Skip to content

feat(coherence): AI-driven conversational entrypoint — talk-first UX/UI POC #2486

Description

@rozagerardo

Context

Hypha's interface today is mouse-led and menu-driven — you arrive and have to already know
where to click. #2478 (AI-assistance
layer architecture) surfaced a delivery-surface / UI-UX redesign thread: the human-machine
chat interface is where Hypha should lead, and both Joachim and Vlad agree a large UI/UX rethink
is needed. This ticket is the concrete, buildable spike for that thread — a working prototype the
team can see and try, not a doc.

The idea: replace the entry point with a talk-first experience. Like walking into a store and
being met by someone who greets you, offers the actions you're likely to want (more so if we
already know you), and otherwise lets you say what you need — the system then acts on the request
and the ongoing conversation rather than making you navigate to it. It is not a wholly original
concept, but doing it well, wired to Hypha's own signals / agreements / governance data, is where
the value is.

What this involves

On a hypha-web feature branch, reusing existing components wherever possible. This is
version 0 / version 1, not a throwaway — simplicity over a spotless solution, but on correct
foundations: the design definitions and abstractions must be right so it can be iterated on top of
(v0 → a good solution) without a rewrite. Three stacked regions:

  1. Persistent interaction bar (top) — always present, so the user feels a fluent, constant
    conversation with the system. Rendered as a moving sound-wave to signal "this is something you
    talk to".
  2. Next-actions strip (below the bar) — suggested next steps we think the user is likely to
    want; adapts to what they've asked and to what we already know about them.
  3. Contextual canvas (main area) — a widget/component surface that reacts to the conversation.
    Mostly contextual information (e.g. signals filtered to a task/objective, related
    agreements, etc.), though widgets stay interactive like any other UI. Every widget must be
    parameterizable so it can be filtered/adapted to what the user asked for.

Plus:

  • Mode toggle (in or beside the bar) — switch between talk-first mode and the classic
    mouse-led UI, for backward compatibility and users who prefer it.
  • Interaction model: user opens a conversation → assistant presents the likely options → user
    says what they need → the actions strip and the canvas widgets update to match. All real
    interaction flows through talking to the system; clicking widgets is secondary.

Design intent / longevity (informs architecture decisions now, not scope now)

Build it so the reusable core could later be extracted as a library — a way to present an
existing UI conversationally, applicable to other pages / other apps, with guidance for a
consumer on how to adapt their components and what to wire up. Not built in this POC, but the
decisions now must not foreclose it: a clear seam between the generic conversational shell
(bar, actions strip, canvas orchestration, conversation → widget-state mapping) and the
Hypha-specific wiring (components, data, personas); widgets integrated via a defined
parameter/filter contract
, not ad-hoc props; config-driven registration of widgets and
actions. Take the time to make the right design/architectural calls here.

Depends on / relates to

Approach / status

  • Discovery + full spec are deliberately deferred to a dedicated, scoped session — not done
    in the session that opened this ticket. That session covers: component inventory for reusable
    parameterizable widgets; whether the current AI SDK setup supports a persistent streaming bar +
    tool-driven canvas updates without a redesign; prior art for conversational-canvas UIs; the
    interaction-model spec; and demo acceptance criteria.
  • This issue exists to capture the objective and framing now so the work is on the board.
  • Ticket workspace (full technical context):
    progress/members/gerroza/tickets/<id>-ai-conversational-entrypoint-ux-poc/ in the context repo.

Open questions (for the scoped session)

  • Where exactly the seam sits between the generic conversational shell and the Hypha-specific
    wiring, and what the widget contract looks like (so a future library extraction stays open).
  • Which existing components are already parameterizable enough to drop into the canvas, and
    which need a thin filter wrapper.
  • Does the current AI SDK setup support an always-on conversation surface (streaming + tool
    calls that mutate the canvas) without a redesign.
  • Voice in scope for the POC, or text-only with the waveform as pure affordance.
  • How much "we already know this person" personalization is realistic for the POC vs. faked
    for the demo.
  • Scope of the mode toggle — a real switch back to the current app, or a stub for the demo.
  • Is coherence the right home for this, or does it need its own app-shell track.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions