From 3885abd7fec1d29d9cba2a1c41b0e38cd769f567 Mon Sep 17 00:00:00 2001
From: Brad Stein <Brad.Stein@gmail.com>
Date: Mon, 2 Feb 2026 12:38:11 -0300
Subject: [PATCH] atlasbot: strengthen evidence and units

---
 atlasbot/engine/answerer.py  | 23 ++++++++++++++++++++---
 atlasbot/llm/prompts.py      |  4 +++-
 atlasbot/snapshot/builder.py |  2 ++
 3 files changed, 25 insertions(+), 4 deletions(-)

diff --git a/atlasbot/engine/answerer.py b/atlasbot/engine/answerer.py
index 2dc7eed..a8e7fd9 100644
--- a/atlasbot/engine/answerer.py
+++ b/atlasbot/engine/answerer.py
@@ -163,6 +163,7 @@ class AnswerEngine:
         lexicon_ctx = _lexicon_context(summary)
         key_facts: list[str] = []
         metric_facts: list[str] = []
+        facts_used: list[str] = []
 
         started = time.monotonic()
         reply = ""
@@ -385,6 +386,7 @@ class AnswerEngine:
                             "top_scored": scored_preview,
                         },
                     )
+                facts_used = list(dict.fromkeys(key_facts)) if key_facts else list(dict.fromkeys(metric_facts))
                 snapshot_context = "ClusterSnapshot:\n" + "\n".join([chunk["text"] for chunk in selected])
                 if key_facts:
                     snapshot_context = "KeyFacts:\n" + "\n".join(key_facts) + "\n\n" + snapshot_context
@@ -417,6 +419,8 @@ class AnswerEngine:
             runbook_fix = _needs_runbook_fix(reply, runbook_paths)
             runbook_needed = _needs_runbook_reference(normalized, runbook_paths, reply)
             needs_evidence = _needs_evidence_fix(reply, classify)
+            if classify.get("question_type") in {"open_ended", "planning"} and metric_facts:
+                needs_evidence = True
             resolved_runbook = None
             if runbook_paths and (runbook_fix or runbook_needed):
                 resolver_prompt = prompts.RUNBOOK_SELECT_PROMPT + "\nQuestion: " + normalized
@@ -640,7 +644,7 @@ class AnswerEngine:
             reply = await self._dedup_reply(reply, plan, call_llm, tag="dedup")
 
             scores = await self._score_answer(normalized, reply, plan, call_llm)
-            claims = await self._extract_claims(normalized, reply, summary, call_llm)
+            claims = await self._extract_claims(normalized, reply, summary, facts_used, call_llm)
         except LLMLimitReached:
             if not reply:
                 reply = "I started working on this but hit my reasoning limit. Ask again with 'Run limitless' for a deeper pass."
@@ -736,13 +740,24 @@ class AnswerEngine:
         question: str,
         reply: str,
         summary: dict[str, Any],
+        facts_used: list[str],
         call_llm: Callable[..., Any],
     ) -> list[ClaimItem]:
         if not reply or not summary:
             return []
         summary_json = _json_excerpt(summary)
-        prompt = prompts.CLAIM_MAP_PROMPT + "\nQuestion: " + question + "\nAnswer: " + reply
-        raw = await call_llm(prompts.CLAIM_SYSTEM, prompt, context=f"SnapshotSummaryJson:{summary_json}", model=self._settings.ollama_model_fast, tag="claim_map")
+        facts_used = [line.strip() for line in (facts_used or []) if line and line.strip()]
+        facts_block = ""
+        if facts_used:
+            facts_block = "\nFactsUsed:\n" + "\n".join([f"- {line}" for line in facts_used[:12]])
+        prompt = prompts.CLAIM_MAP_PROMPT + "\nQuestion: " + question + "\nAnswer: " + reply + facts_block
+        raw = await call_llm(
+            prompts.CLAIM_SYSTEM,
+            prompt,
+            context=f"SnapshotSummaryJson:{summary_json}",
+            model=self._settings.ollama_model_fast,
+            tag="claim_map",
+        )
         data = _parse_json_block(raw, fallback={})
         claims_raw = data.get("claims") if isinstance(data, dict) else None
         claims: list[ClaimItem] = []
@@ -1745,6 +1760,8 @@ def _best_runbook_match(candidate: str, allowed: list[str]) -> str | None:
 
 
 def _resolve_path(data: Any, path: str) -> Any | None:
+    if path.startswith("line:"):
+        return path.split("line:", 1)[1].strip()
     cursor = data
     for part in re.split(r"\.(?![^\[]*\])", path):
         if not part:
diff --git a/atlasbot/llm/prompts.py b/atlasbot/llm/prompts.py
index 174e6d7..a47f38e 100644
--- a/atlasbot/llm/prompts.py
+++ b/atlasbot/llm/prompts.py
@@ -187,7 +187,9 @@ CLAIM_SYSTEM = (
 )
 
 CLAIM_MAP_PROMPT = (
-    "Return JSON with claims list; each claim: id, claim, evidence (list of {path, reason})."
+    "Return JSON with claims list; each claim: id, claim, evidence (list of {path, reason}). "
+    "If FactsUsed is provided, prefer evidence paths of the form line:<exact line> from FactsUsed. "
+    "Otherwise use SnapshotSummaryJson paths."
 )
 
 FOLLOWUP_SYSTEM = (
diff --git a/atlasbot/snapshot/builder.py b/atlasbot/snapshot/builder.py
index 8e13ae4..c31b972 100644
--- a/atlasbot/snapshot/builder.py
+++ b/atlasbot/snapshot/builder.py
@@ -1364,6 +1364,8 @@ def _append_hottest(lines: list[str], summary: dict[str, Any]) -> None:
             value = _format_rate_bytes(entry.get("value"))
         else:
             value = _format_float(entry.get("value"))
+            if value and key in {"cpu", "ram", "disk"}:
+                value = f"{value}%"
         if node:
             label = node
             if hardware: