#5778 [GUI] Scope opt-in side questions, deeper research and private workflow skills

open · @zapabob · 1 comentários · No GitHub

Setup & installModels & quantsSecurityDocumentationWindows

Descrição

### Area

[GUI] (the desktop app)

### What do you want to do?

Please agree the first small upstream scope for optional host workflows before implementation. This is a scope discussion, not a request to merge the fork wholesale or change upstream defaults.

When a long task is running, I want to ask a short side question, perform a deeper source-based investigation, and retain a reusable procedure after a successful workflow. The current app already has helpers, /research and skills; the proposed additions build on those paths.

There are three separable candidate slices:

- **/btw side questions:** an explicit command starts a dedicated tool-free child using a bounded snapshot of the conversation, with its own cancellation. It must not steer, stop or write into the main conversation. Use the selected model route and show the child's result. A single-slot engine may queue inference; concurrent GPU execution is not promised.
- **/deepresearch:** an explicit deeper mode of /research that gathers several distinct searches and relevant pages, reports sources and missing evidence, and respects existing tool/approval choices. Keep the current /research behaviour. A selectable RustXNG provider could be discussed separately; the fork's Windows prototype uses pinned RustXNG 0.31.1 Full on loopback. Bundling, licence/source notices, reproducibility, platform support and resource limits require upstream agreement. Paid providers still require the person's own opt-in and credentials.
- **Private workflow skills:** independently opt-in learning from host-observed successful completed tool workflows, and opt-in application of matching skills on later turns, including goal/plan/research paths. Fresh profiles keep learning and automatic application off. Bound stored instructions and prompt size, redact private values, honour workspace scope, expose applied skills, allow deletion, and preserve all tool permissions. A learned instruction must not grant itself authority or silently enable tools.

These would be separate PRs after scope agreement, starting with whichever slice is useful to maintainers. No new engine feature, default or automatic network access is requested here.

### What do you do today instead?

I use the experimental implementation in the zapabob fork. Reference code is pinned to [427f77ca225595b68cee750c5a367d8e6cee49f4](https://github.com/zapabob/Strata/commit/427f77ca225595b68cee750c5a367d8e6cee49f4):

- [Side-question validation](https://github.com/zapabob/Strata/blob/427f77ca225595b68cee750c5a367d8e6cee49f4/apps/host/server/chat-btw.mjs) and [Harness child plugin](https://github.com/zapabob/Strata/blob/427f77ca225595b68cee750c5a367d8e6cee49f4/apps/host/server/agent/plugins/strata-btw.mjs).
- [Research/turn integration](https://github.com/zapabob/Strata/blob/427f77ca225595b68cee750c5a367d8e6cee49f4/apps/host/server/agent/agent-bridge.mjs) and [RustXNG lifecycle](https://github.com/zapabob/Strata/blob/427f77ca225595b68cee750c5a367d8e6cee49f4/apps/host/server/rustxng-runtime.mjs).
- [Workflow learning](https://github.com/zapabob/Strata/blob/427f77ca225595b68cee750c5a367d8e6cee49f4/apps/host/server/skill-learning.mjs) and [bounded skill context](https://github.com/zapabob/Strata/blob/427f77ca225595b68cee750c5a367d8e6cee49f4/apps/host/server/skill-context.mjs).

I inspected these sources for this proposal. The fork's [qualification record](https://github.com/zapabob/Strata/blob/427f77ca225595b68cee750c5a367d8e6cee49f4/docs/HAKUA_RELEASE.md) describes its historical tests and limits; those results were not rerun as upstream feature validation here. General model quality, GPU concurrency, full-window behaviour and cross-platform feature support are not established by this proposal.

Searched open/closed issues and PRs on 2026-11-11 for btw, deepresearch, research, RustXNG, skills, workflow and subagents. #5254 concerns visibility of ordinary helper output and #4749 helper replies; this proposal concerns an explicitly requested, tool-free side question. #4521 concerns skill installation, rather than learning from completed workflows. Please redirect this discussion if an existing design already covers the scope.

No site

Links install, modelos, releases.