@young1lin/dsh-gpt-sub 0.1.1 → 0.1.3

package/CHANGELOG.md

@@ -3,6 +3,32 @@
 记录用户可感知的变化；格式参考 [Keep a Changelog](https://keepachangelog.com/zh-CN/1.1.0/)。
 版本号见 `package.json`。
 
+## [0.1.3] - 2026-10-08
+
+### 新增
+
+- Codex 设置面板接入 DSH 中英文语言服务，按钮、提示、倒计时、文件浏览、确认与重置结果跟随语言切换；有效期按所选语言格式化。
+
+### 调整
+
+- 设置导航统一命名为「Codex」，使用跟随主题的 Codex 图标替代默认齿轮，插件卸载时还原宿主图标。
+- 改善重置条目的标题和状态/有效期排版，避免操作按钮挤压换行，并适配窄面板布局。
+- 重置次数列表改为点击懒加载：打开/重开面板、切换语言、后台轮询与普通配额刷新不查询列表；首次点击「查询重置次数」后才读取，后续仅显式刷新或确认使用后更新。
+
+### 修复
+
+- 保留使用重置次数后的结果提示，避免随列表刷新立即清空；已显示的提示可随语言更新。
+
+## [0.1.2] - 2026-10-05
+
+### 新增
+
+- 接管 `openai-codex` 路由的自动压力、上下文溢出及手动 `/compact` 压缩，使用所选模型的 Codex 原生 `compaction_trigger` 协议。
+- 将加密 checkpoint 持久化到 Session，支持恢复、后续对话和连续压缩；原生请求失败时保留原历史，其他提供方继续使用 DSH basic 压缩。
+- 增加原生请求超时、响应大小和保留用户消息预算配置。原生压缩需要宿主 `dsh-llm-pi-ai` 提供 `llm-pi-ai/stream` 扩展；缺失时 `openai-codex` 自动回退 DSH basic 压缩，其余功能不受影响。已压缩的 Session 必须使用同一模型。
+- 增加本地 Web profile 的压缩后端包映射，使现有预设加载原生后端；普通 `dsh web` 启动使用本地构建。
+- 开发依赖改为 registry 版本（DSH `0.2.1-alpha.1` 系列），CI 无需本地 Harness 源码检出即可安装、构建、测试与发布。
+
 ## [0.1.1] - 2026-10-01
 
 ### 修复

package/README.md

@@ -4,15 +4,7 @@ Direct ChatGPT/Codex subscription access for DeepSeek Harness (DSH).
 
 Published on npm as **`@young1lin/dsh-gpt-sub`** (public, MIT).
 
-This plugin owns no transport. pi-ai already ships an `openai-codex` provider
-that speaks the Codex wire format — it sets `store`, derives
-`chatgpt-account-id` from the access token's own JWT claim, and speaks the
-codex-responses protocol — but it authenticates only through OAuth, and
-`dsh-llm-pi-ai` runs no login flow and holds no OAuth store. Naming a
-credential on the route grafts an api-key method beside the provider's own,
-and that credential is the seam this plugin fills: it keeps a live Codex
-access token in the harness credential store, refreshing it from
-`~/.codex/auth.json` before it expires.
+The plugin refreshes the Codex CLI subscription token, routes Codex egress, and uses Codex native context compaction for the `openai-codex` provider. Normal conversation requests use pi-ai; native encrypted checkpoints are stored in the Session and restored into subsequent Responses requests.
 
 The result needs no loopback port, no shared local key, and no external proxy
 binary. The `codex` CLI still owns the login.
@@ -40,7 +32,7 @@ into the profile and mounts the bundled `cordis.patch.yml` automatically — no
 manual yml editing:
 
 ```powershell
-# Prereqs: Node ≥ 22.13 (the pnpm-managed toolchain needs 22.13+), dsh 0.2.0-rc.2 or compatible 0.2.x (required by this release's DSH peers),
+# Prereqs: Node ≥ 22.19, DSH 0.2.1-alpha.1 with the llm-pi-ai/stream extension,
 # pnpm on PATH (corepack enable or npm i -g pnpm)
 dsh plugin --profile web add @young1lin/dsh-gpt-sub
 dsh web    # restart the host (the host half is module code)
@@ -67,13 +59,15 @@ into the profile and list it in `dsh.profile.bundles`:
 
 The bundle ships its own `cordis.patch.yml` with two rows:
 
-- an **insert** for the `gpt-sub` plugin itself (token sync, proxy routing, quota panel), and
+- an **insert** for the `gpt-sub` plugin itself (token sync, proxy routing, quota panel),
 - an **edit** (no `insert:`) of the dormant `llm-pi-ai` row `@deepseek-ai/dsh-base` mounts,
-  seeding the provider route:
+  seeding the provider route.
+
+The token plugin's configuration is:
 
