v0.10.2

更新内容

• 新增流式性能指标：请求日志记录首字延迟（TTFT）与流式生成耗时，便于观察上游响应性能。
• 增强图像意图检测：剥离 Codex 的 XML 信封（<environment_context>、<thread_title> 等），正确识别真实用户图像意图；修复复数形式（posters/images/avatars）匹配失败的问题。
• Responses 系统轮识别：将 thread_title 轮次识别为系统轮，避免标题生成等元请求误触发图像桥接。
• 管理界面中文本地化：将 Fallback 策略、Dry Run、API Key 等界面文案统一为中文。

Changes

• Added streaming performance metrics: request logs now record time-to-first-token (TTFT) and streaming generation duration for upstream performance visibility.
• Improved image-intent detection: strips Codex XML envelopes (<environment_context>, <thread_title>, etc.) to correctly identify real user image intent; fixed plural forms (posters/images/avatars) failing to match.
• Responses system-turn recognition: treats thread_title turns as system turns so title-generation and other meta requests no longer trigger the image bridge.
• Admin UI Chinese localization: unified UI copy (Fallback policies, Dry Run, API Key, etc.) into Chinese.

验证 / Verification

• Full test suite: python -m pytest tests/ -q, 764 passed.

v0.10.1

更新内容

• 加固 Responses 原生探测：更稳健地判断上游是否支持原生 Responses API，避免误判导致不必要的降级或请求失败。
• 完善 Responses 请求的 fallback 处理：原生 Responses 调用失败时，更可靠地回退到 OpenAI Chat 路径。
• 修复工具历史（tool-history）相关错误：处理 Responses 请求中工具调用历史的序列化/回放异常，减少 4xx 失败。
• 同步 README 与使用说明书：补充 0.10.0 特性、配置选项与完整端点清单。

Changes

• Hardened the native Responses capability probe so upstream support is detected more reliably, avoiding misclassification that caused unnecessary fallback or failures.
• Improved fallback handling for Responses requests: native Responses failures now fall back to the OpenAI Chat path more reliably.
• Fixed tool-history related errors: handles serialization/replay anomalies in Responses tool-call history, reducing 4xx failures.
• Synced README and usage docs with 0.10.0 features, config options, and the full endpoint inventory.

验证 / Verification

• Full test suite: python -m pytest tests/ -q, 753 passed.

v0.10.0

更新内容

• 新增入站请求体大小限制（默认 32 MiB，可配置），防止超大 JSON body 耗尽网关内存。
• 新增上游连接池复用：Anthropic 与原生 Responses 直连上游复用 TCP/TLS 连接，降低固定延迟；引用计数保证在飞的流式响应不会被提前关闭。
• 新增上游地址校验（SSRF 防护）：拒绝 file://、gopher:// 等非 http(s) 协议、云元数据端点与保留地址；私网地址默认放行以兼容自建推理服务（llama.cpp / vLLM / Ollama）。
• Codex 兼容图像生成（/responses 与 /images/generations），支持批量执行、幂等重试、压缩预览与图像统计。
• 可配置请求日志 payload 捕获与结构化密钥脱敏。
• 管理员登录尝试限流，以及图像结果下载的网络/大小防护。
• 生产环境启动默认禁用 Uvicorn reload（除非显式启用）。
• Provider 模型发现遵循分页模型列表后再替换已存储模型。

Changes

• Request body size limit (default 32 MiB, configurable) to prevent oversized JSON bodies from exhausting gateway memory.
• Shared upstream connection pool: Anthropic and native Responses adapters reuse TCP/TLS connections to direct upstreams, reducing fixed latency; reference counting ensures in-flight streaming responses are never closed early.
• Upstream URL validation (SSRF guard): rejects non-http(s) schemes (file://, gopher://), cloud metadata endpoints, and reserved addresses; private addresses allowed by default for self-hosted inference (llama.cpp / vLLM / Ollama).
• Codex-compatible image generation via /responses and /images/generations, with batch execution, idempotent retries, compressed previews, and image statistics.
• Configurable request-log payload capture and structured secret redaction.
• Admin login attempt throttling, and network/size safeguards for image-result downloads.
• Production startup now disables Uvicorn reload unless explicitly enabled.
• Provider model discovery follows paginated model lists before replacing stored models.

