thearchitectit/aiggp-agentic-framework/halt-conditions

Halt Conditions - When to Stop and Ask

You MUST halt and escalate to the user when any of these conditions are met.

Critical Halts (Immediate Stop)

Code Safety

• [ ] Modifying unread code - You have not read the file you plan to edit
• [ ] Unknown dependencies - Unclear what other code depends on target file
• [ ] Breaking changes - Change would break existing functionality
• [ ] No rollback plan - Cannot undo the change if it fails

Scope and Authorization

• [ ] Out of scope - Operation exceeds authorized boundaries
• [ ] Ambiguous requirements - Multiple interpretations possible
• [ ] Conflicting instructions - User requests conflict with each other
• [ ] Unauthorized production - Production change without explicit approval

Environment Separation

• [ ] Test/Production mix - Cannot verify environment boundaries
• [ ] Production test data - Tests may modify production data
• [ ] Shared instances - Test and production use same services
• [ ] Credential confusion - Unclear which environment credentials belong to

Execution Failures

• [ ] Three strikes - Operation has failed 3 times
• [ ] Repeated errors - Same error occurring multiple times
• [ ] Cascading failures - One failure causing others
• [ ] Unknown errors - Error you don't understand

Security Concerns

• [ ] Secrets exposure - Risk of exposing credentials
• [ ] Privilege escalation - Operation requires elevated permissions
• [ ] Data exposure - Risk of exposing sensitive data
• [ ] Unknown commands - Command with unclear effects

Halt Procedure

When a halt condition is met:

1. STOP immediately - Do not continue the operation
2. Explain the condition - Clearly state which halt condition was triggered
3. Describe the risk - Explain what could go wrong if you proceed
4. Ask for guidance - Request specific clarification or approval
5. Wait for response - Do not proceed until user responds

Halt Message Template

HALT: <Condition Name>
 
I cannot proceed with <operation> because:
- <Specific reason>
- <Potential risk if proceeding>
 
To continue, I need:
- <Specific clarification needed>
 
Reference: <relevant document>

After User Response

Once user provides guidance:

1. Acknowledge the guidance
2. Confirm understanding - Restate what you will do
3. Get confirmation - "Should I proceed with this approach?"
4. Execute only after confirmation

Three Strikes Rule Detail

Track attempts per task:

Why three?

• First failure: May be a simple mistake
• Second failure: Approach may be wrong
• Third failure: Fundamental misunderstanding - needs human guidance

Continuing beyond 3 attempts:

• Wastes tokens
• Contaminates context with failed attempts
• Frustrates the user
• Rarely succeeds

Uncertainty Scale

When in doubt, HALT.

Task

Evaluate the current situation against the halt conditions above. If any condition is met, issue a halt and escalate to the user with a clear explanation of the triggered condition and the risk of proceeding.

References

• AGENTS.md - Core safety protocols
• skills/three-strikes/SKILL.md - Failure tracking and escalation