, |
| 361 | "prompt": [{"type": "text", "text": "run the first command"}], |
| 362 | }), |
| 363 | ); |
| 364 | |
| 365 | let permission = agent.wait_for_agent_request("session/request_permission"); |
| 366 | let params = &permission["params"]; |
| 367 | assert_eq!(params["sessionId"], session_id.as_str()); |
| 368 | let title = params["toolCall"]["title"].as_str().expect("title"); |
| 369 | assert!( |
| 370 | title.contains("run_command") && title.contains("echo sigit-first"), |
| 371 | "the dialog must show the tool and its arguments, got: {title}" |
| 372 | ); |
| 373 | assert_eq!( |
| 374 | params["toolCall"]["rawInput"]["command"], "echo sigit-first", |
| 375 | "full arguments must travel as rawInput" |
| 376 | ); |
| 377 | let option_ids: Vec<&str> = params["options"] |
| 378 | .as_array() |
| 379 | .expect("options") |
| 380 | .iter() |
| 381 | .map(|option| option["optionId"].as_str().unwrap_or_default()) |
| 382 | .collect(); |
| 383 | assert_eq!(option_ids, ["allow_once", "allow_session", "reject_once"]); |
| 384 | |
| 385 | agent.respond( |
| 386 | permission["id"].clone(), |
| 387 | json!({"outcome": {"outcome": "cancelled"}}), |
| 388 | ); |
| 389 | |
| 390 | let response = agent.wait_for_response(prompt_id); |
| 391 | assert_eq!(response["result"]["stopReason"], "cancelled"); |
| 392 | |
| 393 | // ── Turn 2: history must be repaired; then approve once ───────────── |
| 394 | let prompt_id = agent.request( |
| 395 | "session/prompt", |
| 396 | json!({ |
| 397 | "sessionId": session_id, |
| 398 | "prompt": [{"type": "text", "text": "run the second command"}], |
| 399 | }), |
| 400 | ); |
| 401 | |
| 402 | let permission = agent.wait_for_agent_request("session/request_permission"); |
| 403 | agent.respond( |
| 404 | permission["id"].clone(), |
| 405 | json!({"outcome": {"outcome": "selected", "optionId": "allow_once"}}), |
| 406 | ); |
| 407 | |
| 408 | let response = agent.wait_for_response(prompt_id); |
| 409 | assert_eq!(response["result"]["stopReason"], "end_turn"); |
| 410 | |
| 411 | // ── What the endpoint saw ──────────────────────────────────────────── |
| 412 | let requests = endpoint.requests.lock().unwrap(); |
| 413 | assert_eq!(requests.len(), 3, "expected exactly three completions"); |
| 414 | |
| 415 | // The client explicitly requests automatic tool selection. Some |
| 416 | // OpenAI-compatible gateways do not honour OpenAI's implicit `auto` |
| 417 | // default, which otherwise lets the model describe a tool action in prose |
| 418 | // and end the turn without a tool call. |
| 419 | assert_eq!(requests[0]["tool_choice"], "auto"); |
| 420 | |
| 421 | // Request 2 replays the full history: the cancelled round's tool call |
| 422 | // must be answered by a `role: "tool"` message, not left dangling. |
| 423 | let messages = requests[1]["messages"].as_array().expect("messages"); |
| 424 | let call_position = messages |
| 425 | .iter() |
| 426 | .position(|message| message["tool_calls"][0]["id"] == "call_1") |
| 427 | .expect("cancelled turn's assistant tool call in replayed history"); |
| 428 | let repair = &messages[call_position + 1]; |
| 429 | assert_eq!(repair["role"], "tool", "dangling tool call not closed out"); |
| 430 | assert_eq!(repair["tool_call_id"], "call_1"); |
| 431 | assert!( |
| 432 | repair["content"] |
| 433 | .as_str() |
| 434 | .unwrap_or_default() |
| 435 | .contains("cancelled"), |
| 436 | "repair message should say the turn was cancelled: {repair}" |
| 437 | ); |
| 438 | |
