development
claude/code-feature-parity-q003hm
claude/elegant-carson-l1menh
claude/sigit-acp-local-chat-cx6380
claude/sigit-cloud-agent-expansion-reox0a
claude/tool-permission-system
claude/zen-feynman-0u78dk
development
feature/agent-tools-multiedit-glob-todos-remember
feature/background-commands
feature/commit-coauthor-attribution
feature/headless-mode
feature/init-command
feature/load-local-model-explicitly
feature/session-persistence-compaction
feature/sigit-code-cloud
feature/subagent-tool
feature/tool-permission-system
feature/tui-repo-tabs
feature/tui-tabs
main
release/v1.3.1
| 1 | //! End-to-end stdio MCP transport test against the real binary. |
| 2 | //! |
| 3 | //! Spawns `sigit` in ACP mode wired to a scripted OpenAI-compatible SSE |
| 4 | //! endpoint (the same harness as `acp_permissions.rs`) and to a temp |
| 5 | //! `SIGIT_CONFIG_DIR` whose `mcp.toml` configures three stdio servers, all |
| 6 | //! backed by the `mcp_stdio_stub` helper binary: |
| 7 | //! |
| 8 | //! - `stub` — healthy; exposes one `echo` tool and prefixes replies with the |
| 9 | //! `STUB_PREFIX` env var from `[server.env]`, proving env propagation. |
| 10 | //! - `dying` — completes the discovery handshake, then exits. Its tool is |
| 11 | //! offered to the model, but calling it must fail with an in-band error. |
| 12 | //! - `deadone` — exits(1) at spawn; discovery must record it unavailable and |
| 13 | //! offer no tools for it. |
| 14 | //! |
| 15 | //! The scripted model calls `mcp__stub__echo`, then `mcp__dying__echo`, then |
| 16 | //! finishes. The test asserts the tool round-trip, the dead-child error, and |
| 17 | //! the `/mcp` listing (command lines shown, dead server flagged). |
| 18 | |
| 19 | use std::collections::VecDeque; |
| 20 | use std::io::{BufRead, BufReader, Read, Write}; |
| 21 | use std::net::TcpListener; |
| 22 | use std::process::{Child, ChildStdin, Command, Stdio}; |
| 23 | use std::sync::mpsc::{Receiver, channel}; |
| 24 | use std::sync::{Arc, Mutex}; |
| 25 | use std::time::{Duration, Instant}; |
| 26 | |
| 27 | use serde_json::{Value, json}; |
| 28 | |
| 29 | const TIMEOUT: Duration = Duration::from_secs(60); |
| 30 | |
| 31 | // ── Scripted OpenAI-compatible endpoint ───────────────────────────────────── |
| 32 | |
| 33 | fn sse_body(events: &[Value]) -> String { |
| 34 | let mut body = String::new(); |
| 35 | for event in events { |
| 36 | body.push_str("data: "); |
| 37 | body.push_str(&event.to_string()); |
| 38 | body.push_str("\n\n"); |
| 39 | } |
| 40 | body.push_str("data: [DONE]\n\n"); |
| 41 | body |
| 42 | } |
| 43 | |
| 44 | fn sse_tool_call(id: &str, name: &str, arguments: &str) -> String { |
| 45 | sse_body(&[json!({ |
| 46 | "choices": [{"delta": {"tool_calls": [{ |
| 47 | "index": 0, |
| 48 | "id": id, |
| 49 | "function": {"name": name, "arguments": arguments}, |
| 50 | }]}}] |
| 51 | })]) |
| 52 | } |
| 53 | |
| 54 | fn sse_text(text: &str) -> String { |
| 55 | sse_body(&[json!({"choices": [{"delta": {"content": text}}]})]) |
| 56 | } |
| 57 | |
| 58 | /// Serves one scripted SSE response per request and records each request body. |
| 59 | struct FakeEndpoint { |
| 60 | port: u16, |
| 61 | requests: Arc<Mutex<Vec<Value>>>, |
| 62 | } |
| 63 | |
| 64 | fn start_fake_endpoint(responses: Vec<String>) -> FakeEndpoint { |
| 65 | let listener = TcpListener::bind("127.0.0.1:0").expect("bind fake endpoint"); |
| 66 | let port = listener.local_addr().unwrap().port(); |
| 67 | let requests: Arc<Mutex<Vec<Value>>> = Arc::default(); |
| 68 | let recorded = Arc::clone(&requests); |
| 69 | let queue = Mutex::new(VecDeque::from(responses)); |
| 70 | |
| 71 | std::thread::spawn(move || { |
| 72 | for stream in listener.incoming() { |
| 73 | let Ok(mut stream) = stream else { continue }; |
