Bump Desktop to 0.12.4 and the separately published GoodBuddy Agent to 0.11.18. Add the approved bilingual release notes for HTML previews, long-run and queue reliability, remote terminal-state convergence, and faster OpenCode follow-up on Linux ARM.
This candidate uses the standard six-platform x64/arm64 matrix and does not request a LoongArch experimental preview.
Native capability loading previously cold-started a temporary Runtime and disposed it immediately, so Linux ARM users paid the OpenCode startup cost again on their first request. Native inventory loading now uses the cached execution Runtime, preserving normal idle eviction and lifecycle cleanup.
Release note: 优化 Linux ARM 上切换 OpenCode 后的响应速度；原生能力清单与后续对话现在复用同一 Runtime，避免重复冷启动。
Agent replies that produced HTML could only be read as source, so users had
to copy the markup into an external file to see the result. Completed Agent
and Subagent output now renders complete HTML documents, single-root HTML
replies, and `html` / `htm` code blocks as an isolated static preview in
place. Each preview keeps a persistent notice that scripts and network are
disabled and offers the same icon actions as Mermaid diagrams: a source
toggle and a full-screen preview. The new "render HTML in conversations"
switch under Platform Features / General is on by default, applies to the
open conversation immediately, and falls back to HTML source when disabled.
Previews cannot run scripts, reach the network, submit forms, open windows,
or load remote media; author stylesheets and inline styles are preserved so
the intended layout survives, and remote resource loading is refused by the
preview CSP instead.
Direct-model replies no longer fail once they grow large. The byte caps on
streaming responses, single stream blocks, non-streaming replies, model tool
arguments, tool request context, aggregate tool results, and accumulated
answer text are removed, so users stop seeing "模型接口流式响应超过安全限制"
on long answers. Response length is now governed by the request timeout,
cancellation, and context compression. Image generation responses keep their
decodable-image validation.
A message sent after a failed or interrupted run stayed in the pending queue
instead of running. The conversation is now released whenever a run reaches
any terminal state, not only when the renderer reports readiness again, and a
dispatched queue item whose run never starts releases its reservation and
returns to pending. Later messages dispatch without extra user action.
Release note: Agent 回复中的 HTML 现在可直接在会话内预览，支持查看源码和全屏预览，默认开启并可在“平台功能 / 通用设置”关闭；直连模型的长回复不再因大小限制中断；上一次运行失败或被中断后，新发送的消息会立即执行而不再排队等待。
Long Runtime sessions could silently stop retaining tool and Subagent activity, truncate Run history details, and reject workspace previews over 256 KB. Run history now persists in Main SQLite without the former fixed record or detail caps, surfaces possible loss from legacy caches, and renders long histories in batches.
Local and remote workspace previews now load UTF-8-safe 256 KiB pages on demand instead of rejecting larger text files. The legacy migration, clear-data path, IPC contracts, tests, and feature documentation are updated with the resulting behavior.
Release note: 长时间 Runtime 运行现在会完整保留工具和 Subagent 活动；运行记录迁移到 SQLite，并提示旧版可能已丢失的内容。超过 256KB 的本地及远程文本文件可通过分页面预览继续加载。
After a managed SSH disconnect the renderer kept showing the remote
Agent as processing even though the recovered task had already been
persisted as completed. Local streaming copies shadowed the persisted
terminal state in mergePersistedConversations, and recovery completion
never asked the renderer to re-read conversations.
- Prefer the persisted terminal state over a local streaming message
- Settle active runs whose messages already reached a persisted terminal
state and release the conversation queue once
- Refresh conversations on window focus and visibility changes
- Publish a conversation change before recovery reports completed/failed
OpenCode status checks could cold-start throwaway servers, unrelated
conversations could block each other, and model, Continue, or remote paths
could replay or duplicate work under failure and backlog conditions. Settings
could also lose edits or overwrite unrelated model fields.
