Endpoint Probation — AI skill for Claude Code

Endpoint Probation

**English** · [简体中文](README.zh-CN.md)

Every AI endpoint starts on probation. Trust is earned with evidence you ran yourself — not with a price list, a screenshot, or the model telling you its name.

**Endpoint Probation** is an agent skill for Claude Code, Codex, Cursor and other coding agents. It decides whether an AI API endpoint — a cheap relay, a reseller, a team gateway, an aggregator, a self-hosted proxy, or an official model alias that suddenly feels different — deserves to carry real work, and it switches your tools over only with a backup and a one-line rollback.

It comes with `probation.py`, a dependency-free Python helper that runs a fixed probe battery against the candidate and a trusted reference, flags failures, reconciles the bill, lints the probation card and backs up agent configuration. A built-in `selftest` rehearses everything against two local fake endpoints, so you can see what each kind of cheating looks like before spending a cent.

Why

From the outside every endpoint looks the same: a base URL, a key, a model name. Behind it, any of these can be happening:

• a 200 OK whose stream carries a rate-limit error, which your client shows as the model's answer;
• the flagship name on the label, a cheaper model behind it — especially on long or expensive requests;
• the start of long context silently trimmed, so your agent forgets files it just read;
• a hidden multiplier, padded token counts, or a cache discount the relay keeps;
• fine at 10:00, unusable at the evening peak; gone with your balance next week;
• prompts, code and keys passing through an operator you know nothing about.

Most people judge all of that by "it answered my hello". This skill replaces that with a card, six trials and one verdict.

Features

• Frozen criteria. Pass thresholds are written from the job (tool, sensitivity, longest prompt, working hours) before any test runs; edits after results start a new probation.
• Six trials. Plu