diff --git a/packages/tui/CHANGELOG.md b/packages/tui/CHANGELOG.md
index 99b3e79cf..f28c9c390 100644
--- a/packages/tui/CHANGELOG.md
+++ b/packages/tui/CHANGELOG.md
@@ -7,7 +7,207 @@ This changelog covers important changes that directly affect MCode TUI users. It
 
 Internal refactors, test-only changes, and changes limited to other product surfaces are excluded.
 
-## 0.5.5 · 2026-09-25
+## 0.6.5 · 2026-10-09
+
+### Fixed
+
+- The output-speed readout no longer spikes at the start of a turn. Output buffered upstream and
+  released in one burst is now excluded, so a burst arriving within the first tenth of the request
+  no longer reports a rate far above what the model actually produced.
+
+## 0.6.4 · 2026-10-08
+
+### Fixed
+
+- While you are scrolled back through a long conversation, output that changes or removes text
+  already in terminal scrollback no longer moves the viewport to the top of the session. Without
+  recent input, the screen repaints in place and the full history rebuild waits for your next
+  keystroke; readers following the latest output stay at the bottom.
+- Settled one-time feedback, such as the previous run's duration note, finished `!` shell blocks,
+  and earlier run errors, now stays in the transcript once later output follows it instead of being
+  removed from the middle of the conversation, so later rows no longer shift.
+- `cron once --after` accepts `w` (weeks), alone or combined with other units such as `1w2d`.
+
+## 0.6.3 · 2026-10-08
+
+### Added
+
+- Customize the main Agent's system prompt at startup with `--system-prompt` or
+  `--system-prompt-file`, or extend it with `--append-system-prompt` or
+  `--append-system-prompt-file`. Available in the interactive TUI, `exec`, `exec review`, and ACP.
+- Replacement and appended prompts can be combined. They apply only to the current process, preserve
+  runtime rules and project instructions, and do not affect Task child Agents. A missing or empty
+  prompt file stops the launch instead of running without it.
+- Start a new Session in lightweight conversation mode with `--mode lightweight`. It sends a small
+  conversational system prompt and no tool schemas, and omits workspace instructions, Skills, memory
+  blocks, and MCP schemas, so it cannot inspect or change local files. The mode is fixed when a root
+  Session is created and cannot be combined with `--session` or `--continue`; child Agents,
+  `/compact`, and ACP stay on the standard context. Reopening a lightweight Session keeps its mode
+  and shows `Lightweight` in the status line.
+- Each model request attempt now has a first-response bound and a stream idle bound, both 300
+  seconds by default. A request that passes a bound is aborted and reported as a retryable timeout,
+  so the normal model-request retry runs instead of the turn waiting on a provider that never
+  answers. Override the bounds with `MCODE_LLM_FIRST_EVENT_TIMEOUT_MS` and
+  `MCODE_LLM_STREAM_IDLE_TIMEOUT_MS`, where `0` disables a bound.
+
+### Improved
+
+- A single model request carries at most 20 images by default. Older images in the request are
+  replaced with a text note while the stored history keeps every image, and a provider that allows
+  fewer images per request can set `max_images_per_request` on the model (`api.mistral.ai` defaults
+  to 8). When compaction is rejected for too many images, it retries with attachments replaced by
+  text.
+- Deterministic provider rejections, such as an image limit or an HTTP 400 `invalid_request_error`,
+  are now reported immediately with the upstream reason instead of being retried, and suggest
+  `/compact` or `/new` instead of `/retry`.
+
+### Fixed
+
+- While you are scrolled back, output that shrinks the layout, such as a task list collapsing when a
+  reply ends, no longer moves the viewport to the top of the session. The released rows are padded
+  and the history rebuild waits for your next keystroke.
+- After the agent asks you a question, the turn now waits for your answer instead of ending as a
+  retryable empty-response failure.
+- `/retry` during a running response now says to stop the running turn first, instead of reporting
+  that there is no failed response to retry.
+- Clearing or pausing a Goal while a response is still running now says that the current response
+  keeps running and to press Esc to interrupt it, instead of implying the turn already stopped.
+
+## 0.6.2 · 2026-10-02
+
+### Fixed
+
+- In a `/btw` side session, pressing Esc on an empty Composer no longer suggests editing the latest
+  message; `/edit` is unavailable there.
+- Switching between a side session and the main conversation no longer resets the main run's
+  Running/Loading timer to zero.
+
+## 0.6.1 · 2026-10-01
+
+### Added
+
+- `/doctor` and `/feedback` now work inside a `/btw` side session, so you can diagnose or report a
+  failed side response without leaving it.