| 439 | // Request 3 carries the approved call's real output. |
| 440 | let messages = requests[2]["messages"].as_array().expect("messages"); |
| 441 | let result = messages |
| 442 | .iter() |
| 443 | .find(|message| message["role"] == "tool" && message["tool_call_id"] == "call_2") |
| 444 | .expect("tool result for the approved call"); |
| 445 | assert!( |
| 446 | result["content"] |
| 447 | .as_str() |
| 448 | .unwrap_or_default() |
| 449 | .contains("sigit-approved"), |
| 450 | "the approved command's output should reach the endpoint: {result}" |
| 451 | ); |
| 452 | |
| 453 | drop(agent); |
| 454 | let _ = std::fs::remove_dir_all(&scratch); |
| 455 | } |
| 456 | |
| 457 | #[test] |
| 458 | fn successful_config_option_changes_are_rendered_as_system_status_cards() { |
| 459 | let endpoint = start_fake_endpoint(vec![]); |
| 460 | |
| 461 | let scratch = |
| 462 | std::env::temp_dir().join(format!("sigit_acp_config_messages_{}", std::process::id())); |
| 463 | let config_dir = scratch.join("config"); |
| 464 | let cwd = scratch.join("cwd"); |
| 465 | std::fs::create_dir_all(&config_dir).unwrap(); |
| 466 | std::fs::create_dir_all(&cwd).unwrap(); |
| 467 | std::fs::write( |
| 468 | config_dir.join("credentials.toml"), |
| 469 | "access_token = \"tok_test\"\nemail = \"dev@sigit.si\"\n", |
| 470 | ) |
| 471 | .unwrap(); |
| 472 | |
| 473 | let mut agent = spawn_agent(endpoint.port, &config_dir); |
| 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 id = agent.request( |
| 488 | "session/set_config_option", |
| 489 | json!({ |
| 490 | "sessionId": session_id, |
| 491 | "configId": "sigit-local-inference", |
| 492 | "value": "local-inference-off", |
| 493 | }), |
| 494 | ); |
| 495 | let (response, updates) = agent.wait_for_response_with_updates(id); |
| 496 | assert!( |
| 497 | response["result"]["configOptions"].is_array(), |
| 498 | "config response should refresh picker options: {response}" |
| 499 | ); |
| 500 | assert!( |
| 501 | !updates |
| 502 | .iter() |
| 503 | .any(|update| update["sessionUpdate"] == "agent_message_chunk"), |
| 504 | "Local Inference picker changes are passive UI state, not chat text: {updates:?}" |
| 505 | ); |
| 506 | let status = updates |
| 507 | .iter() |
| 508 | .find(|update| { |
| 509 | update["sessionUpdate"] == "tool_call" |
| 510 | && update["kind"] == "think" |
| 511 | && update["status"] == "completed" |
| 512 | }) |
| 513 | .expect("Local Inference change should render as a system-style status card"); |
| 514 | assert_eq!( |
| 515 | status["title"], |
| 516 | "Local inference is off. siGit Code Cloud tiers are highlighted; pick one from Model." |
| 517 | ); |
| 518 | |
| 519 | let id = agent.request( |
| 520 | "session/set_config_option", |
| 521 | json!({ |
| 522 | "sessionId": session_id, |
| 523 | "configId": "sigit-model", |
| 524 | "value": "sigit-cloud:mini", |
| 525 | }), |
| 526 | ); |
| 527 | let (response, updates) = agent.wait_for_response_with_updates(id); |
| 528 | assert!( |
| 529 | response["result"]["configOptions"].is_array(), |
| 530 | "config response should refresh picker options: {response}" |
| 531 | ); |
| 532 | assert!( |
| 533 | !updates |
| 534 | .iter() |
| 535 | .any(|update| update["sessionUpdate"] == "agent_message_chunk"), |
| 536 | "successful model picker changes must not be rendered as assistant chat text: {updates:?}" |
| 537 | ); |
| 538 | let status = updates |
| 539 | .iter() |
| 540 | .find(|update| { |
| 541 | update["sessionUpdate"] == "tool_call" |
| 542 | && update["kind"] == "think" |
