Skip to content

Kaeso 0.5.3. Released 2 November 2026, with 4 changes.

Latest
Released
Changes
4 changes
In this release
Changed and Fixed
The release on GitHub (opens in a new tab)

Kaeso 0.5.3 widens what the gateway accepts from Codex and Claude Code. Codex requests now reach Groq, Moonshot and Cohere models, and Claude Code requests reach OpenAI models, with the fields a provider refuses left out or adapted and named in response headers. Fresh installs start with KPMR's model data as of 2 November 2026, which removes openai/gpt-5.2-chat from OpenRouter, and running installs get the same data from KPMR on their own. There are no breaking changes in this release.

Contributors: @dvnold

Full changes: v0.5.2...v0.5.3 (opens in a new tab)

Changed

Gateway

  • A fresh install starts with KPMR's model data as of 2 November 2026, 09:30 UTC: OpenRouter no longer lists openai/gpt-5.2-chat, and five of its prices and three of its limits changed. Running installs get the same data from KPMR on their own. (#587 (opens in a new tab))

Fixed

  • Codex requests reach Groq models. Groq refuses prompt_cache_key, which Codex sends on every request, so Kaeso leaves a Responses app's key out for Groq and names it in the kaeso-left-out header; a Chat app's own key is refused before sending, as Groq would refuse it. (#584 (opens in a new tab))
  • Claude Code reaches OpenAI models. Its metadata.user_id is 150 characters long, and OpenAI takes at most 128 in safety_identifier, so an Anthropic app's longer id reaches OpenAI as its SHA-256 hash, the same for the same id, and the kaeso-adapted header names it. Ids that fit, and every other provider, are unchanged. (#585 (opens in a new tab))
  • Codex requests reach Moonshot and Cohere models. Their profiles refuse parallel_tool_calls, which Codex sends as true on every request; true only allows parallel calls, so Kaeso leaves it out and names it in the kaeso-left-out header, while false is still refused. (#586 (opens in a new tab))

More releases