🛡️

Safety, Trust, and Auditability

Approve actions, preview changes, and track agent activity

Overview

Vault Copilot is designed with safety, trust, and transparency as core principles. Every operation the AI performs is visible, reviewable, and requires your approval. You maintain complete control over what happens in your vault through permission gates, preview mechanisms, and comprehensive audit trails.

Safety Mechanisms

Permission Gates

All vault-modifying operations require approval:

  • Read operations: Approve which notes AI can access
  • Write operations: Confirm note creation and updates
  • Delete operations: Explicit confirmation required
  • External calls: Approve MCP and API requests

Operation Preview

See exactly what will happen before approving:

AI wants to: Update note "Project Alpha"
Changes:
  - Frontmatter: status: "in progress" → "completed"
  - Add section: "Final Summary"
  - Update tags: +#completed

[Preview] [Approve] [Deny]

Granular Permissions

Control operations at different levels:

  • Per-operation approval: Approve each action individually
  • Batch approval: Approve related operations together
  • Auto-approve patterns: Set rules for trusted operations
  • Always deny patterns: Block specific operation types

Execution Budgets

Limit the scope of AI operations:

  • Max operations per session: Cap total actions (e.g., 10 operations)
  • Max file modifications: Limit writes per conversation
  • Max external calls: Restrict API calls per session
  • Token budgets: Limit AI processing costs

Trust Building

Transparent Operations

Every operation is fully visible:

  • Operation description: Plain language explanation
  • Parameters: All inputs and outputs shown
  • Affected files: List of files that will change
  • Change preview: Diff view of modifications

Predictable Behavior

AI operations follow consistent patterns:

  • Type safety: All parameters validated
  • Schema compliance: Operations match definitions
  • Error handling: Graceful failures with explanations
  • Rollback capability: Undo recent changes

Human-in-the-Loop

You’re always in control:

  • Review before execution: Nothing happens without approval
  • Interrupt capability: Stop operations mid-execution
  • Manual override: Take control at any point
  • Abort sessions: End conversations immediately

Audit Trails

Conversation Logs

Complete record of all interactions:

  • Message history: All prompts and responses
  • Context sent: What data was shared with AI
  • Operations performed: Every action taken
  • Timestamps: When events occurred
  • Session metadata: Model, agent, settings used

Operation Logs

Detailed record of vault modifications:

2026-02-12 10:15:23 - Operation: update_note
  File: projects/project-alpha.md
  Changes: 
    - Frontmatter: status updated
    - Section added: "Summary"
  Status: Approved and executed
  User: You

Tracing and Diagnostics

Advanced debugging and review:

  • SDK traces: Full GitHub Copilot SDK logs
  • Tool calls: All tool invocations and results
  • Model reasoning: Thinking process (on supported models)
  • Performance metrics: Response times, token usage

Export and Review

Access your audit data:

  • Export logs: Save as JSON or Markdown
  • Search history: Find specific operations
  • Filter by date: Review actions in time range
  • Session playback: Review conversation flow

Safety Features

Preview Mechanisms

Multiple ways to preview changes:

Inline Preview

See changes directly in approval prompt:

- status: in progress
+ status: completed

Side-by-Side Diff

Compare before and after:

Before              | After
--------------------|--------------------
status: in progress | status: completed
                    | summary: Project done

Full-screen detailed view:

  • Line-by-line changes
  • Syntax highlighting
  • Expand/collapse sections
  • Accept/reject individual changes

Rollback Capabilities

Undo recent changes:

  • Single operation rollback: Undo last action
  • Batch rollback: Undo entire conversation’s changes
  • Time-based rollback: Revert to state before timestamp
  • Selective rollback: Choose which operations to undo

Safe Defaults

Conservative default settings:

  • Manual approval required: No auto-approval by default
  • Preview enabled: Always show changes before applying
  • Execution budgets: Reasonable limits set
  • Read-only mode: Available for exploration without risk

Error Handling

Graceful Failures

When operations fail:

  • Clear error messages: Explain what went wrong
  • Recovery suggestions: How to fix or work around
  • Partial completion: Success status of batch operations
  • No side effects: Failed operations don’t corrupt vault

Validation

All operations validated before execution:

  • Parameter validation: Type and format checking
  • File existence: Verify files before modifying
  • Permission checks: Ensure operation is allowed
  • Conflict detection: Identify competing changes

Safe Execution

Protection during operation:

  • Atomic operations: All or nothing execution
  • File locking: Prevent concurrent modifications
  • Backup creation: Automatic backups before major changes
  • Checkpointing: Save state before risky operations

Trust Indicators

Operation Confidence

AI indicates confidence in operations:

High confidence: "Update status field" ✅
Medium confidence: "Infer project completion from recent notes" ⚠️
Low confidence: "Guess which project this relates to" ❌

Review Recommendations

Suggestions for when to review carefully:

  • High-impact operations: Deleting files, bulk updates
  • Uncertain operations: Low-confidence inferences
  • First-time operations: New types of actions
  • Cascading changes: Operations affecting multiple files

Execution Summary

After operations, receive summary:

Session Summary:
✅ 3 notes updated successfully
✅ 1 note created
⚠️ 1 operation skipped (permission denied)
❌ 0 failures

Time: 2 minutes
Operations: 5 total (4 approved, 1 denied)
Files modified: 4

Best Practices

Start Conservative

Begin with strict safety settings:

  1. Enable preview for all operations
  2. Require manual approval
  3. Set low execution budgets
  4. Review every operation

Build Trust Gradually

As you get comfortable:

  1. Identify safe operation patterns
  2. Create auto-approve rules for trusted actions
  3. Increase execution budgets
  4. Batch-approve related operations

Regular Audits

Periodically review:

  • Recent operations: What has AI done?
  • Denied operations: What did you block?
  • Error patterns: Common failure modes?
  • Usage patterns: How is AI being used?

Safety Checklist

Before approving operations:

  • ✅ Understand what will happen
  • ✅ Review affected files
  • ✅ Check change preview
  • ✅ Verify intent matches expectation
  • ✅ Consider reversibility
  • ✅ Approve or deny

Advanced Safety

Operation Sandboxing

Test operations safely:

  • Dry-run mode: Simulate without executing
  • Test vault: Separate vault for experiments
  • Preview-only sessions: Explore without risk
  • Snapshot/restore: Save vault state before risky operations

Custom Safety Rules

Define your own safety policies:

safety_rules:
  - operation: delete_note
    require: explicit_confirmation
    warn: "Deleting notes is permanent"
  
  - operation: update_frontmatter
    auto_approve: true
    if: field in ["tags", "status"]
  
  - operation: create_note
    budget: 5
    per: session

Multi-User Scenarios

For shared vaults:

  • User attribution: Track who approved operations
  • Permission levels: Different users, different access
  • Approval workflows: Require multiple approvals
  • Notification systems: Alert team of changes

Getting Started

  1. Review Settings → Safety & Permissions
  2. Enable preview for all operations
  3. Set conservative execution budgets
  4. Try safe operations (read-only queries)
  5. Gradually approve write operations
  6. Review audit logs regularly
  7. Adjust settings based on experience