Instruction file imported from LiuChang-ao/XI-Resource-Server (.cursor/rules/01-security.mdc). Copyright stays with the author.

Security Rules (MUST FOLLOW)

Transport security

• All HTTP and WSS MUST use TLS.
• Use WSS on port 443 for maximum compatibility.
• Reject plaintext WS/HTTP in production mode.

Agent Authentication (MVP -> Future)

MVP

• Agent authenticates with agent_id + agent_token (pre-shared secret) via:
  • WSS connection query/header, or first Register message.
• Tokens must never be logged in plaintext.
• Store token hashes server-side (bcrypt/argon2).

Future (recommended)

• Upgrade to mTLS:
  • per-agent client certificate, server enforces allowlist + revocation.

Authorization

• Agent can only:
  • request jobs
  • renew lease for assigned jobs
  • report status for its jobs
• Server must validate:
  • job_id belongs to the agent (ASSIGNED/RUNNING) before accepting renew/status.

OSS access

• NEVER place permanent OSS AK/SK on agents.
• Use ONE of:
  • presigned URLs with short expiry (10-30 min), OR
  • STS temporary credentials with least-privilege policy per job/prefix.
• Each job should only grant:
  • read access to inputs/{job_id}/... (or specified keys)
  • write access to jobs/{job_id}/{attempt_id}/...

Secrets management

• Cloud secrets live in environment variables or a secret manager.
• Provide .env.example with placeholders only.
• Agent config can include token, but must support:
  • reading from Windows Credential Manager or encrypted file (optional later).
  • never print token in logs.

Replay / tampering defenses (MVP)

• Each agent->server message should include:
  • timestamp and request_id
• Server may reject messages with large clock skew (configurable).

Hard prohibitions

• Do not implement any "open port on workstation" feature.
• Do not add any "upload file to ECS then forward" logic.
• Do not add debug endpoints without auth.