实现一次性 Pod Runner 后端

README.md

@@ -1,20 +1,33 @@
-# Gitea microVM runner
+# Gitea dynamic runner
 
-为 Gitea Actions 按需启动 Cloud Hypervisor microVM。适合 kind、嵌套容器和其他不应
-在常驻 Kubernetes runner 中执行的 CI 工作负载。
+为 Gitea Actions 按需创建一次性执行环境。对 workflow 提供两种稳定的 runner
+接口：
+
+```yaml
+runs-on: [self-hosted, pod]
+```
+
+```yaml
+runs-on: [self-hosted, vm]
+```
+
+`pod` 使用动态 Kubernetes Pod，`vm` 使用动态 Cloud Hypervisor microVM。每个环境
+只执行一个 job，并在 job 结束后连同本地状态一起销毁。完整的设计约束见
+[`docs/design-principles.md`](docs/design-principles.md)。
 
 组件：
 
 - `controller`：接收 Gitea `workflow_job` webhook，将指定 label 的 queued job
   发布到 NATS JetStream。
-- `worker`：在虚拟化宿主机领取任务，限制本机并发，并启动一次性 microVM。
+- `worker`：领取任务、限制并发，并通过 Pod 或 microVM backend 创建一次性环境。
 - `microvm-runner-launch`：为每个任务创建 COW disk、NoCloud seed 和 TAP，运行
   Cloud Hypervisor，退出后完整清理。
 - `guest-runner`：在 guest 中领取一次性 runner registration token，注册 ephemeral
   runner，执行一个 job 后关机。
-- `jwt-broker`：运行在 Kubernetes runner 外层 Pod 中，以可被 SPIRE attestation
-  的 PID 获取固定 `aud=zot` JWT-SVID；DinD job 通过受限 HTTP endpoint 获取短期
-  token。broker 不记录响应、不缓存 token，也不接受调用方指定 audience。
+- `pod-worker`：在 Kubernetes 中创建一次性 privileged Pod；Pod 内的 workflow 使用
+  host executor，Docker、BuildKit 和 kind 等工具由 pipeline 按需 setup。
+- `jwt-broker`：早期共享 Kubernetes runner 的过渡实验；目标架构不部署它，每个
+  动态 Pod 或 VM 直接取得自己的 SPIFFE 身份。
 
 消息流使用一个 `WorkQueuePolicy` stream。相同 runner label 的所有 worker 共享同一
 durable consumer；扩容只需要增加 worker 或提高单机 capacity。

config/runner.yaml

@@ -0,0 +1,13 @@
+runner:
+  capacity: 1
+  timeout: 3h
+  shutdown_timeout: 1m
+  hooks:
+    job_started: /usr/local/libexec/gitea-job-started
+
+host:
+  workdir_parent: /workspace
+
+container:
+  require_docker: false
+  valid_volumes: []

container/runner.Dockerfile

@@ -0,0 +1,8 @@
+FROM ghcr.io/spiffe/spire-agent:1.15.3@sha256:41b0dcd8b258a69db9e2768292a060766fb76fd866e4bc925849981ea1b825ff AS spire
+
+FROM docker.io/gitea/runner:2@sha256:66d80966792e621c9761c47919644198d35fd1c297e9a01e69ed3c1ae37db0c7
+USER root
+COPY --from=spire /opt/spire/bin/spire-agent /opt/spire/bin/spire-agent
+COPY config/runner.yaml /etc/gitea-runner/config.yaml
+COPY scripts/gitea-job-started /usr/local/libexec/gitea-job-started
+RUN chmod 0755 /usr/local/libexec/gitea-job-started

