v87: add isolated auto-updating Debian RTX 3060 Ti worker

Register the explicitly owned 192.168.0.9 peer with scoped text inference, independent fixed-kernel CPU transport and local safe-tensor validation. Add a non-login account, protected system services, RAM/temperature budgets, boot autostart, SHA-verified runtime/source and deferred root-managed updates. Qualify advisory models independently of CPU transport; preserve accepted V100 training and existing Windows limits. Route available advisory work and serving fallbacks to the node; report failures and telemetry in the dashboard.

Validation: 1389 passed, 65 skipped; additional release/updater checks 13 passed. Ruff, release SHA256 and shell syntax passed. Physical GPU residency, networking and boot execution are checked by operator-run installers, not claimed from mocked tests.

Author: Mantiee

pyproject.toml

@@ -3,7 +3,7 @@ name = "rlms"
 authors = [
     {name = "Alex Zhang", email = "altzhang@mit.edu"},
 ]
-version = "0.1.3+v100.86"
+version = "0.1.3+v100.87"
 description = "Recursive Language Models."
 readme = "README.md"
 license = "MIT"

rlm/v100/competition.py

@@ -411,6 +411,7 @@ def helper_client(
             "max_vram_gib": profile["resources"]["max_vram_gib"],
             "helper_batch_tokens": profile["resources"].get("helper_batch_tokens", 64),
             "helper_duty_percent": profile["resources"].get("helper_duty_percent", 65),
+            "dedicated_host": profile["resources"].get("dedicated_host", False),
             "on_demand": profile["resources"].get("on_demand", False),
             "allow_cpu_offload": profile["resources"].get("allow_cpu_offload", False),
             "require_adaptive_windows_guard": profile["resources"].get(
@@ -436,7 +437,9 @@ def helper_client(
         if profile.get("resources", {}).get("device") in ("cpu", "remote")
         else "model",
         activity_context={
-            "device": profile.get("resources", {}).get("device", "local GPU"),
+            "device": profile.get("resources", {}).get(
+                "worker_name", profile.get("resources", {}).get("device", "local GPU")
+            ),
             "endpoint": settings["base_url"],
             "model_version": settings["model_version"],
             "target": profile["server"]["model"],

rlm/v100/drones.py

@@ -16,6 +16,7 @@
 
 from rlm.v100.activity import ActivityLog
 from rlm.v100.common import atomic_json, load_profile
+from rlm.v100.remote_nodes import selected_node
 
 
 @contextmanager
@@ -174,9 +175,8 @@ def execute(root: Path, job: dict) -> dict:
     if preferences(root).get("remote_helper_enabled", True) is False:
         client = accepted_master(root, job["branch"])
     else:
-        profile = load_profile(
-            selected_helper(root, "planner" if job["kind"] == "critic" else "fast"), root
-        )
+        chosen_path = selected_helper(root, "planner" if job["kind"] == "critic" else "fast")
+        profile = load_profile(chosen_path, root)
         if not remote_profile(profile):
             raise ValueError(
                 "Resident research requires an explicitly selected remote or accepted master"
@@ -186,15 +186,25 @@ def execute(root: Path, job: dict) -> dict:
         try:
             client.identity()
         except requests.RequestException as error:
+            node = selected_node(root)
+            alternative = None
+            if node is not None and node.resolve() != chosen_path.resolve():
+                try:
+                    alternative = helper_client(load_profile(node, root), root, job["branch"])
+                    alternative.identity()
+                except requests.RequestException:
+                    alternative = None
+            client = (
+                alternative if alternative is not None else accepted_master(root, job["branch"])
+            )
             ActivityLog(root, job["branch"], "resident-" + job["kind"]).write(
                 "steps",
-                "local-master-fallback",
+                "owned-node-fallback" if alternative is not None else "local-master-fallback",
                 {
                     "reason": str(error)[:300],
-                    "scope": "Remote transport unavailable; accepted local master only, no changed-identity fallback",
+                    "scope": "Unavailable transport replaced before generation by a qualified owned helper or the protected accepted master; no changed-identity fallback",
                 },
             )
-            client = accepted_master(root, job["branch"])
     client.research_tool_steps = 6
     client.activity_actor = "resident-" + job["kind"]
     client.research_tool_names = {
@@ -358,7 +368,10 @@ def service(root: Path, stop: threading.Event) -> None:
                         (kind in ("researcher", "critic", "benchmark", "income")) == llm
                         for _, kind in active.values()
                     )
-                    if used >= (min(1, budget["cpu_slots"]) if llm else budget["cpu_slots"]):
+                    llm_slots = min(1, budget["cpu_slots"]) + int(
+                        not gpu_disabled(root) and selected_node(root) is not None
+                    )
+                    if used >= (llm_slots if llm else budget["cpu_slots"]):
                         continue
                     job = dict(row)
                     if job["kind"] == "python":
@@ -385,7 +398,9 @@ def service(root: Path, stop: threading.Event) -> None:
                     "cpu_slots": budget["cpu_slots"],
                     "budget": budget,
                     "cpu_affinities": affinities,
-                    "rtx_slots": 0 if gpu_disabled(root) else 1,
+                    "rtx_slots": 0
+                    if gpu_disabled(root)
+                    else 1 + int(selected_node(root) is not None),
                     "local_model_slots": 1 if gpu_disabled(root) else 0,
                 },
             )

rlm/v100/income_work.py

@@ -31,6 +31,31 @@ class MasterNotReady(RuntimeError):
     """Transient serving startup, shutdown or unavailable protected receipt."""
 