| 74 | let mut reader = BufReader::new(match stream.try_clone() { |
| 75 | Ok(clone) => clone, |
| 76 | Err(_) => continue, |
| 77 | }); |
| 78 | let mut content_length = 0usize; |
| 79 | loop { |
| 80 | let mut line = String::new(); |
| 81 | if reader.read_line(&mut line).unwrap_or(0) == 0 { |
| 82 | break; |
| 83 | } |
| 84 | let line = line.trim(); |
| 85 | if line.is_empty() { |
| 86 | break; |
| 87 | } |
| 88 | if let Some(length) = line.to_ascii_lowercase().strip_prefix("content-length:") { |
| 89 | content_length = length.trim().parse().unwrap_or(0); |
| 90 | } |
| 91 | } |
| 92 | let mut body = vec![0u8; content_length]; |
| 93 | if reader.read_exact(&mut body).is_err() { |
| 94 | continue; |
| 95 | } |
| 96 | if let Ok(request) = serde_json::from_slice::<Value>(&body) { |
| 97 | recorded.lock().unwrap().push(request); |
| 98 | } |
| 99 | let payload = queue |
| 100 | .lock() |
| 101 | .unwrap() |
| 102 | .pop_front() |
| 103 | .unwrap_or_else(|| sse_text("out of scripted responses")); |
| 104 | let response = format!( |
| 105 | "HTTP/1.1 200 OK\r\ncontent-type: text/event-stream\r\n\ |
| 106 | content-length: {}\r\nconnection: close\r\n\r\n{}", |
| 107 | payload.len(), |
| 108 | payload |
| 109 | ); |
| 110 | let _ = stream.write_all(response.as_bytes()); |
| 111 | } |
| 112 | }); |
| 113 | |
| 114 | FakeEndpoint { port, requests } |
| 115 | } |
| 116 | |
| 117 | // ── ACP client over the binary's stdio ────────────────────────────────────── |
| 118 | |
| 119 | struct AgentUnderTest { |
| 120 | child: Child, |
| 121 | stdin: ChildStdin, |
| 122 | incoming: Receiver<Value>, |
| 123 | next_id: u64, |
| 124 | } |
| 125 | |
| 126 | fn spawn_agent(port: u16, config_dir: &std::path::Path, cwd: &std::path::Path) -> AgentUnderTest { |
| 127 | let mut child = Command::new(env!("CARGO_BIN_EXE_sigit")) |
| 128 | .current_dir(cwd) |
| 129 | .env("OPENAI_BASE_URL", format!("http://127.0.0.1:{port}")) |
| 130 | .env("OPENAI_API_KEY", "test-key") |
| 131 | .env("SIGIT_MODEL", "scripted-model") |
| 132 | .env("SIGIT_CONFIG_DIR", config_dir) |
| 133 | // MCP stays ON (that's what we test), but the baked-in official |
| 134 | // server must not phone home from CI. |
| 135 | .env("SIGIT_MCP_OFFICIAL", "off") |
| 136 | .env_remove("SIGIT_MCP") |
| 137 | // MCP tools are mutating and would otherwise wait at the permission |
| 138 | // gate; permissions have their own test. |
| 139 | .env("SIGIT_PERMISSIONS", "allow") |
| 140 | .env_remove("SIGIT_LOCAL_INFERENCE") |
| 141 | .stdin(Stdio::piped()) |
| 142 | .stdout(Stdio::piped()) |
| 143 | .stderr(Stdio::null()) |
| 144 | .spawn() |
| 145 | .expect("spawn sigit in ACP mode"); |
| 146 | |
| 147 | let stdout = child.stdout.take().unwrap(); |
| 148 | let (message_tx, incoming) = channel(); |
| 149 | std::thread::spawn(move || { |
| 150 | for line in BufReader::new(stdout).lines() { |
| 151 | let Ok(line) = line else { break }; |
| 152 | if let Ok(message) = serde_json::from_str::<Value>(&line) |
| 153 | && message_tx.send(message).is_err() |
| 154 | { |
| 155 | break; |
| 156 | } |
| 157 | } |
| 158 | }); |
| 159 | |
| 160 | let stdin = child.stdin.take().unwrap(); |
| 161 | AgentUnderTest { |
| 162 | child, |
| 163 | stdin, |
| 164 | incoming, |
| 165 | next_id: 0, |
| 166 | } |
| 167 | } |
| 168 | |
| 169 | impl AgentUnderTest { |
| 170 | fn send(&mut self, message: Value) { |
| 171 | let mut line = message.to_string(); |
| 172 | line.push('\n'); |
| 173 | self.stdin |
| 174 | .write_all(line.as_bytes()) |
| 175 | .expect("write to agent stdin"); |
| 176 | self.stdin.flush().expect("flush agent stdin"); |
| 177 | } |
| 178 | |
| 179 | fn request(&mut self, method: &str, params: Value) -> u64 { |
| 180 | self.next_id += 1; |
| 181 | let id = self.next_id; |
