fix(agent): 部署 agent 改异步，避免网关 504 + 回滚

POST /api/heicode/agents 原本在请求里同步阻塞 ~30s 等 AM 起 agent。Azure 网关 ~20s 超时 → 504 → 请求 context 被取消 → AM 调用中止 → 部署回滚（agent 建不出）。改为：先把 agent 存为 Pending 立即返回；AM 启动放到 context.Background() 的后台 goroutine（脱离请求 context），成功回填 subdomain/runtime_id/status，失败标记 status=failed 并吊销已铸的模型 key。客户端经已合并的列表刷新看 Pending→running。 go build ./... 通过；controller vet 干净。影响面：仅 Manager(HM) 部署路径。客户端契约：deploy 现在立即返回 Pending（原为阻塞后 running 或 504）；AM 失败在列表里表现为 status=failed（原为同步 RUNTIME_UNAVAILABLE）。客户端本就轮询列表等 running。不改计费/密钥/审计。 Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
2026-06-08 16:43:46 +08:00
parent b3d864b29b
commit a1a5b6037b
1 changed files with 85 additions and 36 deletions
@@ -177,8 +177,36 @@ func HeicodeDeployAgent(c *gin.Context) {
 	env["AGENT_ACCESS_TOKEN"] = agentAccessToken
 	env["HEICODE_AGENT_ID"] = deploymentID

-	// Ask AM to start the agent with the template definition (.md) + env injected.
-	result, err := amStartTemplateAgent(c.Request.Context(), amStartArgs{
+	// Persist the agent as Pending and return immediately; start it in AM in the
+	// background. Starting an agent in AM takes ~30s — longer than the Azure
+	// gateway's ~20s timeout — so doing it inside the request returned 504, which
+	// cancelled the request context, aborted the AM call and rolled the deploy
+	// back. Now the client gets the Pending agent at once and polls
+	// GET /api/heicode/agents (which refreshes live status from AM), seeing
+	// Pending -> running once AM is up (or -> failed if AM start failed).
+	now := agentNow()
+	nowMs := time.Now().UnixMilli()
+	row := model.AgentDeployment{
+		DeploymentID:   deploymentID,
+		UserID:         strconv.Itoa(userID),
+		TemplateID:     req.TemplateID,
+		AccessToken:    sealAgentToken(agentAccessToken),
+		BindingIDsJSON: string(bindingIDsJSON),
+		Status:         "Pending",
+		ModelTokenID:   modelTokenID,
+		CreatedAtText:  now,
+		UpdatedAtText:  now,
+		CreatedAtMs:    nowMs,
+		UpdatedAtMs:    nowMs,
+	}
+	if err := model.DB.Create(&row).Error; err != nil {
+		common.SysLog("HeicodeDeployAgent persist: " + err.Error())
+		revokeAgentModelToken(modelTokenID)
+		agentError(c, "DEPLOYMENT_PERSIST_FAILED", "failed to persist agent")
+		return
+	}
+
+	startTemplateAgentAsync(deploymentID, modelTokenID, amStartArgs{
 		ManagerDeploymentID: deploymentID,
 		UserID:              strconv.Itoa(userID),
 		TemplateKey:         tpl.TemplateKey,
@@ -187,43 +215,64 @@ func HeicodeDeployAgent(c *gin.Context) {
 		Env:                 env,
 		CallbackURL:         agentRuntimeCallbackURL(),
 	})
-	if err != nil {
-		revokeAgentModelToken(modelTokenID) // don't leak the minted key
-		agentError(c, "RUNTIME_UNAVAILABLE", "failed to start agent: "+err.Error())
-		return
-	}

