Skip to content

Automate browser based workflows with AI

SKY-15418 Make output policy block reasons countable (#8403)

Andrew Neilson146e654a068b
diff --git a/skyvern/forge/sdk/copilot/agent.py b/skyvern/forge/sdk/copilot/agent.py
index fec767934..2d6176ea1 100644
--- a/skyvern/forge/sdk/copilot/agent.py
+++ b/skyvern/forge/sdk/copilot/agent.py
@@ -3478,6 +3478,7 @@ async def _translate_to_agent_result(
prior_workflow_yaml=chat_request.workflow_yaml,
output_policy_diagnostics=inline_diagnostics,
require_full_workflow_test=chat_request.product_action == "test_end_to_end",
+ evaluated_reason_codes=list(inline_raw_verdict.reason_codes),
)
# REPLACE_WORKFLOW bypasses the update_workflow tool guardrail, so
# policy and post-emission rejects run here before YAML processing.
@@ -3745,6 +3746,7 @@ async def _translate_to_agent_result(
prior_workflow_yaml=chat_request.workflow_yaml,
output_policy_diagnostics=output_policy_diagnostics,
require_full_workflow_test=direct_test_handoff,
+ evaluated_reason_codes=list(raw_output_policy_verdict.reason_codes),
)
final_user_response = str(user_response)
@@ -4248,6 +4250,19 @@ def _output_policy_diagnostics_from_guardrail_exception(exc: BaseException) -> d
return {key: data[key] for key in keys if key in data}
+def _output_policy_reason_codes_from_guardrail_exception(exc: BaseException) -> list[OutputPolicyReason]:
+ diagnostics = _output_policy_diagnostics_from_guardrail_exception(exc)
+ if diagnostics is None:
+ return list(_output_policy_verdict_from_guardrail_exception(exc).reason_codes)
+ reason_codes: list[OutputPolicyReason] = []
+ for raw_reason in diagnostics.get("raw_reason_codes") or []:
+ try:
+ reason_codes.append(OutputPolicyReason(str(raw_reason)))
+ except ValueError:
+ continue
+ return reason_codes or list(_output_policy_verdict_from_guardrail_exception(exc).reason_codes)
+
+
def _unapproved_credential_reference_reply() -> str:
# "Credentials UI" is a credential_prompt_reason() text marker the FE credential card keys off, so
# this reply must keep it verbatim. One sentence, no candidate enumeration: the card renders the
@@ -4272,6 +4287,7 @@ def _build_output_policy_blocked_result(
prior_workflow_yaml: str | None,
output_policy_diagnostics: dict[str, Any] | None = None,
require_full_workflow_test: bool = False,
+ evaluated_reason_codes: list[OutputPolicyReason] | None = None,
) -> AgentResult:
# A blocker turn whose signal owns final rendering never ships a proposal;
# steering-only blockers should still flow through normal output-policy
@@ -4282,9 +4298,15 @@ def _build_output_policy_blocked_result(
ctx.last_workflow if ctx.last_workflow is not None and ctx.last_workflow_yaml and not blocker_active else None
)
preserved_workflow_yaml = ctx.last_workflow_yaml if preserved_workflow is not None else None
+ output_policy_reasons = list(verdict.reason_codes if evaluated_reason_codes is None else evaluated_reason_codes)
structured = StructuredContext.from_json_str(prior_global_llm_context)
structured.decisions_made.append(
- "output-policy blocked final output: " + ", ".join(reason.value for reason in verdict.reason_codes)
+ "output-policy blocked final output: " + ", ".join(reason.value for reason in output_policy_reasons)
+ )
+ LOG.warning(
+ "copilot output policy blocked final output",
+ log_code="copilot_output_policy_block",
+ **{"copilot.output_policy_reasons": [reason.value for reason in output_policy_reasons]},
)
add_saved_draft_copy = False