| 543 | && update["status"] == "completed" |
| 544 | }) |
| 545 | .expect("cloud tier change should render as a system-style status card"); |
| 546 | assert!( |
| 547 | status["title"] |
| 548 | .as_str() |
| 549 | .unwrap_or_default() |
| 550 | .starts_with("Switched to siGit Code Cloud"), |
| 551 | "unexpected cloud switch status title: {status}" |
| 552 | ); |
| 553 | |
| 554 | drop(agent); |
| 555 | let _ = std::fs::remove_dir_all(&scratch); |
| 556 | } |
| 557 | |
| 558 | #[test] |
| 559 | fn a_slow_commands_progress_reaches_the_client_while_it_runs() { |
| 560 | let scratch = std::env::temp_dir().join(format!("sigit_acp_progress_{}", std::process::id())); |
| 561 | let config_dir = scratch.join("config"); |
| 562 | let cwd = scratch.join("cwd"); |
| 563 | std::fs::create_dir_all(&config_dir).unwrap(); |
| 564 | std::fs::create_dir_all(&cwd).unwrap(); |
| 565 | |
| 566 | // The command has to run for a few seconds through `sh -c` / `cmd /C`. |
| 567 | // `sleep` is neither a cmd builtin nor on a Windows runner's PATH (Git's |
| 568 | // usr\bin, which carries the MSYS coreutils, is deliberately left off it), |
| 569 | // and the Windows job hung on it rather than failing fast. `ping` ships in |
| 570 | // System32, waits a second between echoes, and never reads stdin. |
| 571 | // |
| 572 | // Four seconds against the one-second assertion below is deliberate |
| 573 | // padding. A busy runner only ever widens the gap being measured — it |
| 574 | // cannot make the command finish sooner — so the one way a correct agent |
| 575 | // could fail here is this thread being descheduled for seconds between |
| 576 | // reading the two updates off the channel. Three seconds of slack covers |
| 577 | // that without weakening the assertion, which is separating four seconds |
| 578 | // from the 164µs the blocking version produced. |
| 579 | #[cfg(unix)] |
| 580 | let command = "sleep 4"; |
| 581 | #[cfg(windows)] |
| 582 | let command = "ping -n 5 127.0.0.1"; |
| 583 | |
| 584 | // `SIGIT_PERMISSIONS=allow` skips the gate, so the only thing between the |
| 585 | // two updates below is the command itself. |
| 586 | let endpoint = start_fake_endpoint(vec![ |
| 587 | sse_tool_call( |
| 588 | "call_1", |
| 589 | "run_command", |
| 590 | &json!({"command": command, "cwd": cwd}).to_string(), |
| 591 | ), |
| 592 | sse_text("done"), |
| 593 | ]); |
| 594 | |
| 595 | let mut agent = spawn_agent_with_permissions(endpoint.port, &config_dir, Some("allow")); |
| 596 | |
| 597 | let id = agent.request( |
| 598 | "initialize", |
| 599 | json!({"protocolVersion": 1, "clientCapabilities": {}}), |
| 600 | ); |
| 601 | agent.wait_for_response(id); |
| 602 | |
| 603 | let id = agent.request("session/new", json!({"cwd": cwd, "mcpServers": []})); |
| 604 | let session_id = agent.wait_for_response(id)["result"]["sessionId"] |
| 605 | .as_str() |
| 606 | .expect("session id") |
| 607 | .to_string(); |
| 608 | |
| 609 | let prompt_id = agent.request( |
| 610 | "session/prompt", |
| 611 | json!({ |
| 612 | "sessionId": session_id, |
| 613 | "prompt": [{"type": "text", "text": "run the slow command"}], |
| 614 | }), |
| 615 | ); |
| 616 | |
| 617 | let started = agent.wait_for_update("the tool call starting", |update| { |
| 618 | update["sessionUpdate"] == "tool_call" && update["toolCallId"] == "call_1" |
| 619 | }); |
| 620 | let in_progress_at = Instant::now(); |
| 621 | assert_eq!( |
| 622 | started["status"], "in_progress", |
