prompt injection uses a TON of tokens.

prompt injection uses a TON of tokens. If I have plugins they are used globally, and as such they're injected into every prompt. It means if I'm using local models (or heck even just doing API cost) theres a ton of tokens I don't always need. We need a way to tie plugins to sessions, or specific workflows. A way to enable or disable them situationally instead of just globally. A local model can NEVER be used with this app unless it has minimum 64k context window, but 128k is better. Even with that, session length must be kept short so as not to overwhelm the models.

---
| Field | Value |
| --- | --- |
| App version | 1.0.17 |
| OS | macOS 26.5.2 |
| Theme | GitHub |
| Path | /chat |
| Tenure | Week 14 |

Start at the /chat entry point and trace how plugins are included in prompts and how sessions or workflows are represented. Define how plugins can be enabled or disabled per session or workflow, then verify that unused plugins are excluded from prompts and that the chat still works with shorter model context windows.

Written by the indexing model from the issue text.