fallback_user_response: str | None = None
@@ -4390,6 +4412,7 @@ def _build_output_policy_blocked_result(
reason_code="output_policy_block",
terminal_reason="output_policy_block",
)
+ output_policy_outcome = output_policy_outcome.model_copy(update={"output_policy_reasons": output_policy_reasons})
proposal_tested = ctx.last_full_workflow_test_ok if require_full_workflow_test else ctx.last_test_ok
return _make_agent_result(
ctx,
@@ -5081,6 +5104,7 @@ async def _run_copilot_turn_impl(
prior_workflow_yaml=chat_request.workflow_yaml,
output_policy_diagnostics=_output_policy_diagnostics_from_guardrail_exception(exc),
require_full_workflow_test=chat_request.product_action == "test_end_to_end",
+ evaluated_reason_codes=_output_policy_reason_codes_from_guardrail_exception(exc),
)
except CopilotTurnHalt as exc:
LOG.info(
diff --git a/skyvern/forge/sdk/copilot/output_policy.py b/skyvern/forge/sdk/copilot/output_policy.py
index e2c3843ea..8e7e2f29c 100644
--- a/skyvern/forge/sdk/copilot/output_policy.py
+++ b/skyvern/forge/sdk/copilot/output_policy.py
@@ -30,6 +30,7 @@ from skyvern.forge.sdk.copilot.workflow_credential_utils import (
workflow_credential_ids_from_parsed,
workflow_credential_origins_from_parsed,
)
+from skyvern.forge.sdk.schemas.copilot_turn_outcome import OutputPolicyReason
WORKFLOW_PRESENT_SENTINEL = object()
_CREDENTIAL_ID_RE = re.compile(r"\bcred_[A-Za-z0-9][A-Za-z0-9_-]*\b")
@@ -174,20 +175,6 @@ class CopilotOutputKind(StrEnum):
WORKFLOW_RUN_RESULT = "workflow_run_result"
-class OutputPolicyReason(StrEnum):
- RAW_SECRET_LEAK = "raw_secret_leak"
- UNAPPROVED_CREDENTIAL_REFERENCE = "unapproved_credential_reference"
- CREDENTIAL_SCOPE_BROADENED = "credential_scope_broadened"
- UNBACKED_WORKFLOW_DELIVERY_CLAIM = "unbacked_workflow_delivery_claim"
- MISSING_PROPOSAL_STATE = "missing_proposal_state"
- PERSISTENCE_STATE_MISMATCH = "persistence_state_mismatch"
- OUTPUT_POLICY_CONTEXT_MISSING = "output_policy_context_missing"
- INTERNAL_BLOCK_TAXONOMY_LEAK = "internal_block_taxonomy_leak"
- INTERNAL_CLASSIFIER_VOCAB_LEAK = "internal_classifier_vocab_leak"
- SELF_PRESCRIPTIVE_PHRASE_LEAK = "self_prescriptive_phrase_leak"
- WORKFLOW_YAML_IN_REPLY = "workflow_yaml_in_reply"
-
-
@dataclass
class OutputPolicyVerdict:
allowed: bool = True
diff --git a/skyvern/forge/sdk/schemas/copilot_turn_outcome.py b/skyvern/forge/sdk/schemas/copilot_turn_outcome.py
index 8db842510..69ca35bef 100644
--- a/skyvern/forge/sdk/schemas/copilot_turn_outcome.py
+++ b/skyvern/forge/sdk/schemas/copilot_turn_outcome.py
@@ -23,6 +23,20 @@ class ResponseKind(StrEnum):
RECOVER = "recover"
+class OutputPolicyReason(StrEnum):
+ RAW_SECRET_LEAK = "raw_secret_leak"
+ UNAPPROVED_CREDENTIAL_REFERENCE = "unapproved_credential_reference"
+ CREDENTIAL_SCOPE_BROADENED = "credential_scope_broadened"
+ UNBACKED_WORKFLOW_DELIVERY_CLAIM = "unbacked_workflow_delivery_claim"
+ MISSING_PROPOSAL_STATE = "missing_proposal_state"
+ PERSISTENCE_STATE_MISMATCH = "persistence_state_mismatch"
+ OUTPUT_POLICY_CONTEXT_MISSING = "output_policy_context_missing"
+ INTERNAL_BLOCK_TAXONOMY_LEAK = "internal_block_taxonomy_leak"
+ INTERNAL_CLASSIFIER_VOCAB_LEAK = "internal_classifier_vocab_leak"
+ SELF_PRESCRIPTIVE_PHRASE_LEAK = "self_prescriptive_phrase_leak"
+ WORKFLOW_YAML_IN_REPLY = "workflow_yaml_in_reply"
+
+
_UNSAFE_IDENTIFIER_RE = re.compile(r"[^A-Za-z0-9_\- ]")
@@ -65,6 +79,7 @@ class TurnOutcome(BaseModel):
response_kind: ResponseKind
reason_code: str = ""
+ output_policy_reasons: list[OutputPolicyReason] = Field(default_factory=list)
actuation_obligation_key: str = ""
normalized_reply_signature: str = ""
tool_calls: list[str] = Field(default_factory=list)
@@ -78,3 +93,16 @@ class TurnOutcome(BaseModel):
idempotency_digest: str | None = None
unresolved_runtime_failure: UnresolvedRuntimeFailure | None = None
connected_account_choices: list[ConnectedAccountChoice] | None = None
+
+ @field_validator("output_policy_reasons", mode="before")
+ @classmethod
+ def _drop_unknown_output_policy_reasons(cls, value: object) -> object:
+ if not isinstance(value, list):
+ return value
+ known_reasons: list[OutputPolicyReason] = []
+ for raw_reason in value:
+ try:
+ known_reasons.append(OutputPolicyReason(raw_reason))
+ except (TypeError, ValueError):
+ continue
+ return known_reasons
diff --git a/tests/unit/test_copilot_output_policy.py b/tests/unit/test_copilot_output_policy.py
index 1f9f68a8f..91448347a 100644
--- a/tests/unit/test_copilot_output_policy.py
+++ b/tests/unit/test_copilot_output_policy.py
@@ -1609,6 +1609,79 @@ def test_output_policy_credential_block_asks_for_credential_confirmation(
assert "I could not safely return" not in result.user_response
+def test_output_policy_block_turn_outcome_persists_and_logs_every_reason(monkeypatch: pytest.MonkeyPatch) -> None:
+ log = MagicMock()
+ monkeypatch.setattr(agent_module, "LOG", log)
+ reasons = [
+ OutputPolicyReason.PERSISTENCE_STATE_MISMATCH,
+ OutputPolicyReason.OUTPUT_POLICY_CONTEXT_MISSING,
+ ]
+
+ result = agent_module._build_output_policy_blocked_result(
+ _ctx(),
+ OutputPolicyVerdict(reason_codes=reasons),
+ prior_global_llm_context="{}",
+ prior_workflow_yaml="title: Prior",
+ )
+
+ assert result.turn_outcome is not None
+ assert result.turn_outcome.output_policy_reasons == reasons
+ log.warning.assert_any_call(
+ "copilot output policy blocked final output",
+ log_code="copilot_output_policy_block",
+ **{"copilot.output_policy_reasons": [reason.value for reason in reasons]},
+ )
+
+
+def test_unblocked_turn_outcome_has_empty_output_policy_reasons() -> None:
+ outcome = TurnOutcome(response_kind=ResponseKind.ANSWER)
+
+ assert outcome.output_policy_reasons == []
+ assert outcome.model_dump(mode="json")["output_policy_reasons"] == []
+
+
+def test_turn_outcome_drops_unknown_historical_output_policy_reasons() -> None:
+ outcome = TurnOutcome.model_validate(
+ {
+ "response_kind": "clarify",
+ "reason_code": "output_policy_block",
+ "terminal_reason": "output_policy_block",
+ "output_policy_reasons": [
+ "internal_tool_instruction_leak",
+ OutputPolicyReason.RAW_SECRET_LEAK.value,
+ ],
+ }
+ )
+
+ assert outcome.response_kind is ResponseKind.CLARIFY
+ assert outcome.terminal_reason == "output_policy_block"
+ assert outcome.output_policy_reasons == [OutputPolicyReason.RAW_SECRET_LEAK]
+
+
+def test_output_policy_fallback_replacement_keeps_turn_outcome_reasons(
+ monkeypatch: pytest.MonkeyPatch,
+) -> None:
+ ctx = _ctx()
+ ctx.last_test_anti_bot = "challenge"
+ reasons = [OutputPolicyReason.PERSISTENCE_STATE_MISMATCH]
+ monkeypatch.setattr(
+ agent_module,
+ "evaluate_output_policy",
+ lambda **_kwargs: OutputPolicyVerdict(reason_codes=[OutputPolicyReason.RAW_SECRET_LEAK]),
+ )
+
+ result = agent_module._build_output_policy_blocked_result(
+ ctx,
+ OutputPolicyVerdict(reason_codes=reasons),
+ prior_global_llm_context="{}",
+ prior_workflow_yaml="title: Prior",
+ )
+
+ assert result.turn_outcome is not None
+ assert result.turn_outcome.output_policy_reasons == reasons