| 182 | self.send(json!({"jsonrpc": "2.0", "id": id, "method": method, "params": params})); |
| 183 | id |
| 184 | } |
| 185 | |
| 186 | /// Wait for the response to our request `id`, collecting the raw JSON of |
| 187 | /// every `session/update` notification that arrives before it. |
| 188 | fn wait_for_response_collecting_updates(&mut self, id: u64) -> (Value, String) { |
| 189 | let deadline = Instant::now() + TIMEOUT; |
| 190 | let mut updates = String::new(); |
| 191 | loop { |
| 192 | let remaining = deadline.saturating_duration_since(Instant::now()); |
| 193 | match self.incoming.recv_timeout(remaining) { |
| 194 | Ok(message) if message["id"] == id && message.get("method").is_none() => { |
| 195 | assert!( |
| 196 | message.get("error").is_none(), |
| 197 | "request {id} failed: {message}" |
| 198 | ); |
| 199 | return (message, updates); |
| 200 | } |
| 201 | Ok(message) => { |
| 202 | if message["method"] == "session/update" { |
| 203 | updates.push_str(&message["params"].to_string()); |
| 204 | updates.push('\n'); |
| 205 | } |
| 206 | } |
| 207 | Err(_) => panic!("timed out waiting for response to request {id}"), |
| 208 | } |
| 209 | } |
| 210 | } |
| 211 | |
| 212 | fn wait_for_response(&mut self, id: u64) -> Value { |
| 213 | self.wait_for_response_collecting_updates(id).0 |
| 214 | } |
| 215 | } |
| 216 | |
| 217 | impl Drop for AgentUnderTest { |
| 218 | fn drop(&mut self) { |
| 219 | let _ = self.child.kill(); |
| 220 | let _ = self.child.wait(); |
| 221 | } |
| 222 | } |
| 223 | |
| 224 | // ── The round-trip ────────────────────────────────────────────────────────── |
| 225 | |
| 226 | #[test] |
| 227 | fn stdio_mcp_discovery_call_and_dead_child() { |
| 228 | let stub = env!("CARGO_BIN_EXE_mcp_stdio_stub"); |
| 229 | |
| 230 | let endpoint = start_fake_endpoint(vec![ |
| 231 | sse_tool_call("call_1", "mcp__stub__echo", r#"{"text":"hello"}"#), |
| 232 | sse_tool_call("call_2", "mcp__dying__echo", r#"{"text":"gone"}"#), |
| 233 | sse_text("done"), |
| 234 | ]); |
| 235 | |
| 236 | let scratch = std::env::temp_dir().join(format!("sigit_mcp_stdio_{}", std::process::id())); |
| 237 | let config_dir = scratch.join("config"); |
| 238 | let cwd = scratch.join("cwd"); |
| 239 | std::fs::create_dir_all(&config_dir).unwrap(); |
| 240 | std::fs::create_dir_all(&cwd).unwrap(); |
| 241 | |
| 242 | // TOML literal strings (single quotes) keep Windows backslashes intact. |
| 243 | let mcp_toml = format!( |
| 244 | r#"official = false |
| 245 | |
| 246 | [[server]] |
| 247 | name = "stub" |
| 248 | command = '{stub}' |
| 249 | |
| 250 | [server.env] |
| 251 | STUB_PREFIX = "pfx:" |
| 252 | |
| 253 | [[server]] |
| 254 | name = "dying" |
| 255 | command = '{stub}' |
| 256 | args = ["--exit-after-list"] |
| 257 | |
| 258 | [[server]] |
| 259 | name = "deadone" |
| 260 | command = '{stub}' |
| 261 | args = ["--fail"] |
| 262 | "# |
| 263 | ); |
| 264 | std::fs::write(config_dir.join("mcp.toml"), mcp_toml).unwrap(); |
| 265 | |
| 266 | let mut agent = spawn_agent(endpoint.port, &config_dir, &cwd); |
| 267 | |
| 268 | let id = agent.request( |
| 269 | "initialize", |
| 270 | json!({"protocolVersion": 1, "clientCapabilities": {}}), |
| 271 | ); |
| 272 | agent.wait_for_response(id); |
| 273 | |
| 274 | let id = agent.request("session/new", json!({"cwd": cwd, "mcpServers": []})); |
| 275 | let session_id = agent.wait_for_response(id)["result"]["sessionId"] |
| 276 | .as_str() |
| 277 | .expect("session id") |
| 278 | .to_string(); |
| 279 | |
| 280 | // ── /mcp: listing shows command lines and flags the dead server ───── |
| 281 | let prompt_id = agent.request( |
| 282 | "session/prompt", |
| 283 | json!({ |
| 284 | "sessionId": session_id, |
| 285 | "prompt": [{"type": "text", "text": "/mcp"}], |