This update makes status checks passive, keeps per-conversation ordering while
allowing concurrency, preserves native long-running work with bounded display
capture and backpressure, narrows retries to known pre-dispatch failures, and
aligns remote Ask permissions without reducing Execute. It also keeps
Embedding updates narrow, improves settings and browser guidance, removes the
unused legacy Main gateway state, and records the validated Desktop 0.12.3 and
Agent 0.11.17 release metadata.
Release note: 改善 OpenCode 切换、多会话长任务、模型重试、远程 Ask/Execute 与设置保存可靠性，并完成 Desktop 0.12.3 / Agent 0.11.17 发布准备。
Adds the full documentation set for the planned desktop pet feature under
docs/features/desktop-pet: README, PRD, user stories, logic design, UI
design, and technical design. The pet is scoped as a lightweight always-on
entry point for quick drafting, drag-and-drop staging, and task status at a
glance; it does not introduce a second Ask/Execute channel, task execution
path, or independent task-state store. Design only, no implementation yet.
Model connections can now save bounded JSON request headers and top-level body fields. Direct requests and Continue support both; local OpenCode and DeepSeek Harness use only native header support, while the managed remote OpenCode gateway applies both.
Runtime, protocol, and authentication fields keep precedence. Custom values are ordinary connection settings, not credential storage. This migrates Runtime settings from v20 to v21 and updates the DeepSeek control and managed Runtime ACP capabilities.
Release note: 模型连接详情新增自定义请求头和请求体；直连与 Continue 支持两者，本地 OpenCode 和 DeepSeek Harness 仅应用原生支持的请求头，受管远程 OpenCode 支持两者，界面会明确提示生效范围。
Reframe the feature section around the broader desktop product rather than only Agent Runtimes. Clarify official x64 and arm64 coverage, the separate LoongArch preview, and that compatibility claims are not vendor certification.
The built-in browser was limited to direct-model tools, and its Agent switch also blocked manual Browser workbar actions. It is now a request-scoped GoodBuddy capability assignable to direct models, managed OpenCode, and Continue, while manual browser controls remain available when Agent access is off.
Existing browser enablement migrates to capability settings version 6. DeepSeek Harness remains unsupported.
Release note: 内置浏览器现在可分配给直连模型、OpenCode 和 Continue；关闭 Agent 浏览器能力不会影响用户在浏览器工作栏中的手动操作。
The workbar tabs, close controls, and add button now use the shared rounded control styling again. The terminal panel and its toolbar remain square as requested.
Release note: 恢复工作栏页签、关闭按钮和加号入口的圆角外观；终端面板与终端工具栏继续保持直角。
Local embedded OpenCode serialized otherwise independent conversations, and GoodBuddy could stop long Runtime tasks after fixed tool, activity, model-call, or accumulated display-output limits.
Different conversations can now run concurrently while requests in the same conversation stay ordered. Dynamic MCP tools remain isolated to the current request. OpenCode, Continue, managed ACP, and the Agent model gateway no longer stop native execution at GoodBuddy-defined call counts or prompt-wide output totals; display copies truncate and transport queues apply backpressure instead. Prompt-wide quota fields are removed from the current contracts.
Per-call model settings, provider request timeouts, cancellation, and same-conversation ordering remain unchanged.
Release note: 修复本机 OpenCode 不同会话互相等待的问题，并移除 GoodBuddy 对单次 Runtime 工具、活动和模型调用次数的人为上限；长输出现在只截断展示或施加传输背压，不再取消原生任务。
Terminal paste previously relied on the Renderer browser Clipboard API, which could be denied by Electron and leave the toolbar action unusable. Text clipboard access now goes through the trusted Main/Preload bridge, and the workbar tabs plus terminal toolbar use straight corners for a continuous docked appearance.
