Human Gate Confirmation Mechanism

Human Gate is the core mechanism of DesireCore's security system—forcing a pause and requesting your confirmation before the agent executes high-risk operations.

What Operations Trigger Human Gate

When the agent executes tasks, the following types of operations automatically trigger Human Gate:

Operation Type	Examples	Risk Level
Send Message	Send email, send message to others	Medium - High
Modify File	Write or delete files	Medium
Execute Command	Run scripts or system commands	Medium - High
External Call	Call third-party APIs or services	Medium - High
Financial Operation	Involving payment, transfer, etc.	High - Critical
Publish/Public	Publish content to external platforms	High
Delete Operation	Delete data or resources	High
Permission Change	Modify access permissions	Critical

Not All Operations Need Confirmation

Read-only operations like reading files, analyzing data, and generating drafts usually don't need confirmation. The agent autonomously completes these low-risk operations to improve efficiency.

Confirmation Dialog

When Human Gate is triggered, you will see a confirmation dialog:

┌──────────────────────────────────────────────────────┐
│  ⚠️ Needs Your Confirmation                    [Cancel] │
├──────────────────────────────────────────────────────┤
│                                                        │
│  Operation type: Write file                            │
│  Impact scope: docs/review-report-XX-2025-001.md      │
│  Risk level: 🟡 Medium                                 │
│                                                        │
│  ──────────────────────────────────────────────────    │
│                                                        │
│  Source trace:                                         │
│  According to your instruction "Help me review this procurement contract," the agent plans │
│  to generate a review report and save it to the documents directory.                      │
│                                                        │
│  ▼ View details                                        │
│  ┌──────────────────────────────────────────────┐      │
│  │  File content preview:                        │      │
│  │  # Contract Review Report                     │      │
│  │  ## I. Basic Information                      │      │
│  │  Party A: XX Technology Co., Ltd.             │      │
│  │  ...                                          │      │
│  └──────────────────────────────────────────────┘      │
│                                                        │
│  ☐ Allow and remember (don't ask again for similar operations) │
│                                                        │
│  ┌──────────┐  ┌──────────┐  ┌──────────┐            │
│  │  ✓ Allow  │  │  ✗ Deny  │  │  ✎ Modify  │            │
│  └──────────┘  └──────────┘  └──────────┘            │
│                                                        │
└──────────────────────────────────────────────────────┘

Information Included in Confirmation Dialog

Information	Explanation
Operation Type	What operation the agent wants to perform
Impact Scope	What resources this operation will affect
Risk Level	Risk degree of the operation (Low/Medium/High/Critical)
Source Trace	Why this operation is being performed, originating from which task
Details	Specific operation content (expandable for viewing)

Three Choices

Approve

Confirm execution of this operation:

You: [Click ✓ Allow]
Agent: "Okay, executing now..."

If you check "Allow and remember," similar operations next time (e.g., writing files in the same directory) will be automatically allowed without popping up confirmation.

Deny

Block this operation:

You: [Click ✗ Deny]
Agent: "Okay, I won't execute this operation.
        I will adjust the plan: change to displaying report content in conversation,
        instead of writing to file. Continue?"

The agent will adjust subsequent behavior based on your reason for denial.

Modify

You agree to execute but want to modify some parameters:

You: [Click ✎ Modify]

[Edit panel opens]
File path: docs/review-report-XX-2025-001.md
→ Change to: docs/drafts/review-report-XX-2025-001-draft.md

You: "Save to drafts directory and mark 'draft' in the filename."

Risk Levels and Confirmation Frequency

Different risk level operations trigger confirmation at different frequencies:

Risk Level	Indicator	Confirmation Strategy	Examples
Low	🟢	Auto-allowed by default	Read files, search information
Medium	🟡	Ask by default, can be set to auto-allow	Write files, run safe commands
High	🟠	Always ask, can check "remember"	Send emails, call external APIs
Critical	🔴	Always ask, cannot skip	Delete data, financial operations, permission changes

Critical Risk Operations Cannot Be Auto-Approved

Operations involving finance, deletion, permission changes, and other critical risks will require confirmation every time, even if you check "remember." This is to protect you from unrecoverable losses.

Managing "Remember" Rules

Rules you previously created through "Allow and remember" can be managed in settings:

You: "List all auto-allow rules."

Agent: "Currently have 3 auto-allow rules:

        1. ✅ Allow writing .md files in docs/ directory
           Created: 3 days ago

        2. ✅ Allow executing git related commands
           Created: 1 week ago

        3. ✅ Allow calling translation API
           Created: 2 weeks ago

        You can edit or delete these rules."

You can delete any rule at any time, making the agent revert to asking every time.

Trust is Gradual

When first using the agent, it's recommended to maintain more confirmations. As your trust in it increases, gradually expand the scope of auto-allows. This is like mentoring a new person in reality—first watch them work, confirm there are no problems, then slowly let go.

Next Step

After the agent completes the task, it generates a detailed execution receipt. Go to Execution Receipts to learn how to interpret receipt information.