v1.4.12
AgentGate 1.4.12 — ChatGPT+Codex merge & gpt-5.6 / ChatGPT+Codex 合并客户端与 gpt-5.6
What's new
- ChatGPT+Codex merged client support. The merged ChatGPT+Codex desktop app (gpt-5.6 protocol) puts tool definitions in
inputasadditional_toolsinstead of the top-leveltoolsfield. The gateway now hoists them intotoolsbefore conversion, so tool calling works on Chat Completions / Anthropic / Gemini upstreams. - gpt-5.6 model family. Added
gpt-5.6,gpt-5.6-sol,gpt-5.6-terra, andgpt-5.6-lunato the OpenAI catalog with pricing; default model isgpt-5.6-terra; Codex recommended mappings cover the family (-lunamaps to mini-tier). - Merged desktop app restart. The Codex restart tool handles the merged desktop app (main process "ChatGPT") and still falls back to standalone Codex.app.
Fixes
- Fake channel tags in replies. Strips literal
<commentary>/<context_addition>tags emitted by third-party models while keeping the inner text; streaming-safe across chunk boundaries. - Redaction panic on multibyte values. Key redaction no longer panics when a slice boundary falls inside a CJK character.
- Request log truncation.
raw_requeststorage limit raised from 50KB to align with the 1MB per-field cap so large tool definitions are not cut off.
本次更新
- 支持 ChatGPT+Codex 合并后的新版客户端。 合并版桌面端(gpt-5.6 协议)把工具定义放在
input数组的additional_tools里;网关在转换前提升合并进tools,Chat Completions / Anthropic / Gemini 上游工具调用恢复正常。 - gpt-5.6 模型家族。 内置目录补齐
gpt-5.6/gpt-5.6-sol/gpt-5.6-terra/gpt-5.6-luna(含价格),默认模型升级到gpt-5.6-terra,Codex 推荐映射覆盖全家族。 - 合并版桌面端重启适配。 Codex 客户端重启工具识别主进程 "ChatGPT",并兼容旧的独立 Codex.app。
修复
- 回复里出现假 channel 标签。 剥离第三方模型输出的字面
<commentary>/<context_addition>标签并保留正文,流式跨 chunk 半截标签也能处理。 - 日志脱敏遇多字节字符崩溃。 密钥脱敏切片收敛到字符边界,不再 panic。
- 请求日志截断放宽。
raw_request落库上限与单字段 1MB 对齐,大体积工具定义不再被 50KB 截断误导排查。
Why it matters / 为什么重要
Without hoisting additional_tools, third-party models saw no tools and could only fake <tool_call> text — the merged ChatGPT+Codex client would look "broken" on non-OpenAI providers. This release restores real tool calling and tracks the gpt-5.6 model lineup.
若网关不把 additional_tools 提升进 tools,第三方模型收不到任何工具、只能在正文输出假 <tool_call>;合并版客户端在非 OpenAI 上游会像"坏了"。本版恢复真实工具调用,并跟上 gpt-5.6 模型线。