Release note: 修复终端工具栏粘贴可能因剪贴板权限被拒绝而失败的问题，并将工作栏页签与终端工具栏调整为直角样式。
Managed SSH Ask previously required bubblewrap and applied a process-level filesystem sandbox, while Execute launched directly. Ask and Execute now share the same direct signed Runtime startup; Ask remains read-only through the Agent's existing tool permission mediation, which only permits native read requests and rejects other tool kinds.
Release note: 远程 Ask 与 Execute 现在都可直接启动签名 Runtime，无需 Host 安装 bubblewrap；Ask 仍在工具权限层仅允许读取操作。
Remote Hosts without the Ask-only bubblewrap tool could report matching Agent and Runtime versions but fail project creation because the Agent hid the Runtime ACP capability entirely. The Agent now advertises a verified Runtime independently of bubblewrap, so Execute can start directly while Ask keeps its existing read-only bubblewrap boundary.
Release note: 修复远程 Host 缺少 bubblewrap 时 Execute 项目无法创建的问题；更新到 Agent 0.11.16 后，Execute 不再受 Ask 专用系统工具影响。
Direct model conversations previously stopped after the first transient network error or request timeout, including TLS setup ECONNRESET failures. Text requests now retry up to three times with cancellable backoff before visible output, while partial streamed responses are never replayed.
Release note: 直连模型遇到临时网络故障或请求超时时，会在尚未输出内容的情况下自动重试最多 3 次，并支持随时取消。
The topbar repeated a textual state after the Runtime or model name even
though it already displayed a compact status indicator. It now uses green,
yellow, and gray dots for ready, connecting, and unavailable states while
keeping detailed status available to assistive technology and tooltips.
Release note: 顶栏 Runtime 状态改用绿、黄、灰三色标记，不再重复显示“就绪”“连接中”或“不可用”文字。
The published Agent 0.11.14 predates the registered-Runtime activation change included in the Desktop candidate. Bump the Agent lock to 0.11.15 and the unpublished Desktop recovery to 0.12.2 so the compound package minimum Desktop version, bilingual notes, and feature documentation describe one coordinated release.
Keep v0.12.0 and v0.12.1 immutable and unpublished, and document that Agent tags also require the exact main candidate workflow to pass before tagging.
Release note: 托管 Linux Host 使用 Agent 0.11.15，已登记 Runtime 激活时不再重复检查完整 OpenCode 二进制。
The Unix private-endpoint relay regression combines socket authentication, SQLite journaling, attach relay framing, and two protocol requests. Give only that end-to-end case a bounded 30-second allowance so a loaded serial CI runner does not fail its previous 5-second default.
Document the mandatory Desktop release order in BUILD.md: push the candidate to both main branches, wait for validators and the production bundle on that exact SHA, and only then create the annotated tag.
Managed Python installation rejected valid case-distinct terminfo symlinks in the pinned Astral archives on Linux x64 and arm64. Archive uniqueness now follows the target filesystem: Linux preserves case-distinct paths, while every target still rejects exact duplicates and unsafe entries.
Move the unpublished 0.12.0 bilingual release notes to the immutable-tag recovery version 0.12.1 and record the failed tag evidence.
Release note: 修复 Linux x64/ARM64 上托管 Python 因官方归档包含大小写不同路径而无法安装的问题。
The complete-snapshot boundary test performs 501 synchronous SQLite source and document writes. It passed functionally but could exceed its 10-second case timeout on the serial Linux CI runner, so give this I/O-heavy case a bounded 30-second allowance without changing product behavior or global test limits.
The Linux routing test still referenced the obsolete bin/opencode name. With intentional path-only Runtime selection, that stale fixture was no longer masked by automatic discovery; use the locked opencode-ai 1.18.9 bin/opencode.exe entrypoint on every platform.
Keep the approved 0.12.0 notes and bilingual feature matrices explicit that local OpenCode paths are not prevalidated and fail only at actual launch.
The Linux release validator still exhausted process and socket timing with two Vitest workers, causing unrelated OpenCode and private-endpoint tests to fail. Run test files with one worker so process-backed integration tests remain deterministic within the existing CI job budget.
