ForcePilot/docker/sandbox_provisioner/app.py

986 lines
38 KiB
Python
Raw Normal View History

from __future__ import annotations
2026-03-05 22:50:38 +08:00
import logging
import os
import threading
2026-03-05 22:50:38 +08:00
import time
from contextlib import asynccontextmanager
from dataclasses import dataclass
2026-03-05 22:50:38 +08:00
from pathlib import Path
from urllib import request
from fastapi import FastAPI, HTTPException
from pydantic import BaseModel
2026-03-05 22:50:38 +08:00
logger = logging.getLogger(__name__)
class CreateSandboxRequest(BaseModel):
sandbox_id: str
thread_id: str
class SandboxResponse(BaseModel):
sandbox_id: str
sandbox_url: str
status: str | None = None
class DeleteSandboxResponse(BaseModel):
ok: bool
sandbox_id: str
2026-03-05 22:50:38 +08:00
class TouchSandboxResponse(BaseModel):
ok: bool
sandbox_id: str
status: str | None = None
class ListSandboxesResponse(BaseModel):
sandboxes: list[SandboxResponse]
count: int
@dataclass(slots=True)
class SandboxRecord:
sandbox_id: str
sandbox_url: str
status: str | None = None
class MemoryProvisionerBackend:
def __init__(self):
self._lock = threading.Lock()
self._records: dict[str, SandboxRecord] = {}
self._url_template = os.getenv("MEMORY_SANDBOX_URL_TEMPLATE", "http://agent-sandbox:8000")
def _url_for(self, sandbox_id: str) -> str:
template = self._url_template
if "{sandbox_id}" in template:
return template.format(sandbox_id=sandbox_id)
return template
def create(self, sandbox_id: str, thread_id: str) -> SandboxRecord:
del thread_id
with self._lock:
existing = self._records.get(sandbox_id)
if existing is not None:
return existing
record = SandboxRecord(
sandbox_id=sandbox_id,
sandbox_url=self._url_for(sandbox_id),
status="Running",
)
self._records[sandbox_id] = record
return record
def discover(self, sandbox_id: str) -> SandboxRecord | None:
with self._lock:
return self._records.get(sandbox_id)
def list(self) -> list[SandboxRecord]:
with self._lock:
return list(self._records.values())
def delete(self, sandbox_id: str) -> None:
with self._lock:
self._records.pop(sandbox_id, None)
2026-03-05 22:50:38 +08:00
def wait_for_sandbox_ready(sandbox_url: str, timeout_seconds: int = 30) -> bool:
deadline = time.time() + timeout_seconds
opener = request.build_opener(request.ProxyHandler({}))
while time.time() < deadline:
try:
with opener.open(f"{sandbox_url.rstrip('/')}/v1/sandbox", timeout=3) as response:
status_code = getattr(response, "status", 200)
if status_code == 200:
return True
except Exception:
pass
time.sleep(1)
return False
class LocalContainerProvisionerBackend:
def __init__(self):
import docker
from docker.errors import DockerException
self._docker = docker
self._lock = threading.Lock()
self._container_port = int(os.getenv("SANDBOX_CONTAINER_PORT", "8080"))
self._sandbox_image = os.getenv(
"SANDBOX_IMAGE",
"enterprise-public-cn-beijing.cr.volces.com/vefaas-public/all-in-one-sandbox:latest",
)
self._network = os.getenv("DOCKER_NETWORK")
self._threads_host_path = os.getenv("DOCKER_THREADS_HOST_PATH")
self._skills_host_path = os.getenv("DOCKER_SKILLS_HOST_PATH")
self._container_prefix = os.getenv("DOCKER_SANDBOX_PREFIX", "yuxi-sandbox")
self._sandbox_host = os.getenv("DOCKER_SANDBOX_HOST", "host.docker.internal")
self._health_timeout_seconds = int(os.getenv("SANDBOX_HEALTH_TIMEOUT_SECONDS", "30"))
try:
self._client = docker.from_env()
self._client.ping()
except DockerException as exc:
raise RuntimeError(f"docker backend unavailable: {exc}") from exc
self._resolve_host_paths()
@staticmethod
def _validate_thread_id(thread_id: str) -> str:
candidate = str(thread_id or "").strip()
if not candidate:
raise ValueError("thread_id is required")
