atlas-iac/testing/tests/test_hermes_execution_pool.py
Hermes Agent 5d963c9354 hermes: make pool lease recovery and release isolation safe
Independent review t_5975c06a blocked this branch on a P1: a Kanban write that
failed while a lease expired left a `lease_failed` row that was invisible to
every pass, immortal to garbage collection, and fatal to the coordinator. It
poisoned `reconcile()` forever with a conflicting-duplicate primary key,
produced a spurious capability `block_task` from `dispatch()`, and -- because
startup maintenance ran unguarded before the port bound, against a store on a
PVC -- crash-looped the coordinator with no automatic recovery.

`lease_failed` is now a retryable state that every maintenance pass drains, and
a row only reaches a terminal state on authoritative evidence about its exact
Kanban run, so nothing is collected before its outcome is known and nothing is
silently dropped. Each row, task, and board is processed in isolation, and a
coordinator-side fault is never converted into a Kanban mutation. Startup runs
through the same guarded cycle as the steady-state loop.

The wire protocol and the durable store are now separate modules, and the
maintenance passes moved out of the coordinator, so each file stays under the
managed line ceiling with room for the recovery logic.

Also closes three consequential handoff risks the same review raised:

* mediator-N pinned itself hard to worker-N while sharing a ReadWriteOnce
  claim, so a drain or preemption that moved only the lower-priority worker
  deadlocked the ordinal on Multi-Attach until an operator deleted a Pod. The
  shared workspace is now ReadWriteMany (as the hermes-chat tenant workspaces
  already are on the same class), colocation is a preference, and the mediator
  shares the worker's preemption priority, so each Pod reschedules on its own.
* the broker permits only branch creation, so a retry that added commits could
  never submit and the run's work was discarded with the failure. Submission
  now targets a fresh attempt- or content-scoped ref in the same reviewed
  namespace -- never an update -- and is idempotent under replay. A refused
  submission downgrades the result and says why instead of unwinding the run.
* the provider CLIs were reinstalled into an emptyDir on every Pod start inside
  the 10m Flux health window for the whole hermes app. They now install once
  per pinned version onto a durable volume, re-verified against the real
  binaries and time-bounded, and the best-effort pool no longer gates the
  health of the app its dependents wait on.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
2026-08-17 16:31:15 +00:00

486 lines
20 KiB
Python

"""Adversarial contracts for the fenced three-node Hermes execution pool."""
from __future__ import annotations
import json
import subprocess
import sys
import threading
import time
from pathlib import Path
import pytest
import yaml
ROOT = Path(__file__).parents[2]
HERMES = ROOT / "services/hermes"
SCRIPTS = HERMES / "scripts"
sys.path[:0] = [str(SCRIPTS), str(HERMES / "scm-common/scripts")]
import execution_pool_coordinator as coordinator # noqa: E402
import execution_pool_client as pool_client # noqa: E402
import execution_pool_protocol as protocol # noqa: E402
import execution_pool_scm as scm # noqa: E402
import execution_pool_worker as worker # noqa: E402
KEY = b"k" * 32
def binding(**values):
result = {
"board": "atlas",
"task_id": "t_deadbeef",
"run_id": "run-1234",
"worker_ordinal": 0,
"attempt": 1,
}
result.update(values)
return result
def assignment_payload(**values):
result = {
"context": "Implement the bounded objective.",
"assignee": "cli-auto",
"repo_url": "https://scm.bstein.dev/atlas/titan-iac.git",
"branch": "feature/hermes-safe-pool",
"base_branch": "main",
"max_runtime_seconds": 3600,
"deadline_unix": int(time.time()) + 3600,
}
result.update(values)
return result
def test_envelope_binds_run_ordinal_attempt_and_digest():
signed = protocol.sign_envelope(KEY, "heartbeat", binding(), {"note": "active"})
verified = protocol.verify_envelope(KEY, signed, expected_kind="heartbeat")
assert verified["payload_digest"] == protocol.payload_digest({"note": "active"})
for name, value in binding().items():
assert verified[name] == value
