Before each new task in a long session, re-runs the last N passing checkpoints captured by verify-orchestrator as a smoke suite, catching regressions caused by recent edits before they compound.
Installs into .claude/skills of the current project.
Are you the author of Regression Sentinel?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/rmazrim-regression-sentinel)
---
name: Regression Sentinel
description: Before each new task in a long session, re-runs the last N passing checkpoints captured by verify-orchestrator as a smoke suite, catching regressions caused by recent edits before they compound.
metadata:
source: skills/regression-sentinel/regression-sentinel.md
---
# Regression Sentinel
## Prerequisites & Dependencies
- Python 3.10+ (for checkpoint orchestration)
- Access to verification output from `verify-orchestrator` (exit codes, assertion outputs)
- No external runtime dependencies
## Execution Steps
1. **Capture checkpoints**: After each task completion where `verify-orchestrator` returns `VERIFIED`, persist the verification result as a checkpoint:
- Task ID or description
- Verification status (`VERIFIED`)
- Exit code and output snippet
- Timestamp
2. **Store checkpoints**: Write checkpoint data to `regression-checkpoints.json` at the repository root:
```json
[
{"task_id": "task-1", "status": "VERIFIED", "exit_code": 0, "output": "...", "timestamp": "2026-09-11T10:00:00Z"},
{"task_id": "task-2", "status": "VERIFIED", "exit_code": 0, "output": "...", "timestamp": "2026-09-11T10:30:00Z"}
]
```
3. **Before each new task**: Read the latest N checkpoints (default N=5) from `regression-checkpoints.json` and execute the same verification commands as smoke tests.
4. **Classify results**:
- All checkpoints pass → proceed with new task as normal
- Any checkpoint fails → halt the new task and surface the regression with diff highlighting
- No checkpoints exist → first task in session; skip smoke suite
5. **Regression report**: On detection, generate a diff report showing:
- What changed between the last VERIFIED state and current state
- Which checkpoint(s) failed and why
- Recommended remediation steps
## Checkpoint Workflow Example
```
Task 1 completes → verify-orchestrator → VERIFIED → save checkpoint
Task 2 begins → regression-sentinel reads 1 checkpoint → runs same test → passes → proceed
Task 3 begins → regression-sentinel reads 2 checkpoints → one fails → halt, show diff
```
## Windows PowerShell Notes
- Checkpoints stored as `regression-checkpoints.json` with UTF8NoBOM encoding
- Smoke test commands extracted from prior `verify-orchestrator` output
- Use `Get-ChildItem` for mtime-based checkpoint pruning (remove older than N days)