ECC
ECC 功能与快速上手：把 Skills 与 Hooks 接入安全代理循环
在临时仓库选择一个客户端，观察 skills、hooks、memory，运行 AgentShield 并验证回滚。
你将学会
- Run a bounded ECC task loop
- Inspect hook, MCP, and memory behavior
- Build a security and rollback baseline
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
- Start with one harness, one harmless task, and a pinned ECC revision.
- Inspect hooks, MCP, and memory as executable state with explicit permissions.
- Use AgentShield, manifests, and rollback to make the baseline repeatable.
Choose one harness and one task
ECC 的最小实验是一个固定版本、一个客户端和一个无害任务。记录技能选择、生命周期钩子、MCP、记忆写入与权限，再用 AgentShield 扫描生成文件，最后验证卸载和回滚。
Record the package or commit, host, shell, model, and repository permissions. The same skill name can have different lifecycle behavior across Claude Code, Codex, Cursor, OpenCode, Gemini, Zed, and Copilot, so the host is part of the test fixture.
Exercise skills, hooks, and memory separately
Run a planning command and inspect which skill or specialist agent is selected. Then execute one test command and observe lifecycle hooks before and after the tool call. Finally, inspect whether memory or instincts were written, what fields they contain, and where they persist.
Keep the first run read-only and deny network access unless the task requires it. Hooks and MCP servers are executable configuration; a prompt that looks like documentation can still cause a shell command, file write, or outbound request. Human approval should remain in the loop for destructive actions.
Use the security surface as a feature
Run the documented AgentShield scan against the disposable project and review findings for prompts, hooks, permissions, secrets, and agent files. GateGuard-style safeguards can block risky commands, but a team still needs a clear reason, an approval path, and a way to reproduce the blocked event.
Redact prompts, tokens, and source snippets before sharing logs. Keep a manifest of generated files and MCP endpoints so uninstall can remove every surface instead of leaving a hidden hook behind.
Turn the experiment into a repeatable baseline
Replay the same task after pinning an ECC upgrade. Compare selected skills, hook events, context size, memory writes, test outcomes, and network destinations. If a client adapter lacks a capability, mark it as advisory rather than assuming parity.
Finish by running the documented uninstall or cleanup flow, removing generated files, revoking temporary credentials, and verifying the host behaves normally. A successful quickstart leaves an evidence receipt and a rollback, not a permanently enabled black box.
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
Create a disposable repository and choose one supported harness.
- 2
Run a planning/test task while recording skills, hooks, and memory writes.
- 3
Scan with AgentShield and review permissions, secrets, and endpoints.
- 4
Replay after an upgrade, then uninstall and verify a clean rollback.
可复制示例
npx ecc-universal setup
# inspect generated files before enabling production work
find . -path '*/.claude/*' -o -path '*/.codex/*'
# security review
npx -y ecc-agentshield scan --path .常见问题
Can I enable every ECC skill at once?
Prefer a selective subset first; smaller context and fewer permissions make behavior and review easier to understand.
What should I do when an adapter lacks a hook?
Mark the behavior as advisory, document the parity gap, and keep a human checklist until the host supports an equivalent event.
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