+
+## 0.6.0 · 2026-10-01
+
+### Added
+
+- `/retry` now works inside a `/btw` side session. When a side-session model request fails and can
+  be retried, run `/retry` there to resend the side session's last message without switching back to
+  the main conversation.
+
+### Improved
+
+- The official default model is now `MiniMax-M3.1-Flash-Preview` when no model has been selected.
+- When a Goal is marked complete, the agent now ends with one final reply that summarizes what was
+  done and where the deliverables are, instead of stopping right after the status update.
+- Safety refusals from the model provider are no longer retried, including on BYOK. The original
+  error is shown and identified as a refusal from the model provider.
+- TLS record verification errors during a model request now trigger the automatic retry like other
+  transient network failures.
+
+### Fixed
+
+- In regular mode, the screen no longer jumps back to the opening logo and replays the whole
+  conversation while a run is in progress. Rows that can still change stay on screen above the
+  footer until they are final.
+- The `tok/s` speed is no longer inflated when the provider hides thinking text; generation time now
+  starts when thinking starts.
+- Fixed repeated failures after context compaction during an automatic continuation.
+- Context compaction through OpenAI-compatible Chat Completions providers no longer sends an empty
+  `tools` list, which some backends rejected with HTTP 400.
+
+## 0.5.10 · 2026-09-30
+
+### Fixed
+
+- Your own messages now appear exactly as typed in the conversation, pending instructions, and the
+  transcript viewer. Text such as `__pycache__`, `*`, and a leading `>` is no longer reformatted as
+  Markdown, and trailing blank lines no longer add empty rows.
+- The TUI no longer stays stuck showing a run as in progress after Runtime has already finished it.
+  Enter now sends a new message instead of steering the finished run, and queued messages are no
+  longer held back.
+
+## 0.5.9 · 2026-09-30
+
+### Added
+
+- The build-mode status line now reports `background=N`, the number of active background tasks in
+  the current root session. Bash still owned by its foreground tool call is not counted, and the
+  count holds at `1` after a turn settles until a fresh task list arrives, so a command that has
+  just moved to the background is not reported as `0`.
+- `worktreeRefreshBeforeCreate` in `config.yaml`, on by default, fetches the selected upstream
+  before a new worktree is created.
+- The OpenAI Codex OAuth entry in model settings is enabled by default; set `beta.codexOAuth: false`
+  to hide it.
+
+### Improved
+
+- BYOK requests now apply byte limits to large media and accumulated history context, reducing
+  failures caused by oversized requests.
+- Long conversations load their file list and history navigation with less data, making browsing
+  them faster.
+- The MiniMax API-key route now uses the same official model catalog as the signed-in route.
+
+### Fixed
+
+- Stopping a conversation with Esc also cancels the background Bash commands and subagents it owns,
+  and their completion notices no longer restart the conversation. Their results stay available to
+  the next turn. Switching sessions and `/clear` continue to leave background work running.
+- Editing and rewinding a message now works after an interrupt that left only a background reminder
+  behind.
+- Terminal errors from a failed run are cleared when the next run starts, so a successful retry no
+  longer shows the earlier error again.
+- The thinking toggle for M3 models is sent as a thinking setting instead of a generic reasoning
+  effort.
+
+## 0.5.8 · 2026-09-28
+
+### Added
+
+- Added M3.1-Flash-Preview to the built-in fallback model catalog.
+
+### Improved
+
+- Plugin Hook messages appear in the active conversation and restored history, making Hook results
+  visible without adding them to model context.
+
+### Fixed
+
+- Restoring a session with `--session` or `--continue` clears terminal scrollback left by the
+  previous process before replaying history, avoiding duplicate content. Normal startup preserves
+  shell history.
+- Improved output queuing on slow terminals and waits for pending output before exiting, suspending,
+  or opening an external editor.
+
+## 0.5.6 · 2026-09-28
+
+### Added
+
+- Custom provider setup can import models from `/models` with the API key entered during setup.
+  Search and select a model, then save the imported models after a successful connection test; retry
+  discovery or enter a model ID manually when needed.
+
+### Fixed
+
+- Token speed now uses actual provider output tokens and the generation window, excluding the wait
+  for the first token and tool execution. Streaming text estimates are no longer shown; confirmed
+  speed remains visible while waiting for subsequent requests or tools.
+- History reloads preserve generation timing. Older messages without timing are excluded from speed
+  statistics, and speeds below 10 token/s retain one decimal place.
+- Bash tool descriptions are displayed as plain text instead of receiving shell command syntax
+  highlighting; actual commands retain syntax highlighting.