验证 / Verification

• Full test suite: python -m pytest tests/ -q, 745 passed.

v0.9.12

更新内容 / Updates

• 新增思考参数透传与请求日志展示，支持 reasoning_effort、chat_template_kwargs 和 enable_thinking。

• Added thinking-parameter passthrough and request-log display for reasoning_effort, chat_template_kwargs, and enable_thinking.

• 请求日志新增输出 Tokens、TPS、总耗时和生成耗时指标。

• Added completion-token, TPS, total-duration, and generation-duration metrics to request logs.

• 优化流式请求性能统计，并兼容 Responses API 的嵌套 reasoning 配置。

• Improved streaming performance metrics and Responses API nested reasoning configuration compatibility.

验证 / Verification

• JavaScript syntax: node --check app/web/static/app.js passed.
• GitHub Actions tests and packaging completed successfully.
• Windows、Linux 和 macOS ARM64 packages uploaded successfully.

Docker

docker pull ghcr.io/pricening/llm-aio-gateway:latest

v0.9.11

更新内容 / Updates

• 将提供商选项与上游 HTTP Headers 拆开，旧的 extra_headers 会自动迁移。

• Split provider options from upstream HTTP headers, with automatic migration from legacy extra_headers.

• 历史工具调用缺少 reasoning 时自动关闭 thinking，避免 DeepSeek 等后端拒绝续写。

• Disabled thinking when historical tool calls lack reasoning content, so backends such as DeepSeek can continue the tool loop.

• Codex 的 ambient / system 后台轮次不再进入生图桥接。

• Kept Codex ambient and system background turns out of the image-generation bridge.

• Responses 能力为 unknown 时遵守短暂退避，不再立刻重复探测。

• Honored the transient backoff for unknown Responses capability instead of probing again immediately.

• 仅 previous_response_id 会阻止跨提供商 fallback；普通 function_call_output 仍可回退。

• Blocked cross-provider fallback only for previous_response_id; ordinary function_call_output can still fall back.

• 模型测试失败时展示上游 HTTP 状态、响应和调用栈，便于排查 WAF / 协议错误。

• Surfaced upstream HTTP status, response body, and traceback in model-test failures for WAF and protocol diagnosis.

验证 / Verification

• JavaScript syntax: node --check app/web/static/app.js passed.
• Full test suite: 658 passed.
• GitHub Actions packaging, release, and GHCR jobs completed successfully.
• Windows, Linux, and macOS ARM64 packages uploaded successfully.

Docker

docker pull ghcr.io/pricening/llm-aio-gateway:latest

v0.9.10

更新内容 / Updates

• 将 Responses 工具 schema 规范为 Chat Completions 可接受的 object 根，补齐空 required，并合并根级 anyOf/oneOf。

• Normalized Responses tool schemas for Chat Completions by forcing an object root, adding an empty required list, and flattening root anyOf/oneOf unions.

• 历史工具参数若不是合法 JSON 对象，会包装后再发给 llama.cpp，避免上游拒绝。

• Wrapped invalid historical tool arguments as JSON objects before sending them to llama.cpp.

• 流式回复若只有 ... 占位内容，会在发给客户端前重试，并丢弃后续真实输出前的占位前缀。

• Retried placeholder-only stream completions such as ... before they reach the client, and dropped placeholder prefixes when real output followed.

验证 / Verification

• JavaScript syntax: node --check app/web/static/app.js passed.
• Full test suite: 650 passed.
• GitHub Actions packaging, release, and GHCR jobs completed successfully.
• Windows, Linux, and macOS ARM64 packages uploaded successfully.

Docker

docker pull ghcr.io/pricening/llm-aio-gateway:latest

v0.9.9

更新内容 / Updates

• 保留 Anthropic thinking / redacted_thinking 签名，并按官方协议发送 budget_tokens。

• Preserved Anthropic thinking and redacted_thinking signatures, and sent budget_tokens per the official protocol.

• 不再把 reasoning 复制到可见文本，避免思考内容被当成最终回复。

• Stopped copying reasoning into visible text so thinking content is no longer treated as the final answer.

• 仅在明确 forced tool_choice 时关闭 thinking，工具存在不再一律禁用。

• Disabled thinking only for an explicit forced tool_choice, not merely because tools are present.