+ assert result.turn_outcome.model_dump(mode="json")["output_policy_reasons"] == [reason.value for reason in reasons]
+
+
def test_output_policy_block_preserves_already_gated_workflow_proposal() -> None:
ctx = _ctx()
ctx.last_workflow = SimpleNamespace(name="draft")
@@ -1919,13 +1992,13 @@ workflow_definition:
assert ctx.consecutive_tool_tracker == ["update_workflow", "update_workflow"]
-def test_inline_replace_workflow_rejects_raw_secret_before_processing(monkeypatch) -> None:
+def test_inline_replace_output_policy_turn_outcome_keeps_raw_reasons_before_processing(monkeypatch) -> None:
process_mock = MagicMock()
monkeypatch.setattr("skyvern.forge.sdk.copilot.tools.workflow_update._process_workflow_yaml", process_mock)
result = _fake_run_result(
{
"type": "REPLACE_WORKFLOW",
- "user_response": "Here is the workflow.",
+ "user_response": "Here is the workflow. Saw [copilot:nudge] in the trace.",
"workflow_yaml": """
workflow_definition:
blocks:
@@ -1950,6 +2023,11 @@ workflow_definition:
assert agent_result.updated_workflow is None
assert agent_result.clear_proposed_workflow is False
assert agent_result.proposal_disposition == "no_proposal"
+ assert agent_result.turn_outcome is not None
+ assert agent_result.turn_outcome.output_policy_reasons == [
+ OutputPolicyReason.RAW_SECRET_LEAK,
+ OutputPolicyReason.INTERNAL_BLOCK_TAXONOMY_LEAK,
+ ]
process_mock.assert_not_called()
@@ -2179,6 +2257,59 @@ def test_sdk_output_guardrail_records_raw_soft_reason_alongside_hard_block() ->
assert "internal_block_taxonomy_leak" in raw_reason_codes
+def test_translate_output_policy_block_turn_outcome_keeps_raw_soft_reason(
+ monkeypatch: pytest.MonkeyPatch,
+) -> None:
+ log = MagicMock()
+ monkeypatch.setattr(agent_module, "LOG", log)
+ leak = f"Use the key {_FAKE_OPENAI_KEY} to authenticate. Saw [copilot:nudge] in the trace."
+
+ agent_result = asyncio.run(
+ agent_module._translate_to_agent_result(
+ _fake_run_result({"type": "REPLY", "user_response": leak}),
+ _ctx(),
+ global_llm_context=None,
+ chat_request=_chat_request(),
+ organization_id="org-1",
+ )
+ )
+
+ assert agent_result.turn_outcome is not None
+ assert agent_result.turn_outcome.output_policy_reasons == [
+ OutputPolicyReason.RAW_SECRET_LEAK,
+ OutputPolicyReason.INTERNAL_BLOCK_TAXONOMY_LEAK,
+ ]
+ log.warning.assert_any_call(
+ "copilot output policy blocked final output",
+ log_code="copilot_output_policy_block",
+ **{
+ "copilot.output_policy_reasons": [
+ OutputPolicyReason.RAW_SECRET_LEAK.value,
+ OutputPolicyReason.INTERNAL_BLOCK_TAXONOMY_LEAK.value,
+ ]
+ },
+ )
+
+
+def test_output_policy_guardrail_turn_outcome_recovers_every_raw_reason() -> None:
+ output_info = {
+ "allowed": False,
+ "reason_codes": [OutputPolicyReason.RAW_SECRET_LEAK.value],
+ "raw_reason_codes": [
+ OutputPolicyReason.RAW_SECRET_LEAK.value,
+ OutputPolicyReason.INTERNAL_BLOCK_TAXONOMY_LEAK.value,
+ ],
+ }
+ exc = SimpleNamespace(guardrail_result=SimpleNamespace(output=SimpleNamespace(output_info=output_info)))
+
+ reasons = agent_module._output_policy_reason_codes_from_guardrail_exception(exc)
+
+ assert reasons == [
+ OutputPolicyReason.RAW_SECRET_LEAK,
+ OutputPolicyReason.INTERNAL_BLOCK_TAXONOMY_LEAK,
+ ]
+
+
def test_evaluate_output_policy_skips_new_detectors_on_ask_question() -> None:
classifier_leak = "RequestPolicy blocked this turn, safe_reason_code=request_policy_clarification."
self_prescriptive_leak = "Please send 'continue debugging' next and I'll keep going."