 
+def owned_advisor(root: Path, branch: str, reason: str):
+    """Use an independently qualified advisory node while protected serving is unavailable."""
+    from rlm.v100.competition import helper_client
+    from rlm.v100.mission_chat import preferences
+    from rlm.v100.remote_nodes import selected_node
+
+    path = selected_node(root) if preferences(root).get("remote_helper_enabled", True) else None
+    if path is None:
+        raise MasterNotReady(reason)
+    client = helper_client(load_profile(path, root), root, branch)
+    client.timeout = 5
+    try:
+        client.identity()
+    except requests.RequestException as error:
+        raise MasterNotReady(reason) from error
+    client.timeout = 120
+    client.request_deadline = time.monotonic() + 150
+    client.enable_thinking = False
+    client.sampling_args = dict(client.sampling_args) | {"max_tokens": 2048}
+    client.research_token_ceiling = 2048
+    client.research_thinking_allowed = False
+    client.activity_actor = "income-owned-advisor"
+    return client
+
+
 def accepted_master(root: Path, branch: str):
     """Reuse a protected serving master, never launch a model or use a candidate."""
     from rlm.v100.competition import helper_client
@@ -65,7 +90,9 @@ def accepted_master(root: Path, branch: str):
     paths.append(run / f"profile-{state.get('context_window')}.json")
     path = next((candidate for candidate in paths if candidate.is_file()), None)
     if path is None:
-        raise MasterNotReady("Accepted master profile not ready; collected evidence retained")
+        return owned_advisor(
+            root, branch, "Accepted master profile not ready; collected evidence retained"
+        )
     startup = run / f"server-{state.get('context_window')}.startup.json"
     # The initial server is stopped when continuous learning starts its own
     # protected server. Its old startup file must not block the live profile.
@@ -74,16 +101,18 @@ def accepted_master(root: Path, branch: str):
         and startup.exists()
         and json.loads(startup.read_text()).get("stage") != "ready"
     ):
-        raise MasterNotReady("Accepted V100 serving startup not ready; evidence retained")
+        return owned_advisor(
+            root, branch, "Accepted V100 serving startup not ready; evidence retained"
+        )
     profile = load_profile(path, root)
     client = helper_client(profile, root, branch)
     client.timeout = min(10, client.timeout)
     # Full protected receipt and endpoint check precedes any generation. A
     # candidate temporarily occupying the same endpoint is never used.
     try:
         assert_served_expert(client, profile, root)
-    except (FileNotFoundError, ProcessLookupError, requests.RequestException) as error:
-        raise MasterNotReady("Accepted serving process or endpoint not ready") from error
+    except (FileNotFoundError, ProcessLookupError, requests.RequestException):
+        return owned_advisor(root, branch, "Accepted serving process or endpoint not ready")
     client.timeout = 40
     client.request_deadline = time.monotonic() + 60
     client.enable_thinking = False

rlm/v100/live_status.py

@@ -423,6 +423,47 @@ def snapshot(root: Path, mission: dict, report: dict, gpu: dict) -> dict:
             if remaining and not rtx["stale"]:
                 rtx["state"] = "cooldown"
     actors.append(rtx)
+    owned = root / "research/owned-node/status.json"
+    if (root / "research/owned-node.json").exists():
+        try:
+            value = (
+                json.loads(owned.read_text())
+                if owned.exists()
+                else {"phase": "waiting for controller service"}
+            )
+            age = time.time() - value.get("updated", 0)
+            actors.append(
+                {
+                    "id": "debian-3060ti",
+                    "label": "Debian RTX 3060 Ti / CPU",
+                    "state": value.get("phase", "unknown"),
+                    "detail": value.get(
+                        "reason",
+                        "Dedicated owned advisory inference and fixed-kernel CPU trials; remote metrics require local validation",
+                    )
+                    + (
+                        "; CPU: " + json.dumps(value["cpu_errors"])
+                        if value.get("cpu_errors")
+                        else ""
+                    ),
+                    "metrics": value.get("resources", {}),
+                    "stale": age > 90,
+                    "model": value.get("model"),
+                    "receipt": str(owned),
+                    "jobs": [],
+                }
+            )
+        except (OSError, ValueError) as error:
+            actors.append(
+                {
+                    "id": "debian-3060ti",
+                    "label": "Debian RTX 3060 Ti / CPU",
+                    "state": "unavailable",
+                    "detail": str(error)[:300],
+                    "stale": True,
+                    "jobs": [],
+                }
+            )
     for identity, label, relative in (
         ("knowledge", "Wikipedia / RAG", "research/knowledge/import-status.json"),
         ("knowledge-index", "Live Wikipedia index", "research/knowledge/index-progress.json"),