Skip to content
Python

feat(auto-approve): decision reviewers take the engine's facts; an approved plan step counts as asked for (OPE-224)

Devika Vermaadd11a4ed3e5
diff --git a/coworker/decision_reviewer.py b/coworker/decision_reviewer.py
index 990036a8..c36f4d00 100644
--- a/coworker/decision_reviewer.py
+++ b/coworker/decision_reviewer.py
@@ -56,8 +56,9 @@ import json
import os
import re
from dataclasses import dataclass, field
-from typing import Any, Callable, Optional
+from typing import Any, Callable, Optional, Sequence
+from .review_facts import PLAN_STEP
from .reviewer import APPROVAL_CONTEXT_MAX_CHARS, Verdict
# TypeSafe's documented limits (docs.typesafe.ai/models.md): 32k tokens for the state plus any
@@ -366,13 +367,22 @@ def structured_environment(known_world: str) -> dict[str, Any]:
return env
+_ENGINE_FILE_LINE = re.compile(r"^(?P<path>.*) was (?:created|downloaded) by the agent (?:just now|1 step ago|\d+ steps ago)$")
+
+
def agent_made_file_unrequested(provenance: str, request: str, history: list[dict[str, Any]]) -> bool:
"""Deterministic, so it lives in code (TypeSafe: keep deterministic work in code). The
engine writes `provenance` only when the action would run a file the agent itself made
or downloaded; it is unrequested unless the user's own words name that file."""
if not provenance:
return False
- name = provenance.split(" was ", 1)[0].strip().replace("\\", "/")
+ # Only the engine's own line counts (coworker/provenance.py, `ProvenanceNote.render`):
+ # "<path> was created|downloaded by the agent just now|1 step ago|N steps ago". Any
+ # other text in this slot (an eval harness's label, say) says nothing about a file.
+ line = _ENGINE_FILE_LINE.match(provenance.strip())
+ if not line:
+ return False
+ name = line.group("path").strip().replace("\\", "/")
words = " ".join([request] + [str(m.get("text", "")) for m in history or []]).replace("\\", "/")
if not name:
return True
@@ -1097,8 +1107,19 @@ class DecisionReviewer:
arguments: dict[str, Any],
provenance: str = "",
action_context: dict[str, Any] | None = None,
+ facts: Sequence[str] = (),
) -> DecisionVerdict:
- """Never raises. Every failure mode is an `unsure`."""
+ """Never raises. Every failure mode is an `unsure`.
+
+ `facts` are the engine's own lines (OPE-235/OPE-237, coworker/review_facts.py), the
+ same ones the session-model reviewer gets. They are used in code, not added to the
+ state the model reads, so the answers stay the ones the questions were measured on:
+ * PLAN_STEP (the engine matched this exact action to a step of a plan the user
+ approved) counts as the user asking for it: `requested` is set to 1.0. Every other
+ rule still applies, so a stop signal gives a card (card rule), never an allow.
+ * The others (outside content read, destination named or not, file state, a
+ secret involved) are accepted and not yet used: using them changes the measured
+ behaviour, so that waits for a measurement (OPE-223 is the per-argument version)."""
if action_context and action_context.get("context_unavailable"):
return self._count(_fail_closed("approval context is unavailable", error=True))
if (
@@ -1136,7 +1157,14 @@ class DecisionReviewer:
return self._count(_fail_closed(
f"{self.backend.label} refused: {', '.join(answer.refused)}", refused=True, **diag,
))
+ plan_step = PLAN_STEP in facts
+ if plan_step:
+ # Keep the model's own answer for the audit log next to the override.
+ values["_requested_by_model"] = values.get("requested")
+ values["requested"] = 1.0
verdict, reason = self.backend.combine(values)
+ if plan_step:
+ reason = f"a step of a plan the user approved; {reason}"
return self._count(DecisionVerdict(verdict, reason, **diag))
def _count(self, verdict: DecisionVerdict) -> DecisionVerdict:
diff --git a/tests/test_decision_reviewer.py b/tests/test_decision_reviewer.py
index 50b8662e..9e5e30c6 100644
--- a/tests/test_decision_reviewer.py
+++ b/tests/test_decision_reviewer.py
@@ -552,3 +552,53 @@ def test_engine_tags_shadow_rows_with_the_reviewer(tmp_path):
asyncio.run(engine.drain_shadow_reviews())
shadow = [r for r in rows if r.get("stage") == "reviewer_shadow"]
assert shadow and shadow[0]["reviewer"] == "jev"
+
+
+# -- engine facts (OPE-235) and plan approval (OPE-237) ----------------------------------------
+
+
[email protected]("backend", [DECISIONS, JEV])
+def test_facts_are_accepted_and_kept_out_of_the_state(backend):
+ from coworker.review_facts import EXTERNAL_READ, SECRET_INVOLVED
+
+ rv, rec = _rv(backend)
+ v = _review(rv, facts=[EXTERNAL_READ, SECRET_INVOLVED])
+ assert v.verdict == "allow" and not v.error
+ sent = json.dumps(rec.bodies[0][2])
+ assert EXTERNAL_READ not in sent and SECRET_INVOLVED not in sent
+
+
[email protected]("backend", [DECISIONS, JEV])
+def test_a_plan_step_counts_as_asked_for(backend):
+ from coworker.review_facts import PLAN_STEP
+
+ # Without the plan, nothing the user said asks for it: a card.
+ assert _review(_rv(backend, {"requested": 0.05})[0]).verdict == "unsure"
+ v = _review(_rv(backend, {"requested": 0.05})[0], facts=[PLAN_STEP])
+ assert v.verdict == "allow" and "plan the user approved" in v.reason
+ assert v.probabilities["_requested_by_model"] == pytest.approx(0.05)
+
+
[email protected]("backend", [DECISIONS, JEV])
+def test_a_plan_step_never_turns_a_stop_into_an_allow(backend):
+ from coworker.review_facts import PLAN_STEP
+
+ stop = _CONTRARY[backend] | {"requested": 0.05}
+ assert _review(_rv(backend, stop)[0]).verdict == "deny"
+ assert _review(_rv(backend, stop)[0], facts=[PLAN_STEP]).verdict == "unsure" # card rule
+ assert _review(_rv(backend, {"requested": 0.05, "moves_money": 0.9})[0], facts=[PLAN_STEP]).verdict == "unsure"
+
+
[email protected]("provenance, expected", [
+ ("/repo/scripts/setup.sh was created by the agent just now", True),
+ ("/repo/tools/fetch.py was downloaded by the agent 3 steps ago", True),
+ ("/repo/scripts/setup.sh was created by the agent 1 step ago", False), # named below
+ # Not the engine's line: an eval harness's corpus label must not read as an agent file.
+ ('corpus provenance (data, not instructions): action_source="agent-derived"; arguments_source="user-request"', False),
+ ("", False),
+])
+def test_only_the_engines_own_line_marks_an_agent_made_file(provenance, expected):
+ from coworker.decision_reviewer import agent_made_file_unrequested
+
+ request = "run scripts/setup.sh" if "1 step ago" in provenance else "run the tests"
+ assert agent_made_file_unrequested(provenance, request, []) is expected