+
+## 0.5.5 · 2026-09-27
+
+### Added
+
+- Terminal titles show the current state, session name, and MCode, for example
+  `Needs approval | Fix login | MCode`, and follow session renames and switches. Notifications for
+  completed turns, failures, permission requests, and questions identify the session and suppress
+  duplicates; completion waits for the session's queue to finish. Configure them with
+  `tui.terminalTitle` and `tui.notifications` in `config.yaml`.
 
 ### Improved
 
@@ -15,10 +215,164 @@ Internal refactors, test-only changes, and changes limited to other product surf
   total timeout. Explicit background tasks also default to one hour when no timeout is specified.
   Foreground calls may return a task ID after 60 seconds without resetting the command timer.
 - Reduced the dependencies loaded at startup to keep the startup bundle smaller.
-
 - Improved automatic context compaction reliability: large-context models reserve room for replies
   and start compacting earlier, and adaptive-thinking models at high reasoning effort no longer fail
   repeatedly when reasoning consumes the output budget.
+- Update downloads follow the same proxy settings as the TUI.
+- Memory tool results shown to the model are limited to 16 KiB, with a preview of the beginning and
+  end and guidance for reading more.
+- Long sessions repeat less work when indexing and replaying history.
+
+## 0.5.4 · 2026-09-24
+
+### Added
+
+- Press `Ctrl+S` in `/model` to add or remove a favorite. Favorites appear in a separate group at
+  the top, retain their insertion order, and persist across sessions.
+
+### Fixed
+
+- Scrolling back through chat remains detached from the latest output when the viewport grows or
+  content and the footer shrink.
+
+## 0.5.3 · 2026-09-24
+
+### Added
+
+- Type `@` to select an installed Plugin. Its source identity is preserved when sending, queueing,
+  restoring drafts, and editing history, so plugins with the same name cannot replace one another.
+
+### Improved
+
+- Bash tools describe the selected shell and the current turn's background capabilities. Managed
+  foreground commands wait up to 60 seconds before returning a task ID, with one command deadline
+  covering foreground and background execution.
+- Large Bash results preserve the beginning and end of the output, report process outcomes and
+  output persistence separately, and keep task receipts bounded.
+- `/theme` uses the left and right arrow keys to select light, automatic, or dark appearance. Custom
+  theme watching reconciles once at startup to catch edits that could otherwise be missed.
+- Workspace Skill directories (`.agents/skills`, `.claude/skills`, and `.minimax/skills`) support
+  symbolic links to other locations and watch linked directories for additions and changes.
+
+### Fixed
+
+- Pressing `Esc` before the model has produced an answer or used a tool restores the submitted
+  message to the composer, preserving attachments and text entered while stopping.
+- Closing autocomplete, history search, settings, or multiline input restores the released chat area
+  correctly. Reading earlier messages keeps your position while output arrives, and sending a
+  message resumes following the latest output.
+- Editing or rewinding history clears todos from discarded turns and stale duration indicators.
+  Entering the complete `/rewind` command while editing still runs the local command.
+- Large-file edit previews now bound diff size and computation time. Omitted previews show a clear
+  notice instead of incorrectly reporting no changes.
+- Goal tool calls outside a bound Goal turn return a recoverable error without ending the user turn.
+- Windows managed updates validate the installed package through its Node entry point, including
+  installation paths containing spaces and shell metacharacters.
+
+## 0.5.2 · 2026-09-22
+
+### Added
+
+- `/theme` opens a theme panel with live preview. Pick from the built-in `minimax`, `midnight`,
+  `graphite`, and `aurora` themes, or write your own theme file under `tui/themes/*.json`. Each
+  theme provides both a dark and a light palette, and the appearance follows your terminal's
+  background until you lock it to light or dark.
+- Custom theme files support partial overrides, color aliases, syntax palettes, and hot reload, so
+  editing a theme you are already using takes effect without restarting MCode.
+- The theme choice is stored in `tui/tui-settings.json` alongside `tuiMode` and can also be set as
+  `<theme-id>/light` or `<theme-id>/dark` to pin one appearance. An unrecognized value falls back to
+  the default theme instead of blocking startup.
+
+### Improved
+
+- Long conversations stay responsive by reusing verified history and loading less at runtime.
+- A failed optional Provider preset sync no longer blocks startup as long as the existing
+  configuration stays intact.
+
+### Fixed
+
+- Fixed the transcript jumping back to the start of a conversation when a long turn ended after you
+  had scrolled up. Previously visible content now stays in place and the composer remains at the
+  bottom.
