fix(engine): single live dispatcher per run + unified approval decision channel
OBS-4: checkpoint answers for a run whose dispatcher is still live no longer re-enter _execute_company_mode (re-entry reset live claim registries and the attempt ledger stamped in-flight cards interrupted until the streak limit killed them). The executor keeps a _live_run_dispatchers refcount; task and peer checkpoint resumes deliver the input in place and wake the dispatcher. Runs without a live dispatcher keep the original re-entry resume semantics. OBS-7: approval decisions expressed through the chat/checkpoint route now reach the approval engine instead of being parsed as plain task input (which _ask_user treated as deny, re-escalating until the card died). normalize_escalation_reply maps decision tokens/synonyms (never silently denies free text), escalation_context_for_blocked_tool rebuilds the allowlist context from pause_request.permission_context, and _resume_task_checkpoint applies the grant via apply_deferred_escalation_decision — the same engine path as the UI card. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This commit is contained in:
LZH-YS1998
@@ -0,0 +1,149 @@
|
||||
"""Regression: the unified approval decision channel (OBS-7).
|
||||
|
||||
An approval escalation that outlives its inline wait survives as a
|
||||
``task_user_input`` checkpoint whose payload carries the runtime block's
|
||||
``permission_context``. A reply that expresses a decision must reach the
|
||||
approval engine — ``normalize_escalation_reply`` maps human phrasing to a
|
||||
decision token, ``escalation_context_for_blocked_tool`` rebuilds the
|
||||
allowlist context, and ``apply_deferred_escalation_decision`` persists the
|
||||
grant exactly like the Office UI card click. Non-decision text stays plain
|
||||
task input (never a silent deny).
|
||||
"""
|
||||
from __future__ import annotations
|
||||
|
||||
import unittest
|
||||
|
||||
from opc.core.config import AutonomyConfig
|
||||
from opc.core.models import CompanyMemberSession, Task
|
||||
from opc.layer2_organization.approval import (
|
||||
ApprovalEngine,
|
||||
normalize_escalation_reply,
|
||||
)
|
||||
from opc.layer2_organization.company_mode import CompanyWorkItemExecutor
|
||||
|
||||
|
||||
class _PreferencesStub:
|
||||
def get_autonomy_preferences(self, project_id=None):
|
||||
_ = project_id
|
||||
return {"learned_actions": {}}
|
||||
|
||||
def record_autonomy_feedback(self, **kwargs):
|
||||
_ = kwargs
|
||||
|
||||
|
||||
class _StoreStub:
|
||||
async def record_approval(self, **kwargs):
|
||||
_ = kwargs
|
||||
|
||||
|
||||
class _MemoryStub:
|
||||
def append_autonomy_event(self, event, project=False):
|
||||
_ = (event, project)
|
||||
|
||||
|
||||
def _engine() -> ApprovalEngine:
|
||||
return ApprovalEngine(
|
||||
llm=object(),
|
||||
store=_StoreStub(),
|
||||
preferences=_PreferencesStub(),
|
||||
memory=_MemoryStub(),
|
||||
escalation=None,
|
||||
config=AutonomyConfig(),
|
||||
)
|
||||
|
||||
|
||||
class NormalizeEscalationReplyTests(unittest.TestCase):
|
||||
def test_exact_tokens_pass_through(self) -> None:
|
||||
for token in ("approve_once", "approve_session", "always_project",
|
||||
"always_global", "deny"):
|
||||
self.assertEqual(normalize_escalation_reply(token), token)
|
||||
|
||||
def test_approve_synonyms_map_to_approve_once(self) -> None:
|
||||
for text in ("approve", "Yes", " y ", "同意", "允许"):
|
||||
self.assertEqual(normalize_escalation_reply(text), "approve_once")
|
||||
|
||||
def test_deny_synonyms_map_to_deny(self) -> None:
|
||||
for text in ("no", "Reject", "拒绝"):
|
||||
self.assertEqual(normalize_escalation_reply(text), "deny")
|
||||
|
||||
def test_plain_content_is_not_a_decision(self) -> None:
|
||||
for text in ("", "please use conda instead", "proceed with best judgment"):
|
||||
self.assertEqual(normalize_escalation_reply(text), "")
|
||||
|
||||
|
||||
class EscalationContextForBlockedToolTests(unittest.TestCase):
|
||||
def test_shell_exec_context_carries_command_patterns(self) -> None:
|
||||
engine = _engine()
|
||||
task = Task(id="t1", title="t", project_id="p1", session_id="s1")
|
||||
context = engine.escalation_context_for_blocked_tool(
|
||||
task,
|
||||
tool_name="shell_exec",
|
||||
arguments={"command": "pip install pandas"},
|
||||
)
|
||||
self.assertEqual(context["action_kind"], "tool")
|
||||
self.assertEqual(context["action_name"], "shell_exec")
|
||||
self.assertEqual(context["project_id"], "p1")
|
||||
self.assertTrue(context["allowlist_enabled"])
|
||||
self.assertTrue(context["candidates"])
|
||||
self.assertTrue(
|
||||
any("pip" in str(item) for item in context["candidates"]),
|
||||
context["candidates"],
|
||||
)
|
||||
|
||||
def test_decision_applies_through_deferred_channel(self) -> None:
|
||||
engine = _engine()
|
||||
task = Task(id="t2", title="t", project_id="p1", session_id="s1")
|
||||
task.metadata["session_scope_id"] = "scope-1"
|
||||
context = engine.escalation_context_for_blocked_tool(
|
||||
task,
|
||||
tool_name="shell_exec",
|
||||
arguments={"command": "pip install pandas"},
|
||||
)
|
||||
outcome = engine.apply_deferred_escalation_decision("approve_session", context)
|
||||
self.assertTrue(outcome.get("approved"))
|
||||
deny = engine.apply_deferred_escalation_decision("deny", context)
|
||||
self.assertFalse(deny.get("approved"))
|
||||
|
||||
|
||||
class LiveRunDispatcherRegistryTests(unittest.TestCase):
|
||||
"""OBS-4: checkpoint answers wake a live dispatcher instead of re-entry."""
