test: add timeout support to spec tests (#32292)

## Summary

- Adds a `timeout` field (in seconds) to spec test `__test__.jsonc`
files
- When a test step's process exceeds the timeout, it is killed and the
test fails with a "Test command timed out" panic
- Timeout can be set at the multi-test level (propagates to all
contained tests) or per individual test
- No default timeout — the watchdog thread is only spawned when
`timeout` is explicitly configured
- Reduces CI test job timeout from 240 minutes to 30 minutes

Example usage in `__test__.jsonc`:
```jsonc
{
  "timeout": 60,
  "args": "run long_running.ts",
  "output": "expected.out"
}
```

---------

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>

Author: bartlomieju

.github/workflows/ci.generate.ts

@@ -989,7 +989,7 @@ const buildJobs = buildItems.map((rawBuildItem) => {
           `test ${testMatrix.test_crate} ${testMatrix.shard_label}${buildItem.profile} ${buildItem.os}-${buildItem.arch}`,
         needs: [buildJob],
         runsOn: buildItem.testRunner ?? buildItem.runner,
-        timeoutMinutes: 240,
+        timeoutMinutes: 30,
         defaults,
         env,
         strategy: {

.github/workflows/ci.yml

@@ -352,7 +352,7 @@ jobs:
     needs:
       - build-debug-macos-x86_64
     runs-on: macos-15-intel
-    timeout-minutes: 240
+    timeout-minutes: 30
     defaults:
       run:
         shell: bash
@@ -766,7 +766,7 @@ jobs:
     needs:
       - build-release-macos-x86_64
     runs-on: '${{ !contains(github.event.pull_request.labels.*.name, ''ci-full'') && github.event_name == ''pull_request'' && ''ubuntu-24.04'' || ''macos-15-intel'' }}'
-    timeout-minutes: 240
+    timeout-minutes: 30
     defaults:
       run:
         shell: bash
@@ -1062,7 +1062,7 @@ jobs:
     needs:
       - build-debug-macos-aarch64
     runs-on: macos-14
-    timeout-minutes: 240
+    timeout-minutes: 30
     defaults:
       run:
         shell: bash
@@ -1613,7 +1613,7 @@ jobs:
     needs:
       - build-release-macos-aarch64
     runs-on: '${{ !contains(github.event.pull_request.labels.*.name, ''ci-full'') && github.event_name == ''pull_request'' && ''ubuntu-24.04'' || ''macos-14'' }}'
-    timeout-minutes: 240
+    timeout-minutes: 30
     defaults:
       run:
         shell: bash
@@ -1906,7 +1906,7 @@ jobs:
     needs:
       - build-debug-windows-x86_64
     runs-on: windows-2022
-    timeout-minutes: 240
+    timeout-minutes: 30
     defaults:
       run:
         shell: bash
@@ -2429,7 +2429,7 @@ jobs:
     needs:
       - build-release-windows-x86_64
     runs-on: '${{ !contains(github.event.pull_request.labels.*.name, ''ci-full'') && github.event_name == ''pull_request'' && ''ubuntu-24.04'' || ''windows-2022'' }}'
-    timeout-minutes: 240
+    timeout-minutes: 30
     defaults:
       run:
         shell: bash
@@ -2699,7 +2699,7 @@ jobs:
     needs:
       - build-debug-windows-aarch64
     runs-on: windows-11-arm
-    timeout-minutes: 240
+    timeout-minutes: 30
     defaults:
       run:
         shell: bash
@@ -3127,7 +3127,7 @@ jobs:
     needs:
       - build-release-windows-aarch64
     runs-on: '${{ !contains(github.event.pull_request.labels.*.name, ''ci-full'') && github.event_name == ''pull_request'' && ''ubuntu-24.04'' || ''windows-11-arm'' }}'
-    timeout-minutes: 240
+    timeout-minutes: 30
     defaults:
       run:
         shell: bash
@@ -3581,7 +3581,7 @@ jobs:
     needs:
       - build-release-linux-x86_64
     runs-on: ubuntu-24.04
-    timeout-minutes: 240
+    timeout-minutes: 30
     defaults:
       run:
         shell: bash
@@ -4161,7 +4161,7 @@ jobs:
     needs:
       - build-debug-linux-x86_64
     runs-on: ubuntu-24.04
-    timeout-minutes: 240
+    timeout-minutes: 30
     defaults:
       run:
         shell: bash
@@ -4637,7 +4637,7 @@ jobs:
     needs:
       - build-debug-linux-aarch64
     runs-on: ubuntu-24.04-arm
-    timeout-minutes: 240
+    timeout-minutes: 30
     defaults:
       run:
         shell: bash
@@ -5214,7 +5214,7 @@ jobs:
     needs:
       - build-release-linux-aarch64
     runs-on: '${{ !contains(github.event.pull_request.labels.*.name, ''ci-full'') && github.event_name == ''pull_request'' && ''ubuntu-24.04'' || ''ubuntu-24.04-arm'' }}'
-    timeout-minutes: 240
+    timeout-minutes: 30
     defaults:
       run:
         shell: bash

tests/specs/mod.rs

@@ -62,6 +62,9 @@ struct MultiTestMetaData {
   pub ignore: bool,
   #[serde(default)]
   pub variants: BTreeMap<String, JsonMap>,
+  /// Timeout in seconds for each step. Defaults to 300 (5 minutes).
+  #[serde(default)]
+  pub timeout: Option<u64>,
 }
 
