Fork or move a live agent session to another host

[jonocodes]

What I'm trying to do

Continue an agent conversation on a different host than the one it started on, without losing the transcript. Concrete cases I keep hitting:

• Started a Claude session on my laptop, the task turned out heavier than expected → move it to a beefier dev box
• Started a long session on a remote workstation, leave the desk → hand it back to the laptop
• Want to try "what if Codex on the workstation did this instead of OpenCode on the laptop" without re-prompting from scratch and re-establishing context

How I do it today

My current workaround is Matt Pocock's /handoff skill (aihero.dev/skills-handoff). It compacts the current session into a portable markdown file. I copy the plan content from that file over to the other host (manual copy/paste or scp), start a fresh agent there, and point it at the file: "read this and continue." The plan content carries over, but the full transcript and tool history don't — the new agent reconstructs from the summary.

I've also looked at the native /paseo-handoff skill. It builds the same kind of briefing but launches the new agent via create_workspace + create_agent on the current daemon — no --host parameter, no cross-daemon routing. It reads ~/.paseo/orchestration-preferences.json from the current daemon, which makes sense for within-host handoffs but rules out cross-host. So it doesn't bridge the gap I'm trying to fill.

When neither path fits, I just don't move the session and limp along on the wrong host.

Where Paseo gets in the way

The fork action only operates inside one daemon. Conversation state lives with the host's daemon, and there's no way to rebind or transfer it across daemons. The existing handoff primitives split the difference:

• Matt's /handoff is portable — a file you can move anywhere — but agent-agnostic: the receiving agent starts from the file, with no Paseo-managed continuity
• /paseo-handoff has agent continuity (auto-launches the new agent, manages the workspace) but is daemon-bound: it operates on the current daemon, no cross-host routing

I get portability or continuity, never both. With multiple hosts already configured in the web UI, the cross-host move is the obvious missing piece.

What it would look like

A "Fork to host…" option on the existing fork menu that lists my other connected daemons. Picking one would:

1. Take the current session (provider-native transcript + Paseo-side metadata)
2. Send it to the chosen host
3. On the target host, open it as either a continuation of the same session or a new agent seeded with the full transcript
4. Optionally keep the original session alive on the source host so I can compare or come back

The simplest version is "fork-style: new session on B with the old transcript." I'd be happy with that as a first cut. Live migration (the same agent process moves and keeps running) is nice but not what I'm asking for.

Adjacent

This is the cross-host sibling of the within-daemon workspace move requested in #3270. That's a daemon-local rebind (the agent stays on one daemon, switches workspace). My case adds the daemon-to-daemon transfer on top. If the within-daemon plumbing lands first, the cross-host version can build on it for the local rebind half.

Context

Already running two daemons (laptop + workstation) and managing them from the web UI. Not a hypothetical. Happy to test builds, write up a narrower spec for just the export/import half, or help in whatever way is useful.

One note on the model

I expect this would lean on the existing relay or direct-connection transport between daemons for the transfer itself, with the transcript serialized in a provider-native format on the receiving side (so re-import doesn't need replay logic per provider). I haven't read the source, so this is a guess at the shape — happy to be corrected.

[paseo-bot[bot]]

Your summary-based handoff to a beefier machine gains an automated launch path in merged PR #7370. The source agent can create a fresh agent on a connected destination through MCP and put the handoff briefing in initialPrompt, rather than requiring you to start that agent manually.

Register the destination with paseo host add <connection>, use list_hosts to get its serverId, and pass it to create_workspace and create_agent with the destination workspace and explicit provider/model. The remote child appears in the source agent's Subagents track and reports completion. This still starts a fresh conversation from your briefing; full-transcript transfer, live migration, and a “Fork to host” menu are not included.

This is merged for the next release. The current release, v0.12.0, does not contain it, and the next version number is not decided yet. Watch the releases page if you want it sooner; changes often appear in a beta before stable.

Thanks for separating the summary workaround from the transcript-preserving move you want.