Three-Layer Controllability
Execution records, approval rules, and checkpoint restoration provide visibility, control, and reversibility. The sections below describe how to use them and their limits.
Layer 1: Visible
You can see what the agent is doing.
Tool cards, execution receipts, and activity records show operations and results.
Specific manifestations:
- Pre-operation Confirmation: Operations requiring approval show a confirmation card with their type, parameters, and risk. The current approval mode and permission rules determine whether approval is required
- Source Tracing: Every operation request explains "why this operation is being executed," tracing back to which of your instructions triggered it
- Detail Expansion: You can expand to view complete operation parameters, such as file content diff, full command text
- Execution Receipt: A detailed receipt is generated after each task completion, recording the tool call chain, decision basis, and outputs
Visibility doesn't mean information overload. DesireCore decides the level of detail based on risk level—brief prompts for low-risk operations, detailed displays for high-risk operations.
Layer 2: Controllable
You can intervene and modify at any time.
Seeing isn't enough—you also need to be able to influence the agent's behavior.
Specific manifestations:
- Three-Choice Decision: Faced with agent operation requests, you can choose "Allow," "Reject," or "Modify"
- Allow: Let the agent execute the operation
- Reject: Cancel the operation, agent needs to adjust the plan
- Modify: Open the editing panel, manually adjust operation parameters before executing
- Rule Memory: For operations that support remembered rules, select "Allow and Remember"; later operations matching that rule follow its approval policy
- Permission Rule Management: View and manage all remembered permission rules centrally in settings, supporting editing, disabling, and deletion
- Interrupt Capability: During task execution, you can click the stop button to interrupt at any time
Layer 3: Reversible
Inspect and restore local changes through checkpoints.
Open Rewind and Checkpoint, inspect the differences and recovery scope, then decide whether to restore.
| Content | Recovery boundary |
|---|---|
| Conversation records and local files | Limited to records and recoverable files covered by the actual checkpoint |
| Sent email, payments, publications, and other external actions | Must be handled through the relevant service; local rollback cannot undo them |
| Interrupted tasks | Check outputs, execution records, and commands still running before continuing or retrying |
Restoring a conversation does not restart every step. Before continuing a task with external effects, check whether the action already happened to avoid duplicates.
Three Layers Working Together
During execution, check the following in order:
Visible → You know what happened
↓
Controllable → You decide whether to continue
↓
Reversible → Check checkpoint scope, then restore local changes
Before changing approval settings, check which operations the current mode permits.
DesireCore's security design draws on the human-machine collaboration principles of the NIST AI Security Framework, while making localized adaptations for desktop applications. We believe good security mechanisms should make you feel assured, not constrained.