Skip to content

"Vibe-Trading: Your Personal Trading Agent"

fix(grounding): bind artifact reader rows to unchanged engine evidence

Haozhe Wue29c8fb9b419
diff --git a/agent/src/agent/grounding/evidence.py b/agent/src/agent/grounding/evidence.py
index 065e87b3e..49799cb75 100644
--- a/agent/src/agent/grounding/evidence.py
+++ b/agent/src/agent/grounding/evidence.py
@@ -798,7 +798,11 @@ class _EvidenceMixin:
room = _MAX_GENERIC_EVIDENCE
for name in _BACKTEST_SUMMARY_FILES:
path = directory / name
- if not path.is_file() or self._is_model_written(path):
+ if (
+ not path.is_file()
+ or not path.resolve().is_relative_to(directory.resolve())
+ or self._is_model_written(path)
+ ):
continue
if name not in _MANIFEST_EXEMPT and manifest.get(name) != _file_sha256(path):
continue
@@ -827,7 +831,11 @@ class _EvidenceMixin:
tables = directory / "artifacts"
if tables.is_dir():
for path in sorted(tables.glob("*.csv")):
- if self._is_model_written(path) or f"artifacts/{path.name}" in _BACKTEST_SUMMARY_FILES:
+ if (
+ not path.resolve().is_relative_to(directory.resolve())
+ or self._is_model_written(path)
+ or f"artifacts/{path.name}" in _BACKTEST_SUMMARY_FILES
+ ):
continue
digest = _file_sha256(path)
if digest is not None:
@@ -857,7 +865,9 @@ class _EvidenceMixin:
except OSError:
return True
- def _ingest_engine_table(self, payload: dict[str, Any], call_id: str) -> None:
+ def _ingest_engine_table(
+ self, payload: dict[str, Any], call_id: str, *, tool_name: str = "read_file"
+ ) -> None:
"""Record the rows of a backtest table the model just read back.
Only a table a completed backtest wrote counts, and only while it is
@@ -867,31 +877,48 @@ class _EvidenceMixin:
chance. Price columns are left out, as in a summary file.
Args:
- payload: The ``read_file`` result.
- call_id: The ``read_file`` call.
+ payload: The file or structured artifact reader's result.
+ call_id: The reader's call.
+ tool_name: The reader that showed these rows to the model.
"""
- raw, content = payload.get("path"), payload.get("content")
- if not isinstance(raw, str) or not isinstance(content, str):
+ raw = payload.get("path")
+ if not isinstance(raw, str):
return
try:
path = Path(raw).resolve()
except OSError:
return
known = self._engine_tables.get(str(path))
- if known is None or _file_sha256(path) != known[0]:
+ if known is None or self._is_model_written(path) or _file_sha256(path) != known[0]:
return
scope = known[1]
- truncated = "\n... (truncated)"
- if content.endswith(truncated):
- content = content[: -len(truncated)]
- content = content[: content.rfind("\n") + 1]
- try:
- rows = list(csv.reader(content.splitlines()))
- except csv.Error:
- return
- if len(rows) < 2:
- return
- header = [cell.strip() for cell in rows[0]]
+ if tool_name == "read_run_artifact":
+ columns, records = payload.get("columns"), payload.get("rows")
+ if (
+ not isinstance(columns, list)
+ or not all(isinstance(cell, str) for cell in columns)
+ or not isinstance(records, list)
+ or not all(isinstance(row, list) and len(row) == len(columns) for row in records)
+ ):
+ return
+ header = [cell.strip() for cell in columns]
+ data_rows = [["" if cell is None else str(cell) for cell in row] for row in records]
+ else:
+ content = payload.get("content")
+ if not isinstance(content, str):
+ return
+ truncated = "\n... (truncated)"
+ if content.endswith(truncated):
+ content = content[: -len(truncated)]
+ content = content[: content.rfind("\n") + 1]
+ try:
+ rows = list(csv.reader(content.splitlines()))
+ except csv.Error:
+ return
+ if len(rows) < 2:
+ return
+ header = [cell.strip() for cell in rows[0]]
+ data_rows = rows[1:]
folded = [cell.casefold() for cell in header]
date_index = next(
