Tool Consent & Safety

Caide rejects the notion that an autonomous agent should execute unchecked shell commands or wipe project directories without supervision. We implement a multi-layered safety boundary to ensure you stay in total control.

---

1. Explicit Consent Gates

By default, Caide requires explicit user consent before executing any high-risk action:

Tool Action	Consent Required
Read Files & Search Code	No (Safe)
Edit Existing Files	Configurable (Default: Auto)
Shell Command Execution	Yes (Default: Prompt)
Database Migrations	Yes (Default: Prompt)
MCP External Tools	Yes (Default: Prompt)
Package Installations	Yes (Default: Prompt)

When a tool requires approval, the chat stream renders an interactive Approval Card:

• Highlights the exact command line or SQL query.
• Explains the agent's stated rationale.
• Shows the target working directory.
• Features Approve, Reject, or Edit Command options.

---

2. Shell Command Blocklists

Certain commands pose catastrophic risks to developers' machines (e.g. wiping root, modifying global SSH keys, altering firewall rules). Caide enforces an immutable internal blocklist:

• Recursive deletion commands targeting root or system directories (rm -rf /, rm -rf ~).
• Fork bombs and memory exhaustion loops.
• Modifications to sensitive platform credentials (~/.ssh, ~/.aws/credentials, /etc/shadow).
• Blind privilege escalation (sudo, doas) without explicit interactive terminal handoff.

If the model attempts to invoke a blocked pattern, the turn runner rejects the call immediately and advises the model to use safer alternatives.

---

3. Version Checkpoints & 1-Click Rollback

Before every turn begins and after every tool commit:

1. Caide records a snapshot in its internal Git version tree.
2. If the agent makes an unwanted edit, click Restore Checkpoint in the top navigation bar to immediately restore your project files and database state to the exact condition prior to that turn.