| 623 | "a running tool call must be announced as in_progress: {started}" |
| 624 | ); |
| 625 | |
| 626 | let finished = agent.wait_for_update("the tool call finishing", |update| { |
| 627 | update["sessionUpdate"] == "tool_call_update" |
| 628 | && update["toolCallId"] == "call_1" |
| 629 | && update["status"] == "completed" |
| 630 | }); |
| 631 | |
| 632 | // The command runs between the two updates. If the turn blocked the |
| 633 | // connection's actors, both would only reach the client once it had |
| 634 | // already finished, and the editor would never draw a spinner. |
| 635 | assert!( |
| 636 | in_progress_at.elapsed() >= Duration::from_millis(1_000), |
| 637 | "in_progress arrived only {:?} before completed — the client had no \ |
| 638 | window to show progress in. `{command}` said: {}", |
| 639 | in_progress_at.elapsed(), |
| 640 | finished["rawOutput"] |
| 641 | ); |
| 642 | |
| 643 | let response = agent.wait_for_response(prompt_id); |
| 644 | assert_eq!(response["result"]["stopReason"], "end_turn"); |
| 645 | |
| 646 | drop(agent); |
| 647 | let _ = std::fs::remove_dir_all(&scratch); |
| 648 | } |
| 649 | |
| 650 | #[test] |
| 651 | fn text_from_two_tool_rounds_reaches_the_client_as_separate_paragraphs() { |
| 652 | let endpoint = start_fake_endpoint(vec![ |
| 653 | // `list_directory` is read-only, so it runs without a permission gate |
| 654 | // and the two rounds follow each other straight away. |
| 655 | sse_text_then_tool_call( |
| 656 | "I'll search for this pattern.", |
| 657 | "call_1", |
| 658 | "list_directory", |
| 659 | r#"{"path":"."}"#, |
| 660 | ), |
| 661 | sse_text("Let me broaden the search:"), |
| 662 | ]); |
| 663 | |
| 664 | let scratch = std::env::temp_dir().join(format!("sigit_acp_spacing_{}", std::process::id())); |
| 665 | let config_dir = scratch.join("config"); |
| 666 | let cwd = scratch.join("cwd"); |
| 667 | std::fs::create_dir_all(&config_dir).unwrap(); |
| 668 | std::fs::create_dir_all(&cwd).unwrap(); |
| 669 | |
| 670 | let mut agent = spawn_agent(endpoint.port, &config_dir); |
| 671 | |
| 672 | let id = agent.request( |
| 673 | "initialize", |
| 674 | json!({"protocolVersion": 1, "clientCapabilities": {}}), |
| 675 | ); |
| 676 | agent.wait_for_response(id); |
| 677 | |
| 678 | let id = agent.request("session/new", json!({"cwd": cwd, "mcpServers": []})); |
| 679 | let session_id = agent.wait_for_response(id)["result"]["sessionId"] |
| 680 | .as_str() |
| 681 | .expect("session id") |
| 682 | .to_string(); |
| 683 | |
| 684 | let prompt_id = agent.request( |
| 685 | "session/prompt", |
| 686 | json!({ |
| 687 | "sessionId": session_id, |
| 688 | "prompt": [{"type": "text", "text": "find it"}], |
| 689 | }), |
| 690 | ); |
| 691 | let (response, rendered) = agent.wait_for_prompt(prompt_id); |
| 692 | assert_eq!(response["result"]["stopReason"], "end_turn"); |
| 693 | |
| 694 | assert!( |
| 695 | rendered.contains("pattern.\n\nLet me broaden"), |
| 696 | "the second round must open a new paragraph, got: {rendered:?}" |
| 697 | ); |
| 698 | assert!( |
| 699 | !rendered.contains("pattern.Let me"), |
| 700 | "rounds must not run together into one sentence, got: {rendered:?}" |
| 701 | ); |
| 702 | |
| 703 | drop(agent); |
| 704 | let _ = std::fs::remove_dir_all(&scratch); |
| 705 | } |
| 706 | |
| 707 | /// A model that writes its tool call out as literal `<tool_call>` text instead |
| 708 | /// of using the structured field must still drive the loop. Before recovery |
| 709 | /// the tag was streamed to the client as prose and the turn ended with no tool |
