fix(agent99): route and dedupe alerts before dispatch

This commit is contained in:
ogt
2026-07-10 16:44:01 +08:00
parent ea98503a9a
commit 5c620844fe
7 changed files with 937 additions and 62 deletions

View File

@@ -0,0 +1,61 @@
from pathlib import Path
ROOT = Path(__file__).resolve().parents[3]
SRE_INBOX = ROOT / "agent99-sre-alert-inbox.ps1"
TELEGRAM_INBOX = ROOT / "agent99-telegram-inbox.ps1"
def test_alertmanager_groups_before_agent99_dispatch() -> None:
source = (ROOT / "apps/api/src/api/v1/webhooks.py").read_text(encoding="utf-8")
normalized_at = source.index('"alertmanager_normalized"')
grouping_at = source.index("get_alert_grouping_service().evaluate", normalized_at)
grouped_return_at = source.index(
"if grouping_result and grouping_result.is_grouped", grouping_at
)
dispatch_at = source.index("bridge_alertmanager_to_agent99", grouped_return_at)
assert grouping_at < grouped_return_at < dispatch_at
def test_agent99_bridge_has_fingerprint_single_flight_before_transport() -> None:
source = (ROOT / "apps/api/src/services/agent99_sre_bridge.py").read_text(
encoding="utf-8"
)
bridge_at = source.index("async def bridge_alertmanager_to_agent99")
lock_at = source.index("try_acquire_agent99_sre_single_flight", bridge_at)
dispatch_at = source.index("dispatch_agent99_sre_alert, payload", lock_at)
assert lock_at < dispatch_at
def test_agent99_runtime_prioritizes_structured_route_over_text() -> None:
source = SRE_INBOX.read_text(encoding="utf-8")
route_at = source.index("function Resolve-AgentAlertRoute")
suggested_at = source.index("if ($suggestedMode", route_at)
kind_at = source.index("if ($kind -and $kindModes.ContainsKey", suggested_at)
fallback_at = source.index('if ($text -match "502', kind_at)
assert suggested_at < kind_at < fallback_at
assert '"database_health" = "Status"' in source
assert '"alert_chain_health" = "Status"' in source
assert '"host_recovery" = "Recover"' in source
def test_agent99_runtime_has_local_correlation_dedupe_and_lock() -> None:
source = SRE_INBOX.read_text(encoding="utf-8")
assert '"sre-alert-single-flight.json"' in source
assert "[System.IO.FileShare]::None" in source
assert 'reason = "duplicate_single_flight"' in source
assert "severity_escalation_dispatch" in source
assert "correlationKey = $correlationKey" in source
def test_telegram_ingress_writes_same_structured_route_contract() -> None:
source = TELEGRAM_INBOX.read_text(encoding="utf-8")
assert "function Get-AgentTelegramSuggestedMode" in source
assert 'schemaVersion = "agent99_alert_route_v1"' in source
assert 'routeSource = "telegram_structured_kind"' in source
assert '"database_health" = "Status"' in source
assert '"alert_chain_health" = "Status"' in source