docs/design-principles.md

@@ -0,0 +1,88 @@
+# 动态 Runner 设计原则
+
+## 对 workflow 的接口
+
+Runner 只向 workflow 暴露两个执行环境：
+
+```yaml
+runs-on: [self-hosted, pod]
+```
+
+```yaml
+runs-on: [self-hosted, vm]
+```
+
+- `self-hosted` 是固定前缀。
+- `pod` 表示一次性 Kubernetes Pod，承担常规 CI、镜像构建和 kind 等任务。
+- `vm` 表示一次性 microVM，承担需要独立内核、KVM、systemd 或更强隔离的任务。
+
+执行后端是基础设施选择，不是权限角色。workflow 不需要额外声明由 controller
+维护的 role 或权限 label。
+
+## 一个 job，一个环境
+
+Controller 根据 Gitea `workflow_job` webhook 创建执行环境。每个 Pod 或 VM 注册一个
+ephemeral runner，只执行一个 job；任务结束后注销 runner，并删除计算环境及其全部
+本地状态。
+
+`job_id` 仅用于消息去重、状态追踪、实例关联和失败清理，不进入 workload 身份，也
+不参与资源授权。
+
+Gitea 不保证由某次 `queued` webhook 创建的 runner 一定领取该 webhook 对应的 job。
+因此创建环境时只赋予无业务权限的启动身份。runner 实际领取任务后，controller 根据
+`in_progress` webhook 返回的 `runner_name` 和真实 job 名称绑定业务身份；环境中的
+job-start hook 必须等目标 SVID 可用后才放行 workflow 的第一步。不能依据 queued
+事件提前赋予任务权限。
+
+## 环境只提供运行边界
+
+基础镜像只提供启动 runner 和执行 workflow 所需的最小环境。Docker、BuildKit、
+kind 等工具由 pipeline 按需安装和启动，而不是由 controller 预制成常驻服务。
+
+例如 Pod job 可以在 Pod 内启动仅供本次任务使用的 Docker daemon。该 daemon 及其
+镜像、容器和缓存属于当前 job 的临时状态，随 Pod 一起销毁。Docker 创建的容器不是
+独立的身份边界；需要访问凭据的操作由 Pod 中的 workflow 进程完成，并通过环境变量
+或标准输入把短期凭据交给具体工具。
+
+## Workload 身份
+
+动态 Pod 和 VM 都直接拥有自己的 SPIFFE 身份，不继承常驻 runner 的共享身份：
+
+- Pod 通过 Kubernetes workload attestation 取得身份。
+- VM 通过 VM 内的 SPIRE Agent 取得身份。
+
+SPIFFE ID 由具有业务意义且稳定的 workflow 上下文派生：
+
+```text
+spiffe://ddupan.top/ci/<owner>/<repository>/<job-name>
+```
+
+同一种任务在不同运行中使用相同的逻辑 SPIFFE ID；每次运行取得独立、短期的 SVID。
+Pod 与 VM 是可替换的执行实现，因此默认不写入 SPIFFE ID。
+
+job 名称必须经过确定性的路径规范化。规范化结果必须保留仓库边界，并在发生冲突时
+拒绝创建环境，不能静默地让两个任务共享身份。同一仓库内需要不同权限的任务应使用
+不同的 job 名称；workflow 文件只是编排载体，不进入权限身份。
+
+## Self-service 与授权边界
+
+新增 workflow 或 job 时，controller 自动为它派生身份，不维护第二份任务或角色
+allowlist。能够修改仓库 CI 的主体本来就能修改该仓库已有任务，因此 controller 的
+重复审批不能形成额外的安全边界，只会破坏 self-service。
+
+身份不等于权限。新任务可以立即取得自己的 SPIFFE ID，但默认不会因此获得 Zot、
+OpenBao 或其他资源的特殊权限。资源所有者在资源端按照有意义的 job 身份
+配置授权策略。
+
+## 非目标设计
+
+目标架构不依赖以下机制：
+
+- 多个 job 共享的常驻 Docker daemon。
+- 常驻 runner Pod 的共享 SPIFFE 身份。
+- 为嵌套 CI 容器转发共享身份的 JWT broker。
+- 将 Gitea 数字 job ID 编入 SPIFFE ID。
+- controller 维护的仓库任务权限 allowlist。
+
+仓库中的 `jwt-broker` 是早期方案的实验实现，在 Pod/VM 动态执行环境完成迁移后不应
+部署。

pyproject.toml

@@ -15,6 +15,7 @@ test = ["pytest==8.4.2", "pytest-asyncio==1.2.0"]
 [project.scripts]
 gitea-microvm-controller = "gitea_microvm_runner.controller:main"
 gitea-microvm-worker = "gitea_microvm_runner.worker:cli"
+gitea-pod-worker = "gitea_microvm_runner.pod_worker:cli"
 gitea-spire-jwt-broker = "gitea_microvm_runner.jwt_broker:main"
 
 [tool.pytest.ini_options]

scripts/gitea-job-started

@@ -0,0 +1,19 @@
+#!/bin/sh
+set -eu
+
+socket=${SPIRE_AGENT_SOCKET:-/run/spire/agent-sockets/spire-agent.sock}
+wait_seconds=${SPIFFE_IDENTITY_WAIT_SECONDS:-60}
+deadline=$(( $(date +%s) + wait_seconds ))
+
+while [ "$(date +%s)" -lt "$deadline" ]; do
+    if timeout 2 /opt/spire/bin/spire-agent api fetch jwt \
+        -audience ci-job-ready \
+        -socketPath "$socket" \
+        >/dev/null 2>&1; then
+        exit 0
+    fi
+    sleep 1
+done
+
+echo "timed out waiting for the job SPIFFE identity" >&2
+exit 1

scripts/gitea-microvm-guest-runner

@@ -4,7 +4,7 @@ set -eu
 token_url=${1:?token URL is required}
 instance=${2:?Gitea instance is required}
 runner_name=${3:?runner name is required}
-runner_label=${4:?runner label is required}
+runner_labels=${4:?runner labels are required}
 token_file=/run/gitea-runner-registration-token
 
 cleanup() {
@@ -22,7 +22,7 @@ gitea-runner register \
     --ephemeral \
     --instance "$instance" \
     --name "$runner_name" \
-    --labels "$runner_label:host" \
+    --labels "$runner_labels" \
     --token-file "$token_file"
 rm -f -- "$token_file"
 gitea-runner daemon

scripts/microvm-runner-launch

@@ -21,7 +21,7 @@ vm_timeout=${RUNNER_VM_TIMEOUT:-3h}
 cpus=${RUNNER_VM_CPUS:-4}
 memory=${RUNNER_VM_MEMORY:-3G}
 gitea_instance=${GITEA_INSTANCE:-https://git.ddupan.top}
-runner_label=${RUNNER_LABEL:-kind-microvm}
+runner_labels=${RUNNER_LABELS:-self-hosted:host,vm:host}
 token_url=${RUNNER_TOKEN_URL:-http://172.30.0.1:8787}
 
 vm_dir="$state_root/instances/$instance_id"
@@ -48,7 +48,7 @@ EOF
 cat >"$vm_dir/user-data" <<EOF
 #cloud-config
 runcmd:
-  - [ /usr/local/libexec/gitea-microvm-guest-runner, "$token_url/token/$nonce", "$gitea_instance", "gitea-${instance_id%%-*}", "$runner_label" ]
+  - [ /usr/local/libexec/gitea-microvm-guest-runner, "$token_url/token/$nonce", "$gitea_instance", "gitea-${instance_id%%-*}", "$runner_labels" ]
 EOF
 cloud-localds "$seed" "$vm_dir/user-data" "$vm_dir/meta-data"
 

src/gitea_microvm_runner/controller.py

@@ -13,8 +13,10 @@ from aiohttp import web
 from nats.js.api import DiscardPolicy, RetentionPolicy, StorageType, StreamConfig
 from nats.js.errors import NotFoundError
 
