ProxyWarden/.agent/skills/security-hardening/SKILL.md

Skill: Security Hardening

Когда использовать

Используй этот skill при изменениях, связанных с Tauri security, CSP, secrets, subscription fetching, process execution, PowerShell, temporary files, install/uninstall, file writes, generated configs, service control, logs, diagnostics.

Threat model

ProxyWarden — desktop app that can influence network routing and run elevated Windows operations. Главные риски:

• leaking proxy/subscription credentials;
• unsafe local/network fetches;
• unsafe generated elevated PowerShell scripts;
• unscoped process execution;
• corrupting generated/service configs;
• deleting wrong directories;
• stale component status causing wrong actions;
• XSS/webview compromise amplified by privileged backend commands.

Non-negotiables

• tauri.conf.json must not use "csp": null as final state.
• Do not add broad Tauri shell permissions.
• Do not execute user-controlled strings as commands.
• Do not log full subscription URLs, proxy passwords, userinfo, access tokens, or outbound configs with credentials.
• Do not recursively delete directories based only on fuzzy name matching.
• Do not treat fuzzy-detected services as managed without verification.

CSP guidance

Prefer a restrictive CSP such as:

"security": {
  "csp": "default-src 'self'; img-src 'self' asset: data:; style-src 'self' 'unsafe-inline'; script-src 'self'"
}

Tighten further when possible. If inline styles are removed, remove 'unsafe-inline'.

Secret redaction

For URLs use a parser, not string splitting. Redacted display should include only:

• scheme;
• host;
• port if useful;
• generic path marker if necessary.

Never show:

• username;
• password;
• query string;
• fragment;
• subscription token path;
• full proxy credentials.

Bad:

https://user:password@example.com/...

Good:

https://example.com/...

Subscription fetch hardening

• Add connect/read timeout.
• Accept only http and https unless explicitly designed otherwise.
• Consider blocking loopback/private/link-local/multicast/metadata addresses by default.
• Add explicit allow-local option only if needed.
• Do not follow redirects into blocked address ranges without re-check.
• Avoid storing remote body in logs.

Temp/elevated script hardening

Runtime-generated elevated scripts must:

• use unpredictable names, preferably UUID/random;
• be written to a safe controlled directory where possible;
• set restrictive ACL when practical;
• be generated from static templates with escaped parameters;
• avoid including secrets in command line args;
• be cleaned up best-effort;
• fail closed if path validation fails.

Timestamp-only temp names are not enough.

Atomic writes

For config files used by services:

1. Write to temp file in same directory.
2. Validate temp file if validator exists.
3. Backup current file.
4. Rename temp to final.
5. On failure, preserve backup and return actionable error.

Safe delete checklist

Before recursive delete:

• Is path absolute?
• Is it under expected managed root?
• Does it contain ProxyWarden marker metadata?
• Does service PathName point inside this directory?
• Is it not drive root, user profile root, Desktop, ProgramData root, Windows directory, temp root?
• Is user action explicit?

If answer is unclear, do not delete.

Final report expectations

When touching security-sensitive code, report:

• what threat was addressed;
• what was hardened;
• what remains unverified;
• whether any secrets could appear in logs/UI;
• whether Windows elevated path was tested.

Как отчитываться

Перед финальным ответом применить .agent/skills/communication-reporting/SKILL.md и .agent/checklists/communication.md.

Минимум для нетривиальной задачи:

• короткая сводка;
• таблица файлов Файл / Что изменилось / Зачем;
• важные места без пересказа каждой строки;
• что проверено;
• что не проверено;
• конкретные риски.