View File

@@ -2,9 +2,14 @@ from __future__ import annotations
import json
import pytest
from src.services.agent99_sre_bridge import (
agent99_sre_single_flight_key,
bridge_alertmanager_to_agent99,
build_agent99_sre_alert,
post_agent99_sre_alert,
release_agent99_sre_single_flight,
resolve_agent99_alert_kind,
write_agent99_sre_alert,
)
@@ -37,13 +42,99 @@ def test_provider_freshness_alert_builds_agent99_contract() -> None:
assert payload["source"] == "awoooi-api-alertmanager"
assert payload["kind"] == "provider_freshness_signal"
assert payload["suggestedMode"] == "ProviderFreshness"
assert payload["controlledApply"] is True
assert payload["controlledApply"] is False
assert payload["force"] is True
assert "provider-freshness" in payload["labels"]
assert "no-false-green" in payload["labels"]
assert "token=[redacted]" in payload["labels"]
assert "must-not-leak" not in json.dumps(payload)
assert "do not switch provider" in payload["instruction"]
assert payload["routing"]["routeSource"] == "structured_kind"
assert payload["routing"]["correlationKey"] == payload["routing"]["groupKey"]
@pytest.mark.parametrize(
("alertname", "target", "message", "labels", "expected_kind", "expected_mode"),
[
(
"HostBackupFailed",
"backup",
"snapshot stale",
{},
"backup_health",
"BackupCheck",
),
(
"provider_freshness_signal",
"provider_freshness_signal",
"last_seen missing",
{},
"provider_freshness_signal",
"ProviderFreshness",
),
(
"RebootAutoRecoveryActiveBlocker",
"reboot-auto-recovery-slo",
"all_required_hosts_not_in_10_minute_reboot_window cpu pressure",
{"service": "awoooi"},
"host_recovery",
"Recover",
),
(
"PostgresqlDown",
"postgresql",
"database unavailable deployment awoooi",
{"service": "awoooi"},
"database_health",
"Status",
),
(
"AlertChainBroken",
"alertmanager",
"monitoring pipeline down awoooi",
{"service": "awoooi"},
"alert_chain_health",
"Status",
),
(
"HostDown",
"192.168.0.120",
"host_down vm_not_running",
{"host": "192.168.0.120"},
"host_recovery",
"Recover",
),
(
"KubePodCrashLooping",
"awoooi-api",
"k3s pod crashloopbackoff",
{"service": "awoooi"},
"awoooi_k3s",
"AwoooRepair",
),
],
)
def test_agent99_routing_replay_corpus(
alertname: str,
target: str,
message: str,
labels: dict[str, str],
expected_kind: str,
expected_mode: str,
) -> None:
payload = build_agent99_sre_alert(
alert_id=f"replay-{alertname}",
alertname=alertname,
severity="critical",
namespace="awoooi-prod",
target_resource=target,
message=message,
labels=labels,
fingerprint=f"fp-{alertname}",
)
assert payload["kind"] == expected_kind
assert payload["suggestedMode"] == expected_mode
def test_alert_kind_mapping_for_route_and_performance() -> None:
@@ -108,7 +199,7 @@ def test_post_agent99_sre_alert_to_relay(monkeypatch) -> None:
def read(self, _limit: int) -> bytes:
return b'{"ok":true}'
def __enter__(self) -> "FakeResponse":
def __enter__(self) -> FakeResponse:
return self
def __exit__(self, *_args: object) -> None:
@@ -141,3 +232,155 @@ def test_post_agent99_sre_alert_to_relay(monkeypatch) -> None:
assert "authorization=[redacted]" in seen["payload"]["labels"]
assert "must-not-leak" not in json.dumps(seen["payload"])
assert seen["headers"]["X-agent99-relay-token"] == "relay-token"
def test_agent99_single_flight_key_does_not_expose_fingerprint() -> None:
key = agent99_sre_single_flight_key("private/source/fingerprint")
assert key.startswith("agent99:sre_dispatch:single_flight:")
assert "private" not in key
@pytest.mark.asyncio
async def test_agent99_bridge_suppresses_duplicate_single_flight(monkeypatch) -> None:
async def duplicate(*_args, **_kwargs) -> bool:
return False
dispatched: list[dict[str, object]] = []
monkeypatch.setattr(
"src.services.agent99_sre_bridge.try_acquire_agent99_sre_single_flight",
duplicate,
)
monkeypatch.setattr(
"src.services.agent99_sre_bridge.dispatch_agent99_sre_alert",
lambda payload: dispatched.append(payload) or "relay",
)
result = await bridge_alertmanager_to_agent99(
alert_id="duplicate-alert",
alertname="HostDown",
severity="critical",
namespace="node",
target_resource="192.168.0.120",
message="host_down",
fingerprint="same-fingerprint",
)
assert result["status"] == "suppressed"
assert result["reason"] == "single_flight_duplicate"
assert dispatched == []
@pytest.mark.asyncio
async def test_agent99_bridge_dispatches_single_flight_winner(monkeypatch) -> None:
async def winner(*_args, **_kwargs) -> bool:
return True
dispatched: list[dict[str, object]] = []
monkeypatch.setattr(
"src.services.agent99_sre_bridge.try_acquire_agent99_sre_single_flight",
winner,
)
monkeypatch.setattr(
"src.services.agent99_sre_bridge.dispatch_agent99_sre_alert",
lambda payload: dispatched.append(payload) or "relay",
)
result = await bridge_alertmanager_to_agent99(
alert_id="winner-alert",
alertname="HostDown",
severity="critical",
namespace="node",
target_resource="192.168.0.120",
message="host_down",
fingerprint="winner-fingerprint",
)
assert result["status"] == "dispatched"
assert result["dispatch"] == "relay"
assert result["suggestedMode"] == "Recover"
assert len(dispatched) == 1
@pytest.mark.asyncio
async def test_agent99_bridge_releases_lock_when_transport_fails(monkeypatch) -> None:
async def winner(*_args, **_kwargs) -> bool:
return True
released: list[tuple[str, str]] = []
async def release(fingerprint: str, alert_id: str) -> None:
released.append((fingerprint, alert_id))
def fail_transport(_payload: dict[str, object]) -> str:
raise RuntimeError("relay unavailable")
monkeypatch.setattr(
"src.services.agent99_sre_bridge.try_acquire_agent99_sre_single_flight",
winner,
)
monkeypatch.setattr(
"src.services.agent99_sre_bridge.release_agent99_sre_single_flight",
release,
)
monkeypatch.setattr(
"src.services.agent99_sre_bridge.dispatch_agent99_sre_alert",
fail_transport,
)
result = await bridge_alertmanager_to_agent99(
alert_id="transport-failure",
alertname="HostDown",
severity="critical",
namespace="node",
target_resource="192.168.0.120",
message="host_down",
fingerprint="transport-fingerprint",
)
assert result == {"status": "failed", "reason": "RuntimeError"}
assert released == [("transport-fingerprint", "transport-failure")]
@pytest.mark.asyncio
async def test_agent99_single_flight_uses_redis_nx(monkeypatch) -> None:
seen: dict[str, object] = {}
class FakeRedis:
async def set(self, key, value, *, ex, nx): # type: ignore[no-untyped-def]
seen.update({"key": key, "value": value, "ex": ex, "nx": nx})
return True
monkeypatch.setattr("src.core.redis_client.get_redis", lambda: FakeRedis())
from src.services.agent99_sre_bridge import try_acquire_agent99_sre_single_flight
acquired = await try_acquire_agent99_sre_single_flight(
"source-fingerprint", "alert-1"
)
assert acquired is True
assert str(seen["key"]).startswith("agent99:sre_dispatch:single_flight:")
assert seen["value"] == "alert-1"
assert seen["ex"] == 600
assert seen["nx"] is True
@pytest.mark.asyncio
async def test_agent99_single_flight_release_is_owner_checked(monkeypatch) -> None:
seen: dict[str, object] = {}
class FakeRedis:
async def eval(self, script, key_count, key, owner): # type: ignore[no-untyped-def]
seen.update(
{"script": script, "key_count": key_count, "key": key, "owner": owner}
)
return 1
monkeypatch.setattr("src.core.redis_client.get_redis", lambda: FakeRedis())
await release_agent99_sre_single_flight("source-fingerprint", "alert-1")
assert "redis.call('get'" in str(seen["script"])
assert seen["key_count"] == 1
assert seen["owner"] == "alert-1"

View File

@@ -2,9 +2,9 @@ from __future__ import annotations
from src.services.telegram_gateway import (
_agent99_alert_from_ai_automation_card,
format_aiops_signal_alert_card,
_outbound_source_envelope,
_sanitize_telegram_error,
format_aiops_signal_alert_card,
)
@@ -204,7 +204,8 @@ Authorization: Bearer abcdefghijklmnopqrstuvwxyz
)
assert payload["kind"] == "provider_freshness_signal"
assert payload["suggestedMode"] == "ProviderFreshness"
assert payload["controlledApply"] is True
assert payload["controlledApply"] is False
assert payload["routing"]["routeSource"] == "structured_kind"
assert payload["awoooi"]["alertCategory"] == "source_provider_freshness_triage"
assert payload["awoooi"]["notificationType"] == "ai_automation_alert_card_v1"
assert "component=telegram_gateway" in payload["labels"]