v0.1.18

Groq connections work again, three preset defaults move off models their providers retired, and the retry on a lost language mixture says what went wrong (#1).

Preset defaults

• Groq: Groq shut llama-3.3-70b-versatile, the preset default, down on 2026-08-16, so every Groq connection on it has failed since. The default is now openai/gpt-oss-120b. A connection you already saved on llama-3.3-70b-versatile or llama-3.1-8b-instant moves to it when settings load, as long as its base URL is still Groq's. Measured through OpenRouter pinned to Groq, 12 runs: 13.9/14 on the live check, no false alarms, about 310 ms.
• OpenRouter: the default is now openai/gpt-5.6-luna, because OpenAI shuts gpt-4.1-mini down on 2027-01-14. Its thinking is switched off (reasoning_effort: "none"). Measured 2026-10-20, served by OpenAI, 5 runs: 14.0/14 at $0.33 per 1,000 requests, against $0.56 with thinking on and $0.52 for gpt-4.1-mini. Connections you already saved are left alone, because that model still answers.
• DeepSeek: DeepSeek scheduled deepseek-chat to stop on 2026-07-24, so the default is now deepseek-flash. This one comes from DeepSeek's docs and was not measured: there is no DeepSeek key here.

OpenAI reasoning models

On the OpenAI preset, the o-series and gpt-5 or later models are now sent max_completion_tokens in place of max_tokens. OpenAI's API reference says those models refuse max_tokens, and before this change the field could not be removed, so a connection on one of them could not get any request through. Every other model and endpoint gets the same request as before. This comes from OpenAI's docs and was not measured: there is no OpenAI key here.

Text that switches language clause by clause

Since 0.1.17, a rewrite that comes back in one language is asked for once more. That second request now says the first rewrite was in one language, which was wrong. Measured paired on gemini-2.5-flash through OpenRouter, Expand, six clause-mixed texts, 20 runs: of 59 first replies that lost the mixture, 26 kept it on the second request when it said so, against 10 when the request was sent unchanged. Across the five rewording actions, 10 runs: 267/300 → 282/300 checks, Expand 28/60 → 42/60. No reply mentioned the earlier rewrite. #1 stays open.

How it is known

On the released tree, npm run verify reported 705 passing checks and no failures, in Playwright's Chromium build under xvfb-run. Details and limits are in MODELS.md and COMPATIBILITY.md.