|
||||
|
||||
def _executor(self) -> CompanyWorkItemExecutor:
|
||||
import asyncio
|
||||
|
||||
executor = CompanyWorkItemExecutor.__new__(CompanyWorkItemExecutor)
|
||||
executor._live_run_dispatchers = {}
|
||||
executor._dispatcher_wake = asyncio.Event()
|
||||
return executor
|
||||
|
||||
def test_wake_returns_false_when_no_live_dispatcher(self) -> None:
|
||||
executor = self._executor()
|
||||
self.assertFalse(executor.wake_live_run_dispatcher("run-1"))
|
||||
self.assertFalse(executor._dispatcher_wake.is_set())
|
||||
|
||||
def test_wake_signals_live_dispatcher(self) -> None:
|
||||
executor = self._executor()
|
||||
executor._live_run_dispatchers["run-1"] = 1
|
||||
self.assertTrue(executor.wake_live_run_dispatcher("run-1"))
|
||||
self.assertTrue(executor._dispatcher_wake.is_set())
|
||||
|
||||
def test_run_id_extraction_prefers_first_tagged_task(self) -> None:
|
||||
t1 = Task(id="a", title="a", project_id="p", session_id="s")
|
||||
t2 = Task(id="b", title="b", project_id="p", session_id="s")
|
||||
t2.metadata["delegation_run_id"] = "run-9"
|
||||
self.assertEqual(
|
||||
CompanyWorkItemExecutor._delegation_run_id_for_tasks([t1, t2]),
|
||||
"run-9",
|
||||
)
|
||||
self.assertEqual(
|
||||
CompanyWorkItemExecutor._delegation_run_id_for_tasks([t1]),
|
||||
"",
|
||||
)
|
||||
|
||||
def test_member_session_import_smoke(self) -> None:
|
||||
# Guard against accidental import regressions in the test module.
|
||||
self.assertTrue(CompanyMemberSession)
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
unittest.main()
|
||||
@@ -0,0 +1,155 @@
|
||||
"""Engine-integration regression for OBS-4 + OBS-7.
|
||||
|
||||
A ``task_user_input`` answer for a run whose dispatcher is live must:
|
||||
1. apply an explicit approval decision through the approval engine
|
||||
(OBS-7 — the chat route was input-only, so the blocked tool
|
||||
re-escalated until the attempt ledger killed the card), and
|
||||
2. deliver the input in place and wake the live dispatcher instead of
|
||||
re-entering ``_execute_company_mode`` (OBS-4 — re-entry reset live
|
||||
claim registries and the ledger stamped in-flight cards interrupted).
|
||||
|
||||
When no dispatcher is live the resume must fall through to the legacy
|
||||
re-entry path unchanged.
