Compute environments: let a central session dispatch long-running jobs to remote machines

[ch99q]

Before submitting

• I searched existing issues and discussions and did not find a duplicate.
• I am describing a concrete problem or use case, not just a vague idea.

Area

apps/desktop, apps/server, remote environments

Problem or use case

Today an SSH environment hosts the whole session. T3 Code installs its server on the target and I start sessions there.

I want the opposite split: my sessions live on my laptop or a central server, and agents can use several remote machines purely as compute. Typical jobs are GPU training runs or CPU-heavy processing that run overnight. The targets are a mix of Linux, macOS and Windows.

Proposed solution

A "compute environment" type, added over SSH like today's environments, that:

• Installs a lightweight job runner on the target as a persistent service (systemd, launchd, Windows service).
• Exposes agent tools to submit a job, check status, stream logs, cancel, and fetch outputs.
• Shows a job list per compute environment in the UI, with status and logs.

The session keeps running where it is; only the job runs on the target.

Why this matters

It separates where I work from where compute runs. One session can use many machines, and jobs survive disconnects, laptop sleep and reboots.

Smallest useful scope

Linux targets only, with submit, status and logs tools available to the agent.

Alternatives considered

• Agents running detached jobs over plain SSH (nohup, systemd-run, schtasks). Works, but there is no job tracking, no UI, and behavior differs per OS.
• External schedulers like Slurm, Ray or SkyPilot. Powerful, but heavy to set up for a few machines and not integrated with T3 sessions.

Risks or tradeoffs

• Cross-platform service management adds complexity.
• Job outputs, credentials and environment variables need safe handling.
• Some overlap with existing schedulers; scope should stay small.

Examples or references

Related to #12324 (desktop-managed environments) and #10940 (multiple routes to one environment).

Contribution

Happy to test on Linux, macOS and Windows targets.