signed["attempt"] = 2
with pytest.raises(protocol.ProtocolError, match="authentication"):
protocol.verify_envelope(KEY, signed)
def test_empty_poll_acknowledgement_is_authenticated():
empty = {
"board": "", "task_id": "", "run_id": "",
"worker_ordinal": 2, "attempt": 0,
}
signed = protocol.sign_envelope(KEY, "ack", empty, {"assignment": None})
assert protocol.verify_envelope(KEY, signed, expected_kind="ack")["payload"] == {
"assignment": None
}
@pytest.mark.parametrize(
"mutation,error",
[
({"board": "../atlas"}, "invalid board"),
({"worker_ordinal": 3}, "outside the pool"),
({"expires_at": 1}, "validity window"),
],
)
def test_malformed_or_traversal_bindings_fail_closed(mutation, error):
signed = protocol.sign_envelope(KEY, "heartbeat", binding(), {"note": "active"})
signed.update(mutation)
unsigned = dict(signed)
unsigned.pop("signature")
import hashlib
import hmac
signed["signature"] = hmac.new(
KEY, protocol.canonical_json(unsigned), hashlib.sha256
).hexdigest()
with pytest.raises(protocol.ProtocolError, match=error):
protocol.verify_envelope(KEY, signed)
def test_oversized_and_malformed_payloads_are_rejected():
with pytest.raises(protocol.ProtocolError, match="wire limit"):
protocol.sign_envelope(
KEY, "result", binding(), {"output": "x" * protocol.MAX_WIRE_BYTES}
)
with pytest.raises(protocol.ProtocolError, match="malformed"):
protocol.parse_wire(b"{not-json")
with pytest.raises(protocol.ProtocolError, match="oversized"):
protocol.parse_wire(b"x" * (protocol.MAX_WIRE_BYTES + 1))
def test_key_requires_private_regular_file_and_rejects_symlink(tmp_path):
key = tmp_path / "key"
key.write_bytes(KEY)
key.chmod(0o644)
with pytest.raises(protocol.ProtocolError, match="private"):
protocol.read_key(key)
key.chmod(0o600)
assert protocol.read_key(key) == KEY
link = tmp_path / "link"
link.symlink_to(key)
with pytest.raises(protocol.ProtocolError, match="unavailable"):
protocol.read_key(link)
with pytest.raises(protocol.ProtocolError, match="unavailable"):
protocol.read_key(tmp_path / "missing")
def test_local_signing_boundary_rejects_foreign_or_unassigned_result(monkeypatch):
monkeypatch.setattr(pool_client, "ORDINAL", 0)
boundary = pool_client.ClientBoundary(KEY)
request = {"binding": binding(), "payload": {"note": "active"}}
with pytest.raises(protocol.ProtocolError, match="does not own"):
boundary.heartbeat(request)
boundary.current = binding()
request["binding"] = binding(worker_ordinal=1)
with pytest.raises(protocol.ProtocolError, match="does not own"):
boundary.heartbeat(request)
def test_simultaneous_claim_materialization_has_one_winner(tmp_path):
store = protocol.PoolStore(tmp_path / "pool.db")
barrier = threading.Barrier(8)
outcomes = []
def add(index):
barrier.wait()
try:
outcome = store.add(
binding(task_id=f"t_{index}", run_id=f"run-{index}"),
assignment_payload(),
)
except protocol.ProtocolError:
outcome = False
outcomes.append(outcome)
threads = [threading.Thread(target=add, args=(index,)) for index in range(8)]
for thread in threads:
thread.start()
for thread in threads:
thread.join()
assert outcomes.count(True) == 1
assert store.available_ordinals() == [1, 2]
def test_duplicate_assignment_is_idempotent_but_conflict_is_rejected(tmp_path):
store = protocol.PoolStore(tmp_path / "pool.db")
assert store.add(binding(), assignment_payload()) is True
assert store.add(binding(), assignment_payload()) is False
with pytest.raises(protocol.ProtocolError, match="conflicting"):
store.add(binding(), assignment_payload(context="different"))
with pytest.raises(protocol.ProtocolError, match="conflicting"):
store.add(binding(worker_ordinal=1), assignment_payload())
def test_duplicate_heartbeat_and_result_are_idempotent(tmp_path):
store = protocol.PoolStore(tmp_path / "pool.db")
store.add(binding(), assignment_payload())
store.offer(0)
heartbeat = protocol.sign_envelope(
KEY, "heartbeat", binding(), {"note": "active"}, delivery_id="delivery-1"
)
assert store.heartbeat(heartbeat) == (True, False)
assert store.heartbeat(heartbeat) == (True, True)
result = protocol.sign_envelope(
KEY, "result", binding(), {"structured": {"status": "completed"}}
)
_, duplicate = store.accept_result(result)
assert duplicate is False
_, duplicate = store.accept_result(result)
assert duplicate is True