+- Fixed style-only redraws leaving a stale composer, status bar, or duplicated body text behind in
+  Apple Terminal. Visible-region repaints now erase in place instead of scrolling old lines into the
+  terminal's native scrollback.
+- Fixed failed commands being reported as successful when Bash ran through the Windows PowerShell
+  wrapper; the native exit code is now preserved.
+- Attachment paths are resolved automatically in WSL, so Windows drive letters, UNC paths, and
+  `file://` URLs can be attached directly.
+- Rewriting private configuration files on Windows no longer interrupts startup.
+- Session history writes that wait on a busy store are retried, and stopping a turn while a write
+  waits keeps the turn's stopped state.
+
+## 0.5.1 · 2026-09-21
+
+### Improved
+
+- Reduced repeated history processing and token estimation in long conversations.
+- Provider setup now displays the Base URL and supports editing it with `ctrl+e` before testing.
+  Failed connection tests clearly indicate that changes were not saved. The Z.AI and Zhipu presets
+  label their Coding Plan and API endpoints separately.
+
+### Fixed
+
+- Fixed `/btw` side conversations failing to open while tools are running.
+- Fixed multiline text pasted by terminals without bracketed-paste framing being submitted line by
+  line.
+- Kept the input method candidate window anchored to the composer during redraws, including on
+  Windows terminals, and prevented stale history rows after content changes.
+- Local sessions using custom providers or API keys no longer require a managed login for title
+  handling, and the welcome screen no longer asks you to sign in when the selected custom provider
+  does not need a MiniMax account.
+- Token estimates for Z.AI and Zhipu Coding Plan endpoints no longer fall back to a local
+  approximation after a failed token-count request.
+- Memory edits containing `$&`, `$$`, or similar sequences are inserted literally.
+
+## 0.5.0 · 2026-09-20
+
+### Added
+
+- Interactive sessions support `mcode -m <provider/model>` to select a model for that session.
+- Provider setup supports explicit context and output limits, image input capability
+  (`--support-image`), and testing before saving with `--use`.
+- ACP clients can discover enabled Skills for the attached session's Agent and workspace alongside
+  native slash commands.
+- Fullscreen Transcript supports clicking and dragging the scrollbar. An optional `context-meter`
+  status line item shows remaining context capacity.
+- Each GitHub Release includes a versioned npm installation archive, `minimax-code-X.Y.Z.tar.gz`,
+  with a `.sha256` checksum.
+
+### Improved
+
+- Reduced repeated history processing and token estimation during long conversations.
+- Provider presets import their explicit reasoning effort options.
+- Source builds now use `~/.minimax` as the default data directory, the same as the published npm
+  CLI. Data in the earlier `~/.minimax-code` default is not moved; set `MINIMAX_DATA_DIR` to keep
+  using it.
+- Automatic metrics and diagnostics uploads are off by default and require their own consent; global
+  opt-outs are honored.
+
+### Fixed
+
+- Preserved terminal scrollback during redraws and kept collapsed thinking summaries and the
+  composer visible when content shrinks.
+- Settings updates now reject malformed configuration files without overwriting their contents.
+- Invalid attachment markup is removed from assistant messages even when no valid attachment
+  remains, and path-based attachments no longer show duplicate placeholders.
+- Thinking-enabled custom connections to Mistral, SiliconFlow, DashScope, and Kimi Coding send the
+  system prompt with a role those endpoints accept.
+- Resource files of built-in and active-Agent Skills can be read in every permission mode.
+- Content review failures are reported as such instead of as connectivity problems.
+- Context usage is recalculated after compaction instead of showing the pre-compaction value.
+- Bash deny and ask rules also match commands prefixed by transparent wrappers such as `nohup` or
+  `timeout`.
+- Terminal control sequences in model output are sanitized before display, and reading MCode's
+  private runtime data requires approval.
+- Configuration files that can hold credentials are restricted to owner access.
+- `replace_all` edits keep literal `$` sequences.
+- Session diffs report renamed files under their destination path and keep Unicode paths.
+- A blank `NO_PROXY` falls back to `no_proxy`, and updates work when npm is provided through a pnpm
+  shell shim.
