codex-delegation

name	codex-delegation
description	Intelligent routing between Solara and Codex — decides when to delegate backend tasks to Codex vs handle internally.
tags	["skill"]
triggers	["codex delegation","use codex delegation","run codex delegation"]
tier	stable

Codex Delegation — Dual-AI Routing for SunBiz-Agent

Purpose: Solara and Codex are complementary. This skill determines WHEN and HOW to delegate work to Codex for maximum leverage.

The Dual-AI Split

Solara (Claude)                        Codex (GPT-5.4)
├── Funding workflow orchestration      ├── Backend API implementation
├── Lender intelligence synthesis       ├── Deep debugging with stack traces
├── Operator handoff + comms            ├── Adversarial code review
├── Memory / state management           ├── Parallel backend tasks
├── CASL compliance judgment            ├── Root-cause analysis
└── Call sheet + daily ops              └── Write-capable rescue tasks

Delegation Decision Matrix

Auto-Delegate to Codex (No Ezra approval needed)

Task Type	Why Codex
Pre-ship code review of API changes	Second pair of eyes catches blind spots
Backend bug with stack trace in logs	Codex excels at systematic root-cause
Implementing a new API endpoint	Runs in background while Solara works on ops
Test suite debugging	Codex is strong at test diagnosis
Database query optimization	Codex handles SQL + performance analysis well

Keep in Solara (Never delegate)

Task Type	Why Solara
Funding workflow decisions	Requires domain context + operator relationship
CASL compliance judgment calls	Regulatory + business context needed
Lender relationship intelligence	Solara has the historical context
Operator handoff messages	Must match Ezra's communication style
Memory/state/context files	Solara's domain — Codex has no access
Simple fixes (< 3 files)	Delegation overhead > task effort

How to Delegate

export CLAUDE_PLUGIN_ROOT="/c/Users/User/.claude/codex-plugin"
node "$CLAUDE_PLUGIN_ROOT/scripts/codex-companion.mjs" task --write \
  "Context: SunBiz-Agent — Next.js 14 App Router, Supabase, TypeScript.
   Stack: scripts/integrations/send_gateway.py is the single outbound send path.
   Task: [specific task description]
   Constraint: [any constraints]"

Always inject context: stack, file paths, constraints, what Solara has already done.

Pre-Flight Check

Before delegating, verify Codex CLI is operational:

codex --version 2>&1 | head -1

If command not found: npm i -g @openai/codex@latest If version is stale (symptoms: every model alias rejected, companion loops with no progress): upgrade first.

Failure Recovery — 3-Strike Rule

1. First failure: Retry with more context (inject file contents, narrow scope)
2. Second failure: Switch to --model spark (simpler tasks) or --model gpt-5.4-mini (faster)
3. Third failure: Solara takes over. Log to memory/MISTAKES.md with what Codex struggled with.

Never retry the same prompt 3 times unchanged.

End-of-Task Audit (MANDATORY for big tasks)

After any task with ≥3 commits, ≥5 files touched, or any user-facing change:

node "$CLAUDE_PLUGIN_ROOT/scripts/codex-companion.mjs" review --wait

Present both reviews verbatim to Ezra — Solara's self-review first, then Codex's audit under a clear header. Do not paraphrase or soften Codex findings.

Commands Quick Reference

# Standard review
node "$CLAUDE_PLUGIN_ROOT/scripts/codex-companion.mjs" review --wait

# Adversarial design challenge
node "$CLAUDE_PLUGIN_ROOT/scripts/codex-companion.mjs" adversarial-review --wait "challenge the lender-match scoring logic"

# Delegate a task
node "$CLAUDE_PLUGIN_ROOT/scripts/codex-companion.mjs" task --write "<context + task>"

# Check status
node "$CLAUDE_PLUGIN_ROOT/scripts/codex-companion.mjs" status

# Get results
node "$CLAUDE_PLUGIN_ROOT/scripts/codex-companion.mjs" result

Obsidian Links

• [[skills/systematic-debugging/SKILL.md]] | [[skills/operator-handoff/SKILL.md]]
• [[memory/MISTAKES]] | [[memory/PATTERNS]]