-	now := agentNow()
-	nowMs := time.Now().UnixMilli()
-	row := model.AgentDeployment{
-		DeploymentID:        deploymentID,
-		UserID:              strconv.Itoa(userID),
-		TemplateID:          req.TemplateID,
-		Subdomain:           result.Subdomain,
-		AccessToken:         sealAgentToken(agentAccessToken),
-		BindingIDsJSON:      string(bindingIDsJSON),
-		RuntimeDeploymentID: result.RuntimeID,
-		Status:              firstNonEmpty(result.Status, "running"),
-		ModelTokenID:        modelTokenID,
-		CreatedAtText:       now,
-		UpdatedAtText:       now,
-		CreatedAtMs:         nowMs,
-		UpdatedAtMs:         nowMs,
-	}
-	if err := model.DB.Create(&row).Error; err != nil {
-		common.SysLog("HeicodeDeployAgent persist: " + err.Error())
-		revokeAgentModelToken(modelTokenID)
-		// We started an agent in AM but failed to record it: roll back the
-		// orphan so it does not leak/keep running with no Manager record.
-		if strings.TrimSpace(result.RuntimeID) != "" {
-			if delErr := amDeleteTemplateAgent(c.Request.Context(), result.RuntimeID); delErr != nil {
-				common.SysLog("HeicodeDeployAgent orphan cleanup failed: " + delErr.Error())
+	common.ApiSuccess(c, templateAgentResponse(row))
+}
+
+// startTemplateAgentAsync starts the template agent in AM off the request path
+// and records the outcome. The deploy handler returns the Pending agent at once
+// so it never blocks ~30s on AM (which exceeded the Azure gateway timeout → 504
+// + rollback). Uses context.Background() because the request context is gone
+// once the handler returned; the AM HTTP call is still bounded by amTemplateDo's
+// own client timeout. The client observes Pending -> running (or -> failed) by
+// polling the agent list, which refreshes live status from AM.
+func startTemplateAgentAsync(deploymentID string, modelTokenID int, args amStartArgs) {
+	go func() {
+		result, err := amStartTemplateAgent(context.Background(), args)
+		if model.DB == nil {
+			return
+		}
+		now := agentNow()
+		nowMs := time.Now().UnixMilli()
+		if err != nil {
+			common.SysLog("startTemplateAgentAsync: AM start failed for " + deploymentID + ": " + err.Error())
+			reason := err.Error()
+			if len(reason) > 480 {
+				reason = strings.ToValidUTF8(reason[:480], "")
+			}
+			// Mark failed and revoke the minted model key (the agent never started).
+			_ = model.DB.Model(&model.AgentDeployment{}).
+				Where("deployment_id = ?", deploymentID).
+				Updates(map[string]any{
+					"status":          "failed",
+					"failure_reason":  reason,
+					"updated_at_text": now,
+					"updated_at_ms":   nowMs,
+				}).Error
+			revokeAgentModelToken(modelTokenID)
+			return
+		}
+		res := model.DB.Model(&model.AgentDeployment{}).
+			Where("deployment_id = ?", deploymentID).
+			Updates(map[string]any{
+				"subdomain":             result.Subdomain,
+				"runtime_deployment_id": result.RuntimeID,
+				"status":                firstNonEmpty(result.Status, "running"),
+				"updated_at_text":       now,
+				"updated_at_ms":         nowMs,
+			})
+		if res.Error != nil {
+			common.SysLog("startTemplateAgentAsync: persist result failed for " + deploymentID + ": " + res.Error.Error())
+			return
+		}
+		if res.RowsAffected == 0 && strings.TrimSpace(result.RuntimeID) != "" {
+			// Record was deleted (user cancelled) before AM finished starting — clean
+			// up the orphan agent in AM so it doesn't keep running untracked.
+			if delErr := amDeleteTemplateAgent(context.Background(), result.RuntimeID); delErr != nil {
+				common.SysLog("startTemplateAgentAsync orphan cleanup failed: " + delErr.Error())
 			}
 		}
-		agentError(c, "DEPLOYMENT_PERSIST_FAILED", "failed to persist agent")
-		return
-	}
-	common.ApiSuccess(c, templateAgentResponse(row))
+	}()
 }

 // findUserTemplateAgent loads a deployed template agent owned by the caller.