if any(ch in candidate for ch in ("/", "\\", "\x00")):
raise ValueError("thread_id must be a single safe path segment")
if candidate in {".", ".."} or ".." in candidate:
raise ValueError("thread_id contains invalid path traversal sequence")
return candidate
@staticmethod
def _sanitize_id(value: str) -> str:
sanitized = "".join(ch if ch.isalnum() or ch in "-_" else "-" for ch in value.strip().lower())
return sanitized[:48] or "sandbox"
def _container_name(self, sandbox_id: str) -> str:
return f"{self._container_prefix}-{self._sanitize_id(sandbox_id)}"
def _resolve_host_paths(self) -> None:
if self._threads_host_path and self._skills_host_path:
return
container_id = os.getenv("HOSTNAME", "").strip()
if not container_id:
raise RuntimeError("HOSTNAME is required to infer docker backend host paths")
inspected = self._client.api.inspect_container(container_id)
mounts = inspected.get("Mounts") or []
saves_source = None
for mount in mounts:
destination = (mount.get("Destination") or "").rstrip("/")
if destination == "/app/saves":
saves_source = mount.get("Source")
break
if not saves_source:
raise RuntimeError("cannot infer host path for /app/saves mount")
base = Path(saves_source)
if not self._threads_host_path:
self._threads_host_path = str(base / "threads")
if not self._skills_host_path:
self._skills_host_path = str(base / "skills")
def _host_port_for(self, container) -> int | None:
ports = (container.attrs.get("NetworkSettings") or {}).get("Ports") or {}
bindings = ports.get(f"{self._container_port}/tcp")
if not bindings:
return None
host_port = bindings[0].get("HostPort")
if not host_port:
return None
return int(host_port)
def _sandbox_url(self, host_port: int) -> str:
return f"http://{self._sandbox_host}:{host_port}"
def _to_record(self, container, sandbox_id: str) -> SandboxRecord:
state = (container.attrs.get("State") or {}).get("Status")
host_port = self._host_port_for(container)
sandbox_url = self._sandbox_url(host_port) if host_port is not None else ""
return SandboxRecord(
sandbox_id=sandbox_id,
sandbox_url=sandbox_url,
status=state or "unknown",
)
@staticmethod
def _ensure_user_data_writable(container) -> None:
cmd = (
"sh -lc "
'"mkdir -p /home/gem/user-data/workspace /home/gem/user-data/uploads /home/gem/user-data/outputs '
'&& chmod -R a+rwX /home/gem/user-data"'
2026-03-05 22:50:38 +08:00
)
result = container.exec_run(cmd, user="0:0")
if result.exit_code != 0:
output = (
result.output.decode("utf-8", errors="ignore")
if isinstance(result.output, bytes)
else str(result.output)
)
2026-03-05 22:50:38 +08:00
raise RuntimeError(f"failed to ensure writable thread user-data mount: {output}")
def _get_container(self, sandbox_id: str):
from docker.errors import NotFound
name = self._container_name(sandbox_id)
try:
return self._client.containers.get(name)
except NotFound:
return None
def create(self, sandbox_id: str, thread_id: str) -> SandboxRecord:
with self._lock:
safe_thread_id = self._validate_thread_id(thread_id)
existing = self._get_container(sandbox_id)
if existing is not None:
if existing.status != "running":
existing.start()
existing.reload()
self._ensure_user_data_writable(existing)
record = self._to_record(existing, sandbox_id)
if not record.sandbox_url:
raise RuntimeError(f"sandbox {sandbox_id} has no mapped host port")
if not wait_for_sandbox_ready(record.sandbox_url, timeout_seconds=self._health_timeout_seconds):
raise RuntimeError(f"sandbox {sandbox_id} is not ready at {record.sandbox_url}")
return record
2026-03-14 15:58:35 +08:00
# 检测是否是 Windows 绝对路径 (如 D:/ 或 D:\)
threads_root_str = self._threads_host_path
is_windows_path = len(threads_root_str) >= 2 and threads_root_str[1] == ":"
2026-03-14 15:58:35 +08:00
if is_windows_path:
