You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Start a persistent conversation on a selected device
#789
Superseded by issue #815, which expands this proposal with current implementation references, related issues and acceptance scenarios. Please continue the conversation there. Closing this discussion to keep the proposal in one place; this does not mean the feature has been implemented.
What I would like to do
Start a conversation directly on another device and keep talking to that device from my current phone or computer.
For example: on my phone, choose my Windows desktop, start a new conversation, and send follow-up messages that continue in the same session on the desktop. The desktop uses its own model, tools, credentials, and local environment; the phone is the interface for following the work and answering approvals.
This differs from asking the local Muse to delegate one task with @device: the execution destination belongs to the conversation and remains visible throughout it.
Suggested experience
A device selector when creating a conversation, plus a Chat on this device action in Devices.
A persistent destination in the conversation header and chat list: device name and icon, optionally complemented by a stable device color. Color should not be the only identifier, and online/offline status should remain visually distinct.
Every follow-up goes to the same remote conversation, without asking the local model to reinterpret a delegation instruction each turn.
Changing the destination starts a new conversation rather than silently moving history, files, or permissions.
If the destination is offline, keep the draft and explain the situation. Do not silently run locally or queue automatic execution.
Keep @device available for one-off delegation. A selected conversation destination should remain unambiguous if mentions are used inside it.
Existing code that could be reused
I reviewed the current Hub task path before proposing this:
Android's DeviceMention adds a one-turn instruction asking the local model to delegate through nanomuse-pc task.
Hub tasks already accept a conversation identifier. Android and desktop keep mappings to reusable target sessions, so continuing a remote session does not appear to require a second task protocol.
Existing task events, approvals, cancellation, and desktop streaming provide useful building blocks.
Remote task sessions are excluded from ordinary conversation synchronization. A device selector alone would therefore not solve history recovery and reopening after restart. We should define a remote-conversation representation without requiring users to enable synchronization of all side chats or producing duplicate histories.
These are code observations, not a claim that the complete experience has been tested across clients.
#514 — remote response streaming: directly relevant. The desktop code I inspected already listens to assistant stream events and supports stream: true with partial text. This proposal should reuse that work and validate the receiving clients, including replacement of partial text rather than duplicate messages. It should not introduce a competing streaming mechanism.
#616 — cross-client workflows: complementary. Being able to follow work and answer its approvals remotely is essential here. However, choosing a persistent execution device for a conversation differs from moving an individual failed tool call to another device.
For an initial version, I would keep approvals available to the initiating client using the existing task ownership rules. An account-wide approval inbox could follow the direction discussed in #616, without making its entire scope a prerequisite.
Credential/settings synchronization, automatic fallback to other devices, private-host allowlists, and a judge model would remain separate work. The selected remote device can use its existing configuration.
Design questions for maintainers
Does a conversation explicitly bound to a device fit the intended relationship between the shared main chat, side chats, and Hub task sessions?
Should the initiating client store a persistent remote-conversation entry backed by the existing Hub session mapping, or is there a preferred shared representation for reopening and history recovery?
What should the history/reconnection contract be? In particular, how should a client recover a completed response after disconnecting and avoid executing the same turn twice when delivery is uncertain?
Would name + icon + an optional stable device color be a useful identification pattern for both remote conversations and @device mentions?
The initial goal is explicit device selection, persistent follow-ups, streamed progress where supported, approvals, cancellation, and reliable reopening—not automatic device selection or conversation migration. I would like to agree on the design before implementing it, starting with Android and desktop while keeping the Hub contract usable by other clients.
reacted with thumbs up emoji reacted with thumbs down emoji reacted with laugh emoji reacted with hooray emoji reacted with confused emoji reacted with heart emoji reacted with rocket emoji reacted with eyes emoji
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
What I would like to do
Start a conversation directly on another device and keep talking to that device from my current phone or computer.
For example: on my phone, choose my Windows desktop, start a new conversation, and send follow-up messages that continue in the same session on the desktop. The desktop uses its own model, tools, credentials, and local environment; the phone is the interface for following the work and answering approvals.
This differs from asking the local Muse to delegate one task with
@device: the execution destination belongs to the conversation and remains visible throughout it.Suggested experience
@deviceavailable for one-off delegation. A selected conversation destination should remain unambiguous if mentions are used inside it.Existing code that could be reused
I reviewed the current Hub task path before proposing this:
DeviceMentionadds a one-turn instruction asking the local model to delegate throughnanomuse-pc task.conversationidentifier. Android and desktop keep mappings to reusable target sessions, so continuing a remote session does not appear to require a second task protocol.These are code observations, not a claim that the complete experience has been tested across clients.
Relationship to #514 and #616
#514 — remote response streaming: directly relevant. The desktop code I inspected already listens to assistant stream events and supports
stream: truewith partial text. This proposal should reuse that work and validate the receiving clients, including replacement of partial text rather than duplicate messages. It should not introduce a competing streaming mechanism.#616 — cross-client workflows: complementary. Being able to follow work and answer its approvals remotely is essential here. However, choosing a persistent execution device for a conversation differs from moving an individual failed tool call to another device.
For an initial version, I would keep approvals available to the initiating client using the existing task ownership rules. An account-wide approval inbox could follow the direction discussed in #616, without making its entire scope a prerequisite.
Credential/settings synchronization, automatic fallback to other devices, private-host allowlists, and a judge model would remain separate work. The selected remote device can use its existing configuration.
Design questions for maintainers
@devicementions?The initial goal is explicit device selection, persistent follow-ups, streamed progress where supported, approvals, cancellation, and reliable reopening—not automatic device selection or conversation migration. I would like to agree on the design before implementing it, starting with Android and desktop while keeping the Hub contract usable by other clients.
All reactions