Release gproxy v2.6.5 · LeenHawk/gproxy

v2.6.5

Makes cross-protocol agent tool workflows end-to-end across OpenAI Responses, Claude, Gemini, and Codex-compatible upstreams.

End-to-end protocol transforms. OpenAI Responses, Claude Messages, and Gemini GenerateContent now convert requests, responses, streaming events, usage, and tool lifecycles directly instead of routing through OpenAI Chat.
Approximate agent tool mappings. Shell/Bash, ApplyPatch/TextEditor, WebFetch/WebSearch, and related provider-specific tools now map across protocols when no exact equivalent exists.

Tool round trips. Tool definitions, calls, results, and stop reasons are preserved through non-streaming and streaming conversions, including Codex Shell and ApplyPatch calls. Gemini feedback now correlates each call ID back to its function name, and Responses function-call items use distinct valid fc_* item IDs while retaining their call_* call IDs.
Reasoning continuity. Responses reasoning IDs use valid stable rs_* identifiers, while encrypted reasoning data, Claude signatures, and Gemini thoughtSignature values survive cross-protocol round trips.
Gemini signed content. Signed Gemini text remains visible, and tool continuations retain the signatures required by Gemini upstreams.
Responses lifecycle and history. Gemini-backed Responses streams now emit a complete lifecycle with distinct output indexes, without empty message items. Replayed Claude assistant text is encoded as Responses output_text, including when explicit prompt-cache markers are applied.
Cline buffered responses. Cline's non-streaming success / data envelope is unwrapped to the documented OpenAI Chat Completions shape; its already-standard streaming SSE path remains unchanged.

Codex upstream shaping. Codex-compatible requests now shape agent tools to the upstream wire format at the channel boundary without weakening the shared protocol model.

端到端协议转换。 OpenAI Responses、Claude Messages 与 Gemini GenerateContent 现在直接转换请求、响应、流式事件、用量及完整工具生命周期，不再绕经 OpenAI Chat。
近似代理工具映射。 Shell/Bash、ApplyPatch/TextEditor、WebFetch/WebSearch 以及其他没有严格等价关系的供应商工具现在可以跨协议映射。

工具调用往返。 工具定义、调用、结果及停止原因现在能在非流式和流式转换中完整保留，包括 Codex Shell 与 ApplyPatch 调用。Gemini 工具回填会按 call ID 找回原始函数名；Responses 函数调用 item 使用独立且合法的 fc_* item ID，同时保留原 call_* call ID。
推理上下文连续性。 Responses reasoning ID 统一使用合法且稳定的 rs_* 标识；加密推理数据、Claude signature 与 Gemini thoughtSignature 也能在跨协议往返中保留。
Gemini 签名内容。 带签名的 Gemini 文本不再丢失，工具续接也会保留 Gemini 上游所需的签名。
Responses 生命周期与历史消息。 Gemini 驱动的 Responses 流现在会输出完整生命周期和互不冲突的 output index，也不再生成空 message item。Claude assistant 历史文本会编码为 Responses output_text，应用显式提示缓存标记后也保持合法类型。
Cline 非流式响应。 Cline 非流式响应多出的 success / data envelope 会被解包为其文档声明的 OpenAI Chat Completions 结构；原本已符合规范的流式 SSE 保持不变。