"""Typed failures used to distinguish protocol, policy, and infrastructure errors.""" from __future__ import annotations import json from typing import Mapping from benchkit.contracts import JsonValue, Usage, ensure_jsonable from benchkit.validation import ValidationIssue class EvaluationError(RuntimeError): """Base class for evaluation failures.""" class CompatibilityError(EvaluationError): """The policy and benchmark contracts are incompatible before rollout.""" class ProtocolError(EvaluationError): """An adapter or policy violated the declared evaluation contract.""" class PolicyError(EvaluationError): """The evaluated policy failed to produce a valid decision.""" class PolicyPreflightError(EvaluationError): """A local-readiness failure, not a policy call or an evaluation attempt. Closed codes keep dependency/credential exceptions out of CLI diagnostics. This check does not validate credentials remotely or attest provider access. """ CODES = frozenset({"credentials_missing", "dependency_unavailable", "initialization_failed", "unsupported"}) def __init__(self, code: str) -> None: if not isinstance(code, str) or code not in self.CODES: code = "initialization_failed" super().__init__("Policy runtime preflight: " + code) self.code = code class PolicyCallError(EvaluationError): """Failed act with retained accounting, never a decision or a stop. Callers must supply a sanitized message and receipt metadata only: no raw response/error bodies, credentials, or hidden reasoning. The runner does not serialize this exception's traceback or chained provider exceptions. Receipt metadata is copied and limited to 16 KiB of finite JSON; this is a storage bound, not a provider-identity or accounting attestation. """ MAX_METADATA_BYTES = 16 * 1024 def __init__(self, message: str, *, usage: Usage, metadata: Mapping[str, JsonValue] | None = None, category: str = "policy") -> None: if not isinstance(message, str) or not message.strip() or len(message) > 1024: raise ValueError("policy-call error needs a nonempty sanitized message of at most 1024 characters") if category not in ("policy", "infrastructure"): raise ValueError("policy-call error category must be policy or infrastructure") if not isinstance(usage, Usage): raise TypeError("policy-call error usage must be Usage") if metadata is not None and not isinstance(metadata, Mapping): raise TypeError("policy-call error metadata must be a JSON object") try: receipt = ensure_jsonable(metadata if metadata is not None else {}) encoded = json.dumps(receipt, allow_nan=False, separators=(",", ":")) except (TypeError, ValueError, RecursionError) as exc: raise ValueError("policy-call error metadata must be finite JSON") from exc if len(encoded.encode("utf-8")) > self.MAX_METADATA_BYTES: raise ValueError("policy-call error metadata exceeds 16 KiB") super().__init__(message) self.usage = Usage(**usage.to_dict()) self.metadata = json.loads(encoded) self.category = category class InfrastructureError(EvaluationError): """The simulator, hardware, asset, or transport layer failed.""" class ActionRejected(PolicyCallError): """Deterministic invalid input, with no execution of the rejected control. Only ``issue`` crosses the actor feedback boundary. Sanitized receipt metadata is evaluator-private, and follows PolicyCallError's copy/size checks. Service-side validation uses zero Usage; model parsing must retain the one failed call's real or explicitly unavailable usage. """ def __init__(self, issue: ValidationIssue, *, usage: Usage = Usage(), metadata: Mapping[str, JsonValue] | None = None) -> None: if not isinstance(issue, ValidationIssue): raise TypeError("action rejection requires a ValidationIssue") self.issue = ValidationIssue.from_dict(issue.to_dict()) super().__init__("Action rejected: " + self.issue.code, usage=usage, metadata=metadata, category="policy")