-LABEL = os.environ.get("RUNNER_LABEL", "kind-microvm")
-SUBJECT = os.environ.get("NATS_SUBJECT", f"ci.runner.{LABEL}")
+from .models import IdentityBinding, RunnerRequest
+
+
+SUBJECT_PREFIX = os.environ.get("NATS_SUBJECT_PREFIX", "ci.runner")
 STREAM = os.environ.get("NATS_STREAM", "CI_RUNNER")
 NATS_URL = os.environ.get("NATS_URL", "tls://nats.ad.ddupan.top:4222")
 NATS_USER = os.environ.get("NATS_USER", "ci-producer")
@@ -24,15 +26,9 @@ WEBHOOK_SECRET_FILE = Path(os.environ.get("WEBHOOK_SECRET_FILE", "/run/secrets/g
 
 
 def accepts(payload: object) -> tuple[bool, str | None]:
-    if not isinstance(payload, dict) or payload.get("action") != "queued":
-        return False, None
-    job = payload.get("workflow_job")
-    if not isinstance(job, dict) or LABEL not in job.get("labels", []):
-        return False, None
-    job_id = job.get("id")
-    if not isinstance(job_id, int) or isinstance(job_id, bool):
-        return False, None
-    return True, str(job_id)
+    """Compatibility helper for callers that only need acceptance and identity."""
+    request = RunnerRequest.from_webhook(payload)
+    return (request is not None, str(request.job_id) if request else None)
 
 
 def valid_signature(body: bytes, signature: str) -> bool:
@@ -43,7 +39,7 @@ def valid_signature(body: bytes, signature: str) -> bool:
 async def ensure_stream(js: object) -> None:
     config = StreamConfig(
         name=STREAM,
-        subjects=["ci.runner.*"],
+        subjects=[f"{SUBJECT_PREFIX}.>"],
         retention=RetentionPolicy.WORK_QUEUE,
         storage=StorageType.FILE,
         discard=DiscardPolicy.OLD,
@@ -68,15 +64,25 @@ async def webhook(request: web.Request) -> web.Response:
         payload = json.loads(body)
     except json.JSONDecodeError as error:
         raise web.HTTPBadRequest(text="invalid JSON\n") from error
-    accepted, job_id = accepts(payload)
-    if not accepted:
-        return web.Response(status=204)
-    await request.app["js"].publish(
-        SUBJECT,
-        body,
-        headers={"Nats-Msg-Id": f"gitea-workflow-job-{job_id}"},
-    )
-    return web.Response(status=202, text="queued\n")
+    runner_request = RunnerRequest.from_webhook(payload)
+    if runner_request is not None:
+        await request.app["js"].publish(
+            f"{SUBJECT_PREFIX}.{runner_request.backend}",
+            runner_request.to_json(),
+            headers={"Nats-Msg-Id": f"gitea-workflow-job-{runner_request.job_id}-queued"},
+        )
+        return web.Response(status=202, text="queued\n")
+
+    binding = IdentityBinding.from_webhook(payload)
+    if binding is not None:
+        await request.app["js"].publish(
+            f"{SUBJECT_PREFIX}.{binding.backend}.binding",
+            binding.to_json(),
+            headers={"Nats-Msg-Id": f"gitea-workflow-job-{binding.job_id}-in-progress"},
+        )
+        return web.Response(status=202, text="binding queued\n")
+
+    return web.Response(status=204)
 
 
 async def health(request: web.Request) -> web.Response:

src/gitea_microvm_runner/kubernetes.py

@@ -0,0 +1,89 @@
+"""Small in-cluster Kubernetes API client used by the Pod backend."""
+
+from __future__ import annotations
+
+import json
+from pathlib import Path
+from urllib.parse import quote
+
+from aiohttp import ClientResponseError, ClientSession, TCPConnector
+import ssl
+
+
+class KubernetesClient:
+    def __init__(
+        self,
+        *,
+        api_url: str,
+        token_file: Path,
+        ca_file: Path,
+        namespace: str,
+    ) -> None:
+        self.api_url = api_url.rstrip("/")
+        self.token_file = token_file
+        self.ca_file = ca_file
+        self.namespace = namespace
+        self.session: ClientSession | None = None
+
+    async def __aenter__(self) -> KubernetesClient:
+        context = ssl.create_default_context(cafile=self.ca_file)
+        self.session = ClientSession(
+            connector=TCPConnector(ssl=context),
+            headers={
+                "Authorization": f"Bearer {self.token_file.read_text().strip()}",
+            },
+            raise_for_status=True,
+        )
+        return self
+
+    async def __aexit__(self, *_: object) -> None:
+        if self.session is not None:
+            await self.session.close()
+
+    async def create_pod(self, manifest: dict[str, object]) -> dict[str, object]:
+        response = await self._request("POST", self._pods_path(), json=manifest)
+        return await response.json()
+
+    async def get_pod(self, name: str) -> dict[str, object] | None:
+        try:
+            response = await self._request("GET", f"{self._pods_path()}/{quote(name)}")
+        except ClientResponseError as error:
+            if error.status == 404:
+                return None
+            raise
+        return await response.json()
+
+    async def bind_identity(self, name: str, identity_path: str) -> None:
+        patch = {
+            "metadata": {
+                "labels": {"ci.ddupan.top/identity-bound": "true"},
+                "annotations": {"ci.ddupan.top/spiffe-path": identity_path},
+            }
+        }
+        response = await self._request(
+            "PATCH",
+            f"{self._pods_path()}/{quote(name)}",
+            data=json.dumps(patch),
+            headers={"Content-Type": "application/merge-patch+json"},
+        )
+        response.release()
+
+    async def delete_pod(self, name: str) -> None:
+        try:
+            response = await self._request(
+                "DELETE",
+                f"{self._pods_path()}/{quote(name)}",
+                json={"gracePeriodSeconds": 30, "propagationPolicy": "Background"},
+            )
+            response.release()
+        except ClientResponseError as error:
+            if error.status != 404:
+                raise
+
+    async def _request(self, method: str, path: str, **kwargs: object):
+        if self.session is None:
+            raise RuntimeError("KubernetesClient is not open")
+        return await self.session.request(method, f"{self.api_url}{path}", **kwargs)
+
+    def _pods_path(self) -> str:
+        return f"/api/v1/namespaces/{quote(self.namespace)}/pods"

src/gitea_microvm_runner/models.py

@@ -0,0 +1,176 @@
+"""Validated messages shared by the controller and runner backends."""
+
+from __future__ import annotations
+
+from dataclasses import asdict, dataclass
+import json
+
+
+BACKENDS = frozenset({"pod", "vm"})
+REQUIRED_LABEL = "self-hosted"
+
+
+@dataclass(frozen=True, slots=True)
+class RunnerRequest:
+    """A queued Gitea job that needs one disposable runner environment."""
+
+    job_id: int
+    run_id: int
+    backend: str
+    repository: str
+    job_name: str
+    labels: tuple[str, ...]
+    schema_version: int = 1
+
+    @classmethod
+    def from_webhook(cls, payload: object) -> RunnerRequest | None:
+        if not isinstance(payload, dict) or payload.get("action") != "queued":
+            return None
+
+        job = payload.get("workflow_job")
+        repository = payload.get("repository")
+        if not isinstance(job, dict) or not isinstance(repository, dict):
+            return None
+
+        labels_value = job.get("labels")
+        if not isinstance(labels_value, list) or not all(
+            isinstance(label, str) for label in labels_value
+        ):
+            return None
+        labels = tuple(dict.fromkeys(labels_value))
+        selected = BACKENDS.intersection(labels)
+        if REQUIRED_LABEL not in labels or len(selected) != 1:
+            return None
+
+        job_id = job.get("id")
+        run_id = job.get("run_id")
+        job_name = job.get("name")
+        full_name = repository.get("full_name")
+        if not _positive_int(job_id) or not _positive_int(run_id):
+            return None
+        if not all(_nonempty(value) for value in (job_name, full_name)):
+            return None
+
+        return cls(
+            job_id=job_id,
+            run_id=run_id,
+            backend=next(iter(selected)),
+            repository=full_name.strip(),
+            job_name=job_name.strip(),
+            labels=labels,
+        )
+
+    def to_json(self) -> bytes:
+        document = asdict(self)
+        document["labels"] = list(self.labels)
+        return json.dumps(
+            document,
+            ensure_ascii=False,
+            separators=(",", ":"),
+            sort_keys=True,
+        ).encode()
+
+    @classmethod
+    def from_json(cls, body: bytes) -> RunnerRequest:
+        document = json.loads(body)
+        if not isinstance(document, dict) or document.get("schema_version") != 1:
+            raise ValueError("unsupported runner request")
+        labels = document.get("labels")
+        if not isinstance(labels, list) or not all(
+            isinstance(label, str) for label in labels
+        ):
+            raise ValueError("invalid runner labels")
+        try:
+            request = cls(
+                job_id=document["job_id"],
+                run_id=document["run_id"],
+                backend=document["backend"],
+                repository=document["repository"],
+                job_name=document["job_name"],
+                labels=tuple(labels),
+            )
+        except KeyError as error:
+            raise ValueError(f"missing runner request field: {error.args[0]}") from error
+        if (
+            not _positive_int(request.job_id)
+            or not _positive_int(request.run_id)
+            or request.backend not in BACKENDS
+            or not _nonempty(request.repository)
+            or not _nonempty(request.job_name)
+            or REQUIRED_LABEL not in request.labels
+            or request.backend not in request.labels
+        ):
+            raise ValueError("invalid runner request")
+        return request
+
+
+@dataclass(frozen=True, slots=True)
+class IdentityBinding:
+    """The actual task claimed by an ephemeral runner."""
+
+    job_id: int
+    run_id: int
+    backend: str
+    runner_name: str
+    repository: str
+    job_name: str
+    schema_version: int = 1
+
+    @classmethod
+    def from_webhook(cls, payload: object) -> IdentityBinding | None:
+        if not isinstance(payload, dict) or payload.get("action") != "in_progress":
+            return None
+        job = payload.get("workflow_job")
+        repository = payload.get("repository")
+        if not isinstance(job, dict) or not isinstance(repository, dict):
+            return None
+
+        labels = job.get("labels")
+        if not isinstance(labels, list) or not all(
+            isinstance(label, str) for label in labels
+        ):
+            return None
+        selected = BACKENDS.intersection(labels)
+        if REQUIRED_LABEL not in labels or len(selected) != 1:
+            return None
+        backend = next(iter(selected))
+
+        job_id = job.get("id")
+        run_id = job.get("run_id")
+        runner_name = job.get("runner_name")
+        job_name = job.get("name")
+        full_name = repository.get("full_name")
+        if not _positive_int(job_id) or not _positive_int(run_id):
+            return None
+        if not all(
+            _nonempty(value)
+            for value in (runner_name, job_name, full_name)
+        ):
+            return None
+        if not runner_name.startswith(f"gitea-{backend}-"):
+            return None
+
+        return cls(
+            job_id=job_id,
+            run_id=run_id,
+            backend=backend,
+            runner_name=runner_name.strip(),
+            repository=full_name.strip(),
+            job_name=job_name.strip(),
+        )
+
+    def to_json(self) -> bytes:
+        return json.dumps(
+            asdict(self),
+            ensure_ascii=False,
+            separators=(",", ":"),
+            sort_keys=True,
+        ).encode()
+
+
+def _positive_int(value: object) -> bool:
+    return isinstance(value, int) and not isinstance(value, bool) and value > 0
+
+
+def _nonempty(value: object) -> bool:
+    return isinstance(value, str) and bool(value.strip())

