safety-net

Agent Skills

Curated plugins for Claude Code - skills, hooks, and tools.

Quick Start

# Add the marketplace (once)
/plugin marketplace add https://github.com/andreasasprou/agent-skills

# Install plugins
/plugin install oracle@andreas-agent-skills
/plugin install safety-net@andreas-agent-skills
/plugin install opensrc@andreas-agent-skills

How It Works

Claude Code plugins extend the agent with skills (new capabilities) and hooks (intercept actions).

Type	Purpose	Example
Skill	Add new commands/capabilities	Oracle adds /oracle for consulting GPT-5.2
Hook	Intercept and validate actions	Safety-net blocks destructive commands

---

Plugins

Oracle

Strategic technical advisor with two complementary modes for different types of questions.

Core Question: "Is the truth external, or is it in our code?"

Mode	Model	Best For
Repo	GPT-5.2 xhigh (Codex SDK)	Exploring codebase, finding gaps, code review
Web	5.2 Thinking (default)	Best practices, library comparisons, current docs
Web	GPT-5.2-pro (escalation)	Complex multi-source research

Repo Mode - when the answer is in your codebase:

/oracle "What's causing the race condition in the queue processor?"
/oracle "Audit the blast radius before I refactor the payment service"
/oracle "Review the uncommitted changes for security issues"

Web Mode - when you need external knowledge:

/oracle "Drizzle vs Prisma for heavy read loads - what are teams saying?"
/oracle "Current security gotchas with Google OAuth?"
/oracle "Compare Socket.io vs Ably vs Pusher for a team of 3"

Both in Parallel - when comparing your code against current standards:

/oracle "Is our auth middleware following current OWASP guidelines?"
/oracle "Does our error handling match RFC 7807?"

Features:

• Repo mode: Codex SDK with read-only sandbox, explores files autonomously
• Web mode: @steipete/oracle with browser automation for Deep Research
• Intelligent routing based on question type
• Parallel execution for comprehensive analysis
• Asks clarifying questions when routing is unclear

Second Opinion Workflow:

Best practice: review diffs and tests, not raw code.

1. Primary agent writes code
2. Package context (diff + key files + test results)
3. Oracle reviews (bugs, edge cases, missing tests)
4. Apply fixes, run tests
5. Repeat until critique converges

Response Structure:

### Essential
- **Bottom Line**: Direct answer (1-2 sentences)
- **Action Plan**: Numbered next steps
- **Effort Estimate**: Quick (<1h) | Short (1-4h) | Medium (1-2d) | Large (3d+)

### Expanded (when relevant)
- **Reasoning**, **Trade-offs**, **Dependencies**

### Edge Cases (when applicable)
- **Escalation Triggers**, **Alternatives**, **Gotchas**

---

Safety Net

A PreToolUse hook that intercepts destructive commands before execution. 200+ dangerous patterns across 14 categories.

How it works:

1. Intercepts every Bash tool call
2. Analyzes the command for destructive patterns
3. Blocks (deny) or warns (warn) based on severity
4. Safe commands pass through silently

Coverage:

Category	Blocked Examples
Filesystem	rm -rf /, rm -rf ~, find -delete
Git	git reset --hard, git push --force, git clean -fd, git stash clear
Docker	docker system prune, docker volume prune, docker rm -f, docker-compose down -v
Kubernetes	kubectl delete namespace, kubectl delete --all, kubectl drain, helm uninstall
Terraform	terraform destroy, terraform apply -auto-approve, terraform state rm
AWS	terminate-instances, delete-db-instance, s3 rb --force, s3 rm --recursive
Google Cloud	gcloud projects delete, gcloud compute instances delete, gsutil rm -r
Azure	az group delete, az vm delete, az storage account delete
Databases	DROP DATABASE, DROP TABLE, DELETE without WHERE, FLUSHALL
Pulumi	pulumi destroy, pulumi stack rm --force
Stripe	stripe delete --live, stripe refunds create --live
GitHub CLI	gh repo delete, gh release delete, gh secret delete
System	kill -9 1, dd of=/dev/sda, mkfs, chmod 777, shutdown
APIs	Linear GraphQL mutations, Datadog DELETE requests

Cloud Provider Coverage:

AWS (verb-based classification)

Instead of maintaining allowlists, safety-net classifies AWS commands by verb: