diff --git a/crates/meshloop-adapters/src/harness.rs b/crates/meshloop-adapters/src/harness.rs
index cfc8e1c..35cb041 100644
--- a/crates/meshloop-adapters/src/harness.rs
+++ b/crates/meshloop-adapters/src/harness.rs
@@ -197,6 +197,9 @@ impl HarnessCapabilities for CliHarness {
         // Idempotent against an already-exited/never-tracked process, per ADR 0003.
         if let Some((mut child, _, _)) = registry.remove(&Self::attempt_key(handle)) {
             child.kill_tree();
+        } else if let Some(pid) = handle.pid {
+            // Not ours: a previous Meshloop process started it and is gone (#30).
+            crate::process::kill_orphaned_tree(pid, &self.config.executable);
         }
         Ok(())
     }
diff --git a/crates/meshloop-adapters/src/process.rs b/crates/meshloop-adapters/src/process.rs
index af0ca08..3d1c5c2 100644
--- a/crates/meshloop-adapters/src/process.rs
+++ b/crates/meshloop-adapters/src/process.rs
@@ -103,6 +103,57 @@ impl ProcessView for NullPidIsDead {
     }
 }
 
+/// Kills the process tree of an attempt this process no longer holds a handle for, such as
+/// a harness left running when a previous Meshloop process crashed (ADR 0025 amendment, #30).
+///
+/// PIDs can be reused, so the tree is killed only when `pid` is alive, runs the expected
+/// executable, and on POSIX still leads its own process group (as `spawn_owned` arranges).
+/// Returns whether a kill was issued.
+pub fn kill_orphaned_tree(pid: u32, executable: &std::path::Path) -> bool {
+    if pid == 0 || !orphan_identity_matches(pid, executable) {
+        return false;
+    }
+    kill_process_tree(pid);
+    true
+}
+
+#[cfg(windows)]
+fn orphan_identity_matches(pid: u32, executable: &std::path::Path) -> bool {
+    let Some(stem) = executable.file_stem() else {
+        return false;
+    };
+    let image = format!("{}.exe", stem.to_string_lossy());
+    WindowsProcessView.is_live(&ProcessHint {
+        pid,
+        image_name: Some(image),
+    }) == LiveCheck::Live
+}
+
+#[cfg(not(windows))]
+fn orphan_identity_matches(pid: u32, executable: &std::path::Path) -> bool {
+    let Some(name) = executable
+        .file_name()
+        .map(|n| n.to_string_lossy().into_owned())
+    else {
+        return false;
+    };
+    let Ok(out) = Command::new("ps")
+        .args(["-o", "pgid=,stat=,comm=", "-p", &pid.to_string()])
+        .output()
+    else {
+        return false;
+    };
+    let text = String::from_utf8_lossy(&out.stdout);
+    let mut fields = text.split_whitespace();
+    let (Some(pgid), Some(stat), Some(comm)) = (fields.next(), fields.next(), fields.next()) else {
+        return false;
+    };
+    // Linux truncates comm to 15 bytes; macOS reports a path.
+    let comm = comm.rsplit('/').next().unwrap_or(comm);
+    let name_matches = comm == name || (comm.len() == 15 && name.starts_with(comm));
+    pgid == pid.to_string() && !stat.starts_with('Z') && name_matches
+}
+
 /// Kills a process and all of its descendants (the entire process tree).
 ///
 /// On Windows, executes `taskkill /F /T /PID <pid>` to terminate child compilers,
diff --git a/crates/meshloop-adapters/tests/harness_contract.rs b/crates/meshloop-adapters/tests/harness_contract.rs
index c3e63df..f64d0c8 100644
--- a/crates/meshloop-adapters/tests/harness_contract.rs
+++ b/crates/meshloop-adapters/tests/harness_contract.rs
@@ -190,3 +190,115 @@ fn collect_on_an_unknown_handle_is_unsupported_not_a_panic() {
         Err(HarnessError::ProcessFault { .. })
     ));
 }
+
+fn wait_until_dead(pid: u32) -> bool {
+    use meshloop_engine::ports::{LiveCheck, ProcessHint, ProcessView};
+    let view = meshloop_adapters::process::HostProcessView {};
+    let hint = ProcessHint {
+        pid,
+        image_name: None,
+    };
+    let start = std::time::Instant::now();
+    while start.elapsed() < Duration::from_secs(5) {
+        if view.is_live(&hint) != LiveCheck::Live {
+            return true;
+        }
+        std::thread::sleep(Duration::from_millis(50));
+    }
+    false
+}
+
+#[test]
+fn cancel_kills_an_orphaned_tree_left_by_a_previous_process() {
+    // Simulates recovery after a crash (#30): the attempt's process tree is still running but
+    // this harness instance never started it, so only the recorded pid is known.
