Repository navigation
Releases
0.2.1 — 模型可以在设置面板里配置 / Configure models from the settings panel
0.2.1 — Configure models from the settings panel
English follows. / 中文在下。
Highlights
Models can now be configured from the panel. Every model shows its state and
the button that changes it — no hand-editing cordis.patch.yml:
| State | Button | Effect |
|---|---|---|
| 未加入配置 | 启用 | adds the model to the provider's models array |
| 已启用 | 取消配置 | removes it again |
| no provider block at all | 一键配置 | writes the complete default profile |
| user layer overrides the default | 恢复默认 | unsets it so cordis.patch.yml serves again |
A first install now declares exactly four free models: agnes-2.5-flash and
agnes-3.0-flash for chat, plus agnes-image-2.5-flash and
agnes-video-2.5-flash. The paid Pro models and the other media models ship
unconfigured and enable with one click.
Writes go through ctx.remote.settings.mutate() — the same Remote call the
built-in Models page makes — fenced by the revision the panel last read, so a
concurrent change reports settings/conflict instead of losing an update.
Also in this release
- Fixed: model entries declared
defaultMaxTokens, which is a provider-level
fallback. Inside amodelsentry it is an unknown field, so it was ignored and
output silently fell back to the route default of 32768 instead of the intended
65536. Entries now use the model-levelmaxTokens. - Reliability since 0.1.5: the plugin was rebuilt on a validated schema
contract (an invalid tool schema now fails the plugin at load instead of making
the provider reject every request in the session), tool results are always
lossless JSON, and@deepseek-ai/dsh-toolsis imported dynamically so a missing
host helper degrades instead of throwing inside ESM. - New:
verify-client.mjs— loads the client bundle through a stubbed module
table and exercises the functions the buttons call, including a cross-check that
cordis.patch.ymland the panel declare the same default models.
Note: this is the first release since 0.1.5; 0.2.0 was never released.
Install / upgrade
dsh plugin --profile <name> add git+https://github.com/wangyemen/agnes-ai-for-dsh.gitThen make sure agnes-ai-for-dsh is in the profile's dsh.profile.bundles and
restart DSH. Refresh the page to pick up the new settings panel.
Known behaviour
Image and video models are declared in the provider's models array with only
id and name, which is how the panel tracks them. They therefore also appear
in the model picker, where selecting one as a chat model fails — they are meant
for agnes_image_generation / agnes_video_generation only.
0.2.1 — 模型可以在设置面板里配置了
主要内容
模型现在可以在面板里就地配置。 每个模型都显示状态和对应按钮,不用再手改
cordis.patch.yml:
| 状态 | 按钮 | 效果 |
|---|---|---|
| 未加入配置 | 启用 | 把模型写进 provider 的 models 数组 |
| 已启用 | 取消配置 | 再把它移出去 |
| 完全没有 provider 声明 | 一键配置 | 写入完整默认配置 |
| 用户层覆盖了默认值 | 恢复默认 | 取消覆盖,回落 cordis.patch.yml |
首次安装默认只声明四个免费模型:对话的 agnes-2.5-flash、agnes-3.0-flash,
加上 agnes-image-2.5-flash 与 agnes-video-2.5-flash。付费的 Pro 模型和其余媒体
模型默认不启用,点一下即可加进来。
写入走 ctx.remote.settings.mutate() —— 内置「模型」页面用的同一个 Remote 调用,
并以面板上次读到的 revision 做并发保护:并发改动会返回 settings/conflict
并提示重试,而不是静默丢失其中一次更新。
本版本还包含
- 修复:模型条目里写的是
defaultMaxTokens,而它是 provider 级兜底值;
放在models条目里属于未知字段会被忽略,输出上限实际落到路由默认 32768,
而不是预期的 65536。现在改用模型级的maxTokens。 - 自 0.1.5 起的可靠性改动:插件在校验过的 schema 契约上重建(工具 schema 非法时
在加载时就报错,而不是让供应商拒绝整个会话的每次请求);工具返回值一律是
lossless JSON;@deepseek-ai/dsh-tools改为动态 import,宿主 helper 缺失时优雅
降级而不是在 ESM 里抛错。 - 新增
verify-client.mjs:用桩化的 module table 加载客户端 bundle,直接验证按钮
调用的那些函数,并交叉校验cordis.patch.yml与面板声明的默认模型是否一致。
注意:这是 0.1.5 之后的第一个发布版本,0.2.0 从未发布过。
安装 / 升级
dsh plugin --profile <name> add git+https://github.com/wangyemen/agnes-ai-for-dsh.git然后确认 profile 的 dsh.profile.bundles 里有 agnes-ai-for-dsh,重启 DSH,
刷新页面即可看到新面板。
已知行为
图像和视频模型以只有 id、name 的条目声明在 provider 的 models 数组里,这是
面板跟踪它们的方式。因此它们也会出现在模型选择器中,在那里当对话模型选中会失败 ——
它们只应被 agnes_image_generation / agnes_video_generation 调用。