| 710 | /// calls, so the request silently went unanswered (getsigit/sigit#73). |
| 711 | #[test] |
| 712 | fn a_tool_call_emitted_as_text_is_executed_rather_than_rendered() { |
| 713 | let endpoint = start_fake_endpoint(vec![ |
| 714 | // Split mid-tag, the way a real stream arrives. |
| 715 | sse_body(&[ |
| 716 | json!({"choices": [{"delta": {"content": "Checking the repo: <tool_c"}}]}), |
| 717 | json!({"choices": [{"delta": {"content": "all>command_output<arg_key>task_id</arg_key>"}}]}), |
| 718 | json!({"choices": [{"delta": {"content": "<arg_value>2</arg_value></tool_call>"}}]}), |
| 719 | ]), |
| 720 | sse_text("All done."), |
| 721 | ]); |
| 722 | |
| 723 | let scratch = std::env::temp_dir().join(format!("sigit_acp_inline_{}", std::process::id())); |
| 724 | let config_dir = scratch.join("config"); |
| 725 | let cwd = scratch.join("cwd"); |
| 726 | std::fs::create_dir_all(&config_dir).unwrap(); |
| 727 | std::fs::create_dir_all(&cwd).unwrap(); |
| 728 | |
| 729 | let mut agent = spawn_agent(endpoint.port, &config_dir); |
| 730 | |
| 731 | let id = agent.request( |
| 732 | "initialize", |
| 733 | json!({"protocolVersion": 1, "clientCapabilities": {}}), |
| 734 | ); |
| 735 | agent.wait_for_response(id); |
| 736 | |
| 737 | let id = agent.request("session/new", json!({"cwd": cwd, "mcpServers": []})); |
| 738 | let session_id = agent.wait_for_response(id)["result"]["sessionId"] |
| 739 | .as_str() |
| 740 | .expect("session id") |
| 741 | .to_string(); |
| 742 | |
| 743 | let prompt_id = agent.request( |
| 744 | "session/prompt", |
| 745 | json!({ |
| 746 | "sessionId": session_id, |
| 747 | "prompt": [{"type": "text", "text": "check the repo"}], |
| 748 | }), |
| 749 | ); |
| 750 | let (_response, rendered) = agent.wait_for_prompt(prompt_id); |
| 751 | |
| 752 | // The raw tag must never be shown to the user. |
| 753 | assert!( |
| 754 | !rendered.contains("<tool_call>") && !rendered.contains("<arg_key>"), |
| 755 | "raw tool-call markup reached the client: {rendered:?}" |
| 756 | ); |
| 757 | assert!( |
| 758 | rendered.contains("Checking the repo:"), |
| 759 | "surrounding prose should still stream: {rendered:?}" |
| 760 | ); |
| 761 | |
| 762 | // The loop has to keep going: a second request means the recovered call |
| 763 | // ran and its result was sent back. Before recovery the turn ended here. |
| 764 | let requests = endpoint.requests.lock().unwrap(); |
| 765 | assert!( |
| 766 | requests.len() >= 2, |
| 767 | "expected a follow-up request carrying the tool result, got {}", |
| 768 | requests.len() |
| 769 | ); |
| 770 | let follow_up = &requests[1]; |
| 771 | let messages = follow_up["messages"].as_array().expect("messages"); |
| 772 | assert!( |
| 773 | messages.iter().any(|message| message["role"] == "tool"), |
| 774 | "the follow-up should answer the recovered call: {messages:?}" |
| 775 | ); |
| 776 | assert!( |
| 777 | messages |
| 778 | .iter() |
| 779 | .any(|message| { message["role"] == "assistant" && message["tool_calls"].is_array() }), |
| 780 | "history must record the recovered call, not the raw tag: {messages:?}" |
| 781 | ); |
| 782 | drop(requests); |
| 783 | |
| 784 | drop(agent); |
| 785 | let _ = std::fs::remove_dir_all(&scratch); |
| 786 | } |
| 787 | |
| 788 | #[test] |
| 789 | fn write_todos_reaches_the_client_as_an_acp_plan() { |
| 790 | let endpoint = start_fake_endpoint(vec![ |
| 791 | sse_tool_call( |
| 792 | "call_1", |
| 793 | "write_todos", |
| 794 | &json!({ |
