Hawk's llama-server is the one on its port

README.md

@@ -149,7 +149,7 @@ from the dashboard.
 | `auto_accept_external_dirs` | `true` | Auto-accept external-directory permission prompts (monitored sessions only) |
 | `auto_answer_questions` | `true` | Auto-answer question-tool prompts |
 | `continue_via_attach` | `true` | Inject through the Desktop sidecar so the GUI streams live |
-| `llama_api_port` | `1234` | llama-server API port |
+| `llama_api_port` | `1234` | llama-server API port. Hawk's server is the one listening on this port; llama-server instances you run on other ports are left alone |
 | `llama_bat_path` | `""` | Script that starts llama-server (`.bat` on Windows, shell script on Linux); the watchdog uses it to respawn a dead server. Empty = no respawn |
 | `llama_pid` | `0` | 0 = find llama-server by process name |
 | `gpu_vram_total_mb` | `0` | VRAM shown in charts; 0 = auto-detect |

opencode_hawk/llama_restart.py

@@ -418,6 +418,8 @@ def _respawn(spec) -> int | None:
             env=spec.get("env") or None,
             stdout=fh, stderr=subprocess.STDOUT,
         )
+        from . import probes
+        probes.record_llama_spawn(proc.pid)
         return proc.pid
     except Exception as e:
         coordinator.log_debug("llama_restart respawn: %s" % e)

opencode_hawk/probes.py

@@ -196,26 +196,58 @@ def llama_slots(cfg):
     return None
 
 
+SPAWN_FILE = "llama_spawn.json"
+SPAWN_GRACE_S = 180  # a server Hawk just started counts as Hawk's own until it listens (model loading)
+
+
+def record_llama_spawn(pid) -> None:
+    """Remember the llama-server Hawk just started (see llama_pid)."""
+    try:
+        core.save_json(core.home_dir() / SPAWN_FILE, {"pid": int(pid), "at": time.time()})
+    except Exception as e:
+        core.log_debug("record_llama_spawn: %s" % e)
+
+
 def llama_pid(cfg) -> int | None:
-    """Return the llama-server process id (config override, else auto-detect by
-    image name), or None if not found."""
+    """Hawk's llama-server: the config override, else the process listening on llama_api_port. Other
+    llama-server instances (other ports, started by the user) are never Hawk's: Hawk neither counts them as its
+    server nor stops them. A server Hawk started in the last SPAWN_GRACE_S that is not listening yet still
+    counts (so no second one is started while the model loads). Name lookup only when listeners cannot be
+    read, and then only if exactly one llama-server runs."""
     pid = int(cfg.get("llama_pid") or 0)
     if pid:
         return pid
+    port = int(cfg.get("llama_api_port") or 1234)
+    try:
+        import psutil
+        for conn in psutil.net_connections(kind="tcp"):
+            if conn.status == psutil.CONN_LISTEN and conn.laddr and conn.laddr.port == port and conn.pid:
+                return conn.pid
+        spawn = core.load_json(core.home_dir() / SPAWN_FILE, {}) or {}
+        sp = int(spawn.get("pid") or 0)
+        if sp and time.time() - float(spawn.get("at") or 0) < SPAWN_GRACE_S and psutil.pid_exists(sp):
+            return sp
+        return None
+    except Exception as e:  # listeners unreadable (permissions): fall back to the name, only if unambiguous
+        core.log_debug("llama_pid by port: %s" % e)
+    pids = _llama_pids_by_name()
+    return pids[0] if len(pids) == 1 else None
+
+
+def _llama_pids_by_name() -> list:
     if os.name != "nt":
-        return hawk_linux.find_pid("llama-server")
+        p = hawk_linux.find_pid("llama-server")
+        return [p] if p else []
     try:
         out = subprocess.run(["tasklist", "/FI", "IMAGENAME eq llama-server.exe",
                               "/FO", "CSV", "/NH"],
                              capture_output=True, text=True, encoding="utf-8",
                              errors="replace", timeout=15,
                              creationflags=core.creation_flags()).stdout
-        for m in re.finditer(r'"([^"]+)","(\d+)"', out):
-            if "llama" in m.group(1).lower():
-                return int(m.group(2))
+        return [int(m.group(2)) for m in re.finditer(r'"([^"]+)","(\d+)"', out) if "llama" in m.group(1).lower()]
     except Exception as e:
         core.log_debug("llama_pid: %s" % e)
-    return None
+    return []
 
