v1.3.0: Memory

wisp-science v1.3.0

v1.3.0 focuses on long-running conversations and the model experience: it adds real long-context memory evaluations, a dynamic context-usage gauge, and per-model reasoning-effort controls. It also improves slash commands, global habits, and end-of-turn handling, while fixing streaming responses, repeated automatic compaction, MCP child-process shutdown, and several UI issues.

Downloads

• macOS (.dmg, signed + notarized): choose the Apple Silicon (aarch64) or Intel (x86_64) build; compatible versions can update in-app.
• Windows installers (MSI / NSIS): signed; if SmartScreen still triggers, choose "Run anyway".
• Linux (deb / AppImage): available for both x86_64 and aarch64.

See Assets below.

Highlights

Long Context, Memory & Compaction

• Added a real long-context example spanning about 685k estimated tokens, plus memory evaluation suites that can run against real models.
• Automatic compaction now chooses its target from measured context growth, avoiding repeated compaction after crossing the threshold.
• Context usage is now shown as a dynamic gauge with a clearer breakdown of the active window.
• New habits can be added directly to global memory from Settings.

Models & Reasoning

• The models.dev catalog is baked into the app and provides exact-model context and output ceilings.
• The model picker now supports per-model reasoning-effort controls, with a more consistent menu interaction and visual treatment.
• New DeepSeek profiles default to deepseek-v4-flash.

Agents & Commands

• Typing / now surfaces built-in slash commands in the command picker.
• When an agent reaches its maximum iteration count, it produces a turn summary instead of ending abruptly.
• In-band upstream streaming errors now expose their actual details, and an empty attempt_completion can no longer end a turn without a visible answer.

Stability & UI Fixes

• MCP stdio servers now terminate their full process trees, with a hardened shutdown path during cancellation.
• Fixed transient API turn interruptions, follow-up suggestions lingering after send, and Chinese chat text not following the UI font-size setting.
• Conversation outlines now animate open and closed and unmount hidden content, avoiding duplicate question text for assistive technology.
• Refined the model picker, composer, settings controls, context-usage display, and default font fallback.

Full Changes

All commits since v1.2.0:
v1.2.0...v1.3.0

---

wisp-science v1.3.0

v1.3.0 聚焦长会话与模型体验：加入真实长上下文记忆评估、动态上下文用量仪表和模型级推理强度选择；同时改进斜杠命令、全局习惯与会话结束提示，并修复流式响应、重复自动压缩、MCP 子进程退出及多处界面问题。

下载

• macOS（.dmg，已签名 + 公证）：请选择 Apple Silicon（aarch64）或 Intel（x86_64）对应架构；兼容版本可从应用内更新。
• Windows 安装包（MSI / NSIS）：已签名；如仍触发 SmartScreen，可选择「仍要运行」。
• Linux（deb / AppImage）：提供 x86_64 与 aarch64 两个架构。

见下方 Assets。

核心亮点

长上下文、记忆与压缩

• 新增覆盖约 68.5 万估算 token 的真实长上下文示例，以及可由真实模型运行的记忆评估数据集。
• 自动压缩改为根据实际上下文增长动态选择目标，避免达到阈值后反复压缩。
• 上下文用量升级为动态仪表，更清晰地展示当前窗口的占用明细。
• 设置页可直接将新习惯加入全局记忆。

模型与推理体验

• models.dev 模型目录随应用内置，按精确模型 ID 提供上下文与输出上限。
• 模型选择器支持逐个模型调整推理强度，并统一了相关菜单的交互与视觉。
• 新建 DeepSeek 配置默认使用 deepseek-v4-flash。

智能体与命令

• 输入 / 时，命令选择器现在会展示内置斜杠命令。
• 智能体达到最大迭代次数时会生成回合摘要，不再突兀地结束。
• 上游返回流内错误时展示具体错误；空的 attempt_completion 不再让回合无可见结果地结束。

稳定性与界面修复

• MCP stdio 服务退出时会清理完整进程树，并强化取消期间的关闭流程。
• 修复瞬时 API 回合中断、发送后仍显示后续建议、中文聊天字号未跟随设置等问题。
• 会话大纲开合加入动画，关闭后会卸载内容，避免辅助技术读取重复问题文本。
• 调整模型选择器、输入区、设置控件、上下文用量区域和默认字体回退的视觉细节。

完整变更

自 v1.2.0 以来的全部提交见：
v1.2.0...v1.3.0

What's Changed

• Add headless memory-v1 eval dataset: compaction modes, retrieval, stage loading by @xuzhougeng in #809
• Fix silent task failures and blank completions with DeepSeek/GLM (#798) by @xuzhougeng in #810
• Add live memory eval and a copyable long-context demo by @xuzhougeng in #811
• ui: replace native reasoning-effort select with in-menu option list by @xuzhougeng in #812
• Make chat text follow the UI font size setting by @xuzhougeng in #813
• memory: add global habits directly from settings by @xuzhougeng in #814
• ui: edit reasoning effort per model row in the model picker by @xuzhougeng in #818
• Hide follow-up suggestions as soon as the user sends by @xuzhougeng in #815
• Default new DeepSeek profiles to deepseek-v4-flash by @xuzhougeng in #817
• Animate conversation outline open and close by @xuzhougeng in #816
• docs: name workflow engines (snakemake/nextflow/oxo-flow) in local-env-setup skill by @ShixiangWang in #819
• seed: replace synthetic memory demo with a real recorded CLI session by @xuzhougeng in #820
• Fix transient API turn interruptions by @xuzhougeng in #821
• Bake models.dev catalog for exact-match context/output ceilings by @xuzhougeng in #822
• Summarize turns at the agent iteration limit by @xuzhougeng in #823
• composer: surface built-in slash commands in the / picker by @xuzhougeng in #824
• fix(mcp): terminate stdio process trees by @yeshubiao2017-source in #825
• Fix repeated auto-compaction and replace the fixed 60% target by @xuzhougeng in #826

New Contributors

• @ShixiangWang made their first contribution in #819

Full Changelog: v1.2.0...v1.3.0