| name | reasoning-guard |
| description | Use when the user asks about Reasoning Guard status, clearing the guard, or investigating shallow reasoning-token warnings in Codex. |
Reasoning Guard
Reasoning Guard is a local Codex plugin that adds lifecycle hooks around complex or high-stakes work.
When this skill is relevant:
- Treat low
reasoning_output_tokens values on the 518n - 2 boundary, such as 516, 1034, or 1552, as a warning sign rather than proof of correctness.
- If Reasoning Guard says a previous turn is suspect, redo the reasoning from first principles before relying on the prior answer.
- For complex refactors, migrations, math, proofs, debugging, security, or production changes, verify with an independent check before editing files or giving a final answer.
- If the user intentionally wants to clear the local guard state, they can send
reasoning-guard: allow or reasoning-guard: clear.
The plugin cannot force hidden model reasoning. It can detect suspicious telemetry, request one more pass through a Stop hook, inject slow-path context, and block risky write tools while the previous turn remains suspect.
