#6342 [GUI] agree scope for desktop agent continuity and context controls

open · @Held0fTheWelt · 3 commentaires · Sur GitHub

Setup & installModels & quantsWindows

Description

### Area

[GUI] (the desktop app)

### What do you want to do?

I'd like to agree the scope for several desktop and Harness improvements that I have been trying in my fork. The shared problem is keeping a long-running agent session usable and understandable as the user sends follow-ups, looks at a helper, approaches the context limit, or updates the app.

The related desktop runtime and UI changes are now grouped in draft PR #6394. The focused fixes below have separate drafts.

- **Follow-ups and helpers.** Keep running helpers and partial replies when the user sends a follow-up. Make a helper's status, log, stop, and archive actions available without losing the chat context. This work is in #6394, which supersedes the scope of my closed #6283 and #6275.
- **Update and resume.** Add an explicit pause that checkpoints active work before installation and reconnects eligible helpers to their owning chats after restart. I'd like to confirm how the updater should handle runs that cannot be checkpointed or resumed.
- **Harness continuity and planning.** Keep a small, opt-in working brief with session evidence and a plan step's verification target. In a separate change, detect repeated tool outcomes or stalled turns without ending work that is still making progress. Preserve explicit cancellation and model-switch behavior. The shared implementation is in #6394.
- **Context and memory settings.** Offer an opt-in context optimizer for large histories and an optional system RAM reserve shared by local model profiles. Confirm which limits belong in user settings before proposing either control. For the optimizer and reserve, preserve the current behavior when the new settings are off or unset.

The [fork](https://github.com/Held0fTheWelt/Strata) contains the integrated implementation. I'm looking for maintainer feedback on which of these proposals fits Strata and what scope should be reviewed first. The SSD admission draft is #6393, with related discussion in #5835.

The focused drafts cover the workflow skills (#6383), summary retries (#6385), concurrent tuning inputs (#6386), tuning rechecks (#6387), large-context reply limits (#6390), and the reused Harness stage check (#6391).

### What do you do today instead?

I use my Windows fork for long sessions and inspect helper output and model settings there. The fork has only been tried on my machine for this workflow. I can provide focused regression cases and separate hardware measurements once we agree on the scope.

<details>
<summary>Original field notes from my Windows setup</summary>

I am running on Windows (still), a Ryzen 9 7950X3D with an RX 9700 AI Pro and 64 gigs of Ram, so i am capable to run the Qwen solutions quite fast.
My own speed measurements lifted me up to 100-120 token/s having a prefill of about 1000+.
With my MSI B950 Gaming Wifi, i do believe this is quite the limit i get, but i am still tuning.
So my solution is not tested on a broad range of systems, but in the code there were several harness problems, where old informationen did compete with new information on different parts of the system. The changes can be seen in the history.
I also reworked the reasoning, especially the high and max section, the thinking budget and the tuning part.
Just saying: On my current system Swift 1.5 IQ2_XS with a System Ram to keep free of 12 GB, Strata uses 6912 Mib of Cache, which does a lot on reusing the prompts given.

</details>

Sur le site

Liens install, modèles, releases.