 
 ## 0.4.12 · 2026-09-18
 
diff --git a/packages/tui/CHANGELOG.zh-CN.md b/packages/tui/CHANGELOG.zh-CN.md
index a47385bbb..01cdc5af9 100644
--- a/packages/tui/CHANGELOG.zh-CN.md
+++ b/packages/tui/CHANGELOG.zh-CN.md
@@ -8,14 +8,246 @@ TUI 正式版中用户能直接感知的重要变化。收录范围以 TUI 用
 
 仅内部重构、测试调整，或只影响其他产品界面且不改变 TUI 行为的变化不记录。
 
-## 0.5.5 · 2026-09-25
+## 0.6.5 · 2026-10-09
+
+### 问题修复
+
+- 修复一轮开头输出速度虚高的问题。上游缓冲后一次性吐出的输出现在会被排除，突发输出落在请求前 10% 区间时不再报出远高于模型真实速度的数值。
+
+## 0.6.4 · 2026-10-08
+
+### 问题修复
+
+- 长对话中上滚阅读时，正文输出改变或移除已进入终端滚动历史的文本，不会再把视口拉回会话开头。最近没有输入时，屏幕原地重绘，完整的历史重建延后到下一次按键；跟随最新输出的读者仍停在底部。
+- 上一轮耗时提示、已结束的 `!` Shell 输出块和较早的运行错误等一次性反馈，在后面已有新输出时会保留在对话记录中，不再从对话中间删除，后续内容也不会因此整体错位。
+- `cron once --after` 支持 `w`（周）单位，可单独使用，也可与其他单位组合，例如 `1w2d`。
+
+## 0.6.3 · 2026-10-08
+
+### 新增功能
+
+- 启动时可通过 `--system-prompt`、`--system-prompt-file` 替换主 Agent 的系统提示词，或通过 `--append-system-prompt`、`--append-system-prompt-file` 追加内容；交互式 TUI、`exec`、`exec review` 和 ACP 均支持。
+- 替换与追加可同时使用，只对当前进程生效，保留运行规则和项目指令，不影响 Task 子 Agent。提示词文件缺失或为空时直接终止启动，不会在缺少该内容的情况下继续运行。
+- 可用 `--mode lightweight` 以轻量对话模式创建新会话：只发送精简的对话系统提示词，不发送工具 schema，并省略工作区指令、Skills、memory 块和 MCP schema，因此不能检查或修改本地文件。模式在根会话创建时固定，不能与 `--session` 或 `--continue` 同时使用；子 Agent、`/compact` 与 ACP 仍使用标准上下文。重新打开轻量会话时保留该模式，并在状态栏显示 `Lightweight`。
+- 每次模型请求新增首个响应时限与流空闲时限，默认各 300 秒。超过时限的请求会中止并按可重试超时处理，走常规模型请求重试，不再一直等待始终不响应的 Provider。可用 `MCODE_LLM_FIRST_EVENT_TIMEOUT_MS` 和 `MCODE_LLM_STREAM_IDLE_TIMEOUT_MS` 以毫秒覆盖，设为 `0` 表示关闭。
+
+### 体验改进
+
+- 单次模型请求默认最多携带 20 张图片，请求中较早的图片会替换为文字说明，已保存的历史仍保留全部图片；Provider 允许的单次图片数更少时，可在模型上设置 `max_images_per_request`（`api.mistral.ai` 默认为 8）。上下文压缩因图片过多被拒绝时，会改用以文字替代附件的内容重试。
+- 图片数量超限、HTTP 400 `invalid_request_error` 等确定性拒绝会直接展示上游原因，不再重试，并建议使用 `/compact` 或 `/new`，而不是 `/retry`。
+
+### 问题修复
+
+- 上滚阅读时，正文输出导致布局收缩（例如回复结束时任务列表收起），不会再把视口拉回会话开头。释放出的行会先留白，历史重建延后到下一次按键。
+- Agent 向你提问后，本轮会等待你的回答，不再以可重试的空响应失败结束。
+- 响应仍在运行时使用 `/retry`，会提示先停止当前回合，而不是报告没有可重试的失败响应。
+- 运行中响应时清除或暂停 Goal，提示会说明当前响应仍在运行、需按 Esc 中断，不再让人以为本轮已经停止。
+
+## 0.6.2 · 2026-10-02
+
+### 问题修复
+
+- `/btw` 侧会话中，在空 Composer 按 Esc 不再提示“再按一次编辑最后一条消息”；侧会话本身不支持 `/edit`。