 ```yaml
-- id: llm-pi-ai
-  name: '@deepseek-ai/dsh-llm-pi-ai'
+- id: gpt-sub
+  name: '@young1lin/dsh-gpt-sub'
   config:
     proxyUrl: http://127.0.0.1:7890
     authFile: ~/.codex/auth.json
@@ -144,6 +138,19 @@ provider knows the Codex backend itself.
 | `syncIntervalMinutes` | `10` | How often to re-read auth.json and republish the token. |
 | `bootstrapRetries` | `3` | Connection-level retries per request when the link to the proxy drops. |
 | `stateFile` | `~/.dsh/gpt-sub.json` | Persisted panel overrides (`proxyUrl`, `authFile`); a field present in the file wins over the config's. |
+| `nativeCompactionTimeoutMs` | `300000` | Maximum duration of one native compaction request. |
+| `nativeRetainTokens` | `64000` | Estimated token budget for recent user messages retained beside the encrypted checkpoint; `0` retains none. |
+| `nativeResponseMaxBytes` | `16777216` | Maximum streamed compaction response bytes. |
+
+## Native context compaction
+
+Select `@young1lin/dsh-gpt-sub/compaction` as the backend inside the existing preset's compaction group. This backend uses native compaction for Codex Subscription (`openai-codex`) models, including Luna. Automatic pressure, context overflow, and `/compact` send the actual selected model's full balanced history, current instructions, and tools to the Codex Responses endpoint with a final `compaction_trigger` input item. The backend requires one encrypted `compaction` output item and a completed response; an upstream rejection or cancellation preserves the existing history and reports the failure without switching to text summarization.
+
+The Session stores a replacement message containing the retained user items, opaque checkpoint, exact provider/model identity, and token count. Subsequent requests expand that replacement into native Responses input, including after restoration and before a second compaction. Ciphertext length does not determine automatic compaction pressure. Other providers use DSH basic compaction and its existing configuration.
+
+Native compaction activates only when the host's `dsh-llm-pi-ai` provides the `llm-pi-ai/stream` extension. On adapters without it the plugin loads normally, `openai-codex` keeps DSH basic compaction, and every other feature is unchanged — no companion checkout required. After a native checkpoint is created, keep the same provider and model in that Session; start a new Session to switch models. Keep the plugin enabled when resuming these Sessions.
+
+For a local trial, build the plugin with `pnpm install`, `pnpm run typecheck`, `pnpm test`, and `pnpm run build`. Link the plugin, companion adapter, and `local/compaction-provider` into the Web profile as described in `LOCAL-TRIAL.md`; ordinary `dsh web` then loads these builds in existing presets. The local provider alias re-exports basic compaction's named API and selects the native backend as its default plugin. Cordis patch `name` fields assert existing module names and cannot rename them. Choose either configured Luna model, let a conversation grow, and run `/compact`. A successful request logs `Codex native compaction (<selected-model>)`; send another message to exercise encrypted-context replay.
 
 ## Why `bootstrapRetries` exists
 
@@ -185,13 +192,29 @@ provider route never reads a stale token.
 
 ## Settings panel diagnostics
 
-The `Codex 配额` settings section does more than render the usage bars. It shows
+The **Codex** section follows the language selected in Settings →
+General (English or Chinese), using DSH's locale service like `dsh-mcp-adapter`.
+Buttons, placeholders, countdowns, file browsing, confirmations, and existing
+action feedback follow locale updates. While the panel remains mounted, a
+locale update preserves an edited path or proxy URL; navigating to another
+settings section unmounts the panel and discards unsaved edits.
+Credit expiration dates use the selected language too; upstream diagnostic
+messages, file paths, and account-provided titles remain unchanged.
+
+The navigation row uses the [Codex mark from Lobe Icons](https://unpkg.com/@lobehub/icons-static-svg@1.95.1/icons/codex.svg)
+(MIT, attribution included in [the client bundle](client.js)), not the shell's
+default gear. Its monochrome SVG follows the theme and is restored on plugin
+unload. Since the settings shell currently has no public icon slot, this uses
+the same narrowly scoped, disposable DOM adaptation as `dsh-mcp-adapter`.
+If a future shell changes its markup, the icon adaptation safely does nothing.
+
+The section does more than render the usage bars. It shows
 one bar per rate-limit window the account reports — the 5-hour rolling window
 and the weekly window — each drawn Codex-style as the percent **remaining**: a
 full bar at 100% that drains toward 0% as the allowance is spent, with its own
-reset countdown. The countdown is exact, not rounded to one unit: `4h22m后重置`,
-`10m22s后重置` under an hour (seconds only appear there, where they still change
-fast enough to read), `2d5h后重置` past a day, and a zero component is dropped
+reset countdown. The countdown is exact, not rounded to one unit: `Resets in 4h22m`
+(or `4小时22分后重置` in Chinese), `10m22s` under an hour (seconds only appear there,
+where they still change fast enough to read), `2d5h` past a day, and a zero component is dropped
 rather than padded — `4h`, never `4h0m`. An account with no 5-hour limit (pro
 accounts currently report only a weekly window) shows a note where that bar
 would be, so the absence reads as a fact about the plan rather than a broken
@@ -222,6 +245,13 @@ a capped window without waiting out its timer. Below the bars the panel shows:
   means direct. The line under the input notes whether the current value comes
   from cordis config or a page override.
 
+Reset credit details are lazy-loaded: opening/reopening the panel, switching
+languages, and background quota/status polling do not query the credit list.
+Click **Check reset credits / 查询重置次数** for the first read; after that, only
+the credit list's **Refresh / 刷新** button or a confirmed consumption refreshes
+it. The header badge simply reuses the ordinary quota response, with no extra
+credit-list request.
+
 Rate-limit reset credits (the codex CLI's `rate-limit-reset-credits` wham API): `GET /gpt-sub/reset-credits` lists the account's redeemable credits, `POST /gpt-sub/reset-credits/consume` (body `{ "creditId": "..." }`, optional) consumes one. The host mints a fresh `redeem_request_id` per click, and the panel asks for confirmation first because consuming is destructive.
 
 Host endpoints behind the panel: `GET /gpt-sub/quota` (the bars), `GET /gpt-sub/status`,