Huaweicloud Safety

Huawei Cloud Safety

STOP - Do not answer from general knowledge. Follow the procedure below.

Use this skill before any Huawei Cloud action that may expose secrets, change resources, affect cost, change IAM, or expose a network endpoint.

Rules

1. Never ask the user to paste AK/SK, SK, tokens, passwords, private keys, or credential files.
2. Never read local .hcloud or .huaweicloud credential files into the agent context.
3. Never call APIs or CLI operations that return secret values directly into the agent context.
4. Always classify operations before execution:
   • read-only
   • write
   • secret
   • IAM/security
   • cost/payment
   • public exposure
5. For any non-read-only operation, show the exact planned change and wait for explicit user approval.
6. Redact secrets from command output, errors, JSON, logs, and generated reports.
7. If a command contains adminPass, password, token, or any secret-like field, warn that plaintext can remain in shell history. Prefer local-only input, runtime injection, or user-side execution.

Enforcement Layers

• Hooks: hooks/huaweicloud-safety.mjs blocks risky Codex native hook calls without requiring Python; hooks/huaweicloud-safety.py remains for Hermes/Python hook compatibility.
• MCP wrapper: the Node MCP server applies the same safety policy for Codex/OpenCode paths that do not enforce hooks.
• Skills: this document teaches agents the rule before they act.

Write Execution Boundary

• Default path: use huaweicloud_plan_cli_command to produce a reviewed command block, then let the user or host agent execute it after approval.
• Approved tool path: use huaweicloud_run_approved_command only when the exact planned command has been shown and the user explicitly approved that exact command.
• After a write, verify with huaweicloud_run_readonly_command or another read-only check.

Proactive Hook Checks

Before executing, deploying, or handing generated cloud artifacts to the user, run the matching check when available:

• huaweicloud_hook_check_command for shell or KooCLI command text.
• huaweicloud_hook_check_artifacts for generated policy, IaC, config, workflow, or deployment files.
• huaweicloud_hook_check_deploy_plan for sandbox, preview, FunctionGraph, ECS, CCE, APIG, OBS, IAM, or cost-affecting plans.

If the result is deny, repair the command, artifact, or deployment plan before execution. If the result is warn, show the warning to the user and repair or ask for explicit confirmation.

Static Analysis Boundary

Hook checks inspect the command or plan text that is available before execution. If a command builds cloud parameters through shell variables, string concatenation, subshells, encoded payloads, or generated scripts, first expand the final command into reviewable text and run huaweicloud_hook_check_command on that final form. When the final values cannot be determined, stop and ask the user to review the expanded command or use huaweicloud_plan_cli_command before execution.

Safe Alternatives

• Use redacted profile inspection instead of raw hcloud configure show.
• Use runtime secret injection instead of fetching secret values into chat.
• Use read-only verification after a write.
• Prefer reviewed IaC for durable infrastructure changes.