def test_reused_delivery_and_stale_attempt_cannot_cross_runs(tmp_path):
store = protocol.PoolStore(tmp_path / "pool.db")
store.add(binding(), assignment_payload())
first = protocol.sign_envelope(
KEY, "heartbeat", binding(), {"note": "one"}, delivery_id="delivery-1"
)
store.heartbeat(first)
reused = protocol.sign_envelope(
KEY, "heartbeat", binding(), {"note": "two"}, delivery_id="delivery-1"
)
with pytest.raises(protocol.ProtocolError, match="reused"):
store.heartbeat(reused)
store.finalize(binding(), "finalized")
store.add(binding(run_id="other-run"), assignment_payload())
cross_run = protocol.sign_envelope(
KEY, "heartbeat", binding(run_id="other-run"), {"note": "one"},
delivery_id="delivery-1",
)
with pytest.raises(protocol.ProtocolError, match="reused"):
store.heartbeat(cross_run)
stale = protocol.sign_envelope(
KEY, "result", binding(attempt=2), {"structured": {"status": "completed"}}
)
with pytest.raises(protocol.ProtocolError, match="attempt is stale"):
store.accept_result(stale)
replacement = protocol.sign_envelope(
KEY,
"result",
binding(run_id="replacement-run"),
{"structured": {"status": "completed"}},
)
with pytest.raises(protocol.ProtocolError, match="unknown or stale"):
store.accept_result(replacement)
def test_restart_heartbeat_loss_and_node_replacement_reuse_durable_assignment(tmp_path):
database = tmp_path / "pool.db"
first = protocol.PoolStore(database)
first.add(binding(), assignment_payload())
offered = first.offer(0)
assert offered and offered["attempt"] == 1
# Reopening the SQLite ledger models coordinator restart. An expired lease
# is deliberately not reassigned across ordinals; the StatefulSet replaces
# ordinal 0 with the same PVC and provider session state.
restarted = protocol.PoolStore(database)
with restarted._connect() as connection:
connection.execute("UPDATE assignments SET lease_until=0")
resumed = restarted.offer(0)
assert resumed and resumed["run_id"] == "run-1234"
assert resumed["attempt"] == 1
assert restarted.offer(1) is None
assert len(restarted.active_assignments()) == 1
def test_result_conflict_never_overwrites_first_result(tmp_path):
store = protocol.PoolStore(tmp_path / "pool.db")
store.add(binding(), assignment_payload())
first = protocol.sign_envelope(KEY, "result", binding(), {"returncode": 0})
store.accept_result(first)
conflict = protocol.sign_envelope(KEY, "result", binding(), {"returncode": 1})
with pytest.raises(protocol.ProtocolError, match="conflicting result"):
store.accept_result(conflict)
def test_scm_workspace_is_ordinal_contained_and_rejects_symlink(tmp_path, monkeypatch):
monkeypatch.setattr(scm, "WORKSPACE_ROOT", tmp_path)
monkeypatch.setattr(scm, "ORDINAL", 0)
signed = protocol.sign_envelope(KEY, "assignment", binding(), assignment_payload())
path = scm.workspace_path(signed)
assert path == tmp_path / "runs/atlas/t_deadbeef/run-1234"
path.parent.mkdir(parents=True, exist_ok=True)
path.symlink_to(tmp_path / "elsewhere", target_is_directory=True)
with pytest.raises(protocol.ProtocolError, match="symlink"):
scm.workspace_path(signed)
root_case = tmp_path / "root-case"
root_case.mkdir()
root = tmp_path / "other-root"
root.mkdir()
(root_case / "runs").symlink_to(root, target_is_directory=True)
monkeypatch.setattr(scm, "WORKSPACE_ROOT", root_case)
with pytest.raises(protocol.ProtocolError, match="run root"):
scm.workspace_path(signed)
def test_scm_boundary_rejects_other_repo_branch_and_ordinal(monkeypatch):
monkeypatch.setattr(scm, "ORDINAL", 0)
for changes, message in (
({"repo_url": "https://evil.example/atlas/titan-iac.git"}, "outside Atlas"),
({"branch": "main"}, "reviewed namespace"),
):
signed = protocol.sign_envelope(
KEY, "assignment", binding(), assignment_payload(**changes)
)
with pytest.raises(protocol.ProtocolError, match=message):
scm._binding(signed)
signed = protocol.sign_envelope(
KEY, "assignment", binding(worker_ordinal=1), assignment_payload()
)
with pytest.raises(protocol.ProtocolError, match="ordinal"):
scm._binding(signed)
def test_scm_git_process_has_no_credential_or_ambient_configuration():
environment = scm._git_environment()
assert environment == {
"HOME": "/nonexistent",
"PATH": "/usr/bin:/bin",
"GIT_CONFIG_NOSYSTEM": "1",
"GIT_TERMINAL_PROMPT": "0",
}
assert scm._broker_repo("titan-iac").startswith("http://hermes-scm-broker.")