Full-suite resource contention could make asynchronous renderer assertions and short runtime timing tests fail even though focused runs passed. Limit Vitest to two workers and give the affected UI operations the existing three-second async window.
Also include the approved Linux ARM OpenCode startup note in both feature matrices and release-note languages, and record the successful isolated Linux x64 Host validation of the current-source Runtime activation path.
OpenCode startup on Linux ARM could run a --version probe before launching the configured Runtime and recheck registered remote binary metadata. Execution now trusts the selected path and registered manifest identity after setup-time verification, so switching and cold launch avoid duplicate binary work.
Release note: 优化 Linux ARM 上切换 OpenCode 的启动速度；已安装或已登记的 Runtime 在执行时不再重复启动或检查二进制。
The Linux CI run could query the Appearance tab before the lazy Settings route finished loading. Appearance tests now wait for the Settings heading before interacting with its tabs, matching the stable navigation pattern used elsewhere.
Update the desktop version, approved bilingual release notes, and synchronized English and Chinese feature matrices for the 0.12.0 candidate.
Release preparation now runs local tests, type checks, and lint without a local production build, then requires the main-branch CI build before tagging. The package matrix uses a native Windows arm64 runner and validates managed Python on every target without adding extra license files to the desktop package.
The tool environment page always displayed an unverified capability-dependency state even though it had no Skill or MCP evidence behind it. The page now shows only actual interpreter and companion-tool diagnostics; capability dependency status remains reserved for capability details with real validation results.
Release note: 工具执行环境页不再显示没有实际验证依据的能力依赖占位状态。
Document that options requiring an in-place prerequisite must remain operable, defer persistence until validation succeeds, avoid implicit first-candidate selection, preserve the previous value on cancellation, and show errors beside the action.
DeepSeek Harness previously required a separate Runtime model selection when no administrator-provided connection was available, even if the user had already selected a compatible GoodBuddy model. It now prefers the administrator connection, then follows the compatible default model or first compatible connection across conversations, projects, and messaging channels.
Release note: 修复 DeepSeek Harness 未自动使用已选兼容模型的问题；没有管理员预置连接时，现在会跟随当前默认或首个兼容模型，无需再到 Harness 页面重复指定。
Completed AI responses now expose a bottom copy action that copies their Markdown source without reasoning, tool logs, or citation metadata. Clipboard writes use a validated Preload/Main bridge so sandboxed Electron sessions no longer fail with a renderer clipboard permission error; full-conversation copy uses the same path.
Release note: AI 回复底部现在可以直接复制 Markdown 原文，并修复了沙箱环境下回复和完整会话可能无法写入剪贴板的问题。
The custom Node.js and Python choices could appear unavailable until a candidate was already discovered, and selecting the source implicitly adopted the first candidate. The chooser now stays operable, waits for an explicit validated path, keeps cancellation quiet, and shows validation errors beside the affected runtime.
Candidate discovery now prioritizes active virtual environments and Conda before bounded user/system PATH and common Node/Python locations, while preserving canonical validation and deduplication.
Release note: 修复自定义 Node.js/Python 环境入口不可用的问题，并增加对 PATH、venv、Conda 和常见安装位置的解释器检测。
GoodBuddy can now provide managed Node.js and on-demand Python to local Skills and stdio MCP servers, or validate an explicitly selected custom interpreter. The Capabilities and tools settings page exposes source selection, diagnostics, installation progress, cancellation, and removal while new processes receive immutable PATH snapshots.
Application settings migrate to v10. Local terminal and remote Host environments remain unchanged. Windows x64 native-source Node and Python paths are validated; OSS mirror publication, redistributed license resources, and the remaining native platform probes are recorded as release blockers.
Release note: 新增工具执行环境，可为本机 Skills 和 stdio MCP 使用 GoodBuddy 托管的 Node.js、按需安装的 Python，或经过验证的自定义解释器。
The website, README, and Smart Heartbeat interface used promotional superlatives, generic benefit labels, and personified growth language that obscured the actual product behavior. The English README is now the repository entry point with a separate Chinese version, and both website and in-app copy describe concrete capabilities, reports, suggestions, and execution boundaries.
