Count a timed-out resume as an attempt, it may have launched

README.md

@@ -91,7 +91,9 @@ endless chain of blocked jobs:
 again it is a *new* session id and qualifies on its own, so an outage spanning
 several reset windows still gets retried. What the cap prevents is one
 conversation being resumed into several concurrent copies. A resume that fails
-to launch at all records no attempt and is retried on the next tick.
+to launch at all records no attempt and is retried on the next tick; one that
+times out records an attempt, because the launch may have registered after the
+watchdog stopped waiting for it.
 
 `--max-per-day` is the one guard that cannot be derived from anything else. A
 usage-limit block is self-pacing, because the stated reset time is respected —

resume-stalled-sessions.py

@@ -307,6 +307,11 @@ def resume(job, claude_bin, prompt, dry_run):
             command + [prompt],
             capture_output=True, cwd=cwd, text=True, timeout=180, check=True,
         )
+    except subprocess.TimeoutExpired:
+        # The launch may have registered before we gave up, and a second try
+        # would put two agents on one conversation.
+        log(f"{job['id']}: resume timed out, counting the attempt")
+        return True
     except (OSError, subprocess.SubprocessError) as err:
         log(f"{job['id']}: resume failed: {err}")
         return False