0.5.8 · 控件对比度修复

Pair native setting card backgrounds and borders with Agent custom palettes. Dark and black palettes no longer inherit light Obsidian card surfaces. Validation: 372 tests, typecheck/build and real host CSS contrast reproduction.

0.5.7 · 亮暗色主题与面板分割线修复

0.5.7

Pane dividers now follow the visible reading/chat palette, including mixed light and dark workspaces. Hover and drag feedback stays neutral, while native hit areas remain unchanged. Hidden tabs and unrelated note panes retain the host appearance. Each family plugin works independently and cleans up its divider styles on unload.

The divider segment beside native tabs retains the Obsidian theme; only the content segment adopts the panel palette. Both light and dark host themes are supported.

Qiaomu Agent 0.5.6

0.5.6 — Open local artifacts as files

• Clicking a downloaded file or local image link opens the actual file in its system default application, instead of creating a Markdown note.
• Local file links offer Open file and Show in Finder (or the system file manager) in their context menu. Saved generated images also offer Show in Finder.
• Handles absolute paths, file URLs, home-relative paths, and vault attachments, including Chinese names and spaces. Missing files report an error without falling back to note creation.
• Ordinary note links and the image preview remain available. The new file actions and errors support Chinese and English; mobile keeps desktop file access unavailable.

Qiaomu Agent 0.5.5

0.5.5 — Completed replies fold failed attempts into the work record

• After a reply finishes, intermediate tool failures stay in the collapsed processed-work record alongside commands, error details, and retries. The conversation shows the final reply without repeating failed commands in red.
• A manually expanded live work record resets to collapsed when the reply finishes. Pending approvals and questions remain visible.
• Live tool failures and connection/retry feedback remain available. If execution stops without a final explanation, a short warning points to the failed operations in the work record.

Qiaomu Agent 0.5.4

0.5.4 — Model menus follow settings updates

• Refreshing an Agent's model list updates existing conversation menus immediately, including a menu that is already open.
• Model visibility switches and manually added IDs continue to apply without changing the selected Agent, conversation, or composer draft.
• Model labels and reasoning options read the current catalog instead of a stale per-view copy.

Qiaomu Agent 0.5.3

0.5.3 — Custom models and easier API key setup

• Connect OpenAI-compatible, OpenAI Responses, and Anthropic-compatible endpoints through the host network stack. Desktop requests stream without renderer CORS restrictions and use the system proxy.
• Fetch a provider's model list or add exact model IDs manually. Both coexist; refreshing preserves manual names, enabled selections, and model options. Saving configuration and testing inference are separate actions.
• Paste API keys directly with a dedicated button, reveal/hide drafts, and keep mobile setup scrollable when the keyboard opens. Clipboard access only follows an explicit click.
• Preserve provider settings when a save fails, avoid conflicting model updates, and allow selecting chat transcript text.

Interactive Cloudflare challenges or provider-side blocks still need the provider's help. Mobile/older-host transport may buffer responses. Mobile keyboard layout was simulated; physical iOS verification remains outstanding.

0.5.2 — Independent status bar handling

Hide the Obsidian status bar while this plugin's own view is active. Switch to a regular note, close the view, or disable the plugin to restore the status bar. This works independently with Qiaomu Home disabled.

Validation: project tests and production build passed (Tests 334 passed (334)). Desktop Obsidian QA opened actual plugin views with Home and other Qiaomu plugins disabled, switched to split Markdown notes, returned, closed views and unloaded the plugin. No host errors. Mobile devices were not independently tested.

Reviewed pending UI-polish changes against 0.5.1: palettes, settings styling, compact activity timelines and image prompts are already incorporated. Added the missing single-status regression assertions while retaining newer question-card and prompt-library behavior.

Qiaomu Agent 0.5.1

输入栏更轻，过程展示更安静。

• 模型名调整为 12px，推理档位调整为 11px，去掉重复的「推理」标签。
• 修复设置页 Logo 对齐，统一字号、留白和浅深色层级，改善深色开关对比度。
• 执行过程默认只占一行，可展开详情；审批、提问和失败保持可见。
• 合并开场工作状态，加入支持减少动态效果的轻量 SVG 指示。
• 新增「极简破框海报 · 小小东」与「X 中文信息图」内置提示词，默认插入编辑框。

验证：334 项测试、类型检查及生产构建通过；Obsidian 桌面测试库升级验证通过。未验证移动真机。

Qiaomu Agent 0.5.0

新功能

• Prompt 库：从聊天快捷条、/ 菜单或命令面板一键运行，自动携带当前笔记、选区或阅读上下文。
• 参数表单：支持 {{主题}}、默认值和 Raycast 参数；发送前补全必填信息，也可选择仅填入输入框。
• 自由固定与管理：搜索、分类、启停、固定、编辑，以及 Qiaomu-QuickPrompt / Raycast JSON 导入导出。
• Markdown 存储：自定义 Prompt 保存在 Qiaomu Agent/Prompts，旧版模板和自定义快捷提示词自动迁移并保留固定状态。
• 内置 50 条乔木精选和 31 条 Yao Open Prompts，默认关闭，可按需启用。

本版本整合近期 Agent worktree，保留 0.4.1 的提问卡片、时间线、配色与模型设置改进。

验证与安装

类型检查、330 项测试与生产构建通过。桌面 Obsidian 已验证旧版升级、Prompt 迁移、参数表单与重载；未进行移动真机和跨设备同步测试。

下载同版本的 main.js、manifest.json、styles.css 放入库的 .obsidian/plugins/qiaomu-agent/，保留现有 data.json。要求 Obsidian 1.11.4 或更新版本。

Qiaomu Agent 0.4.1

The agent can ask you before it guesses

• When Codex, Claude Code or a model API needs a decision, such as a style, a scope or a trade-off, the reply shows a question card instead of guessing.
• Each option has a letter. Press A–J or 1–9, or click. You can also type your own answer under 其他…, or 跳过 a question. With several questions, tabs and a small progress ring show where you are. Picking one option moves to the next question; press 提交 (or ⌘/Ctrl+Enter) on the last one.
• Once answered, the card shrinks to a short 已回答 summary and the agent continues. Saved conversations keep the questions and answers where they appeared. Secret answers are never saved.
• Works with Codex (App Server), Claude Code (ACP) and API models (a new ask_user tool), including in read-only mode.

Validation: 319 tests and the production build. Checked live in Obsidian with Codex (a two-question card, answered, and Codex continued) and with DeepSeek over the API (a typed answer). Claude Code is covered by tests against the claude-agent-acp form format; it could not be run live because the machine's ACP login was unavailable.

中文

Agent 需要你拿主意时（比如文章风格、范围、方案取舍），会先在对话里弹出一张问题卡，而不是自己瞎猜。选项带字母编号，可以按字母或数字键直接选，也可以在「其他…」里写自己的答案，或者跳过。问题有好几个时，顶部会显示标签和进度环；每选一项自动进入下一题，最后一题按「提交」（⌘/Ctrl+Enter）。回答后，卡片收成一行「已回答」摘要，Agent 接着干活。保存的对话会保留问答记录，密钥类答案不会保存。Codex、Claude Code 和模型 API 都支持，只读模式下也能用。