# Windows 路径,直接使用,不调用 resolve()
threads_root = Path(threads_root_str)
thread_user_data = threads_root / safe_thread_id / "user-data"
# Windows 路径下无法在 Linux 容器内创建目录,跳过 mkdir
else:
threads_root = Path(threads_root_str).resolve()
thread_user_data = (threads_root / safe_thread_id / "user-data").resolve()
try:
thread_user_data.relative_to(threads_root)
except ValueError as exc:
raise ValueError("thread_id resolved outside threads host root") from exc
thread_user_data.mkdir(parents=True, exist_ok=True)
skills_path_str = self._skills_host_path
is_skills_windows = len(skills_path_str) >= 2 and skills_path_str[1] == ":"
2026-03-14 15:58:35 +08:00
if is_skills_windows:
skills_path = Path(skills_path_str)
else:
skills_path = Path(skills_path_str)
skills_path.mkdir(parents=True, exist_ok=True)
2026-03-05 22:50:38 +08:00
container_name = self._container_name(sandbox_id)
run_kwargs = {
"name": container_name,
"detach": True,
"labels": {
"app": "yuxi-sandbox",
"sandbox-id": sandbox_id,
"thread-id": thread_id,
"managed-by": "yuxi-sandbox-provisioner",
},
"volumes": {
str(thread_user_data): {"bind": "/home/gem/user-data", "mode": "rw"},
str(skills_path): {"bind": "/skills", "mode": "ro"},
2026-03-05 22:50:38 +08:00
},
"tmpfs": "/tmp:size=256m,mode=1777",
"environment": {
"HOME": "/home/gem/user-data",
"TMPDIR": "/tmp",
},
2026-03-05 22:50:38 +08:00
"ports": {f"{self._container_port}/tcp": None},
"security_opt": ["seccomp=unconfined"],
}
if self._network:
run_kwargs["network"] = self._network
container = self._client.containers.run(self._sandbox_image, **run_kwargs)
container.reload()
self._ensure_user_data_writable(container)
record = self._to_record(container, sandbox_id)
if not record.sandbox_url:
raise RuntimeError(f"sandbox {sandbox_id} has no mapped host port")
if not wait_for_sandbox_ready(record.sandbox_url, timeout_seconds=self._health_timeout_seconds):
raise RuntimeError(f"sandbox {sandbox_id} is not ready at {record.sandbox_url}")
return record
def discover(self, sandbox_id: str) -> SandboxRecord | None:
container = self._get_container(sandbox_id)
if container is None:
return None
container.reload()
record = self._to_record(container, sandbox_id)
if not record.sandbox_url:
return None
if not wait_for_sandbox_ready(record.sandbox_url, timeout_seconds=5):
return None
return record
def list(self) -> list[SandboxRecord]:
containers = self._client.containers.list(
all=True, filters={"label": ["app=yuxi-sandbox", "managed-by=yuxi-sandbox-provisioner"]}
)
records: list[SandboxRecord] = []
for container in containers:
labels = container.labels or {}
sandbox_id = labels.get("sandbox-id")
if sandbox_id:
container.reload()
records.append(self._to_record(container, sandbox_id))
return records
def delete(self, sandbox_id: str) -> None:
container = self._get_container(sandbox_id)
if container is None:
return
if container.status == "running":
container.stop(timeout=10)
container.remove(v=True, force=True)
class KubernetesProvisionerBackend:
def __init__(self):
from kubernetes import client, config
self._lock = threading.Lock()
self._namespace = os.getenv("K8S_NAMESPACE", "yuxi-know")
2026-03-05 22:50:38 +08:00
self._sandbox_image = os.getenv(
"SANDBOX_IMAGE",
"enterprise-public-cn-beijing.cr.volces.com/vefaas-public/all-in-one-sandbox:latest",
)
self._container_port = int(os.getenv("SANDBOX_CONTAINER_PORT", "8080"))
self._shared_userdata_pvc = os.getenv("K8S_SHARED_USERDATA_PVC", "shared-userdata-pvc")
self._shared_skills_pvc = os.getenv("K8S_SHARED_SKILLS_PVC", "shared-skills-pvc")
self._allow_shared_pvc_reuse = os.getenv("K8S_ALLOW_SHARED_PVC_REUSE", "false").strip().lower() == "true"
self._ready_timeout_seconds = int(os.getenv("SANDBOX_READY_TIMEOUT_SECONDS", "60"))
