fix(orchestrator): un-deadlock a CEO-rejected coordination root

A coordination root (team=main_pm, product-linked, no repo) the CEO sends back
lands in needs_revision, but the dev dispatcher skips it (not a cell team) and
the closure path only handles paused parents — so it sat in needs_revision
forever. (NOT a foundation-spec gap: the spec already allows needs_revision ->
claimed for any role.)

- _dispatch_revision_coordination_roots: re-spawn the owning PM for a
  needs_revision coordination root so it re-coordinates the revision (registered
  in the dispatch loop after PM closure)
- _readiness_check_role_for_status: widen the dev-owned states (needs_revision,
  verifying) to also accept cell_pm/main_pm for coordination roots — a pure
  widening; normal code tasks stay dev/doc-only
- 16 unit tests (dispatcher decision + readiness widening)
This commit is contained in:
Renn F
2026-06-17 16:50:10 +02:00
parent 748ff7813e
commit eec35c0357
3 changed files with 189 additions and 3 deletions
+45 -3
View File
@@ -2330,7 +2330,7 @@ class AgentOrchestrator:
@staticmethod @staticmethod
def _readiness_check_role_for_status( def _readiness_check_role_for_status(
agent_id: str, role: str, status: str agent_id: str, role: str, status: str, *, is_coordination: bool = False
) -> str | None: ) -> str | None:
"""Verify agent role matches the role expected for the task status. """Verify agent role matches the role expected for the task status.
@@ -2339,7 +2339,10 @@ class AgentOrchestrator:
developer/documenter to defang the bug where QA got developer/documenter to defang the bug where QA got
respawned on a `needs_revision` task via the crash-restart path respawned on a `needs_revision` task via the crash-restart path
and immediately hit ``role 'qa' may not claim from status and immediately hit ``role 'qa' may not claim from status
'needs_revision'`` at the gateway. 'needs_revision'`` at the gateway. A coordination root (no code; product
fan-out owned by a PM) is the exception: it has no dev, so a CEO-rejected
one returns to its PM the dev-owned states also accept the PM roles for
it (a pure widening; nothing currently allowed is blocked).
""" """
role_mismatch: dict[str, str | set[str]] = { role_mismatch: dict[str, str | set[str]] = {
"awaiting_qa": "qa", "awaiting_qa": "qa",
@@ -2354,6 +2357,8 @@ class AgentOrchestrator:
required = role_mismatch.get(status) required = role_mismatch.get(status)
if required is None: if required is None:
return None return None
if is_coordination and status in ("needs_revision", "verifying"):
required = set(required) | {"cell_pm", "main_pm"}
ok = role in required if isinstance(required, set) else role == required ok = role in required if isinstance(required, set) else role == required
if ok: if ok:
return None return None
@@ -2379,7 +2384,9 @@ class AgentOrchestrator:
# the readiness and stuck-detection paths agree. # the readiness and stuck-detection paths agree.
if _branch_is_expected(task) and not task.get("branch_name"): if _branch_is_expected(task) and not task.get("branch_name"):
return f"state={status} but branch_name is unset" return f"state={status} but branch_name is unset"
return self._readiness_check_role_for_status(agent_id, role, status) return self._readiness_check_role_for_status(
agent_id, role, status, is_coordination=_is_coordination_task(task)
)
@staticmethod @staticmethod
async def _readiness_check_git_token(project_slug: str | None) -> str | None: async def _readiness_check_git_token(project_slug: str | None) -> str | None:
@@ -6217,6 +6224,10 @@ Start now: evidence(task_id="{task_id}")
dispatchers = [ dispatchers = [
("pm_work", self._dispatch_pm_work(client)), ("pm_work", self._dispatch_pm_work(client)),
("pm_closure_work", self._dispatch_pm_closure_work(client)), ("pm_closure_work", self._dispatch_pm_closure_work(client)),
(
"revision_coordination",
self._dispatch_revision_coordination_roots(client),
),
("dev_work", self._dispatch_dev_work(client)), ("dev_work", self._dispatch_dev_work(client)),
("qa_work", self._dispatch_qa_work(client)), ("qa_work", self._dispatch_qa_work(client)),
("pr_review_work", self._dispatch_pr_review_work(client)), ("pr_review_work", self._dispatch_pr_review_work(client)),
@@ -6723,6 +6734,37 @@ Start now: evidence(task_id="{task_id}")
await self._route_unassigned_pm_task(client, task) await self._route_unassigned_pm_task(client, task)
async def _dispatch_revision_coordination_roots(
self, client: httpx.AsyncClient
) -> None:
"""Re-spawn the owning PM for a CEO-rejected coordination root.
A coordination root (team=main_pm, product-linked, no repo) the CEO sends
back lands in ``needs_revision``. The dev dispatcher skips it (not a cell
team) and the closure path only handles paused parents, so without this
it would sit in needs_revision forever the deadlock. Respawn its PM so
it re-coordinates the revision. Cell/code needs_revision tasks are left
to the dev dispatcher; this handles only coordination roots.
"""
tasks = await self._fetch_tasks(client, "needs_revision")
for task in tasks:
if self._is_task_handled_this_tick(task.get("id")):
continue
if not _is_coordination_task(task):
continue
owner = task.get("assigned_to") or task.get("claimed_by")
agent_slug = self._resolve_agent_slug(owner) if owner else None
if not agent_slug or self._is_agent_active(agent_slug):
continue
if get_agent_role(agent_slug) not in ("cell_pm", "main_pm"):
continue
await self.spawn_agent(
agent_id=agent_slug,
task_id=task["id"],
initial_prompt=self._get_prompt_for_agent(agent_slug, task),
git_context=self._task_git_context(task),
)
@staticmethod @staticmethod
def _all_descendants_terminal(descendants: list[dict[str, Any]]) -> bool: def _all_descendants_terminal(descendants: list[dict[str, Any]]) -> bool:
"""Every descendant in a closure-complete state?""" """Every descendant in a closure-complete state?"""
