chore(release): 明确生产发布授权并支持集群发布

将发布与上线授权边界调整为一次明确的生产发布指令可连续执行 publish 和 deploy，同时保留仅构建镜像时的停止边界。

为现有发布脚本增加 compose/kubernetes helper 选择和显式 SSH 私钥支持，保持完整 SHA、digest 与生产基线校验不变。

验证：bash -n、ShellCheck、manual-release-test。

AGENTS.md

@@ -52,14 +52,14 @@ docker compose -f docker-compose.yml config --quiet
 2. `main` 不设置受保护分支或 required status；Agent 在获得用户明确授权后，可以合并已验证的 PR，也可以直接提交并推送 `main`。操作前必须获取最新 `origin/main`、确认工作区边界、检查提交差异和敏感信息，并执行与风险匹配的本地验证。
 3. 不得等待、伪造或手工补写已经取消的 CI status，也不得为了满足旧流程擅自恢复 Gitea Actions。Git 提交、合并、Push 和 Tag 授权只覆盖源码历史变更，永远不等于 publish 或 deploy 授权。
 4. 发布只能使用工作区干净、已经提交并属于 `origin/main` 历史的完整 SHA。镜像 Tag 使用完整 SHA，线上只能使用 Registry digest，禁止使用 `latest`。
-5. 本地镜像发布必须由用户明确要求后单独执行：
+5. 本地镜像发布必须由用户明确要求后执行：
 
    ```bash
    ./scripts/publish-release-images.sh --components auto
    ```
 
-   该命令只能构建、冒烟、推送镜像和生成 `dist/releases/<SHA>.json`，不得修改生产。
-6. publish 成功后，Agent 必须报告 manifest、组件、digest 和验证结果并停止。只有用户再次明确确认上线，才可单独执行：
+   该命令只能构建、冒烟、推送镜像和生成 `dist/releases/<SHA>.json`，不得修改生产。用户只要求“构建镜像”“推送镜像”或明确限定为 publish 时，执行到此为止。
+6. 用户明确要求“发布”“上线”或“部署到线上”时，该次授权同时覆盖 publish 和 deploy。publish 成功并核验 manifest、组件、digest 和验证结果后，应直接继续执行下列部署命令，无需再次请求确认：
 
    ```bash
    ./scripts/deploy-production-release.sh dist/releases/<SHA>.json

scripts/deploy-production-release.sh

@@ -3,7 +3,14 @@ set -euo pipefail
 
 root=$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)
 production_host=${AI_GATEWAY_PRODUCTION_HOST:-root@110.42.51.33}
-remote_helper=${AI_GATEWAY_REMOTE_RELEASE_HELPER:-/usr/local/sbin/easyai-ai-gateway-release}
+deploy_mode=${AI_GATEWAY_DEPLOY_MODE:-compose}
+case $deploy_mode in
+  compose) default_remote_helper=/usr/local/sbin/easyai-ai-gateway-release ;;
+  kubernetes) default_remote_helper=/usr/local/sbin/easyai-ai-gateway-cluster-release ;;
+  *) echo 'AI_GATEWAY_DEPLOY_MODE must be compose or kubernetes' >&2; exit 1 ;;
+esac
+remote_helper=${AI_GATEWAY_REMOTE_RELEASE_HELPER:-$default_remote_helper}
+production_ssh_key=${AI_GATEWAY_PRODUCTION_SSH_KEY:-}
 action=deploy
 value=
 
@@ -14,8 +21,8 @@ Usage:
   scripts/deploy-production-release.sh --rollback <historical-git-sha>
   scripts/deploy-production-release.sh --status
 
-This command changes production. Run it only after a separate successful
-publish command and explicit user confirmation.
+This command changes production. Run it only after a successful publish
+command and an explicit user request to release or deploy to production.
 EOF
 }
 
@@ -48,9 +55,19 @@ esac
   echo 'AI_GATEWAY_REMOTE_RELEASE_HELPER must be an absolute safe path' >&2
   exit 1
 }
+ssh_options=(-o BatchMode=yes)
+scp_options=(-q -o BatchMode=yes)
+if [[ -n $production_ssh_key ]]; then
+  [[ $production_ssh_key == /* && -f $production_ssh_key && ! -L $production_ssh_key ]] || {
+    echo 'AI_GATEWAY_PRODUCTION_SSH_KEY must be an absolute regular file' >&2
+    exit 1
+  }
+  ssh_options+=(-i "$production_ssh_key")
+  scp_options+=(-i "$production_ssh_key")
+fi
 
 if [[ $action == status ]]; then
-  exec ssh -o BatchMode=yes "$production_host" "$remote_helper status"
+  exec ssh "${ssh_options[@]}" "$production_host" "$remote_helper status"
 fi
 
 if [[ $action == rollback ]]; then
@@ -58,7 +75,8 @@ if [[ $action == rollback ]]; then
     echo 'rollback release must be a full lowercase Git SHA' >&2
     exit 1
   }
-  ssh -o BatchMode=yes "$production_host" "$remote_helper rollback $value"
+  # shellcheck disable=SC2029 # Both arguments are validated safe tokens above.
+  ssh "${ssh_options[@]}" "$production_host" "$remote_helper" rollback "$value"
   echo "production_rollback=PASS release=$value"
   exit 0
 fi
@@ -76,11 +94,11 @@ status_file=$(mktemp)
 
 cleanup() {
   rm -f "$status_file"
-  ssh -o BatchMode=yes "$production_host" "rm -f $remote_manifest" >/dev/null 2>&1 || true
+  ssh "${ssh_options[@]}" "$production_host" rm -f -- "$remote_manifest" >/dev/null 2>&1 || true
 }
 trap cleanup EXIT HUP INT TERM
 