self._ready_poll_interval_seconds = max(1, int(os.getenv("SANDBOX_READY_POLL_INTERVAL_SECONDS", "2")))
kubeconfig_path = os.getenv("KUBECONFIG_PATH")
if kubeconfig_path:
config.load_kube_config(config_file=kubeconfig_path)
else:
try:
config.load_incluster_config()
except Exception:
config.load_kube_config()
self._core_api = client.CoreV1Api()
self._client = client
@staticmethod
def _validate_thread_id(thread_id: str) -> str:
candidate = str(thread_id or "").strip()
if not candidate:
raise ValueError("thread_id is required")
if any(ch in candidate for ch in ("/", "\\", "\x00")):
raise ValueError("thread_id must be a single safe path segment")
if candidate in {".", ".."} or ".." in candidate:
raise ValueError("thread_id contains invalid path traversal sequence")
return candidate
@staticmethod
def _sanitize_k8s_name(value: str) -> str:
sanitized = "".join(ch if ch.isalnum() else "-" for ch in value.strip().lower())
while "--" in sanitized:
sanitized = sanitized.replace("--", "-")
sanitized = sanitized.strip("-")
return sanitized[:40] or "sandbox"
def _thread_subpath(self, thread_id: str) -> str:
safe_thread_id = self._validate_thread_id(thread_id)
return f"threads/{safe_thread_id}/user-data"
@staticmethod
def _pod_name(sandbox_id: str) -> str:
safe_id = KubernetesProvisionerBackend._sanitize_k8s_name(sandbox_id)
return f"sandbox-{safe_id}"
@staticmethod
def _service_name(sandbox_id: str) -> str:
safe_id = KubernetesProvisionerBackend._sanitize_k8s_name(sandbox_id)
return f"sandbox-{safe_id}"
def _sandbox_url(self, sandbox_id: str) -> str:
service_name = self._service_name(sandbox_id)
return f"http://{service_name}.{self._namespace}.svc.cluster.local:{self._container_port}"
def _read_pvc(self, pvc_name: str):
from kubernetes.client.rest import ApiException
try:
return self._core_api.read_namespaced_persistent_volume_claim(
name=pvc_name,
namespace=self._namespace,
)
except ApiException as exc:
if exc.status == 404:
return None
raise
def _read_pod(self, sandbox_id: str):
from kubernetes.client.rest import ApiException
pod_name = self._pod_name(sandbox_id)
try:
return self._core_api.read_namespaced_pod(name=pod_name, namespace=self._namespace)
except ApiException as exc:
if exc.status == 404:
return None
raise
def _read_service(self, sandbox_id: str):
from kubernetes.client.rest import ApiException
service_name = self._service_name(sandbox_id)
try:
return self._core_api.read_namespaced_service(name=service_name, namespace=self._namespace)
except ApiException as exc:
if exc.status == 404:
return None
raise
def _ensure_shared_pvcs_ready(self) -> None:
for pvc_name in {self._shared_userdata_pvc, self._shared_skills_pvc}:
pvc = self._read_pvc(pvc_name)
if pvc is None:
raise RuntimeError(f"required pvc not found: {pvc_name}")
phase = pvc.status.phase if pvc and pvc.status else None
if phase != "Bound":
raise RuntimeError(f"required pvc is not Bound: {pvc_name} phase={phase}")
def _validate_pvc_configuration(self) -> None:
same_pvc = self._shared_userdata_pvc == self._shared_skills_pvc
if same_pvc and not self._allow_shared_pvc_reuse:
raise RuntimeError(
"K8S_SHARED_USERDATA_PVC and K8S_SHARED_SKILLS_PVC point to the same PVC; "
"set K8S_ALLOW_SHARED_PVC_REUSE=true only if your storage supports this reliably"
)
def _assert_existing_sandbox_matches(self, sandbox_id: str, thread_id: str) -> None:
pod = self._read_pod(sandbox_id)
if pod is None:
return
annotations = (pod.metadata.annotations if pod.metadata else None) or {}
existing_thread_id = annotations.get("thread-id")
if existing_thread_id and existing_thread_id != thread_id:
raise RuntimeError(
f"sandbox {sandbox_id} already exists for thread_id={existing_thread_id}, not {thread_id}"
)
