Chapter 16 · Deployment: CI and Hosted Platforms

Chapter goals: Learn to run agents in GitHub Actions / GitLab CI, use Daytona and Render as deployment targets, and consume deployed agent sessions with @flue/sdk.

16.1 Scenarios for Running Agents in CI

Agents in CI are not long-lived services — they "run one task and exit":

① PR opened → agent reviews code and comments
② Scheduled task → agent patrols issues and labels them
③ Release pipeline → agent generates changelog

16.2 GitHub Actions

# .github/workflows/agent-triage.yml
name: agent-triage
on:
  issues:
    types: [opened]        # Trigger on new issue
  schedule:
    - cron: "0 * * * *"   # Patrol hourly
jobs:
  triage:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: actions/setup-node@v4
        with: { node-version: 22 }
      - run: npm ci
      - run: npx flue run agents/triage.ts "${{ github.event.issue.title }}"
        env:
          # Keys from repo Settings → Secrets; never hardcode
          ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}

# Static validation before actual execution (good habit to save tokens)
      - name: dry-run validate agent definition
        run: npx flue run agents/triage.ts --dry-run

16.3 GitLab CI/CD

# .gitlab-ci.yml — equivalent GitLab pipeline
agent-triage:
  image: node:22
  rules:
    - if: $CI_PIPELINE_SOURCE == "schedule"   # Schedule trigger
  script:
    - npm ci
    # Keys from GitLab CI/CD Variables (masked)
    - npx flue run agents/triage.ts "$ISSUE_TITLE"
  variables:
    ANTHROPIC_API_KEY: $ANTHROPIC_API_KEY

CI running注意点

• Each job is a新环境: Durability state must use an external database for retry to接力;
• Jobs have time limits: split长 tasks into segments or use --output to persist中间结果;
• Logs are观测: enable verbose output in CI for easier debugging.

16.4 Hosted Platforms: Daytona and Render

Platform	Role	Key Points
Daytona	Remote sandbox	Provides isolated container execution environment for the agent
Render	Node service hosting	Long-lived process + env var injection + Postgres add-on

# Render uses render.yaml to declare the service
# services:
#   - type: web
#     name: flue-bot
#     runtime: node
#     buildCommand: npm ci && npm run build
#     startCommand: node dist/main.js
#     envVars:
#       - key: ANTHROPIC_API_KEY
#         sync: false     # Enter in the console; not committed
#       - key: DATABASE_URL
#         fromDatabase: { name: flue-db, property: connectionString }
# databases:
#   - name: flue-db       # Attach Postgres for persistence

16.5 @flue/sdk: Consuming Deployed Sessions

After deployment, the frontend or other services read agent sessions via the SDK:

import { FlueClient } from '@flue/sdk';

const client = new FlueClient({
  baseUrl: 'https://your-flue-app.example.com',
});

// Send a message to a session
const result = await client.dispatch('triage', [
  { role: 'user', content: 'Triage issue #42' },
]);
console.log('Session ID:', result.id);

// Read session history
const session = await client.getSession(result.id);
console.log('Messages:', session.messages.length);

16.6 Chapter Summary

• CI scenarios: PR review, scheduled patrol, release changelog — all "run once and exit";
• GitHub Actions / GitLab CI: inject keys via secrets/variables; use --dry-run for static validation;
• Daytona: remote sandbox provider for isolated execution;
• Render: Node service hosting with Postgres add-on for persistence;
• @flue/sdk: consume deployed agent sessions from any client.

🧪 随堂测验

点击你认为正确的选项。答错时会展示正确答案与原因解析。

1. What type of workload is an agent in CI?

2. Why use --dry-run in CI?

3. What is the role of Daytona in the Flue ecosystem?

4. How does @flue/sdk differ from running flue run locally?

🛠️ Hands-On Exercises

1. Create a GitHub Actions workflow that triggers on issue opened; run a triage agent and post a comment with the result.
2. Add a scheduled cron job that patrols issues daily; verify the agent labels stale issues.
3. Deploy a Flue agent to Render with a Postgres add-on; verify the session persists across restarts.

CI/CD mastered. Next, Capstone: Full-Stack Load Testing — wait, this is Flue! Next: 实战一：GitHub 自动分诊 Agent.