fix(mcp): let run_command return once a server prints its ready line (#3670)

run_command takes wait_for_run's pattern: the call returns (waitEnded: matched) as soon as the output matches a server's ready line, and the tool description and MCP guide tell agents to use it for servers, watchers and apps instead of a long timeout.

Refs #3646
This commit is contained in:
l0ng-ai
2026-11-16 13:10:01 +08:00
committed by GitHub
parent ead2594d72
commit 2b353dfe53
4 changed files with 40 additions and 4 deletions
+13 -4
View File
@@ -128,9 +128,11 @@ pub fn run_command() -> Tool {
&format!(
"Runs a shell command line in a new or reused tab of OpenIDEA's Run tool window, with the shell of \
Settings › Tools › Terminal. The user is asked first (unless the IDE runs tools without \
confirmation). Waits up to `timeout` milliseconds (default {RUN_TIMEOUT_MS}) for the command to finish; a command \
still running then keeps running, and `wait_for_run`, `read_run_output`, `get_run_status`, `send_run_input` and \
`stop_run` follow it up by name. Returns the tab's status (`state`: starting, running, stopping, succeeded, failed, \
confirmation). Waits up to `timeout` milliseconds (default {RUN_TIMEOUT_MS}) for the command to finish, or until its \
output matches `pattern`; a command still running then keeps running, and `wait_for_run`, `read_run_output`, \
`get_run_status`, `send_run_input` and `stop_run` follow it up by name. A server, watcher or app never finishes: to \
start one, pass `pattern` for the line it prints once ready (e.g. `Listening on|Started .* in|ready in`) rather than \
a long `timeout`, so the call returns as soon as it is up. Returns the tab's status (`state`: starting, running, stopping, succeeded, failed, \
crashed, stoppedByAgent, stoppedByUser or startFailed; `exitCode`; `crash` with the signal or exception of a crashed \
process) and its output so far.{OUTPUT}{SEQUENCE}{INTERVENTIONS}"
),
@@ -152,6 +154,7 @@ process) and its output so far.{OUTPUT}{SEQUENCE}{INTERVENTIONS}"
};
let env = call.args.get("envs").filter(|v| !v.is_null()).map(|_| call.str_map("envs")).transpose()?;
let deadline = deadline(&call, RUN_TIMEOUT_MS)?;
let ready = pattern(&call)?;
let host = call.host();
let runs = &host.mcp_agent_runs;
if let Some(run) = runs.get(call.root(), &name).filter(|r| r.running()) {
@@ -185,7 +188,7 @@ process) and its output so far.{OUTPUT}{SEQUENCE}{INTERVENTIONS}"
return Err(IdeError::invalid(format!("`{name}` did not start: {error}")));
}
let from = run.process_offset();
let end = run.wait(deadline, None, from, &call.cancel).await?;
let end = run.wait(deadline, ready.as_ref(), from, &call.cancel).await?;
let mut extra = output_view(&call, &run, from)?;
extra.insert("waitEnded".into(), end.as_str().into());
result(&run, extra)
@@ -209,6 +212,12 @@ process) and its output so far.{OUTPUT}{SEQUENCE}{INTERVENTIONS}"
ParamType::Integer,
"Milliseconds to wait for the command to finish before returning (0 returns once it started); the command \
keeps running after it",
))
.param(Param::optional(
"pattern",
ParamType::String,
"Regular expression that ends the wait when the command's output matches it: the ready line of a server or \
watcher. `waitEnded` is then `matched`",
))
.param(Param::optional("maxLinesCount", ParamType::Integer, MAX_LINES))
.param(Param::optional("truncateMode", ParamType::String, TRUNCATE_MODE))
+25
View File
@@ -594,6 +594,31 @@ fn agent_process(h: &Harness, name: &str) -> ide_core::RunId {
h.host.processes.list().into_iter().filter(|r| r.config_id == id && r.running).map(|r| r.id).max().unwrap()
}
#[cfg(unix)]
#[tokio::test(flavor = "multi_thread")]
async fn run_command_returns_once_a_server_prints_its_ready_line() {
let h = Harness::new();
h.trust().await;
h.host.mcp_requests.set_ready("main", true);
h.host
.global_settings
.set(ide_workspace::SettingsScope::Global, "mcpServer.runWithoutConfirmation", json!(true))
.unwrap();
let _started = run_tool_window(&h);
// A server started with its ready line returns once it prints it, long before the timeout.