def _pod_ready(self, pod) -> bool:
if not pod or not pod.status or pod.status.phase != "Running":
return False
conditions = pod.status.conditions or []
ready_condition = next((c for c in conditions if c.type == "Ready"), None)
if not ready_condition or ready_condition.status != "True":
return False
container_statuses = pod.status.container_statuses or []
sandbox_container = next((c for c in container_statuses if c.name == "sandbox"), None)
return bool(sandbox_container and sandbox_container.ready is True)
def _extract_pod_failure(self, pod) -> str | None:
if not pod or not pod.status:
return None
if pod.status.phase == "Failed":
reason = pod.status.reason or "Unknown"
message = pod.status.message or ""
return f"pod failed: {reason} {message}".strip()
for cond in pod.status.conditions or []:
if cond.type == "PodScheduled" and cond.status == "False":
return f"pod scheduling failed: {(cond.reason or '').strip()} {(cond.message or '').strip()}".strip()
for container_status in pod.status.container_statuses or []:
state = container_status.state
if state and state.waiting:
reason = state.waiting.reason or "Waiting"
message = state.waiting.message or ""
if reason in {
"ImagePullBackOff",
"ErrImagePull",
"CrashLoopBackOff",
"CreateContainerConfigError",
"CreateContainerError",
}:
return f"container {container_status.name} waiting: {reason} {message}".strip()
if state and state.terminated:
reason = state.terminated.reason or "Terminated"
exit_code = state.terminated.exit_code
message = state.terminated.message or ""
return f"container {container_status.name} terminated: {reason} exit={exit_code} {message}".strip()
return None
def _build_pod_spec(self, sandbox_id: str, thread_id: str):
pod_name = self._pod_name(sandbox_id)
safe_thread_id = self._validate_thread_id(thread_id)
user_data_subpath = self._thread_subpath(safe_thread_id)
same_pvc = self._shared_userdata_pvc == self._shared_skills_pvc
if same_pvc:
volumes = [
self._client.V1Volume(
name="shared-storage",
persistent_volume_claim=self._client.V1PersistentVolumeClaimVolumeSource(
claim_name=self._shared_userdata_pvc,
),
)
]
volume_mounts = [
self._client.V1VolumeMount(
name="shared-storage",
mount_path="/home/gem/user-data",
sub_path=user_data_subpath,
),
self._client.V1VolumeMount(
name="shared-storage",
mount_path="/skills",
sub_path="skills",
read_only=True,
),
]
else:
volumes = [
self._client.V1Volume(
name="user-data",
persistent_volume_claim=self._client.V1PersistentVolumeClaimVolumeSource(
claim_name=self._shared_userdata_pvc,
),
),
self._client.V1Volume(
name="skills",
persistent_volume_claim=self._client.V1PersistentVolumeClaimVolumeSource(
claim_name=self._shared_skills_pvc,
read_only=True,
),
),
]
volume_mounts = [
self._client.V1VolumeMount(
name="user-data",
mount_path="/home/gem/user-data",
sub_path=user_data_subpath,
),
self._client.V1VolumeMount(
name="skills",
mount_path="/skills",
sub_path="skills",
read_only=True,
),
]
volume_mounts.append(self._client.V1VolumeMount(name="tmp", mount_path="/tmp"))
volumes.append(
self._client.V1Volume(
name="tmp",
empty_dir=self._client.V1EmptyDirVolumeSource(medium="Memory"),
)
)
init_script = f'set -eu; mkdir -p "/home/gem/user-data/threads/{safe_thread_id}/user-data"; test -d "/skills"'
init_volume_mounts = [
self._client.V1VolumeMount(name="tmp", mount_path="/tmp"),
]
if same_pvc:
init_volume_mounts.extend(
[
self._client.V1VolumeMount(name="shared-storage", mount_path="/home/gem/user-data"),
self._client.V1VolumeMount(name="shared-storage", mount_path="/skills", sub_path="skills"),
]
)
else:
init_volume_mounts.extend(
[
self._client.V1VolumeMount(name="user-data", mount_path="/home/gem/user-data"),