| 795 | "todos": [ |
| 796 | {"content": "Fetch issue screenshot", "status": "completed"}, |
| 797 | {"content": "Implement ACP styling", "status": "in_progress"}, |
| 798 | {"content": "Build and verify", "status": "pending"} |
| 799 | ] |
| 800 | }) |
| 801 | .to_string(), |
| 802 | ), |
| 803 | sse_text("Continuing."), |
| 804 | ]); |
| 805 | |
| 806 | let scratch = std::env::temp_dir().join(format!("sigit_acp_plan_{}", std::process::id())); |
| 807 | let config_dir = scratch.join("config"); |
| 808 | let cwd = scratch.join("cwd"); |
| 809 | std::fs::create_dir_all(&config_dir).unwrap(); |
| 810 | std::fs::create_dir_all(&cwd).unwrap(); |
| 811 | |
| 812 | let mut agent = spawn_agent(endpoint.port, &config_dir); |
| 813 | |
| 814 | let id = agent.request( |
| 815 | "initialize", |
| 816 | json!({"protocolVersion": 1, "clientCapabilities": {}}), |
| 817 | ); |
| 818 | agent.wait_for_response(id); |
| 819 | |
| 820 | let id = agent.request("session/new", json!({"cwd": cwd, "mcpServers": []})); |
| 821 | let session_id = agent.wait_for_response(id)["result"]["sessionId"] |
| 822 | .as_str() |
| 823 | .expect("session id") |
| 824 | .to_string(); |
| 825 | |
| 826 | let prompt_id = agent.request( |
| 827 | "session/prompt", |
| 828 | json!({ |
| 829 | "sessionId": session_id, |
| 830 | "prompt": [{"type": "text", "text": "fix the ACP display"}], |
| 831 | }), |
| 832 | ); |
| 833 | let (response, updates) = agent.wait_for_response_with_updates(prompt_id); |
| 834 | assert_eq!(response["result"]["stopReason"], "end_turn"); |
| 835 | |
| 836 | let plan = updates |
| 837 | .iter() |
| 838 | .find(|update| update["sessionUpdate"] == "plan") |
| 839 | .expect("write_todos should be rendered as an ACP plan update"); |
| 840 | assert_eq!(plan["entries"][0]["content"], "Fetch issue screenshot"); |
| 841 | assert_eq!(plan["entries"][0]["status"], "completed"); |
| 842 | assert_eq!(plan["entries"][1]["status"], "in_progress"); |
| 843 | assert_eq!(plan["entries"][2]["status"], "pending"); |
| 844 | |
| 845 | assert!( |
| 846 | !updates.iter().any(|update| { |
| 847 | update["sessionUpdate"] == "tool_call" && update["title"] == "write_todos" |
| 848 | }), |
| 849 | "write_todos must not show up as a generic tool-call card: {updates:?}" |
| 850 | ); |
| 851 | |
| 852 | drop(agent); |
| 853 | let _ = std::fs::remove_dir_all(&scratch); |
| 854 | } |
| 855 | |
| 856 | /// An off-enum status must not make the whole call vanish. `exec_write_todos` |
| 857 | /// renders it as pending and reports success, so the plan has to carry it the |
| 858 | /// same way; dropping the plan here would leave the client showing a stale list. |
| 859 | #[test] |
| 860 | fn write_todos_with_an_unknown_status_still_reaches_the_client() { |
| 861 | let endpoint = start_fake_endpoint(vec![ |
| 862 | sse_tool_call( |
| 863 | "call_1", |
| 864 | "write_todos", |
| 865 | &json!({ |
| 866 | "todos": [ |
| 867 | {"content": "Fetch issue screenshot", "status": "completed"}, |
| 868 | {"content": "Ship it", "status": "cancelled"} |
| 869 | ] |
| 870 | }) |
| 871 | .to_string(), |
| 872 | ), |
| 873 | sse_text("Continuing."), |
| 874 | ]); |
| 875 | |
| 876 | let scratch = std::env::temp_dir().join(format!("sigit_acp_plan_odd_{}", std::process::id())); |
| 877 | let config_dir = scratch.join("config"); |
| 878 | let cwd = scratch.join("cwd"); |
| 879 | std::fs::create_dir_all(&config_dir).unwrap(); |
| 880 | std::fs::create_dir_all(&cwd).unwrap(); |
| 881 | |
