Gpt Delegation

SKILL: GPT Delegation

Purpose: Intelligent Codex/GPT consultation for complex problems, escalation when stuck

Quick Start

When to Use

• After 2+ failed attempts on same issue
• Architecture decisions needed
• Security concerns
• Ambiguous requirements
• Plan review for large plans (5+ SCs)
• Proactive consultation: Before asking user when confidence < 0.5 (see rubric below)

Quick Reference

# Graceful fallback + delegation (see CRITICAL WARNING below for valid parameters)
if ! command -v codex &> /dev/null; then
  echo "Warning: Codex CLI not installed - falling back to Claude-only analysis"
  return 0
fi
codex exec -m gpt-5.2 -s workspace-write -c reasoning_effort=medium --json "TASK: [goal]
EXPECTED: [outcome], CONTEXT: [attempts, errors, iteration], MUST: Analyze failures, fresh approach, report files"

Core Concepts

Graceful Fallback (CRITICAL)

MANDATORY: All GPT delegation points MUST include graceful fallback.

if ! command -v codex &> /dev/null; then
  echo "Warning: Codex CLI not installed - falling back to Claude-only analysis"
  return 0  # NOT an error, continue with Claude
fi

Key Points:

• Graceful fallback is NOT an error
• Log warning message
• Return success (exit 0) to allow continuation
• Continue with Claude agents

CRITICAL WARNING: Codex CLI Parameters (READ CAREFULLY)

MANDATORY: Use EXACT values below - NO variations allowed

Valid -s modes: read-only | workspace-write | danger-full-access FORBIDDEN: workspace-read, read-write, write (Claude often creates invalid combinations) Valid -c value: reasoning_effort=medium ONLY (NOT high/low) Required: -m gpt-5.2 --json "PROMPT"

Delegation Triggers

Confidence Score Rubric (Centralized)

Formula:

confidence = 1.0 - (architecture_keywords * 0.3) - (multiple_approaches * 0.2) - (uncertainty_markers * 0.2)

Scale: 0.0 - 1.0

Thresholds:

• 0.9-1.0: Proceed autonomously
• 0.5-0.9: Consider consultation
• < 0.5: MUST consult GPT (proactive)

Detection Patterns:

• Architecture: architecture, tradeoff, design, scalability, pattern, choice
• Multiple approaches: could, might, option A/B, either
• Uncertainty: not sure, unclear, depends

Progressive Escalation

Principle: Delegate ONLY after 2nd failure, not first

Attempt 1 (Claude) → Fail
     ↓
Attempt 2 (Claude) → Fail
     ↓
Attempt 3 (GPT Architect) → Success

Expert Mapping

Integration Pattern

Coder Agent (Ralph Loop)

# After 2nd failure
if [ $iteration -ge 2 ] && [ $TEST_RESULT -ne 0 ]; then
  PROMPT="TASK: Fix failing test ${TEST_NAME}
  EXPECTED OUTCOME: All tests pass
  CONTEXT: Previous attempts: ${ATTEMPT_SUMMARY}, Errors: $(cat /tmp/test.log | tail -20)"
  codex exec -m gpt-5.2 -s workspace-write -c reasoning_effort=medium --json "$PROMPT"
fi

/00_plan Command

# Large plan review
if [ $(echo "$PLAN" | grep -c "SC-") -ge 5 ]; then
  PROMPT="Review plan: $PLAN (Clarity, Completeness, Feasibility, Dependencies, Risks)"
  codex exec -m gpt-5.2 -s read-only -c reasoning_effort=medium --json "$PROMPT"
fi

Troubleshooting

Codex CLI Not Installed

Symptom: Warning message about Codex CLI not installed

Solution: Expected behavior. Skill gracefully falls back to Claude-only analysis.

Delegation Not Triggering

Causes: 1) Trigger not met (< 2 failures) 2) Skill not loaded 3) Codex check failing

Debug: Check echo $iteration, verify grep -r "gpt-delegation" .claude/skills/, test command -v codex

GPT Returns Same Approach

Solution: Document previous attempts clearly with errors and stack traces.

Best Practices

• Always include graceful fallback: Codex CLI may not be installed
• Document previous attempts clearly: Include errors, stack traces, code snippets
• Specify expected outcome: What does success look like?
• Use appropriate mode: workspace-write for code changes, read-only for analysis
• Parse output carefully: GPT output format may vary
• Verify suggestions: Don't blindly apply GPT recommendations

Further Reading

Internal: @.claude/skills/gpt-delegation/REFERENCE.md - Complete prompt templates, integration examples, troubleshooting | @.claude/skills/ralph-loop/SKILL.md - Ralph Loop integration with GPT delegation

External: Codex Documentation