 
 def llama_server_alive(cfg) -> bool:
@@ -274,6 +306,7 @@ def _llama_respawn_locked(cfg, timeout_s=None) -> bool:
         else:
             argv = [bat] if os.access(bat, os.X_OK) else ["sh", bat]
         proc = core.spawn_independent(argv, stdout=fh, stderr=subprocess.STDOUT)
+        record_llama_spawn(proc.pid)
     except Exception as e:
         core.log("LLAMA-DOWN: respawn launch failed: %s" % e)
         return False

tests/selftest_coordinator.py

@@ -614,6 +614,44 @@ def mkllama_idle_test(cpu_frac, slots_processing, gate_slots,
         "OK" if ok_m else "MISMATCH",
         "escalation note written to session DB as synthetic user msg + text part"))
 
+    # N00000: Hawk's llama-server is the one on llama_api_port; the user's other instances are never Hawk's
+    import psutil as _ps, tempfile as _tf2, collections as _col
+    from opencode_hawk import probes as _pr
+    _Conn = _col.namedtuple("_Conn", "status laddr pid")
+    _Addr = _col.namedtuple("_Addr", "ip port")
+    _home2 = Path(_tf2.mkdtemp(prefix="hawk-port-"))
+    _old_home2 = os.environ.get("COORD_HOME")
+    os.environ["COORD_HOME"] = str(_home2)
+    _saved2 = (_ps.net_connections, _ps.pid_exists, _pr._llama_pids_by_name)
+    try:
+        _ps.net_connections = lambda kind="tcp": [_Conn(_ps.CONN_LISTEN, _Addr("127.0.0.1", 8080), 111),
+                                                  _Conn(_ps.CONN_LISTEN, _Addr("127.0.0.1", 1234), 222)]
+        ok_p1 = hawk.llama_pid({"llama_api_port": 1234}) == 222
+        _ps.net_connections = lambda kind="tcp": [_Conn(_ps.CONN_LISTEN, _Addr("127.0.0.1", 8080), 111)]
+        ok_p2 = hawk.llama_pid({"llama_api_port": 1234}) is None  # only the user's instance: Hawk's server is down
+        _pr.record_llama_spawn(333)
+        _ps.pid_exists = lambda pid: pid == 333
+        ok_p3 = hawk.llama_pid({"llama_api_port": 1234}) == 333  # just started, still loading: Hawk's own
+        def _no_access(kind="tcp"):
+            raise _ps.AccessDenied()
+        _ps.net_connections = _no_access
+        _pr._llama_pids_by_name = lambda: [111, 222]
+        ok_p4 = hawk.llama_pid({"llama_api_port": 1234}) is None  # cannot tell which: never guess
+        _pr._llama_pids_by_name = lambda: [222]
+        ok_p5 = hawk.llama_pid({"llama_api_port": 1234}) == 222
+    finally:
+        _ps.net_connections, _ps.pid_exists, _pr._llama_pids_by_name = _saved2
+        if _old_home2 is None:
+            os.environ.pop("COORD_HOME", None)
+        else:
+            os.environ["COORD_HOME"] = _old_home2
+        shutil.rmtree(_home2, ignore_errors=True)
+    ok_port = all([ok_p1, ok_p2, ok_p3, ok_p4, ok_p5])
+    results.append(("llama_by_port", "pass", "pass" if ok_port else "fail", ok_port,
+                    "Hawk's llama-server is the one on its port; other instances are left alone", []))
+    print("%-22s expect=%-9s got=%-9s %s %s" % ("llama_by_port", "pass", "pass" if ok_port else "fail",
+                                               "OK" if ok_port else "MISMATCH", [ok_p1, ok_p2, ok_p3, ok_p4, ok_p5]))
+
     # N0000: one instance only: a lock another process holds is refused; free again once that process ends
     import subprocess as _sp1, tempfile as _tf1
     _lockhome = Path(_tf1.mkdtemp(prefix="hawk-lock-"))