safe-action
Measure twice, cut once. Before taking destructive, irreversible, or high-stakes actions, this skill runs a three-part safety pre-flight: risk assessment, reversibility check, and timing awareness.
Combines three AgentUtil services — think (safety checklists), undo (reversibility intelligence), and context (situational timing) — into a single decision workflow.
When to Activate
Use this skill before:
- Deleting anything — repositories, databases, user accounts, files, branches, DNS records
- Deploying to production — code releases, infrastructure changes, migrations
- Mass operations — bulk updates, batch deletes, sending to many recipients
- Permission changes — revoking access, changing roles, modifying security settings
- Infrastructure changes — scaling down, terminating instances, modifying load balancers
- Financial actions — processing refunds, transferring funds, modifying billing
Do NOT use for: read-only operations, local development changes, or actions the user has explicitly marked as low-risk.
Workflow
Step 1: Safety checklist (think)
Get a pre-flight checklist for the action type.
MCP (preferred): If @agentutil/think-mcp is available:
think_check({ action: "database_migration", description: "Drop column users.email in production" })
HTTP fallback:
curl -X POST https://think.agentutil.net/v1/check \
-H "Content-Type: application/json" \
-d '{"action": "database_migration", "description": "Drop column users.email in production"}'
Note the risk_level and any critical severity items in the checklist.
Step 2: Reversibility check (undo)
Assess whether the action can be undone and find safer alternatives.
MCP (preferred): If @agentutil/undo-mcp is available:
undo_check({ platform: "GitHub", resource: "repository", action: "delete" })
HTTP fallback:
curl -X POST https://undo.agentutil.net/v1/check \
-H "Content-Type: application/json" \
-d '{"platform": "GitHub", "resource": "repository", "action": "delete"}'
Check reversibility_level and safer_alternatives.
Step 3: Timing check (context)
Verify this is a good time — not a holiday, after-hours, or during a known outage.
MCP (preferred): If @agentutil/context-mcp is available:
context_check({ action: "deploy", timezone: "America/New_York", platform: "github" })
HTTP fallback:
curl -X POST https://context.agentutil.net/v1/check \
-H "Content-Type: application/json" \
-d '{"action": "deploy", "timezone": "America/New_York", "platform": "github"}'
Check risk_level and any holidays or business events.
Step 4: Synthesize and decide
Combine results into a clear recommendation:
| Scenario | Action |
|---|---|
| All three return low risk | Proceed. Briefly note key checklist items completed. |
Think returns high risk | Pause. Present the critical checklist items to the user. Ask for explicit confirmation. |
Undo shows irreversible | Warn clearly: "This action cannot be undone." Present any safer_alternatives (e.g., archive instead of delete). |
Undo shows time_limited | Warn: "This can only be undone within [recovery_window]." |
Context returns high risk | Warn about timing: "It's [holiday/after-hours/during known incident]. Consider waiting." |
| Multiple services flag risk | Present all findings together. Recommend the safest path. Do NOT proceed without explicit user approval. |
Presentation format
When reporting findings to the user, use this structure:
**Pre-flight check for [action]:**
Safety: [risk_level] — [summary of critical items]
Reversibility: [reversibility_level] — [recovery_mechanism or "no recovery"]
Timing: [risk_level] — [summary or "all clear"]
[Recommendation: proceed / proceed with caution / suggest alternative / strongly advise against]
Data Handling
This skill sends action descriptions, platform names, and timezone identifiers to three external APIs. No user-generated content (documents, messages, credentials) is transmitted — only structured action metadata.
Pricing
Each sub-check costs $0.001-$0.003 via x402 (USDC on Base). A full pre-flight (all three checks) costs ~$0.004-$0.008. Free tiers available for exploration:
- think: GET /v1/actions (free)
- undo: GET /v1/platforms (free)
- context: GET /v1/calendar (free)
Privacy
No authentication required for free endpoints. No personal data collected. Rate limiting uses IP hashing only. Action descriptions are not stored beyond immediate processing.