src/gitea_microvm_runner/pod_worker.py

@@ -0,0 +1,307 @@
+#!/usr/bin/env python3
+"""JetStream worker that creates one disposable Kubernetes Pod per job."""
+
+from __future__ import annotations
+
+import asyncio
+import json
+import logging
+import os
+from pathlib import Path
+import ssl
+from urllib.parse import quote
+import uuid
+
+import nats
+from aiohttp import ClientResponseError
+from nats.errors import TimeoutError
+from nats.js.api import AckPolicy, ConsumerConfig
+
+from .kubernetes import KubernetesClient
+from .models import IdentityBinding, RunnerRequest
+
+
+LOG = logging.getLogger(__name__)
+STREAM = os.environ.get("NATS_STREAM", "CI_RUNNER")
+REQUEST_SUBJECT = os.environ.get("NATS_SUBJECT", "ci.runner.pod")
+BINDING_SUBJECT = os.environ.get("NATS_BINDING_SUBJECT", "ci.runner.pod.binding")
+NATS_URL = os.environ.get("NATS_URL", "tls://nats.ad.ddupan.top:4222")
+NATS_USER = os.environ.get("NATS_USER", "ci-worker")
+NATS_PASSWORD_FILE = Path(os.environ.get("NATS_PASSWORD_FILE", "/run/secrets/nats/password"))
+NATS_CA_FILE = os.environ.get("NATS_CA_FILE", "/etc/ssl/certs/ca-certificates.crt")
+KUBERNETES_API = os.environ.get("KUBERNETES_API", "https://kubernetes.default.svc")
+KUBERNETES_TOKEN_FILE = Path(
+    os.environ.get(
+        "KUBERNETES_TOKEN_FILE",
+        "/var/run/secrets/kubernetes.io/serviceaccount/token",
+    )
+)
+KUBERNETES_CA_FILE = Path(
+    os.environ.get(
+        "KUBERNETES_CA_FILE",
+        "/var/run/secrets/kubernetes.io/serviceaccount/ca.crt",
+    )
+)
+NAMESPACE = os.environ.get("RUNNER_NAMESPACE", "gitea-actions")
+RUNNER_IMAGE = os.environ.get("RUNNER_IMAGE", "docker.io/gitea/runner:2")
+RUNNER_SERVICE_ACCOUNT = os.environ.get("RUNNER_SERVICE_ACCOUNT", "gitea-dynamic-runner")
+RUNNER_TOKEN_SECRET = os.environ.get("RUNNER_TOKEN_SECRET", "gitea-dynamic-runner")
+GITEA_INSTANCE = os.environ.get("GITEA_INSTANCE", "https://git.ddupan.top")
+CAPACITY = int(os.environ.get("RUNNER_CAPACITY", "4"))
+POD_TIMEOUT = int(os.environ.get("RUNNER_POD_TIMEOUT", str(4 * 60 * 60)))
+
+
+def identity_path(repository: str, job_name: str) -> str:
+    parts = repository.split("/")
+    if len(parts) != 2 or not all(parts):
+        raise ValueError("repository must be owner/name")
+    encoded = [quote(part, safe="-._~") for part in (*parts, job_name)]
+    if not job_name.strip() or any(not part for part in encoded):
+        raise ValueError("identity components must not be empty")
+    return "/".join(encoded)
+
+
+def pod_manifest(request: RunnerRequest, pod_name: str) -> dict[str, object]:
+    if request.backend != "pod":
+        raise ValueError("Pod backend only accepts pod requests")
+    return {
+        "apiVersion": "v1",
+        "kind": "Pod",
+        "metadata": {
+            "name": pod_name,
+            "namespace": NAMESPACE,
+            "labels": {
+                "app.kubernetes.io/name": "gitea-dynamic-runner",
+                "app.kubernetes.io/component": "runner",
+                "ci.ddupan.top/backend": "pod",
+            },
+            "annotations": {
+                "ci.ddupan.top/queued-job-id": str(request.job_id),
+                "ci.ddupan.top/queued-run-id": str(request.run_id),
+            },
+        },
+        "spec": {
+            "serviceAccountName": RUNNER_SERVICE_ACCOUNT,
+            "restartPolicy": "Never",
+            "terminationGracePeriodSeconds": 30,
+            "containers": [
+                {
+                    "name": "runner",
+                    "image": RUNNER_IMAGE,
+                    "imagePullPolicy": "IfNotPresent",
+                    "securityContext": {"privileged": True},
+                    "env": [
+                        {"name": "GITEA_INSTANCE_URL", "value": GITEA_INSTANCE},
+                        {
+                            "name": "GITEA_RUNNER_NAME",
+                            "valueFrom": {"fieldRef": {"fieldPath": "metadata.name"}},
+                        },
+                        {
+                            "name": "GITEA_RUNNER_REGISTRATION_TOKEN_FILE",
+                            "value": "/run/secrets/gitea/token",
+                        },
+                        {"name": "GITEA_RUNNER_LABELS", "value": "self-hosted:host,pod:host"},
+                        {"name": "GITEA_RUNNER_EPHEMERAL", "value": "1"},
+                        {"name": "GITEA_RUNNER_ONCE", "value": "1"},
+                        {"name": "CONFIG_FILE", "value": "/etc/gitea-runner/config.yaml"},
+                    ],
+                    "volumeMounts": [
+                        {
+                            "name": "registration-token",
+                            "mountPath": "/run/secrets/gitea",
+                            "readOnly": True,
+                        },
+                        {
+                            "name": "spire-agent-socket",
+                            "mountPath": "/run/spire/agent-sockets",
+                            "readOnly": True,
+                        },
+                    ],
+                }
+            ],
+            "volumes": [
+                {
+                    "name": "registration-token",
+                    "secret": {
+                        "secretName": RUNNER_TOKEN_SECRET,
+                        "items": [{"key": "token", "path": "token"}],
+                    },
+                },
+                {
+                    "name": "spire-agent-socket",
+                    "csi": {"driver": "csi.spiffe.io", "readOnly": True},
+                },
+            ],
+        },
+    }
+
+
+async def heartbeat(message: object, stop: asyncio.Event) -> None:
+    while True:
+        try:
+            await asyncio.wait_for(stop.wait(), timeout=60)