Release note: 优化官网、README 与智能心跳界面的中英文文案，去除空泛宣传和拟人化表达，改为直接说明产品能力、运行记录与建议内容。
Direct model conversations can now execute bounded platform-native processes and delegate one level of programming work without duplicating tools in Agent Runtimes. Process and Subagent activity follows existing workspace, mode, approval, cancellation, and lifecycle boundaries, with chat attribution and complete feature documentation.
Remove the static Renderer bundle baseline gate and its build-only manifests; production builds continue to compile all processes and verify the control-plane installer. Also include the pending documentation link normalization.
Release note: 直连模型现在可在 Execute 模式运行本机命令，并可委派编程 Subagent 完成修改、测试和复核；构建流程不再因静态 Renderer 包体基线阻塞。
electron-builder listed shared, successfully resolved production dependencies as duplicate references during every package build. All desktop packaging entry points now lower only that non-actionable summary to debug while unresolved reference failures remain visible.
Move Renderer-bundled AntV G6 to development dependencies so its transitive graph is not copied into desktop packages.
Release note: 桌面打包不再显示无需处理的重复依赖引用摘要，也不再把已由 Renderer 打包的 AntV G6 依赖树重复复制到安装包中。
The Skills catalog exposed nine product-marketing recipes as independent
capabilities and enabled every marketing package for every Runtime by default.
It now keeps five top-level Skills, loads the nine recipes on demand through
the product-marketing package, and leaves the three marketing Skills disabled
until users enable them. Existing saved Skill settings remain authoritative.
Release note: 内置 Skills 已从 14 个顶层入口精简为 5 个；去 AI 味和长文工具默认启用，产品营销类默认停用，已保存的用户设置保持不变。
Skills and MCP previously occupied separate top-level settings categories. They now share one Capabilities and tools category with accessible horizontal tabs, while the unimplemented tool environment remains hidden until its settings are usable.
Prevent nested PageTabs wrappers from collapsing to zero height in scrollable Flex layouts, add a CSS regression guard, and document the shared rule plus the planned local Node.js and Python environment.
Release note: 设置中心将 Skills 和 MCP 合并到“能力与工具”，并修复水平页签可能被压缩而不可见的问题；尚未实现的工具执行环境暂不显示。
The bundled deai-writing skill now detects additional template patterns,
unsupported attribution, role-slot prose, engineering metaphors, and absolute
claims. Its JSON report also summarizes findings by level and category while
preserving GoodBuddy's cross-platform Python guidance.
Release note: 中文“去 AI 味”审校新增分析套话、模糊归因、角色槽位、工程黑话和绝对化承诺等检查，并提供按分类汇总的扫描结果。
Define repository and feature-scoped documentation rules for PRDs, User Stories, logic, UI, technical design, and progress evidence.
Move feature-owned documents under docs/features, add one entry point per feature, remove the redundant docs index, and update all affected cross-references. Expand the local tool environment design with explicit product, story, logic, UI, technical, and progress documents.
The Browser workbar required users to enter a complete HTTP(S) URL, so submitting a bare domain or host with a port failed URL validation. Enter and Go now prepend http:// before using the existing browser navigation path.
Explicit protocols remain unchanged so the Main browser URL policy continues to reject unsupported schemes with the correct error.
Release note: 浏览器地址栏现在支持直接输入域名或主机端口，回车或点击“前往”时会自动补全 http://。
The primary sidebar resize hit target still reserved four pixels in the flex row, leaving a visible gap between the divider and the content frame. Its opposing negative margins now keep the nine-pixel pointer target layout-neutral while preserving the aligned one-pixel divider.
Release note: 修复主侧栏拖拽分隔线与内容边缘未对齐的问题。
mesalogo