|
||||
"""
|
||||
from __future__ import annotations
|
||||
|
||||
import asyncio
|
||||
import unittest
|
||||
from datetime import datetime
|
||||
from pathlib import Path
|
||||
from tempfile import TemporaryDirectory
|
||||
from unittest.mock import AsyncMock
|
||||
|
||||
from opc.core.config import AutonomyConfig
|
||||
from opc.core.models import ExecutionCheckpoint, Task, TaskStatus
|
||||
from opc.database.store import OPCStore
|
||||
from opc.engine import OPCEngine
|
||||
from opc.layer2_organization.approval import ApprovalEngine
|
||||
from opc.layer2_organization.company_mode import CompanyWorkItemExecutor
|
||||
from opc.layer2_organization.work_item_links import set_linked_work_item_id
|
||||
|
||||
|
||||
class _PreferencesStub:
|
||||
def get_autonomy_preferences(self, project_id=None):
|
||||
_ = project_id
|
||||
return {"learned_actions": {}}
|
||||
|
||||
def record_autonomy_feedback(self, **kwargs):
|
||||
_ = kwargs
|
||||
|
||||
|
||||
class _StoreStub:
|
||||
async def record_approval(self, **kwargs):
|
||||
_ = kwargs
|
||||
|
||||
|
||||
class _MemoryStub:
|
||||
def append_autonomy_event(self, event, project=False):
|
||||
_ = (event, project)
|
||||
|
||||
|
||||
def _approval_engine() -> ApprovalEngine:
|
||||
return ApprovalEngine(
|
||||
llm=object(),
|
||||
store=_StoreStub(),
|
||||
preferences=_PreferencesStub(),
|
||||
memory=_MemoryStub(),
|
||||
escalation=None,
|
||||
config=AutonomyConfig(),
|
||||
)
|
||||
|
||||
|
||||
class CheckpointAnswerLiveDispatcherTests(unittest.IsolatedAsyncioTestCase):
|
||||
async def asyncSetUp(self) -> None:
|
||||
self._tmp = TemporaryDirectory()
|
||||
self.store = OPCStore(Path(self._tmp.name) / "tasks.db")
|
||||
await self.store.initialize()
|
||||
self.engine = OPCEngine(project_id="p")
|
||||
self.engine.store = self.store
|
||||
self.engine.approval_engine = _approval_engine()
|
||||
self.executor = CompanyWorkItemExecutor.__new__(CompanyWorkItemExecutor)
|
||||
self.executor._live_run_dispatchers = {}
|
||||
self.executor._dispatcher_wake = asyncio.Event()
|
||||
self.engine.company_executor = self.executor
|
||||
self.engine._execute_company_mode = AsyncMock(return_value="re-entered")
|
||||
self.engine._execute_single_agent = AsyncMock(return_value="single-agent")
|
||||
|
||||
async def asyncTearDown(self) -> None:
|
||||
await self.store.close()
|
||||
self._tmp.cleanup()
|
||||
|
||||
async def _seed(self) -> ExecutionCheckpoint:
|
||||
task = Task(
|
||||
id="task-1",
|
||||
title="blocked worker",
|
||||
project_id="p",
|
||||
session_id="s",
|
||||
status=TaskStatus.AWAITING_HUMAN,
|
||||
metadata={
|
||||
"delegation_run_id": "run-1",
|
||||
"execution_mode": "company_mode",
|
||||
"work_item_runtime": True,
|
||||
},
|
||||
)
|
||||
set_linked_work_item_id(task, "wi-1")
|
||||
await self.store.save_task(task)
|
||||
checkpoint = ExecutionCheckpoint(
|
||||
checkpoint_id="ckpt-1",
|
||||
project_id="p",
|
||||
session_id="s",
|
||||
checkpoint_type="task_user_input",
|
||||
task_id="task-1",
|
||||
status="pending",
|
||||
payload={
|
||||
"task_id": "task-1",
|
||||
"session_id": "s",
|
||||
"execution_mode": "company_mode",
|
||||
"task_ids": ["task-1"],
|
||||
"prompt": "Tool execution blocked by autonomy policy",
|
||||
"pause_request": {
|
||||
"requires_user_input": True,
|
||||
"permission_context": {
|
||||
"tool_name": "shell_exec",
|
||||
"candidate": "pip install pandas",
|
||||
"resolution": "ask",
|
||||
},
|
||||
},
|
||||
},
|
||||
created_at=datetime.now(),
|
||||
)
|
||||
await self.store.save_execution_checkpoint(checkpoint)
|
||||
return checkpoint
|
||||
|
||||
async def test_live_dispatcher_gets_wake_and_approval_applies(self) -> None:
|
||||
checkpoint = await self._seed()
|
||||
self.executor._live_run_dispatchers["run-1"] = 1
|
||||
|
||||
reply = await self.engine._resume_task_checkpoint(checkpoint, "approve_session")
|
||||
|
||||
self.assertIn("live", reply)
|
||||
self.assertTrue(self.executor._dispatcher_wake.is_set())
|
||||
self.engine._execute_company_mode.assert_not_awaited()
|
||||
self.engine._execute_single_agent.assert_not_awaited()
|
||||
saved = await self.store.get_task("task-1")
|
||||
injected = str(saved.context_snapshot.get("user_supplied_input", ""))
|
||||
self.assertIn("Approval decision applied", injected)
|
||||
self.assertIn("shell_exec", injected)
|
||||
self.assertEqual(saved.status, TaskStatus.PENDING)
|
||||
|
||||
async def test_no_live_dispatcher_falls_through_to_reentry_path(self) -> None:
|
||||
checkpoint = await self._seed()
|
||||
|
||||
reply = await self.engine._resume_task_checkpoint(checkpoint, "please continue")
|
||||
|
||||
self.assertFalse(self.executor._dispatcher_wake.is_set())
|
||||
self.assertEqual(reply, "single-agent")
|
||||
self.engine._execute_single_agent.assert_awaited()
|
||||
saved = await self.store.get_task("task-1")
|
||||
self.assertEqual(
|
||||
saved.context_snapshot.get("user_supplied_input"),
|
||||
"please continue",
|
||||
)
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
unittest.main()
|
||||
Reference in New Issue
Block a user