(index for index, name in enumerate(folded) if name in _CSV_DATE_COLUMNS), None
@@ -901,7 +928,7 @@ class _EvidenceMixin:
)
artifact = _relative_posix(path, self.run_dir.resolve())
room = _MAX_GENERIC_EVIDENCE
- for row in rows[1:]:
+ for row in data_rows:
timestamp = (
row[date_index].strip()
if date_index is not None and date_index < len(row)
@@ -926,7 +953,7 @@ class _EvidenceMixin:
self._evidence.append(
EvidenceRecord(
call_id=call_id,
- tool="read_file",
+ tool=tool_name,
symbol=None,
source="backtest",
timestamp=timestamp,
diff --git a/agent/src/agent/grounding/ledger.py b/agent/src/agent/grounding/ledger.py
index 3c8948fd0..0392bba70 100644
--- a/agent/src/agent/grounding/ledger.py
+++ b/agent/src/agent/grounding/ledger.py
@@ -280,6 +280,8 @@ class GroundingLedger(
self._ingest_market_data(arguments, payload, call_id)
elif tool_name == "read_file" and payload is not None:
self._ingest_engine_table(payload, call_id)
+ elif tool_name == "read_run_artifact" and payload is not None:
+ self._ingest_engine_table(payload, call_id, tool_name=tool_name)
elif payload is not None:
self._ingest_generic_numeric(tool_name, arguments, payload, call_id)
self.persist()
diff --git a/agent/tests/test_enforcement_l6.py b/agent/tests/test_enforcement_l6.py
index 729cf398a..29b1c8352 100644
--- a/agent/tests/test_enforcement_l6.py
+++ b/agent/tests/test_enforcement_l6.py
@@ -45,7 +45,7 @@ _CANONICAL_READ = {
"get_equity_quotes",
"get_equity_orders",
}
-_CANONICAL_WRITE = {"place_equity_order", "cancel_equity_order"}
+_CANONICAL_WRITE = {"place_equity_order", "place_option_order", "cancel_equity_order"}
def test_catalog_is_exactly_the_canonical_set() -> None:
diff --git a/agent/tests/test_grounding_backtest_outputs.py b/agent/tests/test_grounding_backtest_outputs.py
index 946e5ce83..96bfcd306 100644
--- a/agent/tests/test_grounding_backtest_outputs.py
+++ b/agent/tests/test_grounding_backtest_outputs.py
@@ -837,3 +837,87 @@ def test_the_correction_asks_for_the_answer_alone(two_runs: GroundingLedger) ->
assert "do not mention this rejection" in prompt
assert "in the user's language" in prompt
+
+
[email protected]("mode", ["rows", "downsample"])
+def test_structured_artifact_reader_grounds_only_shown_rows(
+ two_runs: GroundingLedger, monkeypatch: pytest.MonkeyPatch, mode: str
+) -> None:
+ from src.tools.run_artifact_tool import read_run_artifact
+
+ monkeypatch.setenv("VIBE_TRADING_ALLOWED_RUN_ROOTS", str(two_runs.run_dir))
+ result = read_run_artifact(
+ str(two_runs.run_dir / "rp"), "target_positions", format=mode,
+ max_rows=2, columns=["timestamp", "000001.SZ"],
+ )
+ payload = json.loads(result)
+ assert payload["returned_rows"] == 2
+ two_runs.ingest_tool_result(
+ tool_name="read_run_artifact", arguments={"run_dir": str(two_runs.run_dir / "rp")},
+ result=result, call_id="structured-weights", success=True,
+ )
+ observed = [r for r in two_runs._evidence if r.call_id == "structured-weights"]
+ assert {r.field for r in observed} == {"000001.SZ"}
+ assert {r.value for r in observed} == {row[1] for row in payload["rows"]}
+ shown = _declared("初始权重 38.97%。", "38.97% | observed | 权重 | rp/artifacts/target_positions.csv")
+ assert two_runs.validate_final_answer(shown).valid
+ unseen = _declared("未读取权重 29.79%。", "29.79% | observed | 权重 | rp/artifacts/target_positions.csv")
+ assert not two_runs.validate_final_answer(unseen).valid
+
+
[email protected]("changed", ["different_bytes", "same_bytes"])
+def test_structured_reader_never_grounds_model_written_table(
+ two_runs: GroundingLedger, monkeypatch: pytest.MonkeyPatch, changed: str
+) -> None:
+ from src.tools.run_artifact_tool import read_run_artifact
+
