---

name: oracle description: Ask the oracle (GPT-5.2 with deep reasoning) strategic technical questions. Use for architecture decisions, complex debugging, security analysis, or when you need a second expert opinion. Supports background execution for long-running queries. allowed-tools: Bash, TaskOutput, KillShell

Oracle - Strategic Technical Advisor

The oracle is a GPT-5.2 reasoning model configured for deep technical analysis. Use it when you need expert-level guidance on complex decisions.

Capabilities & Restrictions

Can do:

• Read files and explore the codebase
• Run shell commands for investigation
• Search the web for current information
• Provide strategic technical advice

Cannot do (by design):

• Edit or write files (read-only sandbox)
• Make changes to the codebase
• Execute destructive commands

This makes oracle safe to consult without risk of unintended modifications.

When to Use Oracle

• Architecture decisions: System design, technology choices, scaling strategies
• Complex debugging: Multi-system issues, race conditions, performance bottlenecks
• Security analysis: Threat modeling, vulnerability assessment, auth patterns
• Trade-off analysis: When multiple valid approaches exist and you need to weigh them
• Code review: Getting a second opinion on significant implementations
• Post-implementation review: Validating decisions after the fact

How to Invoke

Script location: ${CLAUDE_PLUGIN_ROOT}/skills/oracle/scripts/oracle.sh

Foreground (quick questions, <30 seconds expected)

Use Bash tool directly:

${CLAUDE_PLUGIN_ROOT}/skills/oracle/scripts/oracle.sh "What's the best way to implement rate limiting for this API?"

Background (complex questions, longer reasoning)

For questions that require deep reasoning, run in background and poll:

1. Start the query with run_in_background=true:

   ${CLAUDE_PLUGIN_ROOT}/skills/oracle/scripts/oracle.sh "Analyze the security implications of this authentication flow and suggest improvements"
   

2. Continue working on other tasks while oracle thinks

3. Poll for status using TaskOutput with block=false:

   • Check periodically if the oracle has finished
   • When status shows completed, retrieve the result

4. Get result - for long responses, use tail to avoid context flooding:

   # Get last 100 lines of output
   tail -100 /path/to/output
   

Response Format

Oracle responses follow a tiered structure:

Essential (Always Present)

• Bottom Line: 1-2 sentence direct answer
• Action Plan: Numbered concrete next steps
• Effort Estimate: Quick (<1h) | Short (1-4h) | Medium (1-2d) | Large (3d+)

Expanded (When Relevant)

• Reasoning: Why this approach over alternatives
• Trade-offs: What you gain vs sacrifice
• Dependencies: Prerequisites and external factors

Edge Cases (When Applicable)

• Escalation Triggers: When to reconsider
• Alternatives: Backup options
• Gotchas: Common mistakes to avoid

Example Questions

Good oracle questions are:

• Specific and contextualized
• About decisions rather than implementations
• Complex enough to benefit from deep reasoning

Examples:

• "Given our PostgreSQL + Redis stack, should we add Elasticsearch for full-text search or use pg_trgm? We have 10M records, 100 QPS expected."
• "This codebase uses a custom event system. Should we migrate to a standard library or continue extending it?"
• "Review this authentication flow for security issues: [paste code]"