+    let dir = std::env::temp_dir().join(format!("meshloop-test-{}-orphan", std::process::id()));
+    fs::create_dir_all(&dir).unwrap();
+    let pid_file = dir.join("grandchild.pid");
+    let mut cmd = std::process::Command::new(fixture_path());
+    cmd.arg("--spawn-grandchild")
+        .arg(&pid_file)
+        .arg(dir.join("heartbeat.txt"));
+    let orphan = meshloop_adapters::process::spawn_owned(cmd).expect("spawn orphan tree");
+    let leader = orphan.id();
+    let start = std::time::Instant::now();
+    let grandchild: u32 = loop {
+        if let Some(pid) = fs::read_to_string(&pid_file)
+            .ok()
+            .and_then(|s| s.trim().parse().ok())
+        {
+            break pid;
+        }
+        assert!(
+            start.elapsed() < Duration::from_secs(5),
+            "grandchild never started"
+        );
+        std::thread::sleep(Duration::from_millis(50));
+    };
+    // Drop ownership without killing, as a crashed process would.
+    std::mem::forget(orphan);
+
+    let harness = CliHarness::new(CliHarnessConfig {
+        name: "fixture".into(),
+        executable: fixture_path(),
+        version_args: vec!["--version".into()],
+        probe_timeout: Duration::from_secs(10),
+        invoke_args_template: vec![],
+    });
+    let handle = HarnessHandle {
+        attempt_id: AttemptId(4242),
+        pid: Some(leader),
+        pane_id: None,
+    };
+    harness
+        .cancel(&handle)
+        .expect("cancel of an untracked attempt");
+
+    assert!(
+        wait_until_dead(leader),
+        "orphaned leader {leader} survived cancel"
+    );
+    assert!(
+        wait_until_dead(grandchild),
+        "orphaned grandchild {grandchild} survived cancel"
+    );
+    fs::remove_dir_all(&dir).ok();
+}
+
+#[test]
+fn cancel_leaves_an_unrelated_process_with_a_recorded_pid_alone() {
+    // A recycled pid must not be killed: the recorded pid now belongs to another executable.
+    #[cfg(windows)]
+    let mut other = std::process::Command::new("powershell")
+        .args(["-NoProfile", "-Command", "Start-Sleep -Seconds 30"])
+        .spawn()
+        .expect("spawn unrelated process");
+    #[cfg(not(windows))]
+    let mut other = std::process::Command::new("sleep")
+        .arg("30")
+        .spawn()
+        .expect("spawn unrelated process");
+    let harness = CliHarness::new(CliHarnessConfig {
+        name: "fixture".into(),
+        executable: fixture_path(),
+        version_args: vec!["--version".into()],
+        probe_timeout: Duration::from_secs(10),
+        invoke_args_template: vec![],
+    });
+    let handle = HarnessHandle {
+        attempt_id: AttemptId(4243),
+        pid: Some(other.id()),
+        pane_id: None,
+    };
+    harness
+        .cancel(&handle)
+        .expect("cancel of an untracked attempt");
+    std::thread::sleep(Duration::from_millis(300));
+    let still_running = other.try_wait().expect("poll unrelated process").is_none();
+    let _ = other.kill();
+    let _ = other.wait();
+    assert!(
+        still_running,
+        "cancel killed a process that is not this harness"
+    );
+}
diff --git a/crates/meshloop-engine/src/run_loop.rs b/crates/meshloop-engine/src/run_loop.rs
index 753ff17..41b1e21 100644
--- a/crates/meshloop-engine/src/run_loop.rs
+++ b/crates/meshloop-engine/src/run_loop.rs
@@ -711,6 +711,17 @@ impl<'a> RunLoop<'a> {
             };
             match self.attempt_live(&attempt) {
                 LiveCheck::Dead => {
+                    // The harness no longer tracks this attempt, but a previous Meshloop
+                    // process may have left its process tree running: end it before the
+                    // node is retried in a fresh attempt (#30). Cancel is idempotent.
+                    if let Some(h) = attempt
+                        .harness
+                        .as_deref()
+                        .and_then(|n| self.harnesses.get(n))
+                    {
+                        let _ =
+                            h.cancel(&Self::attempt_handle(&attempt, self.resolve_pane(&attempt)));
+                    }
                     let rec = self.append(
                         graph_id,
                         node.id,
diff --git a/docs/architecture/adr/0025-process-tree-ownership.md b/docs/architecture/adr/0025-process-tree-ownership.md
index 3b5650c..a4452d4 100644
--- a/docs/architecture/adr/0025-process-tree-ownership.md
+++ b/docs/architecture/adr/0025-process-tree-ownership.md
@@ -47,10 +47,25 @@ Constraints:
 - `crates/meshloop-adapters/src/process/job.rs`: Win32 Job Object abstraction (`JOB_OBJECT_LIMIT_KILL_ON_JOB_CLOSE`) and POSIX `setpgid(0, 0)` via `spawn_owned`.
 - `crates/meshloop-adapters/tests/process_tree.rs`:
   - `grandchild_dies_on_timeout`: confirms grandchild termination on timeout/kill_tree.
-  - `grandchild_dies_on_parent_abort`: confirms OS kernel terminates grandchildren when parent abruptly aborts (`std::process::abort()`).
+  - `grandchild_dies_on_parent_abort` (Windows only): confirms OS kernel terminates grandchildren when parent abruptly aborts (`std::process::abort()`). POSIX: see the 2026-10-07 recovery amendment.
 - `xtask bench`: metric `conc.orphan_process_count` reports `0` (PASS).
 - 100% passing tests in `cargo test -p meshloop-adapters` (24 passed) and `cargo test --workspace` (131 passed).
 
+## Amendment 2026-10-07: POSIX orphans are reclaimed during recovery
+
+- Status: Accepted
+- Approval evidence: Samuel (human) chose option 1 ("recovery reconciliation") for issue #30 on 2026-10-07, in the AgentFlow/Meshloop setup session.
+- Executor: Claude Code (Opus 5.5)
+
+**Context.** The guarantee that grandchildren die when Meshloop aborts holds only on Windows, where the Job Object closes with the process. A POSIX process group is torn down only when Meshloop calls `kill_tree`. If Meshloop itself aborts or is SIGKILLed, the harness group keeps running; on WSL2 Ubuntu 24.04 the abort fixture left two processes alive for more than 11 minutes. `grandchild_dies_on_parent_abort` is therefore limited to Windows.
+
+**Decision.** Reclaim orphans during recovery, in safe Rust. When `RunLoop` reattaches a `Running` attempt that the current harness instance does not track, it calls `HarnessCapabilities::cancel` before recording `HarnessCrashedOrTimeout` and retrying. For an untracked attempt with a recorded pid, `CliHarness::cancel` calls `process::kill_orphaned_tree`. That function kills the tree only when the pid is alive and runs the harness executable (image name on Windows; `comm` on POSIX, which Linux truncates to 15 bytes), and, on POSIX, still leads its own process group. This guards against pid reuse.
+
+**Alternatives rejected.** `prctl(PR_SET_PDEATHSIG)` is Linux-only, covers only the direct child, and needs `unsafe`. A subreaper supervisor process adds moving parts.
+
+**Consequences.** Orphans keep running between the crash and the next `resume` or `status` of that graph; they are not reclaimed immediately. A harness launched through a shim whose process image differs from the configured executable is not matched, so it is left alone rather than risk killing an unrelated process.
+
+**Verification.** `crates/meshloop-adapters/tests/harness_contract.rs`: `cancel_kills_an_orphaned_tree_left_by_a_previous_process` fails without the untracked-pid branch ("orphaned leader survived cancel") and passes with it. `cancel_leaves_an_unrelated_process_with_a_recorded_pid_alone` covers the pid-reuse guard.
 ## Amendment 2026-10-07: detached spawns do not inherit caller handles
 
 - Status: Accepted