| 882 | let mut agent = spawn_agent(endpoint.port, &config_dir); |
| 883 | |
| 884 | let id = agent.request( |
| 885 | "initialize", |
| 886 | json!({"protocolVersion": 1, "clientCapabilities": {}}), |
| 887 | ); |
| 888 | agent.wait_for_response(id); |
| 889 | |
| 890 | let id = agent.request("session/new", json!({"cwd": cwd, "mcpServers": []})); |
| 891 | let session_id = agent.wait_for_response(id)["result"]["sessionId"] |
| 892 | .as_str() |
| 893 | .expect("session id") |
| 894 | .to_string(); |
| 895 | |
| 896 | let prompt_id = agent.request( |
| 897 | "session/prompt", |
| 898 | json!({ |
| 899 | "sessionId": session_id, |
| 900 | "prompt": [{"type": "text", "text": "plan the work"}], |
| 901 | }), |
| 902 | ); |
| 903 | let (_response, updates) = agent.wait_for_response_with_updates(prompt_id); |
| 904 | |
| 905 | let plan = updates |
| 906 | .iter() |
| 907 | .find(|update| update["sessionUpdate"] == "plan") |
| 908 | .expect("an unknown status must not drop the plan"); |
| 909 | assert_eq!(plan["entries"][1]["content"], "Ship it"); |
| 910 | assert_eq!(plan["entries"][1]["status"], "pending"); |
| 911 | |
| 912 | drop(agent); |
| 913 | let _ = std::fs::remove_dir_all(&scratch); |
| 914 | } |
| 915 | |
| 916 | /// Arguments the converter cannot turn into a plan fall back to the ordinary |
| 917 | /// tool-call card. Without that fallback the call is announced nowhere at all. |
| 918 | #[test] |
| 919 | fn unconvertible_write_todos_falls_back_to_a_tool_call_card() { |
| 920 | let endpoint = start_fake_endpoint(vec![ |
| 921 | // `todos` is absent, so there is no plan to build. `exec_write_todos` |
| 922 | // answers with an error string, and the client still has to see the call. |
| 923 | sse_tool_call("call_1", "write_todos", &json!({"items": []}).to_string()), |
| 924 | sse_text("Continuing."), |
| 925 | ]); |
| 926 | |
| 927 | let scratch = std::env::temp_dir().join(format!("sigit_acp_plan_bad_{}", std::process::id())); |
| 928 | let config_dir = scratch.join("config"); |
| 929 | let cwd = scratch.join("cwd"); |
| 930 | std::fs::create_dir_all(&config_dir).unwrap(); |
| 931 | std::fs::create_dir_all(&cwd).unwrap(); |
| 932 | |
| 933 | let mut agent = spawn_agent(endpoint.port, &config_dir); |
| 934 | |
| 935 | let id = agent.request( |
| 936 | "initialize", |
| 937 | json!({"protocolVersion": 1, "clientCapabilities": {}}), |
| 938 | ); |
| 939 | agent.wait_for_response(id); |
| 940 | |
| 941 | let id = agent.request("session/new", json!({"cwd": cwd, "mcpServers": []})); |
| 942 | let session_id = agent.wait_for_response(id)["result"]["sessionId"] |
| 943 | .as_str() |
| 944 | .expect("session id") |
| 945 | .to_string(); |
| 946 | |
| 947 | let prompt_id = agent.request( |
| 948 | "session/prompt", |
| 949 | json!({ |
| 950 | "sessionId": session_id, |
| 951 | "prompt": [{"type": "text", "text": "plan the work"}], |
| 952 | }), |
| 953 | ); |
| 954 | let (_response, updates) = agent.wait_for_response_with_updates(prompt_id); |
| 955 | |
| 956 | assert!( |
| 957 | !updates |
| 958 | .iter() |
| 959 | .any(|update| update["sessionUpdate"] == "plan"), |
| 960 | "there is no plan to send: {updates:?}" |
| 961 | ); |
| 962 | assert!( |
| 963 | updates.iter().any(|update| { |
| 964 | update["sessionUpdate"] == "tool_call" && update["title"] == "write_todos" |
| 965 | }), |
| 966 | "the call must still be announced as a tool call: {updates:?}" |
| 967 | ); |
| 968 | |
| 969 | drop(agent); |
| 970 | let _ = std::fs::remove_dir_all(&scratch); |
| 971 | } |