Rollback Protocol
HALT
When something goes wrong — regression, broken build, incorrect merge, unexpected behavior — the manager declares HALT.
Problem detected
↓
Manager declares HALT
↓
All agents STOP (no new work)
↓
Investigation phase (user + manager)
↓
Root cause identified
↓
Manager proposes recovery plan
↓
User approves plan
↓
Resume work
When to HALT
| Trigger | Who declares |
|---|---|
| QA reports CRITICAL regression | Manager |
| Code-reviewer finds correctness bug post-merge | Manager |
| Shadow reveals fundamental KB error | Manager |
| Build breaks on main branch | Manager |
| User detects any problem | User directly |
Investigation Phase
During HALT, the user and manager work together:
- Identify the problem. What broke? When? Which commit?
- Scope the damage. What's affected? Just one module or cascading?
- Find root cause. Why did it happen? What was missed?
- Assess options:
- Git revert: Undo the specific commit(s) that caused the problem.
- Fix forward: Write a targeted fix for the issue.
- Shadow block revert: If a shadow merge caused the issue, revert
to the
#else(original) side. - Propose recovery plan. Manager drafts, user approves.
- Resume. Execute the plan, verify fix, then resume normal work.
Rules
No one works during HALT
Only the user and manager investigate. All other agents stop.
- No shortcuts. Don't skip investigation to "just fix it."
- Log the incident. After resolution, add a HALT entry to SYNCLOG:
## HALT --- YYYY-MM-DD
### Problem
{what went wrong}
### Root Cause
{why it happened}
### Resolution
{what was done to fix it}
### Lesson
{what to do differently — feeds back to docs/skills}
- Update process if needed. If the HALT reveals a gap in the framework (missing review check, inadequate test coverage), the lesson becomes a teamdev-developer task.