+            return
+        except asyncio.TimeoutError:
+            await message.in_progress()
+
+
+async def wait_for_pod(client: KubernetesClient, name: str) -> bool:
+    deadline = asyncio.get_running_loop().time() + POD_TIMEOUT
+    while asyncio.get_running_loop().time() < deadline:
+        pod = await client.get_pod(name)
+        if pod is None:
+            raise RuntimeError(f"runner Pod {name} disappeared")
+        status = pod.get("status")
+        phase = status.get("phase") if isinstance(status, dict) else None
+        if phase == "Succeeded":
+            return True
+        if phase == "Failed":
+            return False
+        await asyncio.sleep(2)
+    raise asyncio.TimeoutError(f"runner Pod {name} timed out")
+
+
+async def run_request(message: object, client: KubernetesClient) -> None:
+    try:
+        request = RunnerRequest.from_json(message.data)
+    except (json.JSONDecodeError, UnicodeDecodeError, ValueError) as error:
+        LOG.error("discarding invalid Pod request: %s", error)
+        await message.ack()
+        return
+    if request.backend != "pod":
+        LOG.error("discarding %s request received by Pod worker", request.backend)
+        await message.ack()
+        return
+
+    pod_name = f"gitea-pod-{uuid.uuid4().hex[:12]}"
+    stop = asyncio.Event()
+    pulse = asyncio.create_task(heartbeat(message, stop))
+    try:
+        await client.create_pod(pod_manifest(request, pod_name))
+        if await wait_for_pod(client, pod_name):
+            await message.ack()
+        else:
+            LOG.error("runner Pod %s failed", pod_name)
+            await message.nak(delay=30)
+    except Exception:
+        await message.nak(delay=30)
+        raise
+    finally:
+        stop.set()
+        await pulse
+        await client.delete_pod(pod_name)
+
+
+async def bind_request(message: object, client: KubernetesClient) -> None:
+    try:
+        document = json.loads(message.data)
+        binding = IdentityBinding(**document)
+        if binding.backend != "pod" or not binding.runner_name.startswith("gitea-pod-"):
+            raise ValueError("invalid Pod identity binding")
+        path = identity_path(binding.repository, binding.job_name)
+        await client.bind_identity(binding.runner_name, path)
+    except ClientResponseError as error:
+        if error.status == 404:
+            await message.nak(delay=2)
+            return
+        await message.nak(delay=30)
+        raise
+    except (json.JSONDecodeError, TypeError, ValueError) as error:
+        LOG.error("discarding invalid identity binding: %s", error)
+        await message.ack()
+        return
+    await message.ack()
+
+
+async def consume_requests(subscription: object, client: KubernetesClient) -> None:
+    active: set[asyncio.Task[None]] = set()
+    while True:
+        active = {task for task in active if not task.done()}
+        free = CAPACITY - len(active)
+        if free < 1:
+            await asyncio.wait(active, return_when=asyncio.FIRST_COMPLETED)
+            continue
+        try:
+            messages = await subscription.fetch(batch=free, timeout=5)
+        except TimeoutError:
+            continue
+        for message in messages:
+            task = asyncio.create_task(run_request(message, client))
+            task.add_done_callback(_report_task)
+            active.add(task)
+
+
+async def consume_bindings(subscription: object, client: KubernetesClient) -> None:
+    while True:
+        try:
+            messages = await subscription.fetch(batch=16, timeout=5)
+        except TimeoutError:
+            continue
+        await asyncio.gather(*(bind_request(message, client) for message in messages))
+
+
+def _report_task(task: asyncio.Task[None]) -> None:
+    if not task.cancelled() and (error := task.exception()) is not None:
+        LOG.error("Pod runner task failed", exc_info=(type(error), error, error.__traceback__))
+
+
+async def main() -> None:
+    if CAPACITY < 1:
+        raise ValueError("RUNNER_CAPACITY must be at least 1")
+    context = ssl.create_default_context(cafile=NATS_CA_FILE)
+    nc = await nats.connect(
+        NATS_URL,
+        user=NATS_USER,
+        password=NATS_PASSWORD_FILE.read_text().strip(),
+        tls=context,
+        name="pod-runner-worker",
+    )
+    js = nc.jetstream()
+    requests = await js.pull_subscribe(
+        REQUEST_SUBJECT,
+        durable="pod",
+        stream=STREAM,
+        config=ConsumerConfig(
+            durable_name="pod",
+            filter_subject=REQUEST_SUBJECT,
+            ack_policy=AckPolicy.EXPLICIT,
+            ack_wait=5 * 60,
+            max_ack_pending=max(CAPACITY, 1),
+            max_deliver=5,
+        ),
+    )
+    bindings = await js.pull_subscribe(
+        BINDING_SUBJECT,
+        durable="pod-binding",
+        stream=STREAM,
+        config=ConsumerConfig(
+            durable_name="pod-binding",
+            filter_subject=BINDING_SUBJECT,
+            ack_policy=AckPolicy.EXPLICIT,
+            ack_wait=30,
+            max_ack_pending=64,
+            max_deliver=10,
+        ),
+    )
+    async with KubernetesClient(
+        api_url=KUBERNETES_API,
+        token_file=KUBERNETES_TOKEN_FILE,
+        ca_file=KUBERNETES_CA_FILE,
+        namespace=NAMESPACE,
+    ) as client:
+        try:
+            await asyncio.gather(
+                consume_requests(requests, client),
+                consume_bindings(bindings, client),
+            )
+        finally:
+            await nc.drain()
+
+
+def cli() -> None:
+    logging.basicConfig(level=os.environ.get("LOG_LEVEL", "INFO"))
+    asyncio.run(main())
+
+
+if __name__ == "__main__":
+    cli()

