jlcodes99/cockpit-tools v1.3.49

Cockpit Tools v1.3.49

on GitHub

更新日志（中文）

[1.3.49] - 2026-09-12

变更

• DeepSeek 模型列表支持自定义并逐条控制识图：默认模型更新为官方最新的 deepseek-flash（DeepSeek V4.1 Flash）与 deepseek-v4-pro（旧模型名继续兼容），模型列表改由账号数据决定，可自行新增或删除上游模型，并为每个模型单独开关图片输入（官方 Flash 默认支持识图，自定义模型默认关闭）。模型列表、识图能力和思考档位会同步到本地网关列出与 CDP 注入的模型列表。
• DeepSeek 模型供应商按协议收敛：Responses 下只保留模型目录与逐模型识图开关，供应商级图片开关、图片模型列表与图片兜底模型不再展示；模型目录里的上下文改为预设选择（跟随官方 / 516K / 1M / 自定义），与 Codex 的上下文设置保持一致。
• 并排对比鹈鹕测试结果：开始测试后打开紧凑的独立结果弹框，每个邮箱对应一列画布；进度、失败提示和重试均在对应格子内操作，失败格子提供紧凑的错误、重试和原始回复操作，成功格子直接打开 HTML 预览，详情弹框仍支持在浏览器中打开。结果区移除测试参数说明和账号标签操作。
• 鹈鹕结果可在浏览器打开：新增一次性本地地址（随机端口、随机路径、15 分钟自动过期，最多同时两个）承载结果页，页面注入 CSP 与沙箱锁定脚本，禁用联网、WebRTC 与弹窗类接口；清空测试数据时同步关闭。
• 保留 OpenCode Go 会话标识：供应商网关请求会转发已有的 x-opencode-session，或根据客户端会话标识生成；缺少会话信息时使用不透明兜底标识。

修复

• 修复 DeepSeek 思考强度被重置：切换账号或启动实例时不再把 model_reasoning_effort 重置为 high，Codex 里已选的 low / high / max 保持不变。
• 收敛 CDP 注入的改动范围：模型列表识别收紧为真正的模型描述符，不再改动队列、线程等通用数组；默认模型改为只在用户显式切换模型时回写，减少任务运行中的后端写入。同时修复注入脚本因引用已删除常量而在每次执行末尾报错、导致 CDP 下切换模型状态无法回传的问题。

Changelog (English)

[1.3.49] - 2026-09-12

Changed

• Let DeepSeek model lists be user-defined with per-model image support: the default models are now the official deepseek-flash (DeepSeek V4.1 Flash) and deepseek-v4-pro (legacy names keep working), and the model list comes from account data, so upstream models can be added or removed freely and image input can be toggled per model (official Flash defaults to on, custom models default to off). The model list, image support, and thinking levels flow into both the local gateway listing and the CDP-injected model list.
• Focus the DeepSeek provider dialog on the selected protocol: under Responses only the model catalog and the per-model image switch remain, while the provider-level image toggle, the image model list, and the image routing model are hidden. The context column now uses the same presets as Codex (follow official / 516K / 1M / custom).
• Compare Pelican results side by side: tests open a compact dedicated result dialog with one email column and canvas per account. Progress, failures and retries stay inside each cell; failed cells expose compact error, retry, and raw-reply actions, while successful cells open the HTML preview directly. Browser opening remains available from the detail dialog.
• Open Pelican results in the browser: a one-off local address (random port, random path, expiring after 15 minutes, at most two at a time) serves the result page. The page injects a CSP plus a sandbox lockdown script that disables networking, WebRTC, and dialog APIs, and the listeners close when test data is cleared.
• Preserve OpenCode Go conversation sessions: provider gateway requests now forward an existing x-opencode-session or derive it from the client's session identity, with an opaque fallback when no identity is available.

Fixed

• Preserve the selected DeepSeek thinking level: switching accounts or starting an instance no longer resets the reasoning effort to high, so the low / high / max level chosen in Codex is preserved.
• Narrow the scope of CDP injection changes: model list detection now requires real model descriptors, so queue, thread, and other generic arrays are left untouched, and the default model is written back only when the user explicitly switches models, which reduces backend writes during a running turn. The injected script also no longer throws at the end of every run because of a removed constant, restoring model-switch state reporting over CDP.