+- 在侧会话和主会话之间来回切换时，主任务 Running/Loading 的计时不再被清零。
+
+## 0.6.1 · 2026-10-01
+
+### 新增功能
+
+- `/btw` 侧会话中支持 `/doctor` 和 `/feedback`。侧会话请求出错时，可以直接在侧会话里检查配置或提交反馈，无需先离开侧会话。
+
+## 0.6.0 · 2026-10-01
+
+### 新增功能
+
+- `/btw` 侧会话中支持 `/retry`。侧会话的模型请求失败且可重试时，可以直接在侧会话里运行 `/retry` 重发最后一条消息，无需先回到主会话。
+
+### 体验改进
+
+- 未选择模型时，官方默认模型改为 `MiniMax-M3.1-Flash-Preview`。
+- Goal 标记完成后，Agent 会再写一条最终回复，说明完成了什么、交付物在哪里，不再在状态更新后直接结束。
+- 模型提供方的安全拒答不再重试（BYOK 同样适用），直接展示原始错误，并标识为模型提供方拒答。
+- 模型请求中出现 TLS 记录校验错误时，会和其他临时网络故障一样自动重试。
+
+### 问题修复
+
+- 修复常规模式下运行过程中界面跳回开头 logo、整段对话重新滚动输出的问题。仍可能变化的内容会先保留在底部栏上方，确定后才写入终端历史。
+- 修复 Provider 隐藏 thinking 文本时 `tok/s` 速度虚高的问题；生成时间改为从 thinking 开始计算。
+- 修复自动续跑过程中上下文压缩后反复失败的问题。
+- 通过 OpenAI 兼容的 Chat Completions Provider 压缩上下文时，不再发送空的 `tools` 列表，避免部分后端返回 HTTP 400。
+
+## 0.5.10 · 2026-09-30
+
+### 问题修复
+
+- 你发送的消息在对话、待执行指令和会话记录查看器中都会按原文显示；`__pycache__`、`*`、行首 `>` 等内容不再被当作 Markdown 改写，末尾的空行也不再显示为多余的空白行。
+- 修复 Runtime 已结束运行后 TUI 仍显示运行中的问题；此时按 Enter 会正常发送新消息，不再误当作对已结束运行的 steer，排队消息也不再被卡住。
+
+## 0.5.9 · 2026-09-30
+
+### 新增功能
+
+- 构建模式状态栏新增 `background=N`，表示当前根会话活跃的后台任务数。仍由前台工具调用持有的 Bash 不计入；根 turn 结算后，在新的任务列表返回之前该计数至少为 `1`，避免刚转入后台的命令被显示为 `0`。
+- `config.yaml` 新增 `worktreeRefreshBeforeCreate`，默认开启，创建新 worktree 前先拉取所选上游。
+- 模型设置中的 OpenAI Codex OAuth 入口默认开启；设置 `beta.codexOAuth: false` 可隐藏。
+
+### 体验改进
+
+- BYOK 请求对大体积媒体和累积的历史上下文增加字节上限，减少请求体过大导致的问题。
+- 长对话的文件列表和历史导航读取更少数据，浏览更快。
+- MiniMax API Key 路线改为与登录账号路线共用官方模型列表。
+
+### 问题修复
+
+- 按 Esc 停止对话时，会一并取消该对话持有的后台 Bash 命令和子 Agent，它们的完成通知也不会再重新唤起对话；结果仍保留给下一轮读取。切换会话和 `/clear` 仍保留后台任务继续运行。
+- 中断后仅残留后台提醒时，现在可以正常编辑和撤回消息。
+- 新一轮运行开始时会清除上一轮失败留下的终端错误，重试成功后不再重复显示旧错误。
+- M3 模型的思考开关改为以 thinking 设置发送，不再作为通用推理强度发送。
+
+## 0.5.8 · 2026-09-28
+
+### 新增功能
+
+- 内置兜底模型列表补充 M3.1-Flash-Preview。
+
+### 体验改进
+
+- Plugin Hook 的提示会显示在当前会话和恢复后的历史中，便于了解 Hook 执行结果；这些提示不会加入模型上下文。
+
+### 问题修复
+
+- 通过 `--session` 或 `--continue` 恢复会话时，先清理旧进程遗留的终端滚动历史再回放，避免重复显示历史内容；普通启动保留 Shell 历史。
+- 改进慢终端的输出排队与退出处理，减少输出积压，并在退出、挂起或打开外部编辑器前等待终端输出完成。
+
+## 0.5.6 · 2026-09-28
+
+### 新增功能
+
+- 新建自定义 Provider 时，可使用当前填写的 API Key 从 `/models` 导入模型；搜索并选中模型、连接测试成功后再保存。导入失败时可重试或手动填写模型 ID。