self._client.V1VolumeMount(name="skills", mount_path="/skills", sub_path="skills", read_only=True),
]
)
return self._client.V1Pod(
metadata=self._client.V1ObjectMeta(
name=pod_name,
labels={
"app": "yuxi-sandbox",
"sandbox-id": sandbox_id,
"managed-by": "yuxi-sandbox-provisioner",
},
annotations={"thread-id": safe_thread_id},
),
spec=self._client.V1PodSpec(
restart_policy="Never",
security_context=self._client.V1PodSecurityContext(fs_group=1000),
init_containers=[
self._client.V1Container(
name="prepare-user-data",
image=self._sandbox_image,
command=["sh", "-lc", init_script],
volume_mounts=init_volume_mounts,
)
],
containers=[
self._client.V1Container(
name="sandbox",
image=self._sandbox_image,
ports=[self._client.V1ContainerPort(container_port=self._container_port)],
env=[
self._client.V1EnvVar(name="HOME", value="/home/gem/user-data"),
self._client.V1EnvVar(name="TMPDIR", value="/tmp"),
],
volume_mounts=volume_mounts,
readiness_probe=self._client.V1Probe(
http_get=self._client.V1HTTPGetAction(
path="/v1/sandbox",
port=self._container_port,
),
period_seconds=2,
timeout_seconds=2,
failure_threshold=15,
),
startup_probe=self._client.V1Probe(
http_get=self._client.V1HTTPGetAction(
path="/v1/sandbox",
port=self._container_port,
),
period_seconds=2,
timeout_seconds=2,
failure_threshold=30,
),
)
],
volumes=volumes,
),
)
def _build_service_spec(self, sandbox_id: str):
service_name = self._service_name(sandbox_id)
return self._client.V1Service(
metadata=self._client.V1ObjectMeta(
name=service_name,
labels={
"app": "yuxi-sandbox",
"sandbox-id": sandbox_id,
"managed-by": "yuxi-sandbox-provisioner",
},
),
spec=self._client.V1ServiceSpec(
type="ClusterIP",
selector={"sandbox-id": sandbox_id},
ports=[
self._client.V1ServicePort(
name="http",
port=self._container_port,
target_port=self._container_port,
protocol="TCP",
)
],
),
)
def wait_ready(self, sandbox_id: str) -> SandboxRecord:
deadline = time.monotonic() + self._ready_timeout_seconds
sandbox_url = self._sandbox_url(sandbox_id)
last_failure = None
while time.monotonic() < deadline:
pod = self._read_pod(sandbox_id)
service = self._read_service(sandbox_id)
if pod is None or service is None:
time.sleep(self._ready_poll_interval_seconds)
continue
failure = self._extract_pod_failure(pod)
if failure:
raise RuntimeError(f"sandbox {sandbox_id} failed before ready: {failure}")
if self._pod_ready(pod):
if wait_for_sandbox_ready(sandbox_url, timeout_seconds=3):
return SandboxRecord(
sandbox_id=sandbox_id,
sandbox_url=sandbox_url,
status="Running",
)
last_failure = f"http health check not ready at {sandbox_url}"
time.sleep(self._ready_poll_interval_seconds)
raise RuntimeError(f"sandbox {sandbox_id} timed out waiting ready: {last_failure or 'unknown'}")
def create(self, sandbox_id: str, thread_id: str) -> SandboxRecord:
from kubernetes.client.rest import ApiException
with self._lock:
safe_thread_id = self._validate_thread_id(thread_id)
self._validate_pvc_configuration()
self._ensure_shared_pvcs_ready()
existing_pod = self._read_pod(sandbox_id)
existing_service = self._read_service(sandbox_id)
if existing_pod is not None:
self._assert_existing_sandbox_matches(sandbox_id, safe_thread_id)
if existing_pod is not None and existing_service is not None:
return self.wait_ready(sandbox_id)
try:
self._core_api.create_namespaced_pod(
namespace=self._namespace,
body=self._build_pod_spec(sandbox_id, safe_thread_id),
)
except ApiException as exc:
if exc.status != 409:
raise
try:
self._core_api.create_namespaced_service(
namespace=self._namespace,
body=self._build_service_spec(sandbox_id),
)
except ApiException as exc:
if exc.status != 409:
raise
return self.wait_ready(sandbox_id)
def discover(self, sandbox_id: str) -> SandboxRecord | None:
pod = self._read_pod(sandbox_id)
