Oracle

Escalate to Codex (GPT-5.2) with high reasoning effort for problems requiring deeper analysis. Use for:

• Complex debugging with elusive bugs
• Architectural decisions and tradeoffs
• Security audits and vulnerability analysis
• Code review requiring deep reasoning

Main command

Run Codex in non-interactive mode with high reasoning:

codex exec --sandbox read-only -c model_reasoning_effort='"high"' "$PROMPT"

Where $PROMPT is your analysis request. Codex bundles relevant context from your repo automatically.

Reasoning effort

• high — deep reasoning, ~3x tokens, best for complex problems (oracle default)
• medium — balanced (codex default)
• low — fast, minimal thinking

Options

• --sandbox read-only — analyze without modifying files (recommended)
• --sandbox workspace-write — allow file modifications if needed
• -o /tmp/oracle-response.md — save response to file
• -C <dir> — run in specific directory

Guidelines

• Oracle is read-only — analyzes but doesn't modify code
• Treat outputs as advisory — verify against codebase
• Use for genuinely complex problems (not routine tasks)
• High reasoning uses ~3x tokens but thinks deeper

Examples

# Debug intermittent auth issue
codex exec --sandbox read-only -c model_reasoning_effort='"high"' \
  "The auth flow fails intermittently. Check src/auth/ for race conditions."

# Architectural review
codex exec --sandbox read-only -c model_reasoning_effort='"high"' \
  "Review src/api/ data flow and suggest improvements for scalability."

# Security audit
codex exec --sandbox read-only -c model_reasoning_effort='"high"' \
  "Audit src/handlers/ for OWASP top 10 vulnerabilities."

$ARGUMENTS