jlcodes99/cockpit-tools v1.3.40

Cockpit Tools v1.3.40

on GitHub

更新日志（中文）

[1.3.40] - 2026-09-05

新增

• 常驻提供 Luna Reserve 手动选项：API Service 与 Codex 受管模型列表提供 gpt-reserve，不因暂时没有合格账号而隐藏；不自动切换、不设为默认，也不预设固定压缩阈值。请求保持 gpt-reserve ID，仅在当前 API Key 的账号范围内选择常规额度不可用、备用额度允许且服务端标记为 luna_reserve 的 OAuth 账号；没有合格账号则报错，不改用普通账号或其他模型。用户显式配置的模型访问限制仍生效，关闭可见模型受管后仍由官方控制模型可见性。

变更

• 恢复 Codex OAuth 凭据过期后的单次重试：Codex Alpha Search 收到上游 401 后会刷新当前 Home 凭据并重试一次，与常规 Codex 请求的恢复行为保持一致。

• 账号池暂时无可用账号时自动恢复并重试一次：当账号池存在候选账号但没有账号可被选中时，Cockpit Tools 会重置运行时调度状态并重试原请求；如果重试后仍没有可用账号，响应会明确说明已执行自动恢复。

• 确认关闭第三方 API 路由后立即保存状态：关闭路由时会立即保存“已关闭”，但不会影响当前正在运行的会话；配置会在下次启动 Codex 时生效。

• 按 API Key 作用域隔离 Codex 会话绑定：同一 Codex 会话在不同客户端 API Key 下不会复用另一 Key 的账号绑定，避免受限 Key 被错误路由到不属于自身范围的账号。

• 按客户端版本返回 Codex 模型能力：模型目录会依据 client_version 过滤不兼容的 max/ultra 推理强度，并保留官方模型别名、上下文窗口、优先级、服务层级和可用推理强度；Astra 使用已确认的官方模型能力模板。

• 补齐 Codex API Service 的 HTTP、Responses 与 WebSocket 链路：支持 Codex client model 路由、Responses/WebSocket 流式事件、连接保活与 Ping、上游事件和配额响应头合并、握手限额的 Retry-After，以及本地升级失败时的兼容回退。

• 支持 Codex 多 Agent 与分支会话：兼容 collab_spawn 委派标记、orphan delegation、fork/subagent 会话层级和父子会话身份，保证协作请求能沿正确会话继续执行。

• 完善 Claude/OpenAI 协议转换：修正工具调用与 tool_result 顺序、工具相邻关系、JSON Schema required 严格度、推理文本增量/摘要映射，并保留 Codex→Claude 的 cache-write 用量信息。

• 增强 Codex 认证、调度与配额状态一致性：改进 OAuth 刷新和 401 重试、并发刷新合并、模型级冷却与配额耗尽状态、重试后错误归因、别名/子 Agent 选择和候选账号回退，避免可恢复请求被错误冷却。

• 补齐流式用量与诊断信息：区分首个数据包与有效 TTFT，补充流式模式、WebSocket 响应观察、重试隔离的响应头和请求级诊断信息，使额度与请求失败原因更准确。

• 保持非 Codex 模型目录独立：本次 API Service 同步仅更新 Codex 模型条目及其必要的共享底层能力，不覆盖其他供应商的模型配置。

Changelog (English)

[1.3.40] - 2026-09-05

Added

• Persistent manual Luna Reserve option: API Service and Codex managed catalogs expose gpt-reserve even when no account is currently eligible, without automatically selecting it, changing the default model, or fixing a compaction threshold. Requests retain the gpt-reserve ID and select only OAuth accounts within the current API key's scope whose regular allowance is unavailable, reserve allowance is allowed, and server banner is luna_reserve. If no account qualifies, the request fails without switching to ordinary accounts or other models. Explicit model access restrictions still apply, and disabling managed model visibility returns visibility control to the official client.

Changed

• Restore one retry after an expired Codex OAuth token: Codex Alpha Search now refreshes the selected Home credential and retries once after an upstream 401, matching the normal Codex request recovery behavior.

• Automatically recover temporarily unavailable local-access accounts once before retrying: when an account pool has candidates but none can be selected, Cockpit Tools resets their runtime scheduler state and retries the original request; if the retry still finds no usable account, the response explains that automatic recovery was attempted.

• Persist disabling third-party API routing from the confirmation action: turning routing off now saves the disabled state immediately while leaving the current running session untouched; the change takes effect on the next Codex launch.

• Isolate Codex session affinity by API key scope: the same Codex session no longer reuses an account binding created under another client API key, preventing a restricted key from being routed to an account outside its scope.

• Return Codex model capabilities by client version: the model catalog now filters incompatible max/ultra reasoning efforts by client_version while preserving official model aliases, context windows, priorities, service tiers, and supported reasoning levels; Astra uses the confirmed official capability template.

• Completed the Codex API Service HTTP, Responses, and WebSocket paths: added the Codex client-model route, Responses/WebSocket streaming events, connection keepalive and Ping, upstream event and quota-header merging, handshake Retry-After handling for quota errors, and compatibility fallback when local upgrades fail.

• Supported Codex multi-agent and forked sessions: added compatibility for the collab_spawn delegation marker, orphan delegation, fork/subagent session hierarchy, and parent/child session identity so collaborative requests continue on the correct session.

• Improved Claude/OpenAI protocol translation: corrected tool-call and tool_result ordering, tool adjacency, JSON Schema required strictness, reasoning text delta/summary mapping, and preservation of Codex-to-Claude cache-write usage.

• Strengthened Codex authentication, scheduling, and quota-state consistency: improved OAuth refresh and 401 retries, concurrent refresh merging, model-level cooldown and quota-exhausted state, post-retry error attribution, alias/subagent selection, and candidate-account fallback to avoid cooling accounts for recoverable requests.

• Completed streaming usage and diagnostics: separated first-packet timing from effective TTFT and added stream-mode markers, WebSocket response observation, retry-scoped response headers, and request-level diagnostics for more accurate quota and failure reporting.

• Kept non-Codex model catalogs independent: this API Service sync updates only Codex model entries and the shared low-level capabilities required by Codex, without overwriting other providers' model configuration.