source = (SCRIPTS / "execution_pool_scm.py").read_text()
assert "GITEA_TOKEN" not in source and "GIT_ASKPASS" not in source
def test_activity_is_bounded_sanitized_and_nofollow(tmp_path):
text = coordinator.sanitize_activity(
'route=codex token=super-secret\nAuthorization: Bearer abcdefghijklmnop\n'
'"refresh_token":"sk-ant-oat01-abcdefghijklmnopqrstuvwxyz"'
)
assert "super-secret" not in text
assert "abcdefghijklmnop" not in text
assert "abcdefghijklmnopqrstuvwxyz" not in text
assert len(text.encode()) <= protocol.MAX_ACTIVITY_BYTES
class FakeKanban:
@staticmethod
def worker_log_path(_task, board):
assert board == "atlas"
return str(tmp_path / "worker.log")
envelope = {**binding(), "payload": {"activity": "visible worker activity\n"}}
coordinator._append_activity(FakeKanban, envelope)
assert (tmp_path / "worker.log").read_text() == "visible worker activity\n"
(tmp_path / "worker.log").unlink()
(tmp_path / "worker.log").symlink_to(tmp_path / "target")
with pytest.raises(protocol.ProtocolError, match="symlink"):
coordinator._append_activity(FakeKanban, envelope)
def _documents(path):
return [item for item in yaml.safe_load_all(path.read_text()) if item]
def test_three_node_statefulset_contract_and_cross_worker_isolation():
stateful = _documents(HERMES / "execution-worker-statefulset.yaml")[0]
pod = stateful["spec"]["template"]["spec"]
assert stateful["spec"]["replicas"] == 3
assert stateful["spec"]["podManagementPolicy"] == "Parallel"
assert pod["automountServiceAccountToken"] is False
claims = {
item["metadata"]["name"]: item
for item in stateful["spec"]["volumeClaimTemplates"]
}
assert set(claims) == {"workspace", "tools", "provider-access"}
# The workspace is co-mounted by this ordinal's mediator, so it must be
# ReadWriteMany: a ReadWriteOnce claim deadlocks on Multi-Attach when a drain
# or preemption separates the two Pods. Everything only the worker mounts
# stays single-mounter.
assert claims["workspace"]["spec"]["accessModes"] == ["ReadWriteMany"]
assert claims["tools"]["spec"]["accessModes"] == ["ReadWriteOnce"]
assert claims["provider-access"]["spec"]["accessModes"] == ["ReadWriteOnce"]
assert all(
item["spec"]["storageClassName"] == "astreae" for item in claims.values()
)
anti = pod["affinity"]["podAntiAffinity"]["requiredDuringSchedulingIgnoredDuringExecution"]
assert anti[0]["topologyKey"] == "kubernetes.io/hostname"
spread = pod["topologySpreadConstraints"][0]
assert spread["whenUnsatisfiable"] == "DoNotSchedule"
assert spread["maxSkew"] == 1
def test_worker_placement_prefers_accelerators_and_preserves_exclusions():
stateful = _documents(HERMES / "execution-worker-statefulset.yaml")[0]
pod = stateful["spec"]["template"]["spec"]
assert pod["priorityClassName"] == "scavenger"
affinity = stateful["spec"]["template"]["spec"]["affinity"]["nodeAffinity"]
terms = affinity["requiredDuringSchedulingIgnoredDuringExecution"]["nodeSelectorTerms"]
required = {item["key"]: item for item in terms[0]["matchExpressions"]}
assert required["node-role.kubernetes.io/worker"]["values"] == ["true"]
excluded = set(required["kubernetes.io/hostname"]["values"])
assert {"titan-04", "titan-13", "titan-14", "titan-17", "titan-18", "titan-19", "titan-22", "titan-24"} <= excluded
assert affinity["preferredDuringSchedulingIgnoredDuringExecution"][0]["weight"] == 100
assert affinity["preferredDuringSchedulingIgnoredDuringExecution"][1]["weight"] == 50
worker_container = pod["containers"][0]
assert worker_container["resources"] == {
"requests": {"cpu": "5m", "memory": "128Mi", "ephemeral-storage": "1Gi"},
"limits": {"cpu": "2", "memory": "4Gi", "ephemeral-storage": "8Gi"},
}
mediators = [
item
for item in _documents(HERMES / "execution-mediator.yaml")
if item["kind"] == "Deployment"
]
assert all(
item["spec"]["template"]["spec"]["containers"][0]["resources"][
"requests"
]
== {"cpu": "2m", "memory": "64Mi"}
for item in mediators
)
def test_model_worker_has_no_scm_or_cluster_credential_mount():