| 286 | }), |
| 287 | ); |
| 288 | let (_, listing) = agent.wait_for_response_collecting_updates(prompt_id); |
| 289 | // Raw JSON of the update notifications; escape the path the way JSON does |
| 290 | // so Windows backslashes compare correctly. |
| 291 | let stub_json = serde_json::to_string(stub).unwrap(); |
| 292 | let stub_escaped = stub_json.trim_matches('"'); |
| 293 | assert!( |
| 294 | listing.contains("mcp__stub__echo"), |
| 295 | "/mcp must list the healthy server's tool, got: {listing}" |
| 296 | ); |
| 297 | assert!( |
| 298 | listing.contains(stub_escaped), |
| 299 | "/mcp must show the stdio server's command line, got: {listing}" |
| 300 | ); |
| 301 | assert!( |
| 302 | listing.contains("--exit-after-list"), |
| 303 | "/mcp must include the args in the command line, got: {listing}" |
| 304 | ); |
| 305 | assert!( |
| 306 | listing.contains("unavailable"), |
| 307 | "/mcp must flag the server that died at spawn, got: {listing}" |
| 308 | ); |
| 309 | |
| 310 | // ── One prompt: echo round-trip, then the dead-child call ─────────── |
| 311 | let prompt_id = agent.request( |
| 312 | "session/prompt", |
| 313 | json!({ |
| 314 | "sessionId": session_id, |
| 315 | "prompt": [{"type": "text", "text": "use the stub tools"}], |
| 316 | }), |
| 317 | ); |
| 318 | let response = agent.wait_for_response(prompt_id); |
| 319 | assert_eq!(response["result"]["stopReason"], "end_turn"); |
| 320 | |
| 321 | // ── What the endpoint saw ──────────────────────────────────────────── |
| 322 | let requests = endpoint.requests.lock().unwrap(); |
| 323 | // Slash commands never reach the model, so all three completions belong |
| 324 | // to the tool-calling prompt. |
| 325 | assert_eq!(requests.len(), 3, "expected exactly three completions"); |
| 326 | |
| 327 | // The offered tool specs must include both live servers' echo tools and |
| 328 | // nothing from the server that failed discovery. |
| 329 | let tools = requests[0]["tools"].to_string(); |
| 330 | assert!( |
| 331 | tools.contains("mcp__stub__echo"), |
| 332 | "stub tool missing from specs: {tools}" |
| 333 | ); |
| 334 | assert!( |
| 335 | tools.contains("mcp__dying__echo"), |
| 336 | "dying server's tool missing from specs: {tools}" |
| 337 | ); |
| 338 | assert!( |
| 339 | !tools.contains("mcp__deadone__"), |
| 340 | "a server that failed discovery must contribute no tools: {tools}" |
| 341 | ); |
| 342 | |
| 343 | // The echo call's result must round-trip, carrying the [server.env] |
| 344 | // prefix (proving env vars reached the child). |
| 345 | let messages = requests[1]["messages"].as_array().expect("messages"); |
| 346 | let result = messages |
| 347 | .iter() |
| 348 | .find(|message| message["role"] == "tool" && message["tool_call_id"] == "call_1") |
| 349 | .expect("tool result for the echo call"); |
| 350 | assert_eq!( |
| 351 | result["content"].as_str().unwrap_or_default(), |
| 352 | "pfx:hello", |
| 353 | "echo result should carry the env-var prefix" |
| 354 | ); |
| 355 | |
| 356 | // The call to the server that died after discovery must come back as an |
| 357 | // in-band error string, not hang or crash the agent. |
| 358 | let messages = requests[2]["messages"].as_array().expect("messages"); |
| 359 | let result = messages |
| 360 | .iter() |
| 361 | .find(|message| message["role"] == "tool" && message["tool_call_id"] == "call_2") |
| 362 | .expect("tool result for the dead server's call"); |
| 363 | let content = result["content"].as_str().unwrap_or_default(); |
| 364 | assert!( |
| 365 | content.contains("Error") && content.contains("stdio server 'dying'"), |
| 366 | "dead-child call must fail in-band, got: {content}" |
| 367 | ); |
| 368 | |
| 369 | drop(agent); |
| 370 | let _ = std::fs::remove_dir_all(&scratch); |
| 371 | } |