atlasbot: improve fast fallback and usage filtering

services/comms/atlasbot-deployment.yaml

@@ -16,7 +16,7 @@ spec:
       labels:
         app: atlasbot
       annotations:
-        checksum/atlasbot-configmap: manual-atlasbot-94
+        checksum/atlasbot-configmap: manual-atlasbot-95
         vault.hashicorp.com/agent-inject: "true"
         vault.hashicorp.com/role: "comms"
         vault.hashicorp.com/agent-inject-secret-turn-secret: "kv/data/atlas/comms/turn-shared-secret"

services/comms/scripts/atlasbot/bot.py

@@ -923,7 +923,7 @@ def _nodes_by_arch(inventory: list[dict[str, Any]]) -> dict[str, list[str]]:
         grouped[(node.get("arch") or "unknown")].append(node["name"])
     return {k: sorted(v) for k, v in grouped.items()}
 
-def _node_usage_table(metrics: dict[str, Any]) -> list[dict[str, Any]]:
+def _node_usage_table(metrics: dict[str, Any], *, allowed_nodes: set[str] | None = None) -> list[dict[str, Any]]:
     usage = metrics.get("node_usage") if isinstance(metrics.get("node_usage"), dict) else {}
     per_node: dict[str, dict[str, Any]] = {}
     for metric_name, entries in usage.items() if isinstance(usage, dict) else []:
@@ -935,6 +935,8 @@ def _node_usage_table(metrics: dict[str, Any]) -> list[dict[str, Any]]:
             node = entry.get("node")
             if not isinstance(node, str) or not node:
                 continue
+            if allowed_nodes and node not in allowed_nodes:
+                continue
             per_node.setdefault(node, {})[metric_name] = entry.get("value")
     return [{"node": node, **vals} for node, vals in sorted(per_node.items())]
 
@@ -1139,7 +1141,8 @@ def facts_context(
         if items:
             lines.append(f"- top_restarts_1h: {', '.join(items)}")
 
-    usage_table = _node_usage_table(metrics)
+    allowed_nodes = {node.get("name") for node in inv if isinstance(node, dict) and node.get("name")}
+    usage_table = _node_usage_table(metrics, allowed_nodes=allowed_nodes or None)
     if usage_table:
         lines.append("- node_usage (cpu/ram/net/io):")
         for entry in usage_table:
@@ -3906,6 +3909,31 @@ def _is_quantitative_prompt(prompt: str) -> bool:
     return False
 
 
+def _is_list_prompt(prompt: str) -> bool:
+    q = normalize_query(prompt)
+    if not q:
+        return False
+    if any(phrase in q for phrase in ("list", "names", "name", "show")):
+        return True
+    if any(phrase in q for phrase in ("which nodes", "what nodes", "what are the nodes")):
+        return True
+    return False
+
+
+def _needs_full_fact_pack(prompt: str) -> bool:
+    q = normalize_query(prompt)
+    tokens = set(_tokens(prompt))
+    if _is_quantitative_prompt(prompt) or _is_list_prompt(prompt):
+        return True
+    if tokens & {"workload", "pods", "namespace"}:
+        return True
+    if _NAME_INDEX and tokens & _NAME_INDEX:
+        return True
+    if any(phrase in q for phrase in ("where does", "where is", "where are", "running", "run on", "hosted on", "primary node")):
+        return True
+    return False
+
+
 def _open_ended_fast_single(
     prompt: str,
     *,
@@ -3937,6 +3965,8 @@ def _open_ended_fast_single(
     fallback = _fallback_fact_answer(prompt, context)
     if fallback and (_is_quantitative_prompt(prompt) or not _has_body_lines(reply)):
         reply = fallback
+    if not _has_body_lines(reply):
+        reply = "I don't have enough data in the current snapshot to answer that."
     if state:
         state.update("done", step=_open_ended_total_steps("fast"))
     return _ensure_scores(reply)
@@ -3967,7 +3997,7 @@ def _open_ended_fast(
     )
     selected_meta = _fact_pack_meta(selected_lines)
     selected_pack = _fact_pack_text(selected_lines, selected_meta)
-    if _is_quantitative_prompt(prompt) or not selected_lines:
+    if _needs_full_fact_pack(prompt) or not selected_lines:
         selected_pack = fact_pack
     if state:
         state.total_steps = _open_ended_total_steps("fast")