Repository Public Research preview
1# Changelog23## 1.0.2 — 2026-11-1045- Added `examples/`: two screenshots of real sessions (Claude Code and Codex) using version 1.0.1 installed from the registry, declared in the manifest so the registry page shows them. No change to the code.67## 1.0.1 — 2026-11-1089- Fix: `migrate` left a Chat Completions request invalid when the target model is Responses-only,10 although the documentation said it would be converted. It now converts it (`converted_to:11 "responses"`). Found by installing 1.0.0 from the registry and running it. Regression test added.1213## 1.0.0 — 2026-11-101415First release.1617- Model and rule tables generated from live probes of 29 text models (55 request shapes each, on18 the Responses API and Chat Completions), official prices and official model pages.19- Offline CLI: `models`, `lint`, `fix`, `convert`, `migrate`, `cost`, `compare`, `scaffold`, `guide`.20- 18 scaffold templates (9 approaches × TypeScript and Python). Sixteen were executed against the21 live API; the two `batch` templates were submitted successfully but their batches had not22 completed when this version was released.2324### Measured effect2526An author model was asked to write OpenAI request bodies for 16 tasks, each targeting a specific27model, with and without this skill in its instructions. Every request was sent to the live API; a28request counts when it passes validation (no HTTP 400 or 404).2930| Author model | Without skill | With skill |31|---|---|---|32| `gpt-6-luna` | 10 / 16 | 16 / 16 |33| `gpt-6.1-terra` | 11 / 16 | 16 / 16 |34| `gpt-6.1-sol` | 10 / 16 | 16 / 16 |3536Read these numbers with their limits:3738- The 16 tasks were written by the skill's author and lean towards known pitfalls (sampling39 parameters, effort levels, legacy parameter names, Responses-only models, strict schemas). They40 are not a random sample of work.41- A first run scored 8/16 without and 15/16 with the skill; the single failure in both conditions42 was a broken image URL in the test fixture, not a request error. The fixture was replaced by an43 inline image and the evaluation re-run. The skill text itself was not changed between runs.44- One run per condition; model output varies between runs.45- "Passes validation" is not "does the task well". Output quality was not graded.4647The task list, harness and raw results are in the source repository (`openai/skill-tests/`).48