src/gitea_microvm_runner/worker.py

@@ -2,6 +2,7 @@
 """Capacity-bounded JetStream consumer that launches one ephemeral VM per job."""
 
 import asyncio
+import json
 import logging
 import os
 import secrets
@@ -14,11 +15,13 @@ from aiohttp import web
 from nats.errors import TimeoutError
 from nats.js.api import AckPolicy, ConsumerConfig
 
+from .models import RunnerRequest
+
 LOG = logging.getLogger(__name__)
 CAPACITY = int(os.environ.get("RUNNER_CAPACITY", "1"))
-SUBJECT = os.environ.get("NATS_SUBJECT", "ci.runner.kind-microvm")
+SUBJECT = os.environ.get("NATS_SUBJECT", "ci.runner.vm")
 STREAM = os.environ.get("NATS_STREAM", "CI_RUNNER")
-DURABLE = os.environ.get("NATS_DURABLE", "kind-microvm")
+DURABLE = os.environ.get("NATS_DURABLE", "vm")
 MAX_INFLIGHT = int(os.environ.get("RUNNER_MAX_INFLIGHT", "64"))
 NATS_URL = os.environ.get("NATS_URL", "tls://nats.ad.ddupan.top:4222")
 NATS_USER = os.environ.get("NATS_USER", "ci-worker")
@@ -52,6 +55,17 @@ async def heartbeat(message: object, stop: asyncio.Event) -> None:
 
 
 async def run_one(message: object) -> None:
+    try:
+        request = RunnerRequest.from_json(message.data)
+    except (json.JSONDecodeError, UnicodeDecodeError, ValueError) as error:
+        LOG.error("discarding invalid runner request: %s", error)
+        await message.ack()
+        return
+    if request.backend != "vm":
+        LOG.error("discarding %s request received by VM worker", request.backend)
+        await message.ack()
+        return
+
     instance_id = str(uuid.uuid4())
     nonce = secrets.token_urlsafe(32)
     async with token_lock:
@@ -59,7 +73,18 @@ async def run_one(message: object) -> None:
     stop = asyncio.Event()
     pulse = asyncio.create_task(heartbeat(message, stop))
     try:
-        process = await asyncio.create_subprocess_exec(LAUNCHER, instance_id, nonce)
+        process = await asyncio.create_subprocess_exec(
+            LAUNCHER,
+            instance_id,
+            nonce,
+            env={
+                **os.environ,
+                "RUNNER_JOB_ID": str(request.job_id),
+                "RUNNER_RUN_ID": str(request.run_id),
+                "RUNNER_REPOSITORY": request.repository,
+                "RUNNER_JOB_NAME": request.job_name,
+            },
+        )
         return_code = await process.wait()
         if return_code == 0:
             await message.ack()

tests/test_controller.py

@@ -1,22 +1,88 @@
 import hashlib
 import hmac
+import json
+
+import pytest
 
 from gitea_microvm_runner import controller
+from gitea_microvm_runner.models import IdentityBinding, RunnerRequest
 
 
-def test_accepts_matching_queued_job(monkeypatch):
-    monkeypatch.setattr(controller, "LABEL", "kind-microvm")
-    assert controller.accepts({
+def queued_job(**overrides):
+    job = {
+        "id": 47,
+        "run_id": 12,
+        "name": "publish-image",
+        "labels": ["self-hosted", "pod"],
+    }
+    job.update(overrides)
+    return {
         "action": "queued",
-        "workflow_job": {"id": 47, "labels": ["linux", "kind-microvm"]},
-    }) == (True, "47")
+        "workflow_job": job,
+        "repository": {"full_name": "panxiao81/example"},
+    }
 