service = self._read_service(sandbox_id)
if pod is None or service is None:
return None
failure = self._extract_pod_failure(pod)
if failure:
status = f"Failed:{failure}"
elif self._pod_ready(pod):
status = "Running"
else:
phase = pod.status.phase if pod and pod.status else "Unknown"
status = f"NotReady:{phase}"
return SandboxRecord(
sandbox_id=sandbox_id,
sandbox_url=self._sandbox_url(sandbox_id),
status=status,
)
def list(self) -> list[SandboxRecord]:
from kubernetes.client.rest import ApiException
try:
pod_list = self._core_api.list_namespaced_pod(
namespace=self._namespace,
label_selector="app=yuxi-sandbox",
)
except ApiException:
return []
records: list[SandboxRecord] = []
for pod in pod_list.items:
sandbox_id = (pod.metadata.labels or {}).get("sandbox-id")
if not sandbox_id:
continue
record = self.discover(sandbox_id)
if record is not None:
records.append(record)
return records
def delete(self, sandbox_id: str) -> None:
from kubernetes.client.rest import ApiException
pod_name = self._pod_name(sandbox_id)
service_name = self._service_name(sandbox_id)
for delete_call in (
lambda: self._core_api.delete_namespaced_service(name=service_name, namespace=self._namespace),
lambda: self._core_api.delete_namespaced_pod(name=pod_name, namespace=self._namespace),
):
try:
delete_call()
except ApiException as exc:
if exc.status != 404:
raise
2026-03-05 22:50:38 +08:00
class SandboxIdleReaper:
def __init__(self, backend):
self._backend = backend
self._lock = threading.Lock()
self._last_activity_at: dict[str, float] = {}
self._stop_event = threading.Event()
self._thread: threading.Thread | None = None
self._exec_timeout_seconds = int(os.getenv("SANDBOX_EXEC_TIMEOUT_SECONDS", "180"))
configured_idle_timeout = int(os.getenv("SANDBOX_IDLE_TIMEOUT_SECONDS", "120"))
if 0 < configured_idle_timeout <= self._exec_timeout_seconds:
logger.warning(
"SANDBOX_IDLE_TIMEOUT_SECONDS=%s is <= SANDBOX_EXEC_TIMEOUT_SECONDS=%s; "
"adjusting idle timeout to %s seconds to avoid reaping running commands",
configured_idle_timeout,
self._exec_timeout_seconds,
self._exec_timeout_seconds + 30,
)
configured_idle_timeout = self._exec_timeout_seconds + 30
self._idle_timeout_seconds = configured_idle_timeout
self._check_interval_seconds = max(1, int(os.getenv("SANDBOX_IDLE_CHECK_INTERVAL_SECONDS", "10")))
def touch(self, sandbox_id: str) -> None:
with self._lock:
self._last_activity_at[sandbox_id] = time.time()
def forget(self, sandbox_id: str) -> None:
with self._lock:
self._last_activity_at.pop(sandbox_id, None)
def _seed_existing(self) -> None:
try:
records = self._backend.list()
except Exception as exc: # noqa: BLE001
logger.warning(f"Failed to seed sandbox activity for idle reaper: {exc}")
return
now = time.time()
with self._lock:
for record in records:
self._last_activity_at.setdefault(record.sandbox_id, now)
def _collect_expired_sandbox_ids(self) -> list[str]:
if self._idle_timeout_seconds <= 0:
return []
cutoff = time.time() - self._idle_timeout_seconds
with self._lock:
return [sandbox_id for sandbox_id, last_at in self._last_activity_at.items() if last_at <= cutoff]
def _run(self) -> None:
while not self._stop_event.wait(self._check_interval_seconds):
expired_ids = self._collect_expired_sandbox_ids()
for sandbox_id in expired_ids:
try:
self._backend.delete(sandbox_id)
logger.info(f"Deleted idle sandbox: {sandbox_id}")
self.forget(sandbox_id)
except Exception as exc: # noqa: BLE001
logger.warning(f"Failed to delete idle sandbox {sandbox_id}: {exc}")
def start(self) -> None:
if self._idle_timeout_seconds <= 0:
logger.info("Idle reaper disabled (SANDBOX_IDLE_TIMEOUT_SECONDS <= 0)")
return
self._seed_existing()
self._thread = threading.Thread(target=self._run, name="sandbox-idle-reaper", daemon=True)