+
+### 问题修复
+
+- Token 速度按模型服务返回的实际输出 token 数和生成耗时计算，排除首 token 等待和工具执行时间；不再显示流式文本估算值，等待后续请求或工具时保留已确认的速度。
+- 重新加载历史消息时保留生成计时；缺少计时的旧消息不参与速度统计，低于 10 token/s 时保留一位小数。
+- Bash 工具的自然语言描述不再被当作 Shell 命令高亮；实际命令继续保留语法高亮。
+
+## 0.5.5 · 2026-09-27
+
+### 新增功能
+
+- 终端标题显示当前状态、会话名称和 MCode，例如 `Needs approval | Fix login | MCode`，并随会话重命名和切换更新。回合完成、失败、权限请求和提问的通知会标明所属会话并去重；完成通知会等待该会话的排队消息执行完毕。可在 `config.yaml` 中通过 `tui.terminalTitle` 和 `tui.notifications` 配置。
 
 ### 体验改进
 
 - Bash 命令自动转入后台时，默认及最大总执行时限延长至 1 小时；显式后台任务未指定超时时也默认保留 1 小时。前台等待 60 秒后可返回任务 ID，转入后台不会重新计时。
 - 减少安装包启动时加载的依赖，降低启动依赖体积。
-
 - 改进上下文自动压缩的稳定性：大上下文模型会为回复预留输出空间并更早开始压缩；使用 adaptive thinking 的模型在高推理强度下，不再因思考内容占满输出预算而反复压缩失败。
+- 更新下载遵循与 TUI 相同的代理设置。
+- Memory 工具返回给模型的结果限制在 16 KiB 内，保留首尾预览并提示如何读取更多内容。
+- 长会话在索引和回放历史时减少重复工作。
+
+## 0.5.4 · 2026-09-24
+
+### 新增功能
+
+- 在 `/model` 中按 `Ctrl+S` 收藏或取消收藏模型。收藏模型在顶部单独分组、按添加顺序显示，并跨会话保存。
+
+### 问题修复
+
+- 向上浏览聊天记录后，视口扩大或内容、底部栏缩短时仍保持当前滚动状态。
+
+## 0.5.3 · 2026-09-24
+
+### 新增功能
+
+- 输入 `@` 可选择已安装的 Plugin。发送、排队、恢复草稿和编辑历史消息时保留所选插件的来源，同名插件不会互相替代。
+
+### 体验改进
+
+- Bash 工具按实际 Shell 和本轮后台能力提供指引；托管前台命令最多等待 60 秒后返回任务 ID，前台与后台共用同一命令截止时间。
+- 大量 Bash 输出保留首尾内容，分别报告进程结果与输出持久化状态，并限制后台任务回执长度。
+- `/theme` 使用左右方向键切换浅色、自动和深色外观；自定义主题监听启动时会主动同步，减少保存后未立即生效的情况。
+- 工作区 Skill 目录（`.agents/skills`、`.claude/skills`、`.minimax/skills`）支持指向其他位置的符号链接，并监听链接目录内的新增和修改。
+
+### 问题修复
+
+- 模型尚未输出正文或执行工具时，按 `Esc` 中断会将消息恢复到输入框，并保留附件及中断期间新输入的文字。
+- 修复收起补全、历史搜索、设置等临时面板或多行输入后，聊天区域未恢复、内容错位的问题；阅读较早消息时，新输出不会打断当前位置，发送消息后重新跟随最新输出。
+- 编辑或回退历史后清除已丢弃回合的待办与旧耗时显示；编辑时输入完整 `/rewind` 仍执行本地命令。
+- 大文件编辑的差异预览增加大小与计算时间上限，省略时明确提示，不再误显示为没有改动。
+- 未绑定 Goal 的普通回合调用 Goal 工具时，返回可恢复错误，不再提前结束用户回合。
+- Windows 托管更新通过已安装包的 Node 入口校验版本，支持包含空格与 Shell 特殊字符的安装路径。
+
+## 0.5.2 · 2026-09-22
+
+### 新增功能
+
+- 新增 `/theme` 主题面板，支持实时预览切换。可在 `minimax`、`midnight`、`graphite`、`aurora` 等内置主题中选择，也可以自己在 `tui/themes/*.json` 下编写主题文件。每个主题都提供深色和浅色两套配色，默认跟随终端背景自动切换，也可以锁定明暗。
+- 自定义主题文件支持局部覆盖、颜色别名和语法色板，并支持热重载：正在使用的主题改完保存即生效，无需重启。
+- 主题选择保存在 `tui/tui-settings.json`，与 `tuiMode` 共用同一份文件，也可以写成 `<主题ID>/light` 或 `<主题ID>/dark` 锁定外观。无法识别的值会回退到默认主题，不会阻断启动。
+
+### 体验改进
+