 impl MultiTestMetaData {
@@ -96,6 +99,11 @@ impl MultiTestMetaData {
           }
         }
       }
+      if let Some(timeout) = multi_test_meta_data.timeout
+        && !value.contains_key("timeout")
+      {
+        value.insert("timeout".to_string(), timeout.into());
+      }
       if multi_test_meta_data.ignore && !value.contains_key("ignore") {
         value.insert("ignore".to_string(), true.into());
       }
@@ -179,6 +187,9 @@ struct MultiStepMetaData {
   pub steps: Vec<StepMetaData>,
   #[serde(default)]
   pub ignore: bool,
+  /// Timeout in seconds for each step. Defaults to 300 (5 minutes).
+  #[serde(default)]
+  pub timeout: Option<u64>,
   #[serde(default)]
   pub variants: BTreeMap<String, JsonMap>,
 }
@@ -203,6 +214,9 @@ struct SingleTestMetaData {
   #[allow(dead_code)]
   #[serde(default)]
   pub variants: BTreeMap<String, JsonMap>,
+  /// Timeout in seconds for each step. Defaults to 300 (5 minutes).
+  #[serde(default)]
+  pub timeout: Option<u64>,
 }
 
 impl SingleTestMetaData {
@@ -220,6 +234,7 @@ impl SingleTestMetaData {
       steps: vec![self.step],
       ignore: self.ignore,
       variants: self.variants,
+      timeout: self.timeout,
     }
   }
 }
@@ -584,6 +599,10 @@ fn run_step(
     }
     None => command,
   };
+  let command = match metadata.timeout {
+    Some(secs) => command.timeout(std::time::Duration::from_secs(secs)),
+    None => command,
+  };
   let output = command.run();
 
   let step_output = {

tests/specs/schema.json

@@ -111,6 +111,10 @@
           "repeat": {
             "type": "number"
           },
+          "timeout": {
+            "type": "number",
+            "description": "Timeout in seconds for each step. Defaults to 300 (5 minutes)."
+          },
           "steps": {
             "type": "array",
             "items": {
@@ -154,6 +158,10 @@
             },
             "repeat": {
               "type": "number"
+            },
+            "timeout": {
+              "type": "number",
+              "description": "Timeout in seconds for each step. Defaults to 300 (5 minutes)."
             }
           }
         }, {
@@ -179,6 +187,10 @@
         "repeat": {
           "type": "number"
         },
+        "timeout": {
+          "type": "number",
+          "description": "Timeout in seconds for each step. Defaults to 300 (5 minutes)."
+        },
         "tests": {
           "type": "object",
           "additionalProperties": {

tests/specs/task/signals/__test__.jsonc

@@ -1,4 +1,5 @@
 {
+  "timeout": 60,
   // signals don't really exist on windows
   "if": "unix",
   // this runs a deno task

tests/util/lib/builders.rs

@@ -383,6 +383,20 @@ impl TestContext {
   }
 }
 
+fn kill_process(pid: u32) {
+  #[cfg(unix)]
+  // SAFETY: We're sending SIGKILL to a process we spawned.
+  unsafe {
+    libc::kill(pid as i32, libc::SIGKILL);
+  }
+  #[cfg(not(unix))]
+  {
+    let _ = std::process::Command::new("taskkill")
+      .args(["/F", "/T", "/PID", &pid.to_string()])
+      .output();
+  }
+}
+
 fn curl_fetch(url: &str) -> Vec<u8> {
   let output = std::process::Command::new("curl")
     .args(["--fail", "--silent", "--show-error", "--location", url])
@@ -447,6 +461,7 @@ pub struct TestCommandBuilder {
   args_vec: Vec<String>,
   split_output: bool,
   show_output: bool,
+  timeout: Option<std::time::Duration>,
 }
 
 impl TestCommandBuilder {
@@ -467,6 +482,7 @@ impl TestCommandBuilder {
       args_text: "".to_string(),
       args_vec: Default::default(),
       show_output: false,
+      timeout: None,
     }
   }
 
@@ -553,6 +569,11 @@ impl TestCommandBuilder {
     self
   }
 
+  pub fn timeout(mut self, timeout: std::time::Duration) -> Self {
+    self.timeout = Some(timeout);
+    self
+  }
+
   pub fn stdin<T: Into<Stdio>>(mut self, cfg: T) -> Self {
     self.stdin = Some(StdioContainer::new(cfg.into()));
     self
@@ -753,11 +774,42 @@ impl TestCommandBuilder {
     // and dropping it closes them.
     drop(command);
 
+    // Set up a watchdog thread that kills the process if it exceeds the timeout.
+    // Uses a channel: dropping the sender signals the watchdog to stop. If
+    // recv_timeout hits the deadline first, the process is killed.
+    let timeout_handle = self.timeout.map(|timeout| {
+      let (cancel_tx, cancel_rx) = std::sync::mpsc::channel::<()>();
+      let pid = process.id();
+      let handle =
+        spawn_thread(move || match cancel_rx.recv_timeout(timeout) {
+          Ok(()) | Err(std::sync::mpsc::RecvTimeoutError::Disconnected) => {
+            false
+          }
+          Err(std::sync::mpsc::RecvTimeoutError::Timeout) => {
+            eprintln!(
+              "Test command timed out after {:?}, killing pid {}",
+              timeout, pid
+            );
+            kill_process(pid);
+            true
+          }
+        });
+      (cancel_tx, handle)
+    });
+
     let combined = combined_reader.map(|pipe| {
       sanitize_output(read_pipe_to_string(pipe, self.show_output), &args)
     });
 
     let status = process.wait().unwrap();
+    // Drop the sender to cancel the watchdog, then check if it timed out
+    if let Some((cancel_tx, handle)) = timeout_handle {
+      drop(cancel_tx);
+      let timed_out = handle.join().unwrap_or(true);
+      if timed_out {
+        panic!("Test command timed out");
+      }
+    }
     let std_out_err = std_out_err_handle.map(|(stdout, stderr)| {
       (
         sanitize_output(stdout.join().unwrap(), &args),