Skip to content

v0.2.0-alpha.4 — 辩论与阅读设置

Pre-release
Pre-release

Choose a tag to compare

@xiaomeng2568 xiaomeng2568 released this 04 Oct 11:52
d32df3f

Meldwise 0.2 Alpha 4

这一版开放了辩论模式,也让多模型回答更容易阅读。欢迎试试新的聊法,分享你遇到的问题和使用感受。

新的聊法:辩论

  • A 和 B 先各自回答,再互相审阅,最后由 Judge 整理成一份最终答案。两份初答、两份审阅分别并行进行,Judge 等两份审阅完成后再开始。
  • A、B 需要是不同的模型;Judge 可以使用 A、B,或另一个已配置的模型。每个角色可分别设置模型支持的思考偏好。
  • 完整一轮最多发起 5 次模型请求,费用和额度按各自服务计算。辩论不保证答案一定更好。
  • 辩论历史可以重新打开并继续追问。取消、失败或中断时保留已收到的内容,重试按整轮、原模型快照执行。

更方便地阅读

对比的两份回答、协作的初答和审阅、辩论的两份初答都可以独立收起、展开,并带有轻量过渡动效。收起后仍显示模型与阶段状态,不影响生成或历史内容。协作综合和 Judge 最终答案保持直接展示。

隐私和已有能力

跨服务商共享可见上下文与回答前,需要明确确认;许可只用于当前对话与确切的一组服务商。思考内容、凭据与诊断信息不会作为普通跨模型上下文传递。辩论后续轮次使用用户消息与成功的 Judge 最终回答,中间阶段留在本地历史。

原生公式、代码高亮与 JetBrains Mono、流式输出、取消、本地加密保存和主题设置继续保留。不自动切换模型或服务商,不自动重放失败请求。

仍是 Alpha

  • 对比仍为单轮;协作和辩论只支持整轮手动重试,不支持单阶段续跑。
  • 辩论保持固定五阶段,不提供胜负评分、任意代理循环或工具调用。
  • 代码高亮与数学渲染覆盖常用子集;中文代码注释由系统字体回退。
  • ChatGPT 可能受套餐或模型额度限制;SIWC 兼容性仍为 CONDITIONAL。
  • 更多设备、长时间使用和完整真机 AndroidTest 覆盖仍在继续验证。

本次本地完整回归:Debug 1453/1453、Release 1453/1453;构建、lint、AndroidTest 编译打包与依赖锁检查通过。验证包已由开发者手动验收。

Meldwise 自 v0.1.0-alpha.2 起采用 GPL-3.0-only。第三方组件遵循各自许可证,服务商名称和商标归各自权利人所有。


English

Meldwise 0.2 Alpha 4 opens Debate mode and makes multi-model answers easier to read.

Debate

  • A and B answer independently, cross-review, and a Judge puts together the final answer. Each pair of initial answers and reviews runs in parallel; Judge waits for both reviews to finish.
  • A and B must be different models. Judge may match either participant or use another configured model. Supported reasoning preferences are selected per role.
  • A complete round can make up to 5 model requests, with usage and billing following each provider. Debate does not guarantee a better answer.
  • Reopen and continue Debate conversations. Cancellation, failures and interruption preserve received content; explicit retries start a whole new round with the original model snapshots.

Reading controls

Compare answers, Collaborate initial/review stages and Debate initial answers can each collapse or expand with a lightweight transition. Model and stage status stay visible. Collapsing never stops generation or changes history. Collaborate synthesis and Judge final answers remain directly visible.

Privacy and existing capabilities

Sharing visible context and outputs across providers requires explicit confirmation for the current conversation and exact provider set. Reasoning, credentials and diagnostics are never ordinary downstream model context. Later Debate rounds use user messages and successful Judge final answers; intermediate stages stay local.

Native math, code highlighting and JetBrains Mono, streaming, cancellation, encrypted local state and themes remain available. No automatic provider/model fallback or failed-request replay is introduced.

This remains Alpha software: Compare is one-shot; Collaborate/Debate retry whole rounds only. Debate is a fixed five-stage workflow, without competitive scores, arbitrary agent loops or tools. Math/highlighting cover useful subsets; CJK code comments use platform font fallback. ChatGPT may remain subject to plan/model limits, SIWC compatibility remains CONDITIONAL, and broader device/long-running validation is ongoing.

Local full regression: 1453/1453 Debug and 1453/1453 Release. Builds, lint, AndroidTest compile/package and dependency lock checks pass. The developer manually accepted the validation build.

Meldwise is GPL-3.0-only starting with v0.1.0-alpha.2. Third-party components retain their own licenses; provider names and marks remain the property of their owners.