Craft Agents Files

Features

• Moonshot AI (Kimi K3) support — new connection presets for Moonshot AI (global and CN endpoints) with the full Kimi model family. Kimi K3 brings a 1M-token context window, always-on reasoning, and image input; Kimi K2.6 serves as the fast summarization model. The existing Kimi (Coding) preset gains the new k3 model as well
• Pi SDK updated to 0.81.1 — refreshed model catalogs across all Pi-backed providers, including the Kimi K3 thinking-format and reasoning-effort fixes

Improvements

• ChatGPT web search failover is sturdier: model candidates are bounded, hosted-tool refusals are no longer mistaken for model rejections, and multi-attempt error messages stay readable
• GitHub Copilot sign-in re-enables policy-gated models (Claude, Grok, and others) from your account's live model listing, and the Copilot OAuth flows are now app-owned and hardened with network timeouts and standards-compliant polling
• Faster utility-model calls: the Pi model runtime is cached across summarization and title-generation requests instead of being rebuilt each time
• Settings now show proper provider labels for Minimax, Minimax (CN), and Kimi (Coding) connections instead of a generic backend label
• Kimi (Coding) connections pre-fill the current model lineup (k3, kimi-for-coding, kimi-for-coding-highspeed)

Bug Fixes

• ChatGPT web search works again for ChatGPT-account connections — the search request no longer pins a retired model id and follows your active model instead. Fixes #1023
• Custom OpenAI-compatible endpoints no longer receive the unsupported store parameter, which broke strict providers. Fixes #1022
• WhatsApp: messages on LID-migrated accounts now resolve your own identity correctly, so self-sent messages and mentions behave again. Fixes #1021