Think with Sol. Build with Terra. Scale with Luna.
CodexShift is a Codex skill that routes work by task shape instead of making you choose a model for every prompt. It keeps one implementation owner, adds explicit quality gates, and degrades honestly when model overrides are unavailable.
One model is rarely the best economic choice for every part of a software task. Product proposals benefit from deeper judgment, implementation needs a stable owner with repository context, and repetitive work should be cheap and bounded. Manually switching models and copying context between tasks breaks flow and makes ownership unclear.
CodexShift turns that into one small control loop:
intent -> route -> role contract -> execute -> check -> record -> stop
| Work | Preferred role | Preferred model |
|---|---|---|
| Product proposal and architecture choice | Proposal Architect | Sol |
| Implementation, tests, and integration | Implementation Owner | Terra |
| Repetitive search and bounded bulk work | Mechanical Worker | Luna |
| Security, evidence, eval, and final audit | Final Reviewer | Sol |
| Small direct tasks | Root owner | Terra |
- One primary writer for the same work scope.
- Read-only proposal and review roles.
- Maximum delegation depth of one.
- Explicit checks, retry limits, stop conditions, and route records.
- Human approval before implementation scope changes or consequential external actions.
- No false claim that a running Codex turn can hot-switch its root model.
Ask Codex:
Install the skill from https://github.com/danielwanwx/CodexShift/tree/main/codex-shift
Or install manually:
git clone https://github.com/danielwanwx/CodexShift.git
cp -R CodexShift/codex-shift ~/.codex/skills/codex-shiftStart a new Codex task after installation so the skill catalog refreshes.
Invoke it explicitly:
Use $codex-shift to turn this product idea into an approvable proposal, then stop for my approval.
Use $codex-shift to implement the approved spec, delegate only mechanical inventory work, and run a Sol final review if the change crosses a security boundary.
The skill also supports implicit invocation when the request clearly involves model routing, proposal-to-implementation handoff, multi-agent ownership, or a verifiable engineering loop.
CodexShift uses explicit model overrides only when the active Codex surface exposes them. When it cannot select Sol, Terra, or Luna programmatically, it keeps the same role boundaries on the current model and reports the route as degraded. This limitation is intentional and visible.
codex-shift/
├── SKILL.md
├── agents/openai.yaml
└── references/
├── contracts.md
└── roles.md