+ monkeypatch.setenv("VIBE_TRADING_ALLOWED_RUN_ROOTS", str(two_runs.run_dir))
+ table = two_runs.run_dir / "rp" / "artifacts" / "target_positions.csv"
+ if changed == "different_bytes":
+ table.write_text("timestamp,000001.SZ\n2024-01-02,0.4567\n", encoding="utf-8")
+ two_runs.ingest_tool_result(
+ tool_name="write_file", arguments={"path": str(table)},
+ result=json.dumps({"status": "ok", "path": str(table)}),
+ call_id="model-write-table", success=True,
+ )
+ result = read_run_artifact(str(table.parent.parent), "target_positions")
+ two_runs.ingest_tool_result(
+ tool_name="read_run_artifact", arguments={"run_dir": str(table.parent.parent)},
+ result=result, call_id="read-model-table", success=True,
+ )
+ assert not any(r.call_id == "read-model-table" for r in two_runs._evidence)
+
+
+def test_structured_reader_checks_engine_hash_and_scope(
+ two_runs: GroundingLedger, monkeypatch: pytest.MonkeyPatch
+) -> None:
+ from src.tools.run_artifact_tool import read_run_artifact
+
+ monkeypatch.setenv("VIBE_TRADING_ALLOWED_RUN_ROOTS", str(two_runs.run_dir))
+ table = two_runs.run_dir / "rp" / "artifacts" / "target_positions.csv"
+ table.write_text("timestamp,000001.SZ\n2024-01-02,0.4567\n", encoding="utf-8")
+ result = read_run_artifact(str(table.parent.parent), "target_positions")
+ two_runs.ingest_tool_result(
+ tool_name="read_run_artifact", arguments={"run_dir": str(table.parent.parent)},
+ result=result, call_id="read-changed", success=True,
+ )
+ assert not any(r.call_id == "read-changed" for r in two_runs._evidence)
+ result = read_run_artifact(str(two_runs.run_dir / "ew"), "target_positions")
+ two_runs.ingest_tool_result(
+ tool_name="read_run_artifact", arguments={"run_dir": str(two_runs.run_dir / "ew")},
+ result=result, call_id="read-ew", success=True,
+ )
+ wrong = _declared("风险平价权重 33.33%。", "33.33% | observed | 权重 | rp/artifacts/target_positions.csv")
+ assert not two_runs.validate_final_answer(wrong).valid
+
+
+def test_external_symlink_is_not_registered_as_engine_table(tmp_path: Path) -> None:
+ _write_backtest(tmp_path / "rp", RP)
+ _write_backtest(tmp_path / "ew", EW)
+ target = tmp_path / "rp" / "artifacts" / "target_positions.csv"
+ target.unlink()
+ target.symlink_to(tmp_path / "ew" / "artifacts" / "target_positions.csv")
+ ledger = GroundingLedger(run_dir=tmp_path, user_message="回测风险平价")
+ _backtest(ledger, tmp_path / "rp", "bt-rp")
+ _read(ledger, target, "read-sibling")
+ assert not any(r.call_id == "read-sibling" for r in ledger._evidence)
diff --git a/agent/tests/test_robinhood_live_account_binding.py b/agent/tests/test_robinhood_live_account_binding.py
index b2863d329..f76b0bb9f 100644
--- a/agent/tests/test_robinhood_live_account_binding.py
+++ b/agent/tests/test_robinhood_live_account_binding.py
@@ -92,6 +92,20 @@ def test_gate_stamps_the_mandate_account_on_reads_and_the_order(live_runtime: Pa
assert broker.order_calls[0]["account_number"] == ACCOUNT
+def test_unsupported_option_order_never_reaches_broker_with_valid_mandate(live_runtime: Path) -> None:
+ _write_mandate(live_runtime, _bound_mandate())
+ broker = _GateBroker()
+ spec = replace(_spec(), remote_name="place_option_order")
+ guard = order_guard.LiveOrderGuardTool(broker, spec, broker="robinhood", session_id="s1")
+
+ out = json.loads(guard.execute(symbol="AAPL", side="buy", instrument_type="option", notional_usd=100))
+
+ assert out["status"] == "blocked"
+ assert out["reason"] == "order intent could not be parsed"
+ assert broker.reads == []
+ assert broker.order_calls == []
+
+
def test_gate_refuses_every_order_when_the_mandate_names_no_account(live_runtime: Path) -> None:
_write_mandate(live_runtime, _bound_mandate(account_ref=""))
broker = _GateBroker()