cleanexpo Codex Local — Cheap Code Lens
When to invoke
- Multi-file refactor where a diff-first UX is faster than read-edit-write loops
- Code generation where DeepSeek V4 Flash (81% SWE-bench) is cheaper than Sonnet
- Local code work that can tolerate Gemma's mediocre code quality and zero cost matters more
Default backend (cloud)
OPENAI_BASE_URL=https://openrouter.ai/api/v1 \
OPENAI_API_KEY=$OPENROUTER_API_KEY \
OPENAI_MODEL=${CODEX_DEFAULT_MODEL:-deepseek/deepseek-v4-flash} \
codex <command>
Offline backend (local Gemma)
OPENAI_BASE_URL=http://localhost:11434/v1 \
OPENAI_API_KEY=ollama-local \
OPENAI_MODEL=gemma4:e4b \
codex <command>
Output contract
Codex returns a unified diff. The skill MUST:
- Show the diff to the CEO (no auto-apply on >10-line changes)
- Note the backend used (cloud vs offline)
- Note approximate token cost (cloud only) or "$0 local" (offline)
Trade-off table
| Backend | Cost | Code quality | Privacy | Speed |
|---|---|---|---|---|
| OpenRouter → DeepSeek V4 Flash | $0.28/M | Strong (81% SWE-bench) | Cloud | Fast |
| Local Ollama → Gemma 4 E4B | $0 | Mediocre | Stays on laptop | Slow (~5 tok/s) |
Default to cloud unless the task is sensitive or offline-required.
Pre-merge gate
Codex output must pass: npm run type-check && npm run lint && npm test before merge.
This is non-negotiable per .claude/rules/verification-gate.md.