Files
LZH-YS1998 4b29b89371 refactor: unify tool approval into a single engine and cut prompt storms
Collapse the dual permission stack into one policy. The runtime-side
ToolPermissionResolver (own safe lists, own grant memory, bypassed the
ApprovalEngine whenever it said ALLOW) is deleted; runtime_v2 now consults
ApprovalEngine.predict(), a synchronous fast path reading the same config
and the same persisted allowlist as the async authorize pipeline, so a
grant given anywhere is honored everywhere. permissions.py keeps only a
policy-free adapter; the duplicated permissions_v2 config fields and the
runtime grant persistence loop are removed (stale YAML keys are ignored).

New shell_safety module becomes the single source of truth for shell
classification: flag-audited read-only commands (awk/od/jq/sed -n/diff/
git subcommand table/... auto-allow; find -delete, sort -o, curl -o/-d,
rg --pre still prompt even when the bare name is config-listed),
keyword-aware compound splitting (loop/branch headers no longer poison
grants), expansion-safe $() handling, and fail-closed treatment of
anything unparseable or substitution-bearing.

Grant semantics are rebuilt around derived word-boundary prefixes:
"python3 -c" instead of token bags, interpreter -c/-m kept in the prefix,
bash/eval/sudo never grantable as prefixes, read-only segments exempt
from the every-candidate-must-match rule so a granted command chained
with ls/echo verification passes, and approve-once now records the exact
candidates as a session grant so identical re-runs stop re-prompting.
The authorize heuristic also audits the original command text instead of
the quote-dropping preview (echo "<EOF>" no longer reads as redirection).

Validated live on zz_perm_probe1 (native minimal org): awk/od/ls/cat/
sha256sum ran with zero cards, python3 -c parked once and three different
python3 -c commands then passed via the persisted prefix grant, and an
agent-issued rm -f compound correctly re-prompted showing only the
segments needing approval. Full suite failures are byte-identical to the
pre-change HEAD baseline (27 pre-existing).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-08 18:43:27 +08:00