+- 长对话通过复用已验证的历史并减少运行时加载，响应更流畅。
+- 可选的 Provider 预设同步失败时，只要原有配置未被破坏，仍可继续启动。
+
+### 问题修复
+
+- 修复长对话上滚查看历史后，回合结束时画面跳回会话开头的问题。已显示的内容现在保持原位，输入框仍在底部。
+- 修复 Apple Terminal 中仅样式变化时残留旧输入框、状态栏或重复正文的问题。可见区域重绘现在原地擦除，不再把旧行滚入终端的原生滚动历史。
+- 修复 Bash 通过 Windows PowerShell 包装执行时，失败命令被显示为成功的问题；现在会保留原生命令退出码。
+- WSL 中自动转换附件路径，可直接附加 Windows 盘符、UNC 路径和 `file://` URL。
+- 修复 Windows 下重写私有配置文件导致启动失败的问题。
+- 会话历史写入遇到存储繁忙时会自动重试；等待写入期间停止回合，仍会正确保留已停止状态。
+
+## 0.5.1 · 2026-09-21
+
+### 体验改进
+
+- 进一步减少长对话中重复的历史处理和 Token 估算。
+- Provider 配置显示 Base URL，支持通过 `ctrl+e` 在连接测试前修改；测试失败时明确提示配置尚未保存。Z.AI 与智谱预设分别标注 Coding Plan 和 API 两类端点。
+
+### 问题修复
+
+- 修复工具执行期间 `/btw` 侧会话创建失败的问题。
+- 修复部分终端粘贴多行文本时逐行意外提交的问题。
+- 修复重绘时输入法候选框偏离输入框（包括 Windows 终端），以及历史内容变化后残留旧行的问题。
+- 使用自定义 Provider 或 API Key 的本地会话处理标题时，不再依赖托管账号登录；所选自定义 Provider 不需要 MiniMax 账号时，欢迎页也不再提示登录。
+- Z.AI 与智谱 Coding Plan 端点的 Token 估算不再因计数请求失败而退回本地近似值。
+- Memory 编辑内容中的 `$&`、`$$` 等序列会按原文写入。
+
+## 0.5.0 · 2026-09-20
+
+### 新增功能
+
+- 交互模式支持 `mcode -m <provider/model>`，为当前会话指定模型。
+- Provider 配置支持显式声明上下文和输出上限、图片输入能力（`--support-image`），并通过 `--use` 在保存前测试模型。
+- ACP 客户端可在原生命令之外，发现当前会话 Agent 和工作区中已启用的 Skill。
+- 全屏对话记录支持点击和拖动滚动条；状态栏可选 `context-meter`，显示剩余上下文刻度。
+- 每个 GitHub Release 都附带带版本号的 npm 安装包 `minimax-code-X.Y.Z.tar.gz` 及 `.sha256` 校验文件。
+
+### 体验改进
+
+- 减少长对话中重复的历史处理和 Token 估算。
+- Provider 预设会导入其明确声明的推理强度选项。
+- 源码构建的默认数据目录改为 `~/.minimax`，与已发布的 npm CLI 一致。原先默认的 `~/.minimax-code` 中的数据不会被迁移；如需继续使用，请设置 `MINIMAX_DATA_DIR`。
+- 自动上传指标和诊断信息默认关闭，需要单独同意；同时遵循全局退出设置。
+
+### 问题修复
+
+- 改善重绘时的终端滚动历史保留，并确保内容收起后思考摘要和输入框仍然可见。
+- 设置更新遇到格式错误的配置文件时会拒绝写入，保留原文件内容。
+- 回复中的无效附件标记会被清理，即使没有有效附件也不会残留；基于路径的附件不再显示重复占位符。
+- 对 Mistral、SiliconFlow、DashScope 和 Kimi Coding 开启思考的自定义连接，系统提示词改用这些端点支持的角色发送。
+- 内置 Skill 和当前 Agent 的 Skill 资源文件在所有权限模式下均可读取。
+- 内容审核失败会如实报告，不再误报为网络连接问题。
+- 上下文压缩后重新计算上下文用量，不再显示压缩前的数值。
+- Bash 的拒绝和询问规则同样匹配带 `nohup`、`timeout` 等透明包装前缀的命令。
+- 模型输出中的终端控制序列会在显示前清理；读取 MCode 私有运行数据需要授权。
+- 可能包含凭据的配置文件仅限所有者访问。
+- `replace_all` 编辑保留字面 `$` 序列。
+- 会话 diff 按目标路径报告重命名文件，并保留 Unicode 路径。
+- `NO_PROXY` 为空时回退读取 `no_proxy`；npm 由 pnpm Shell shim 提供时也能正常更新。
 
 ## 0.4.12 · 2026-09-18