diff --git a/server/src/http/mcp/shell_tools/script.rs b/server/src/http/mcp/shell_tools/script.rs
index 0c4d589c..43091286 100644
--- a/server/src/http/mcp/shell_tools/script.rs
+++ b/server/src/http/mcp/shell_tools/script.rs
@@ -70,15 +70,31 @@ pub(crate) async fn tool_run_tests(args: &Value, ctx: &AppContext) -> Result<Str
 
     let sid = story_key(&working_dir);
 
-    // Kill any existing in-flight job for this worktree before starting a new one.
-    {
-        let mut jobs = active_jobs().lock().map_err(|e| e.to_string())?;
-        if let Some(mut old_job) = jobs.remove(&working_dir) {
-            let _ = old_job.child.kill();
-            let _ = old_job.child.wait();
-        }
+    // If a test job is already in flight for this worktree, ATTACH to it
+    // rather than kill+respawn. This makes the agent system_prompt advice
+    // ("if run_tests appears to time out, call run_tests again — it
+    // attaches to the in-flight test job") actually true, and eliminates
+    // the respawn-loop bug where MCP client-side timeouts (~60s) cause
+    // agents to retry, killing the still-running cargo build each time
+    // and never making progress. The original job's poll loop below
+    // updates the CRDT on completion; attached callers just poll the CRDT.
+    let already_running = {
+        let jobs = active_jobs().lock().map_err(|e| e.to_string())?;
+        jobs.contains_key(&working_dir)
+    };
+    if already_running {
+        crate::slog!(
+            "[run_tests] Attaching to in-flight test job for {}",
+            working_dir.display()
+        );
+        return attach_to_in_flight_test_job(&sid).await;
     }
 
+    // Worktrees are isolated and may run cargo tests concurrently — no
+    // cross-worktree serialisation. The only invariant enforced here is
+    // "at most one test job per worktree at a time", which the
+    // already_running check above gives us.
+
     // Spawn the test process with piped stdout/stderr so we can capture output.
     // Pipes are drained in background threads to prevent deadlock when the
     // child fills the 64KB OS pipe buffer.
@@ -233,6 +249,39 @@ pub(crate) async fn tool_run_tests(args: &Value, ctx: &AppContext) -> Result<Str
     }
 }
 
+/// Poll the CRDT `test_jobs` collection for `sid` until the entry transitions
+/// out of "running" or we hit [`TEST_TIMEOUT_SECS`].
+///
+/// Used by `run_tests` to attach to a job that another caller already spawned
+/// for the same worktree, so concurrent callers all observe the same single
+/// `cargo test` run rather than racing to kill+respawn.
+async fn attach_to_in_flight_test_job(sid: &str) -> Result<String, String> {
+    let start = std::time::Instant::now();
+    loop {
+        match crate::crdt_state::read_test_job(sid) {
+            None => {
+                // The job entry disappeared from the CRDT — most likely the
+                // spawning task lost the race between insert/read. Treat as a
+                // transient error so the agent can retry.
+                return Err("In-flight test job vanished from CRDT before completing".to_string());
+            }
+            Some(view) if view.status == "pass" || view.status == "fail" => {
+                return format_crdt_result(&view);
+            }
+            Some(_) => {
+                // Still "running" — wait and re-poll.
+            }
+        }
+
+        if start.elapsed().as_secs() > TEST_TIMEOUT_SECS {
+            return Err(format!(
+                "Attached test job for '{sid}' did not complete within {TEST_TIMEOUT_SECS}s"
+            ));
+        }
+        tokio::time::sleep(std::time::Duration::from_secs(1)).await;
+    }
+}
+
 // ── get_test_result ──────────────────────────────────────────────────────────
 
 /// How long `get_test_result` blocks server-side before returning "running".
@@ -536,6 +585,60 @@ mod tests {
         assert_eq!(parsed["exit_code"], 1);
     }
 
+    #[tokio::test]
+    async fn tool_run_tests_concurrent_calls_attach_to_single_job() {
+        // Bug 903 regression: a second `run_tests` call for the same worktree
+        // while the first is still in flight must ATTACH to that job (i.e.
+        // observe its result) rather than kill+respawn. Pre-fix, every call
+        // killed the prior `cargo test` child and spawned a fresh one,
+        // creating a respawn loop driven by the agent's MCP-timeout retries.
+        let tmp = tempfile::tempdir().unwrap();
+        let script_dir = tmp.path().join("script");
+        std::fs::create_dir_all(&script_dir).unwrap();
+        let script_path = script_dir.join("test");
+        // Slow enough that the second call definitely overlaps the first.
+        std::fs::write(
+            &script_path,
+            "#!/usr/bin/env bash\nsleep 2\necho 'test result: ok. 0 passed'\nexit 0\n",
+        )
+        .unwrap();
+        #[cfg(unix)]
+        {
+            use std::os::unix::fs::PermissionsExt;
+            std::fs::set_permissions(&script_path, std::fs::Permissions::from_mode(0o755)).unwrap();
+        }
+
+        let ctx = test_ctx(tmp.path());
+        crate::crdt_state::init_for_test();
+        crate::db::ensure_content_store();
+
+        let start = std::time::Instant::now();
+        let ctx_a = ctx.clone();
+        let ctx_b = ctx.clone();
+        let t1 = tokio::spawn(async move { tool_run_tests(&json!({}), &ctx_a).await });
+        // Give T1 time to spawn its child + insert into active_jobs before T2
+        // arrives, so T2 deterministically takes the attach path.
+        tokio::time::sleep(std::time::Duration::from_millis(200)).await;
+        let t2 = tokio::spawn(async move { tool_run_tests(&json!({}), &ctx_b).await });
+
+        let r1 = t1.await.unwrap().unwrap();
+        let r2 = t2.await.unwrap().unwrap();
+        let elapsed = start.elapsed().as_secs_f64();
+
+        let p1: serde_json::Value = serde_json::from_str(&r1).unwrap();
+        let p2: serde_json::Value = serde_json::from_str(&r2).unwrap();
+        assert_eq!(p1["passed"], true, "first call must succeed: {p1}");
+        assert_eq!(p2["passed"], true, "second call must succeed: {p2}");
+
+        // If the second call had killed + respawned, total elapsed would be
+        // ~4s (two 2s runs serially). With attach, both calls observe the
+        // SAME single 2s run, so elapsed stays close to 2s.
+        assert!(
+            elapsed < 3.5,
+            "concurrent calls must share one job (~2s), not respawn (~4s); elapsed={elapsed:.2}s"
+        );
+    }
+
     #[tokio::test]
     async fn tool_run_tests_worktree_path_must_be_inside_worktrees() {
         let tmp = tempfile::tempdir().unwrap();