self._thread.start()
logger.info(
"Started sandbox idle reaper with timeout=%ss interval=%ss",
self._idle_timeout_seconds,
self._check_interval_seconds,
)
def shutdown(self) -> None:
self._stop_event.set()
if self._thread is not None:
self._thread.join(timeout=3)
def _build_backend():
backend = (os.getenv("PROVISIONER_BACKEND", "memory") or "memory").strip().lower()
2026-03-05 22:50:38 +08:00
if backend in {"docker", "local"}:
return LocalContainerProvisionerBackend(), backend
if backend == "kubernetes":
return KubernetesProvisionerBackend(), backend
return MemoryProvisionerBackend(), backend
backend_impl, backend_name = _build_backend()
2026-03-05 22:50:38 +08:00
idle_reaper = SandboxIdleReaper(backend_impl)
@asynccontextmanager
async def lifespan(_app: FastAPI):
idle_reaper.start()
try:
yield
finally:
idle_reaper.shutdown()
app = FastAPI(title="Yuxi Sandbox Provisioner", lifespan=lifespan)
@app.get("/health")
def health():
2026-03-05 22:50:38 +08:00
tracked = len(idle_reaper._last_activity_at) # noqa: SLF001
return {
"status": "ok",
"backend": backend_name,
"idle_timeout_seconds": idle_reaper._idle_timeout_seconds, # noqa: SLF001
"idle_check_interval_seconds": idle_reaper._check_interval_seconds, # noqa: SLF001
"tracked_sandboxes": tracked,
}
@app.post("/api/sandboxes", response_model=SandboxResponse)
def create_sandbox(payload: CreateSandboxRequest):
try:
record = backend_impl.create(payload.sandbox_id, payload.thread_id)
2026-03-05 22:50:38 +08:00
except ValueError as exc:
raise HTTPException(status_code=400, detail=str(exc)) from exc
except Exception as exc: # noqa: BLE001
raise HTTPException(status_code=500, detail=str(exc)) from exc
2026-03-05 22:50:38 +08:00
idle_reaper.touch(record.sandbox_id)
return SandboxResponse(
sandbox_id=record.sandbox_id,
sandbox_url=record.sandbox_url,
status=record.status,
)
@app.get("/api/sandboxes/{sandbox_id}", response_model=SandboxResponse)
def get_sandbox(sandbox_id: str):
try:
record = backend_impl.discover(sandbox_id)
except Exception as exc: # noqa: BLE001
raise HTTPException(status_code=500, detail=str(exc)) from exc
if record is None:
raise HTTPException(status_code=404, detail="sandbox not found")
2026-03-05 22:50:38 +08:00
idle_reaper.touch(record.sandbox_id)
return SandboxResponse(
sandbox_id=record.sandbox_id,
sandbox_url=record.sandbox_url,
status=record.status,
)
2026-03-05 22:50:38 +08:00
@app.post("/api/sandboxes/{sandbox_id}/touch", response_model=TouchSandboxResponse)
def touch_sandbox(sandbox_id: str):
try:
record = backend_impl.discover(sandbox_id)
except Exception as exc: # noqa: BLE001
raise HTTPException(status_code=500, detail=str(exc)) from exc
if record is None:
raise HTTPException(status_code=404, detail="sandbox not found")
idle_reaper.touch(sandbox_id)
return TouchSandboxResponse(ok=True, sandbox_id=sandbox_id, status=record.status)
@app.get("/api/sandboxes", response_model=ListSandboxesResponse)
def list_sandboxes():
try:
records = backend_impl.list()
except Exception as exc: # noqa: BLE001
raise HTTPException(status_code=500, detail=str(exc)) from exc
sandboxes = [
SandboxResponse(
sandbox_id=record.sandbox_id,
sandbox_url=record.sandbox_url,
status=record.status,
)
for record in records
]
return ListSandboxesResponse(sandboxes=sandboxes, count=len(sandboxes))
@app.delete("/api/sandboxes/{sandbox_id}", response_model=DeleteSandboxResponse)
def delete_sandbox(sandbox_id: str):
try:
backend_impl.delete(sandbox_id)
except Exception as exc: # noqa: BLE001
raise HTTPException(status_code=500, detail=str(exc)) from exc
2026-03-05 22:50:38 +08:00
idle_reaper.forget(sandbox_id)
return DeleteSandboxResponse(ok=True, sandbox_id=sandbox_id)