let started_at = std::time::Instant::now();
let up = h
.json(
"run_command",
json!({"name": "web", "command": "sleep 1; echo Listening on 8080; sleep 30", "pattern": "Listening on", "timeout": 60000}),
)
.await;
assert_eq!((up["waitEnded"].as_str(), up["state"].as_str()), (Some("matched"), Some("running")), "{up}");
assert!(up["output"].as_str().unwrap().contains("Listening on 8080"), "{up}");
assert!(started_at.elapsed() < std::time::Duration::from_secs(20), "returned after {:?}", started_at.elapsed());
h.json("stop_run", json!({"name": "web", "force": true, "timeout": 10000})).await;
}
#[cfg(unix)]
#[tokio::test(flavor = "multi_thread")]
async fn agent_commands_run_in_named_tabs_and_report_what_the_user_did() {
+1
View File
@@ -114,6 +114,7 @@ Notes on what they do:
As JetBrains IDEs do, the command runs on pipes, not a terminal: programs print plain lines instead of progress bars and redraws. The agent gets that output cleaned of escape sequences and cut to 2000 lines (`maxLinesCount`); beyond that the end stays, or the start or both ends with `truncateMode`. The tab's raw output log stays complete.
- **Servers return once they are up.** The call waits 10 seconds by default for the command to finish. A server, watcher or app never finishes: the agent passes `pattern` with its ready line (`Listening on`, `Started … in`) and gets the answer as soon as the line appears, while the command keeps running.
- **One tab per name.** A name whose command finished runs the next command in the same tab. A name whose command is still running is refused, with that command and its last output: the agent has to stop it with `stop_run` first or pick another name. Nothing is stopped behind your back.
- **Following a command up.** `wait_for_run` waits for the command to end or for a pattern in its output (a server's "listening" line), `read_run_output` reads the output from any point, `get_run_status` tells whether it is running, succeeded, failed, crashed or was stopped by the agent or by you, and `send_run_input` writes text and Enter to its input; Ctrl+C stops it and Ctrl+D ends its input.
- **You can step in.** What you type into the tab's input line, a rerun, a Stop and closing the tab are reported to the agent with their time. A wait ends at once, and the agent is advised to stop what it is doing, tell you it received the intervention and wait for your confirmation.
+1
View File
@@ -113,6 +113,7 @@ stdio 配置大致如下(路径以你机器上的为准):
和 JetBrains 的 IDE 一样,命令通过管道运行,而不是在终端里运行:程序输出的是普通文本行,不会有进度条和屏幕重绘。智能体拿到的输出会去掉控制字符,并截断到 2000 行(`maxLinesCount`);超出部分默认保留结尾,也可以用 `truncateMode` 改为保留开头或两头。标签页的原始输出日志保持完整。
- **服务启动好就返回。** 调用默认最多等 10 秒让命令结束。服务器、监视进程或应用不会结束:智能体传入 `pattern`(它就绪时输出的那一行,如 `Listening on`、`Started … in`),这一行一出现就返回,命令继续运行。
- **一个名称对应一个标签。** 名称对应的命令已结束时,下一条命令在同一标签中运行。命令仍在运行时会被拒绝,并返回该命令和它最后的输出:智能体必须先用 `stop_run` 停止它,或者换一个名称。任何进程都不会被悄悄停掉。
- **跟进命令。** `wait_for_run` 等待命令结束,或等待输出中出现指定文本(例如服务器的“正在监听”一行);`read_run_output` 从任意位置读取输出;`get_run_status` 报告命令是在运行、成功、失败、崩溃,还是被智能体或你停止;`send_run_input` 向命令的输入写入文本和 Enter,Ctrl+C 会停止命令,Ctrl+D 结束它的输入。
- **你可以随时干预。** 你在标签输入行中输入的内容、重新运行、停止和关闭标签,都会连同时间一起告诉智能体。正在进行的等待会立即返回,并建议智能体停下手头的工作,回复你已收到干预,等你确认后再继续。