192 lines
7.3 KiB
Python

"""Runtime-side permission adapter for Native Runtime V2.
This module intentionally contains NO permission policy. All decisions come
from the single ApprovalEngine (``opc/layer2_organization/approval.py``):
its synchronous ``predict()`` is the fast path consulted before every tool
call, and its async ``authorize_tool_call()`` (reached via the runtime's
approval callback) is the escalation path. The adapter only bridges the
engine into the executor and maps tool results back into permission events.
"""
from __future__ import annotations
from typing import Any
from loguru import logger
from opc.core.models import PermissionResolution, PermissionScope, RiskLevel, RuntimePermissionDecision
_CANDIDATE_KEYS = (
"path",
"file_path",
"directory",
"working_directory",
"target_output_dir",
"workspace_path",
"command",
"cmd",
"url",
)
def _candidate(arguments: dict[str, Any] | None) -> str:
for key in _CANDIDATE_KEYS:
value = str((arguments or {}).get(key, "") or "").strip()
if value:
return value
return "*"
def _risk(value: Any, default: RiskLevel) -> RiskLevel:
try:
return RiskLevel(str(value or default.value))
except Exception:
return default
class RuntimePermissionAdapter:
"""Thin, policy-free bridge between the tool executor and ApprovalEngine.
``policy`` is the ApprovalEngine (duck-typed: ``predict`` and
``record_denial``). Without a policy (bare runtimes, unit tests) the
adapter falls back to a static conservative default: unknown tools and
confirmation-required tools ask, everything else runs — matching the
pre-unification behavior of a runtime without an approval callback.
"""
def __init__(self, policy: Any = None, *, guardian: Any = None) -> None:
self.policy = policy
self.guardian = guardian if guardian is not None else getattr(
getattr(getattr(policy, "config", None), "permissions_v2", None), "guardian", None
)
def predicted_decision(
self,
tool: Any,
arguments: dict[str, Any] | None = None,
*,
task: Any = None,
) -> RuntimePermissionDecision:
if self.policy is not None:
try:
return self.policy.predict(tool, arguments, task=task)
except Exception:
logger.opt(exception=True).warning(
"Permission predictor failed; falling back to ask-first default"
)
return RuntimePermissionDecision(
resolution=PermissionResolution.ASK,
scope=PermissionScope.ONCE,
risk_level=RiskLevel.MEDIUM,
rationale="Permission predictor failed; requiring explicit review.",
source="runtime_prediction",
)
if tool is None:
return RuntimePermissionDecision(
resolution=PermissionResolution.ASK,
scope=PermissionScope.ONCE,
risk_level=RiskLevel.HIGH,
rationale="Unknown tool requires manual review.",
source="runtime_prediction",
)
if bool(getattr(tool, "requires_confirmation", False)):
return RuntimePermissionDecision(
resolution=PermissionResolution.ASK,
scope=PermissionScope.ONCE,
risk_level=RiskLevel.MEDIUM,
rationale="Tool is marked as requiring confirmation.",
source="runtime_prediction",
)
return RuntimePermissionDecision(
resolution=PermissionResolution.ALLOW,
scope=PermissionScope.ONCE,
risk_level=RiskLevel.LOW,
rationale="No permission policy configured.",
source="runtime_prediction",
)
def record_denial(self, tool_name: str, arguments: dict[str, Any] | None = None) -> None:
if self.policy is not None and hasattr(self.policy, "record_denial"):
try:
self.policy.record_denial(tool_name, arguments)
except Exception:
logger.opt(exception=True).debug("Failed to record permission denial")
def build_blocked_result(
self,
decision: RuntimePermissionDecision,
*,
tool_name: str,
arguments: dict[str, Any] | None = None,
) -> dict[str, Any]:
action = "reject" if decision.resolution == PermissionResolution.DENY else "require_input"
candidate = _candidate(arguments)
return {
"error": decision.rationale or f"Runtime permission blocked `{tool_name}`.",
"success": False,
"approval": {
"action": action,
"risk_level": decision.risk_level.value,
"policy_source": decision.source,
"scope": decision.scope.value,
"candidate": candidate,
"explanation": decision.rationale,
"metadata": dict(decision.metadata or {}),
},
"permission_context": {
"tool_name": tool_name,
"candidate": candidate,
"resolution": decision.resolution.value,
"risk_level": decision.risk_level.value,
"source": decision.source,
},
}
def decision_from_result(
self,
tool_name: str,
arguments: dict[str, Any] | None,
result: dict[str, Any],
) -> RuntimePermissionDecision:
"""Map a tool result back to the permission decision it reflects.
Pure event classification — grants are persisted by ApprovalEngine at
decision time, never here.
"""
approval = dict(result.get("approval", {}) or {})
action = str(approval.get("action", "") or "").strip().lower()
if action in {"require_input", "escalate"}:
return RuntimePermissionDecision(
resolution=PermissionResolution.ASK,
scope=PermissionScope.ONCE,
risk_level=_risk(approval.get("risk_level"), RiskLevel.MEDIUM),
rationale=str(result.get("error", "") or "Awaiting explicit permission."),
source="approval_engine",
metadata=approval,
)
if action == "reject":
self.record_denial(tool_name, arguments)
return RuntimePermissionDecision(
resolution=PermissionResolution.DENY,
scope=PermissionScope.ONCE,
risk_level=_risk(approval.get("risk_level"), RiskLevel.HIGH),
rationale=str(result.get("error", "") or "Permission denied."),
source="approval_engine",
metadata=approval,
)
human_reply = str(approval.get("human_reply") or result.get("human_reply") or "").strip().lower()
scope = {
"approve_session": PermissionScope.SESSION,
"always_project": PermissionScope.PROJECT,
"always_global": PermissionScope.GLOBAL,
}.get(human_reply, PermissionScope.ONCE)
return RuntimePermissionDecision(
resolution=PermissionResolution.ALLOW,
scope=scope,
risk_level=RiskLevel.LOW,
rationale="Tool execution allowed.",
source="approval_engine",
metadata=approval,
)