ECC
ECC 部署：避免重复钩子并安全上线 AI 编程代理
选择单一安装路径，在临时仓库审计生成文件、运行 AgentShield，并准备升级与回滚。
你将学会
- Select and document one ECC installation boundary
- Audit generated hooks, MCP entries, and memory
- Operate upgrades and rollback with least privilege
开始前需要
- Basic Git and command-line usage
- Comfort reading a project README
You can explain the project, run its documented first step, and decide what to verify before adopting it.
先看结论
- Pick one install path per harness and keep an inventory of generated files.
- Test hooks and memory in a disposable repository before granting broad permissions.
- Pin upgrades, run AgentShield, and maintain a rollback and ownership plan.
Choose the deployment boundary
ECC 的部署边界是代理宿主和项目上下文。先选择一个客户端与一种安装方式，在无生产密钥的临时仓库观察 skills、hooks、MCP 和 memory，再运行 AgentShield，最后用 canary 项目验证升级和回滚。
Do not combine a native plugin install with a full manual copy unless the project explicitly documents that migration. Duplicate hooks can run twice, duplicate rules can conflict, and a stale generated file can make a later security review misleading. Keep one install method and record the package version and target paths.
Deploy into a disposable repository first
Run the installer in a throwaway repository with no production credentials. Inspect every generated skill, agent, command, rule, hook, MCP entry, and settings file before enabling it for a real project. Verify which files are project-local and which live in the user profile, then check permissions and ownership so a team member can remove or update them deliberately.
Exercise the full loop with a harmless task: plan a documentation change, run a test command, request a fresh-context review, and observe which hooks write memory or telemetry. Capture command output and paths, but redact tokens, private prompts, and source code before sharing a deployment receipt.
Harden operations and upgrades
Treat hooks, MCP servers, and persisted memory as executable configuration. Restrict destructive shell actions behind human approval, keep credentials in the host's secret store, and review outbound domains. Run `npx -y ecc-agentshield scan --path .` from the official package channel, then address findings before making ECC a team default.
Pin the npm package or repository revision and stage upgrades in a canary project. Compare generated files, hook events, context size, and test outcomes between revisions. If a hook fails, disable only that hook, preserve the log, and fall back to a manual checklist while you investigate rather than bypassing all policy controls.
Define a rollback and ownership plan
A production-ready deployment has an uninstall path, an owner for each rule or skill, and a retention policy for memory and instincts. Keep a manifest of installed components and their source revisions. When a project leaves ECC, remove generated files through the documented uninstall flow and verify that no duplicate plugin or hook remains in the host configuration.
ECC's cross-client adapters do not promise identical behavior. Validate the exact editor, model, shell, and repository permissions your team uses, and keep a human approval gate for dependency upgrades, data deletion, security work, and any command that changes remote state.
如何选择
| 比较维度 | 方案 A | 方案 B |
|---|---|---|
| Best when | You need predictable behavior and easy auditing | You need adaptive optimization and have reliable telemetry |
| Main risk | May leave performance on the table | Can become difficult to explain or debug |
实施步骤
- 1
Select a supported harness and pin the ECC package/revision.
- 2
Run guided setup in a disposable repository and inspect generated files.
- 3
Exercise a harmless task, audit hooks/MCP/memory, and run AgentShield.
- 4
Canary upgrades, document uninstall, and retain human approval for risky actions.
可复制示例
# Choose one documented setup path
npx ecc-universal setup
# Or guided installation for supported harnesses
npx ecc-universal install --guided
# Scan the disposable project before wider rollout
npx -y ecc-agentshield scan --path .常见问题
Can I install the Claude plugin and run the full installer too?
Avoid stacking methods unless the release's migration guide says otherwise; the README warns that duplicate hooks and copied configuration can fire twice.
What should be backed up before an upgrade?
Keep the component manifest, package/repository revision, generated files, hook settings, memory policy, and a recent harmless-run receipt with secrets removed.
资料来源
- ECC README (captured 2026-08-31)来源核查 2026-08-31
- ECC repository来源核查 2026-08-31