Merge PR #1767 into reviewed maintenance batch

agent/src/agent/grounding/evidence.py

@@ -29,6 +29,12 @@ _TIMESTAMP_FIELDS = ("trade_date", "date", "datetime", "timestamp", "time", "ind
 
 _MAX_GENERIC_EVIDENCE = 2_000
 
+#: Top-level fields of a code-running tool's envelope (backtest, bash, ...). They
+#: describe the process, not the market, so they never become observed evidence.
+_ENVELOPE_FIELDS = frozenset(
+    {"exit_code", "returncode", "elapsed_seconds", "duration_seconds"}
+)
+
 # Only these CSV columns count; Volume, Adj Close etc. are ignored so the
 # contradiction check gains no values it would be willing to accept.
 _CSV_PRICE_COLUMNS = {
@@ -1157,6 +1163,8 @@ class _EvidenceMixin:
                 for key, item in value.items():
                     if str(key).casefold() in timestamp_fields:
                         continue
+                    if not path and str(key).casefold() in _ENVELOPE_FIELDS:
+                        continue
                     visit(
                         item,
                         f"{path}.{key}" if path else str(key),

agent/tests/test_grounding_code_tool_envelope.py

@@ -0,0 +1,64 @@
+"""Operational envelope fields of code-running tools are not financial evidence."""
+
+from __future__ import annotations
+
+import json
+from pathlib import Path
+
+import pytest
+
+from src.agent.grounding import GroundingLedger
+
+pytestmark = pytest.mark.unit
+
+
+def _ledger(tmp_path: Path, tool: str, payload: dict) -> GroundingLedger:
+    ledger = GroundingLedger(run_dir=tmp_path, user_message="Analyze AAPL.US")
+    ledger.ingest_tool_result(
+        tool_name=tool,
+        arguments={"run_dir": str(tmp_path)},
+        result=json.dumps(payload),
+        call_id="call-1",
+        success=True,
+    )
+    return ledger
+
+
+@pytest.mark.parametrize("tool", ["backtest", "bash"])
+@pytest.mark.parametrize("field", ["exit_code", "elapsed_seconds"])
+def test_envelope_bookkeeping_is_not_observed_evidence(
+    tmp_path: Path, tool: str, field: str
+) -> None:
+    ledger = _ledger(
+        tmp_path,
+        tool,
+        {
+            "status": "ok",
+            "exit_code": 0,
+            "elapsed_seconds": 12.5,
+            "stdout": "done",
+            "stderr": "",
+        },
+    )
+
+    assert [
+        r for r in ledger._evidence if r.field in {"exit_code", "elapsed_seconds"}
+    ] == []
+
+    claim = "12.5" if field == "elapsed_seconds" else "0.0"
+    result = ledger.validate_final_answer(
+        f"The run reported {claim}.\n\n```figures\n{claim} | observed | {field} | {tool}\n```"
+    )
+    assert result.valid is False
+
+
+def test_explicit_numeric_result_of_a_tool_still_grounds(tmp_path: Path) -> None:
+    ledger = _ledger(
+        tmp_path,
+        "bash",
+        {"status": "ok", "exit_code": 0, "result": {"sma_20": 187.42}},
+    )
+
+    assert any(
+        r.field == "result.sma_20" and r.value == 187.42 for r in ledger._evidence
+    )