From fd0d03b0dd799cee365cb649942d3296bfaeec06 Mon Sep 17 00:00:00 2001
From: Brad Stein <Brad.Stein@gmail.com>
Date: Mon, 30 Mar 2026 17:38:20 -0300
Subject: [PATCH] atlasbot: enforce longhorn-qualified fallback matches

---
 services/comms/scripts/atlasbot/bot.py        | 21 +++++++++++-
 .../scripts/tests/test_atlasbot_modes.py      | 34 +++++++++++++++++++
 2 files changed, 54 insertions(+), 1 deletion(-)

diff --git a/services/comms/scripts/atlasbot/bot.py b/services/comms/scripts/atlasbot/bot.py
index cc3ba003..d9883747 100644
--- a/services/comms/scripts/atlasbot/bot.py
+++ b/services/comms/scripts/atlasbot/bot.py
@@ -2125,6 +2125,9 @@ def _doc_intent(query: str) -> bool:
             "triage",
             "recover",
             "remediate",
+            "longhorn",
+            "astreae",
+            "asteria",
         )
     )
 
@@ -4061,6 +4064,7 @@ def _fallback_fact_answer(prompt: str, context: str) -> str:
     hottest_intent = any(word in q for word in ("hottest", "highest", "most", "top", "busiest"))
     metric = _detect_metric(q)
     include_hw, _exclude_hw = _detect_hardware_filters(q)
+    wants_longhorn = any(word in tokens for word in ("longhorn", "astreae", "asteria"))
 
     if hottest_intent and metric in {"cpu", "ram", "net", "io"}:
         hottest_val = _find_value(f"hottest_{metric}")
@@ -4099,6 +4103,18 @@ def _fallback_fact_answer(prompt: str, context: str) -> str:
                 return f"Not ready nodes: {match.group(1)}."
 
     if count_intent and include_hw:
+        if wants_longhorn:
+            for hw in include_hw:
+                for fact, key, val in parsed_facts:
+                    key_tokens = set(_tokens(key or fact))
+                    if "longhorn" not in key_tokens:
+                        continue
+                    if hw not in key_tokens:
+                        continue
+                    nodes = _extract_titan_nodes(val or fact)
+                    if nodes:
+                        return f"{hw} longhorn nodes: {len(nodes)}."
+            return ""
         counts_line = _find_value("nodes_by_hardware_count")
         if counts_line:
             counts = _parse_counts(counts_line)
@@ -4112,7 +4128,6 @@ def _fallback_fact_answer(prompt: str, context: str) -> str:
                 return f"{hw} nodes: {len(items)}."
 
     if list_intent and include_hw:
-        wants_longhorn = "longhorn" in tokens
         if "control" in q:
             cp_by_hw = _find_value("control_plane_by_hardware")
             if cp_by_hw:
@@ -4206,6 +4221,10 @@ def _fallback_fact_answer(prompt: str, context: str) -> str:
     best_fact = ""
     best_score = -1
     for fact in facts:
+        if wants_longhorn:
+            fact_tokens = set(_tokens(fact))
+            if not ({"longhorn", "astreae", "asteria"} & fact_tokens):
+                continue
         key_match = re.match(r"^([\w\s/.-]+):\s*(.+)$", fact)
         if not key_match:
             key_match = re.match(r"^([\w\s/.-]+)=\s*(.+)$", fact)
diff --git a/services/comms/scripts/tests/test_atlasbot_modes.py b/services/comms/scripts/tests/test_atlasbot_modes.py
index f2886386..15f7d536 100644
--- a/services/comms/scripts/tests/test_atlasbot_modes.py
+++ b/services/comms/scripts/tests/test_atlasbot_modes.py
@@ -191,3 +191,37 @@ class AtlasbotModeTests(TestCase):
         self.assertNotIn("titan-14", reply)
         self.assertNotIn("titan-18", reply)
         self.assertIn("Confidence:", reply)
+
+    def test_longhorn_query_uses_kb_context_when_snapshot_only_has_generic_rpi4(self):
+        fact_lines = [
+            "rpi4: titan-12, titan-13, titan-14, titan-15, titan-17, titan-18, titan-19",
+        ]
+        kb_detail = (
+            "Atlas KB (retrieved):\n"
+            "- metis (knowledge/metis.md)\n"
+            "rpi4 armbian longhorn: titan-13/15/17/19"
+        )
+
+        with (
+            mock.patch.object(self.bot, "_fact_pack_lines", return_value=fact_lines),
+            mock.patch.object(self.bot, "kb_retrieve", return_value=kb_detail),
+            mock.patch.object(self.bot, "_ollama_call", side_effect=AssertionError("LLM should not be called")),
+        ):
+            reply = self.bot.open_ended_answer(
+                "which nodes in titan are the rpi4 longhorn nodes?",
+                inventory=[],
+                snapshot=None,
+                workloads=[],
+                history_lines=[],
+                mode="smart",
+                allow_tools=True,
+            )
+
+        self.assertIn("titan-13", reply)
+        self.assertIn("titan-15", reply)
+        self.assertIn("titan-17", reply)
+        self.assertIn("titan-19", reply)
+        self.assertNotIn("titan-12", reply)
+        self.assertNotIn("titan-14", reply)
+        self.assertNotIn("titan-18", reply)
+        self.assertIn("Confidence:", reply)