stateful = _documents(HERMES / "execution-worker-statefulset.yaml")[0]
pod = stateful["spec"]["template"]
containers = {item["name"]: item for item in pod["spec"]["containers"]}
worker = containers["execution-worker"]
environment = {item["name"] for item in worker["env"]}
assert not {"GITEA_TOKEN", "OPENAI_API_KEY", "ANTHROPIC_API_KEY", "KUBECONFIG"} & environment
mounts = {item["mountPath"] for item in worker["volumeMounts"]}
assert "/vault/secrets" not in mounts
assert "/pool-access" not in mounts
assert "/provider-access" in mounts
assert not any(item["name"] == "vault-auth-token" for item in worker["volumeMounts"])
assert pod["metadata"]["annotations"]["vault.hashicorp.com/agent-inject-containers"] == "stage-worker-access"
assert set(containers) == {"execution-worker"}
assert "execution-pool-key" not in str(pod["metadata"]["annotations"])
mediators = [
item for item in _documents(HERMES / "execution-mediator.yaml")
if item["kind"] == "Deployment"
]
assert len(mediators) == 3
for mediator in mediators:
spec = mediator["spec"]["template"]["spec"]
assert spec["automountServiceAccountToken"] is False
privileged = spec["containers"][0]
assert {mount["mountPath"] for mount in privileged["volumeMounts"]} >= {
"/pool-access", "/scm-state", "/opt/scm", "/workspace"
}
assert "/provider-access" not in {
mount["mountPath"] for mount in privileged["volumeMounts"]
}
def test_worker_service_account_has_no_kubernetes_permissions():
documents = _documents(HERMES / "execution-worker-rbac.yaml")
assert len(documents) == 1
account = documents[0]
assert account["kind"] == "ServiceAccount"
assert account["automountServiceAccountToken"] is False
def test_coordinator_remains_single_state_owner_and_workers_do_not_mount_home():
agent = _documents(HERMES / "agent-deployment.yaml")[0]
assert agent["kind"] == "Deployment"
assert agent["spec"]["replicas"] == 1
assert agent["spec"]["strategy"]["type"] == "Recreate"
assert any(
item["name"] == "home"
and item["persistentVolumeClaim"]["claimName"] == "hermes-agent-home"
for item in agent["spec"]["template"]["spec"]["volumes"]
)
worker_text = (HERMES / "execution-worker-statefulset.yaml").read_text()
assert "hermes-agent-home" not in worker_text
assert "kanban.db" not in worker_text
def test_worker_protocol_preserves_switchyard_fallback_and_visible_evidence():
source = (SCRIPTS / "execution_pool_worker.py").read_text()
coordinator_source = (SCRIPTS / "execution_pool_coordinator.py").read_text()
server_source = (SCRIPTS / "execution_pool_server.py").read_text()
assert "cli_lane_runner.select_route" in source
assert 'alternate = "claude" if route.provider == "codex" else "codex"' in source
assert "codex_thread_id" in source and "claude_session_id" in source
assert "final_activity" in source
assert "worker_ordinal" in coordinator_source and "provider_sessions" in coordinator_source
assert "coordinator.reconcile" in server_source
def test_retention_gc_removes_only_clean_terminal_workspace(tmp_path, monkeypatch):
monkeypatch.setattr(worker, "ROOT", tmp_path)
monkeypatch.setattr(worker, "RETENTION_SECONDS", 3600)
workspace = tmp_path / "runs/atlas/t_deadbeef/run-1234"
workspace.mkdir(parents=True)
subprocess.run(["git", "init", "-q", str(workspace)], check=True)
subprocess.run(["git", "-C", str(workspace), "config", "user.email", "test@example.com"], check=True)
subprocess.run(["git", "-C", str(workspace), "config", "user.name", "Test"], check=True)
(workspace / "tracked").write_text("safe\n")
subprocess.run(["git", "-C", str(workspace), "add", "tracked"], check=True)
subprocess.run(["git", "-C", str(workspace), "commit", "-qm", "initial"], check=True)
state_file = tmp_path / "session-state/atlas/t_deadbeef/run-1234.json"
state_file.parent.mkdir(parents=True)
state_file.write_text(
json.dumps({"terminal_at": time.time() - 7200, "workspace": str(workspace)})
)
assert worker.garbage_collect() == 1
assert not workspace.exists()
assert not state_file.exists()