@@ -73,3 +73,60 @@ def test_unmapped_status_allows_any_role() -> None:
) )
is None is None
), f"role={role} status={status} should not be rejected by this gate" ), f"role={role} status={status} should not be rejected by this gate"
# ---------------------------------------------------------------------------
# Coordination roots — a CEO-rejected coordination root returns to its PM (#5).
# ---------------------------------------------------------------------------
def test_pm_on_needs_revision_coordination_allowed() -> None:
"""A coordination root in needs_revision belongs to its PM, not a dev."""
for role in ("main_pm", "cell_pm"):
assert (
AgentOrchestrator._readiness_check_role_for_status(
agent_id="main-pm",
role=role,
status="needs_revision",
is_coordination=True,
)
is None
)
def test_pm_on_verifying_coordination_allowed() -> None:
assert (
AgentOrchestrator._readiness_check_role_for_status(
agent_id="main-pm", role="main_pm", status="verifying", is_coordination=True
)
is None
)
def test_dev_on_needs_revision_coordination_still_allowed() -> None:
"""Widening is additive — developer is still accepted."""
assert (
AgentOrchestrator._readiness_check_role_for_status(
agent_id="be-dev-1",
role="developer",
status="needs_revision",
is_coordination=True,
)
is None
)
def test_qa_on_needs_revision_coordination_still_blocked() -> None:
"""The widening only adds the PM roles — QA is still a misroute."""
reason = AgentOrchestrator._readiness_check_role_for_status(
agent_id="be-qa", role="qa", status="needs_revision", is_coordination=True
)
assert reason is not None
def test_pm_on_needs_revision_noncoordination_still_blocked() -> None:
"""A normal (code) needs_revision task is still dev/doc-only for a PM."""
reason = AgentOrchestrator._readiness_check_role_for_status(
agent_id="be-pm", role="cell_pm", status="needs_revision", is_coordination=False
)
assert reason is not None
@@ -0,0 +1,87 @@
"""_dispatch_revision_coordination_roots — un-deadlock a CEO-rejected root (#5).
A coordination root (team=main_pm, product-linked, no repo) the CEO sends back
lands in needs_revision. The dev dispatcher skips it (not a cell team) and the
closure path only handles paused parents, so without this dispatcher it sits
forever. This re-spawns its owning PM so it re-coordinates the revision.
"""
from __future__ import annotations
from typing import Any
from unittest.mock import AsyncMock, MagicMock
import pytest
import roboco.runtime.orchestrator as orch_mod
from roboco.runtime.orchestrator import AgentOrchestrator
def _orch(
tasks: list[dict[str, Any]], *, slug: str, active: bool
) -> tuple[AgentOrchestrator, AsyncMock]:
"""A bare orchestrator with its dispatch helpers mocked; returns (orch, spawn)."""
orch = object.__new__(AgentOrchestrator)
spawn = AsyncMock()
object.__setattr__(orch, "_fetch_tasks", AsyncMock(return_value=tasks))
object.__setattr__(
orch, "_is_task_handled_this_tick", MagicMock(return_value=False)
)
object.__setattr__(orch, "_resolve_agent_slug", MagicMock(return_value=slug))
object.__setattr__(orch, "_is_agent_active", MagicMock(return_value=active))
object.__setattr__(orch, "_get_prompt_for_agent", MagicMock(return_value="p"))
object.__setattr__(orch, "_task_git_context", MagicMock(return_value=None))
object.__setattr__(orch, "spawn_agent", spawn)
return orch, spawn
def _task() -> dict[str, Any]:
return {
"id": "t1",
"status": "needs_revision",
"assigned_to": "u1",
"team": "main_pm",
}
@pytest.mark.asyncio
async def test_respawns_pm_for_rejected_coordination_root(
monkeypatch: pytest.MonkeyPatch,
) -> None:
monkeypatch.setattr(orch_mod, "_is_coordination_task", lambda _t: True)
orch, spawn = _orch([_task()], slug="main-pm", active=False)
await orch._dispatch_revision_coordination_roots(MagicMock())
spawn.assert_awaited_once()
call = spawn.await_args
assert call is not None
assert call.kwargs["agent_id"] == "main-pm"
assert call.kwargs["task_id"] == "t1"
@pytest.mark.asyncio
async def test_skips_non_coordination_needs_revision(
monkeypatch: pytest.MonkeyPatch,
) -> None:
# A normal (code) needs_revision task → left to the dev dispatcher.
monkeypatch.setattr(orch_mod, "_is_coordination_task", lambda _t: False)
orch, spawn = _orch([_task()], slug="be-dev-1", active=False)
await orch._dispatch_revision_coordination_roots(MagicMock())
spawn.assert_not_awaited()
@pytest.mark.asyncio
async def test_skips_when_pm_already_active(
monkeypatch: pytest.MonkeyPatch,
) -> None:
monkeypatch.setattr(orch_mod, "_is_coordination_task", lambda _t: True)
orch, spawn = _orch([_task()], slug="main-pm", active=True)
await orch._dispatch_revision_coordination_roots(MagicMock())
spawn.assert_not_awaited()
@pytest.mark.asyncio
async def test_skips_non_pm_owner(monkeypatch: pytest.MonkeyPatch) -> None:
# A coordination root owned by a non-PM role → role guard skips it.
monkeypatch.setattr(orch_mod, "_is_coordination_task", lambda _t: True)
orch, spawn = _orch([_task()], slug="be-dev-1", active=False)
await orch._dispatch_revision_coordination_roots(MagicMock())
spawn.assert_not_awaited()