 
-def test_rejects_other_actions_labels_and_boolean_id(monkeypatch):
-    monkeypatch.setattr(controller, "LABEL", "kind-microvm")
-    assert controller.accepts({"action": "completed", "workflow_job": {"id": 1, "labels": ["kind-microvm"]}}) == (False, None)
-    assert controller.accepts({"action": "queued", "workflow_job": {"id": 1, "labels": ["host"]}}) == (False, None)
-    assert controller.accepts({"action": "queued", "workflow_job": {"id": True, "labels": ["kind-microvm"]}}) == (False, None)
+def test_accepts_matching_queued_job():
+    assert controller.accepts(queued_job()) == (True, "47")
+
+
+def test_rejects_other_actions_labels_and_boolean_id():
+    completed = queued_job()
+    completed["action"] = "completed"
+    assert controller.accepts(completed) == (False, None)
+    assert controller.accepts(queued_job(labels=["self-hosted", "other"])) == (False, None)
+    assert controller.accepts(queued_job(labels=["self-hosted", "pod", "vm"])) == (False, None)
+    assert controller.accepts(queued_job(id=True)) == (False, None)
+
+
+def test_runner_request_contains_stable_identity_context():
+    request = RunnerRequest.from_webhook(queued_job())
+    assert request is not None
+    assert request.backend == "pod"
+    assert request.repository == "panxiao81/example"
+    assert request.job_name == "publish-image"
+    assert request.job_id == 47
+    assert json.loads(request.to_json()) == {
+        "backend": "pod",
+        "job_id": 47,
+        "job_name": "publish-image",
+        "labels": ["self-hosted", "pod"],
+        "repository": "panxiao81/example",
+        "run_id": 12,
+        "schema_version": 1,
+    }
+
+
+def test_runner_request_requires_complete_identity_context():
+    assert RunnerRequest.from_webhook(queued_job(run_id=None)) is None
+    assert RunnerRequest.from_webhook(queued_job(name=" ")) is None
+
+
+def test_runner_request_json_round_trip_and_validation():
+    request = RunnerRequest.from_webhook(queued_job())
+    assert request is not None
+    assert RunnerRequest.from_json(request.to_json()) == request
+    with pytest.raises(ValueError, match="unsupported"):
+        RunnerRequest.from_json(b'{"schema_version":2}')
+
+
+def test_identity_binding_uses_actual_runner_assignment():
+    payload = queued_job(runner_name="gitea-pod-6c47d03d")
+    payload["action"] = "in_progress"
+    binding = IdentityBinding.from_webhook(payload)
+    assert binding is not None
+    assert binding.backend == "pod"
+    assert binding.runner_name == "gitea-pod-6c47d03d"
+    assert binding.repository == "panxiao81/example"
+    assert binding.job_name == "publish-image"
+    assert json.loads(binding.to_json())["job_id"] == 47
+
+
+def test_identity_binding_rejects_runner_from_another_pool():
+    payload = queued_job(runner_name="gitea-vm-6c47d03d")
+    payload["action"] = "in_progress"
+    assert IdentityBinding.from_webhook(payload) is None
 
 
 def test_signature_accepts_gitea_and_prefixed_forms(tmp_path, monkeypatch):

tests/test_pod_worker.py

@@ -0,0 +1,43 @@
+import pytest
+
+from gitea_microvm_runner.models import RunnerRequest
+from gitea_microvm_runner import pod_worker
+
+
+def request() -> RunnerRequest:
+    return RunnerRequest(
+        job_id=47,
+        run_id=12,
+        backend="pod",
+        repository="panxiao81/example",
+        job_name="publish image",
+        labels=("self-hosted", "pod"),
+    )
+
+
+def test_identity_path_is_meaningful_and_uri_safe():
+    assert pod_worker.identity_path("panxiao81/example", "publish image") == (
+        "panxiao81/example/publish%20image"
+    )
+    with pytest.raises(ValueError):
+        pod_worker.identity_path("invalid", "test")
+
+
+def test_pod_manifest_is_disposable_host_runner(monkeypatch):
+    monkeypatch.setattr(pod_worker, "NAMESPACE", "ci")
+    manifest = pod_worker.pod_manifest(request(), "gitea-pod-abcd")
+    assert manifest["metadata"]["name"] == "gitea-pod-abcd"
+    assert manifest["metadata"]["annotations"]["ci.ddupan.top/queued-job-id"] == "47"
+    spec = manifest["spec"]
+    assert spec["restartPolicy"] == "Never"
+    container = spec["containers"][0]
+    assert container["securityContext"] == {"privileged": True}
+    env = {item["name"]: item for item in container["env"]}
+    assert env["GITEA_RUNNER_LABELS"]["value"] == "self-hosted:host,pod:host"
+    assert env["GITEA_RUNNER_EPHEMERAL"]["value"] == "1"
+    assert env["GITEA_RUNNER_ONCE"]["value"] == "1"
+    assert all(item["name"] != "DOCKER_HOST" for item in container["env"])
+    volumes = {item["name"]: item for item in spec["volumes"]}
+    assert volumes["spire-agent-socket"]["csi"]["driver"] == "csi.spiffe.io"
+    assert "runner-config" not in volumes
+    assert "docker" not in volumes