• 普通 400 不再被缓存为 Responses unsupported，避免一次校验错误长时间关掉原生 Responses。

• Stopped caching generic 400 errors as unsupported Responses capability so a request-level validation error cannot disable native Responses.

• HTTP 管理页复制失败时回退到 textarea，修复非安全上下文复制。

• Fell back to a textarea copy path when the HTTP admin page cannot use the clipboard API.

验证 / Verification

• JavaScript syntax: node --check app/web/static/app.js passed.
• Full test suite: 642 passed.
• GitHub Actions packaging, release, and GHCR jobs completed successfully.
• Windows, Linux, and macOS ARM64 packages uploaded successfully.

Docker

docker pull ghcr.io/pricening/llm-aio-gateway:latest

v0.9.8

更新内容 / Updates

• 提供商新增“强制 Chat Completions”选项，可跳过 Responses 探测，适配 llama.cpp / Qwen 等本地后端。

• Added a provider option to force Chat Completions and skip Responses capability detection for llama.cpp / Qwen backends.

• 将多条 system / developer 指令合并为一条前置 system，避免 Chat 模板因后置或多条 system 报错。

• Collapsed late and duplicate system / developer instructions into a single leading system message for Chat templates.

• 修复 force_chat_completions 更新时未按 0/1 入库的问题。

• Persisted force_chat_completions as an integer so the provider setting survives updates.

• 生图日志改为记录真实后端身份，并把 planner fallback 与 image fallback 分开，避免误标 degraded。

• Logged the real image backend identity and split planner fallback from image fallback so recovered planners are not marked as image degradation.

• 管理页请求详情补充生图后端、回退和批次统计的中英文字段。

• Added bilingual image-backend, fallback, and batch-count fields to request detail views.

验证 / Verification

• JavaScript syntax: node --check app/web/static/app.js passed.
• Full test suite: 632 passed.
• GitHub Actions packaging, release, and GHCR jobs completed successfully.
• Windows, Linux, and macOS ARM64 packages uploaded successfully.

Docker

docker pull ghcr.io/pricening/llm-aio-gateway:latest

v0.9.7

更新内容 / Updates

• 修复 Responses 能力探测对普通模型不存在 404 的误判，避免错误写入 unsupported 能力缓存。

• Fixed false negative Responses capability caching for ordinary model-not-found 404 responses.

• 修复任意 validation 422 被误判为 Responses 协议不支持的问题。

• Fixed arbitrary validation 422 errors being incorrectly treated as Responses protocol incompatibility.

• 收窄 Responses fallback 和能力缓存更新条件，仅在明确协议/端点不支持时记录 unsupported。

• Narrowed Responses fallback and capability-cache updates so unsupported is recorded only for explicit protocol/endpoint incompatibility.

• 增加相关回归测试，覆盖模型不存在、任意验证错误和协议不支持场景。

• Added regression tests covering model-not-found errors, arbitrary validation errors, and explicit protocol incompatibility.

验证 / Verification

• JavaScript syntax: node --check app/web/static/app.js passed.
• Focused Responses tests: 33 passed.
• Full test suite: 624 passed.
• GitHub Actions packaging, release, and GHCR jobs completed successfully.
• Windows, Linux, and macOS ARM64 packages uploaded successfully.

Docker

docker pull ghcr.io/pricening/llm-aio-gateway:latest

v0.9.6

更新内容 / Updates

• 修复图像生成流式响应中的状态和元数据处理问题，提升生图请求日志与统计的准确性。

• Fixed state and metadata handling for image-generation streaming responses, improving request-log and statistics accuracy.

• 改进生图桥接请求在 fallback 场景下的目标模型、提供商和尝试记录。

• Improved final model, provider, and attempt tracking for image-bridge requests that use fallback.

• 优化生图意图识别，减少普通 Responses 请求被误判为生图请求的情况。

• Improved image-generation intent detection to reduce false positives for ordinary Responses requests.

• 增加相关回归测试并修复若干小问题。

• Added regression coverage and fixed several additional small issues.

验证 / Verification

• JavaScript syntax: node --check app/web/static/app.js passed.
• Full test suite: 618 passed.
• GitHub Actions test, package, release, and GHCR jobs completed successfully.
• Windows, Linux, and macOS ARM64 packages uploaded successfully.

Docker image:

docker pull ghcr.io/pricening/llm-aio-gateway:latest