-if ssh -o BatchMode=yes "$production_host" "$remote_helper status" >"$status_file" 2>/dev/null; then
+if ssh "${ssh_options[@]}" "$production_host" "$remote_helper" status >"$status_file" 2>/dev/null; then
   node "$root/scripts/release-manifest.mjs" validate "$status_file" >/dev/null
   current_sha=$(node "$root/scripts/release-manifest.mjs" get "$status_file" sourceSha)
   [[ $current_sha == "$base_sha" ]] || {
@@ -95,6 +113,7 @@ else
   }
 fi
 
-scp -q "$manifest" "$production_host:$remote_manifest"
-ssh -o BatchMode=yes "$production_host" "$remote_helper deploy $remote_manifest"
+scp "${scp_options[@]}" "$manifest" "$production_host:$remote_manifest"
+# shellcheck disable=SC2029 # remote_manifest is derived from a validated full SHA.
+ssh "${ssh_options[@]}" "$production_host" "$remote_helper" deploy "$remote_manifest"
 echo "production_deploy=PASS release=$source_sha"

scripts/publish-release-images.sh

@@ -6,7 +6,14 @@ cd "$root"
 
 components=auto
 production_host=${AI_GATEWAY_PRODUCTION_HOST:-root@110.42.51.33}
-remote_helper=${AI_GATEWAY_REMOTE_RELEASE_HELPER:-/usr/local/sbin/easyai-ai-gateway-release}
+deploy_mode=${AI_GATEWAY_DEPLOY_MODE:-compose}
+case $deploy_mode in
+  compose) default_remote_helper=/usr/local/sbin/easyai-ai-gateway-release ;;
+  kubernetes) default_remote_helper=/usr/local/sbin/easyai-ai-gateway-cluster-release ;;
+  *) echo 'AI_GATEWAY_DEPLOY_MODE must be compose or kubernetes' >&2; exit 1 ;;
+esac
+remote_helper=${AI_GATEWAY_REMOTE_RELEASE_HELPER:-$default_remote_helper}
+production_ssh_key=${AI_GATEWAY_PRODUCTION_SSH_KEY:-}
 registry=${AI_GATEWAY_IMAGE_REGISTRY:-registry.cn-shanghai.aliyuncs.com/easyaigc}
 platform=linux/amd64
 platform_probe_image=${AI_GATEWAY_PLATFORM_PROBE_IMAGE:-docker.m.daocloud.io/library/alpine:3.22}
@@ -18,8 +25,8 @@ Usage:
   scripts/publish-release-images.sh [--components auto|api|web|all]
 
 This command only builds, smoke-tests, and pushes immutable images. It never
-changes production. The generated manifest must be deployed separately with
-scripts/deploy-production-release.sh after explicit confirmation.
+changes production. When the user requested a production release, continue
+with scripts/deploy-production-release.sh without requesting confirmation again.
 
 Environment:
   AI_GATEWAY_PRODUCTION_HOST       SSH target used for read-only release status
@@ -60,6 +67,14 @@ esac
   echo 'AI_GATEWAY_REMOTE_RELEASE_HELPER must be an absolute safe path' >&2
   exit 1
 }
+ssh_options=(-o BatchMode=yes -o ConnectTimeout=8)
+if [[ -n $production_ssh_key ]]; then
+  [[ $production_ssh_key == /* && -f $production_ssh_key && ! -L $production_ssh_key ]] || {
+    echo 'AI_GATEWAY_PRODUCTION_SSH_KEY must be an absolute regular file' >&2
+    exit 1
+  }
+  ssh_options+=(-i "$production_ssh_key")
+fi
 [[ $registry =~ ^[A-Za-z0-9.-]+(:[0-9]+)?(/[A-Za-z0-9._-]+)+$ ]] || {
   echo 'AI_GATEWAY_IMAGE_REGISTRY has an invalid format' >&2
   exit 1
@@ -114,8 +129,8 @@ if [[ -n $status_file_override ]]; then
   [[ -f $status_file_override ]] || fail 'AI_GATEWAY_RELEASE_STATUS_FILE does not exist'
   cp "$status_file_override" "$status_file"
   has_status=1
-elif ssh -o BatchMode=yes -o ConnectTimeout=8 "$production_host" \
-  "$remote_helper status" >"$status_file" 2>/dev/null; then
+elif ssh "${ssh_options[@]}" "$production_host" \
+  "$remote_helper" status >"$status_file" 2>/dev/null; then
   has_status=1
 fi