diff --git a/cc-ci-plan/launch-supervisor.py b/cc-ci-plan/launch-supervisor.py
index dc8b287..eef0549 100644
--- a/cc-ci-plan/launch-supervisor.py
+++ b/cc-ci-plan/launch-supervisor.py
@@ -143,9 +143,24 @@ def _gate():
     # a live `opencode run … -s <sid> --attach` proc, or a log touched within the stall window.
     pids = lu._run_pids(sid)
     idle = lu._session_idle_min()
-    if pids or (idle is not None and idle < lu.STALL_MIN):
-        via = f"{len(pids)} live run proc(s)" if pids else f"log idle {idle:.0f}m < {lu.STALL_MIN:.0f}m"
-        return False, sid, f"upgrader run progressing ({via}) — leaving it"
+    advancing = idle is not None and idle < lu.STALL_MIN
+    # PROGRESS REQUIRES THE SESSION TO BE ADVANCING — a live proc alone is not enough. A run walled
+    # by the provider keeps its process alive and spinning while emitting nothing (2026-08-07: 3
+    # days, zero output). Treating "proc exists" as progress is what deadlocked the gate.
+    if advancing:
+        return False, sid, f"upgrader run progressing (session advanced {idle:.0f}m ago) — leaving it"
+    if pids and lu._billing_blocked():
+        # NEVER kill it: it may resume when the wall lifts, and its context is the run's state.
+        # Surface it loudly instead — this is an operator-actionable condition, not self-healing.
+        return False, sid, (
+            f"run BLOCKED on a provider billing/usage wall ({len(pids)} proc(s) alive, session idle "
+            f"{idle:.0f}m) — NOT killing (resumable once the wall lifts); operator action required"
+        )
+    if pids:
+        log(
+            f"note: {len(pids)} live run proc(s) but session idle {idle:.0f}m ≥ "
+            f"{lu.STALL_MIN:.0f}m — treating as stalled, not progressing"
+        )
     # The per-run watchdog owns PROMPT recovery (resume on proc-death/stall) and is the single writer
     # while it lives. Defer to it — it gives up (exits its tmux) only after MAX_RESUMES fail, i.e. the
     # run is stuck in a way a bare resume can't fix (e.g. disk-full). THEN the supervisor takes over.
diff --git a/cc-ci-plan/launch-upgrader.py b/cc-ci-plan/launch-upgrader.py
index 3de72a5..f637f4f 100644
--- a/cc-ci-plan/launch-upgrader.py
+++ b/cc-ci-plan/launch-upgrader.py
@@ -417,25 +417,55 @@ def _limit_retry_after():
         return 0
 
 def _run_pids(sid=None):
-    """PIDs of live `opencode run` procs for THIS session (via /proc scan — never matches self)."""
+    """PIDs of live `opencode run` procs for THIS session (via /proc scan — never matches self).
+
+    Matches on FLAG VALUES (`--title <SESSION>` / `-s|--session <sid>`), never a substring of the
+    whole cmdline. The substring form caused the 2026-08-07 three-day deadlock: an agent's kickoff
+    PROMPT is passed as an argv element, and the supervisor's prompt text contains the literal
+    "cc-ci-upgrader", so the supervisor's own (billing-hung) agent matched as a live upgrader run —
+    the hourly gate then read its own corpse as "run progressing" and no-opped ~60 times."""
     me, out = os.getpid(), []
     for p in os.listdir("/proc"):
         if not p.isdigit() or int(p) == me:
             continue
         try:
-            cl = open(f"/proc/{p}/cmdline", "rb").read().split(b"\0")
+            cl = [c for c in open(f"/proc/{p}/cmdline", "rb").read().split(b"\0") if c]
         except Exception:
             continue
-        joined = b" ".join(cl)
-        if not (b"opencode" in joined and b"run" in cl and b"--attach" in cl):
+        if not cl or b"opencode" not in cl[0] or b"run" not in cl or b"--attach" not in cl:
             continue
-        # Scope to THIS managed session only: a fresh run carries `--title <SESSION>`, a resumed
-        # run carries `-s <sid>`. Without this, the report watchdog would kill the idle upgrader
-        # run (and vice-versa) since both are `opencode run … --attach`.
-        if SESSION.encode() in joined or (sid and sid.encode() in joined):
+
+        def _flag(names):
+            for i, a in enumerate(cl[:-1]):
+                if a in names:
+                    return cl[i + 1]
+            return None
+
+        title = _flag((b"--title",))
+        s_val = _flag((b"-s", b"--session"))
+        if title == SESSION.encode() or (sid and s_val == sid.encode()):
             out.append(int(p))
     return out
 
+
+def _billing_blocked(window_bytes=4000):
+    """True when the tail of this run's log shows a provider billing/usage wall.
+
+    Such a process must NEVER be killed: it may resume when the wall lifts, and killing it can
+    lose in-flight work. It is also NOT progress — it can spin for days emitting nothing (observed
+    2026-08-07..10), so the gate reports it as BLOCKED and leaves recovery to the operator."""
+    try:
+        with open(LOG_FILE, "rb") as f:
+            f.seek(0, os.SEEK_END)
+            f.seek(max(0, f.tell() - window_bytes))
+            tail = f.read().decode(errors="replace").lower()
+    except OSError:
+        return False
+    return any(
+        s in tail
+        for s in ("spending limit", "insufficient balance", "usage limit", "usagelimiterror")
+    )
+
 def _completed():
     # Done only when the MODEL signs off with DONE_MARKER as its FINAL word: the marker in the LAST
     # assistant TEXT (prose) message. This guards THREE false-positives that each abandoned a run: