oracle

# Oracle - Second Opinion Model

Invokes OpenAI's GPT-5.2 extra high reasoning model via CLI for complex analysis tasks. It excels at debugging, code review, architecture analysis, and finding better solutions.

**Prerequisite:** Codex CLI installed and authenticated (`codex login`).

**Trade-offs:** Slower and more expensive than the main agent, but significantly better at complex reasoning. Use deliberately, not for every task.

## Invocation

Use `codex exec --profile oracle` to run the reasoning model:

```bash
codex exec --profile oracle "Review @src/auth/jwt.ts for security vulnerabilities"
```

### Examples

```bash
# Security review
codex exec --profile oracle "Review @src/auth/jwt.ts for security vulnerabilities. Provide specific fixes."

# Debugging
codex exec --profile oracle "Find why memory leak in @src/DataFetcher.tsx. Component doesn't clean up on unmount."

# Architecture analysis
codex exec --profile oracle "Analyze how @src/services/payment.ts and @src/services/order.ts interact. Propose refactoring plan."

# Complex bug investigation
codex exec --profile oracle "Bug: Users see stale data after updates. Check @src/cache/invalidation.ts for race conditions."
```

## Prompting for Reasoning Models

Reasoning models work differently from completion models. Follow these guidelines:

**Keep prompts simple and direct:**

- State the goal clearly without excessive context
- Let the model explore and discover relevant information
- Avoid step-by-step instructions - the model reasons on its own

**Focus on WHAT, not HOW:**

- Bad: "First read the file, then analyze each function, then check for..."
- Good: "Review this code for security vulnerabilities"

**Use `@` syntax for file references:**

- Include relevant files directly: `@src/auth/login.ts`
- The model will read and understand the code

**Be specific about expected output:**

- "Provide specific fixes" > "Is this correct?"
- "List all race conditions" > "Are there any bugs?"

## Example Prompts

##