sordino-openai

name	sordino-openai
description	Use when working in a Codex project that routes OpenAI traffic through Sordino for local PII masking.

Sordino OpenAI

• The enable skill starts (or reuses) a per-project proxy and writes the route into $CODEX_HOME/config.toml. The SessionStart hook is verify-only: it warms the proxy and emits a neutral onboarding/warn note — it prints no route URL and writes no routing.
• Routing is enabled by the enable skill writing $CODEX_HOME/config.toml: a custom [model_providers.sordino] provider block + model_provider = "sordino". Do NOT set any *_base_url key by hand — the skill owns the provider block.
• The provider uses HTTP POST + SSE (supports_websockets = false, NOT WebSocket) and the Responses wire API (wire_api = "responses"); Codex speaks only Responses, so POST /v1/responses create traffic (including SSE streams) is masked on requests and unmasked on responses.
• Requires an OpenAI API key exported as OPENAI_API_KEY (sk-...). A ChatGPT-subscription login will NOT work — the custom provider hard-errors on a missing env key.
• The masking enforce/verify hooks (SessionStart + UserPromptSubmit) fire only on codex > 0.140, and require a one-time hook-trust review + a codex restart to activate.
• What masking means: PII in requests is replaced with deterministic tokens like [EMAIL_ADDRESS_a1b2] that you and OpenAI see; the user sees the real values locally. It hides data from the provider, not from the user — so a token is a stable stand-in for something the user can read, and you should never tell the user their own data is hidden, redacted, or that you can't access it.