ZiyaandZiyaZhang
5e9bb51724
* feat(plugins): hide Ralph and session search from default discovery * fix: keep hidden bundled plugins explicitly activatable --------- Co-authored-by: ZiyaZhang <199893125+ZiyaZhang@users.noreply.github.com>
3.0 KiB
description, kind
| description | kind |
|---|---|
| Run explicitly requested repeated work with fresh agents. | package-bundle |
@deepseek-ai/dsh-experimental-ralph-bundle
English | 中文
Summary
Add the experimental Ralph tool to every Agent as a global tool for human-requested repeated work with fresh child agents. It ships switched off in Web and Desktop and is hidden from the default plugin lists.
Table of Contents
- Use this package
- Understand the implementation
- Further Exploration
- Model Experience
- Known Limitations and Deferred Work
- Dev Note
Use this package
In Plugins → Add plugin, enter @deepseek-ai/dsh-experimental-ralph-bundle, choose Install, then Enable now. DSH uses the shipped copy without downloading another one. Alternatively, add the name to dsh.profile.bundles in the target profile's package.json. HMR applies the selection live when enabled; otherwise restart the application. The enabled bundle appears in Plugins; switching it off removes its profile layer and hides its card again. A separately installed profile dependency remains listed while disabled.
With HMR enabled, enabling or disabling the bundle registers or removes its global tools in the running Host, so every Agent sees the change on its next request, in every preset including minimal.
Understand the implementation
Implementation details — click to expand
cordis.patch.yml inserts the Host group optional-ralph, which isolates its own workflowEngine and registers tool-ralph in the global tool layer. Later profile patches can replace the tool's configuration or disable it by row id; an override alone does not select the bundle. The profile composer owns patch ordering and errors.
Further Exploration
Model Experience
Indirectly, through the capability implementation, which owns model context and results.
KV Cache effect
The layer adds no request content directly; the capability implementation owns cache effects from tools, prompts, and results. Toggling the bundle changes the tool list of live Sessions, which invalidates their cached request prefix once.
Known Limitations and Deferred Work
- Ralph requires explicit human intent to run a loop and can consume substantial model tokens across rounds. Its workflow engine is isolated from other workflow tools.
Dev Note
Maintainer context — click to expand
None.