An agent at rest stops costing anything
A workspace whose agent has finished its turn and is waiting for nobody in particular still holds a container's CPU and memory. Grove notices when an agent has been at rest past a cooldown, suspends its runtime in place, and wakes it the moment a message arrives. It works like a laptop lid: close it when nobody is using it, and everything is exactly where you left it when you open it again.
stateDiagram-v2
direction LR
[*] --> Active
Active --> Idle: turn ends, nothing pending
Idle --> Active: transcript moves
Idle --> Frozen: at rest past the cooldown
Frozen --> Thawed: a message arrives
Thawed --> Active: the message is delivered
When a workspace counts as idle
Idle is decided from the agent's own activity, and it is conservative by construction. Anything that suggests work in flight keeps the workspace awake.
- The agent must be at rest. Its main session reads as idle or waiting, with no question pending, no sub-agent running and no queued message.
- A pending question is never idle. An agent asking a person something reads as blocked, so it is never suspended while it waits for the answer.
- The transcript must be still for the whole cooldown, fifteen minutes by default. Any new line resets the clock.
- A paused or already frozen workspace is skipped, and so is one whose session cannot be read.
- Each candidate is confirmed against a fresh reading before it is frozen, so a stale dashboard view can never freeze an agent that just started working.
Freeze and thaw
Freezing is not pausing. A pause drops the worktree and keeps the branch, while a freeze keeps the worktree and the agent's process exactly as they are.
- A container is paused in place. Its memory is kept and it uses no CPU, so a thaw is instant and the agent continues mid-session.
- A host workspace is marked frozen and left as it is. An agent at rest on the host costs a tmux server and an idle process, and there is nothing to suspend that would pay for the thaw.
- The next message wakes it. Sending a message, from the web composer,
grove message, MCP or a peer's mail, thaws the workspace first and then delivers the message. - A frozen workspace reads as idle, not offline. A paused container looks stopped to Docker, and Grove reports it as idle so the dashboard never suggests a respawn that would throw away a session the next message simply wakes.
- A dirty worktree is no reason to refuse. Nothing is discarded by a freeze, so uncommitted work stays exactly where it was.
How it runs
One host-wide idle check rides the daemon's existing watch scheduler, so it adds no loop of its own. It checks roughly ten times per cooldown, which bounds how late a freeze can land. Each freeze and thaw is recorded on the workspace and emitted as a span when telemetry is on. Remote runtimes refuse to freeze until their worker can push the branch first, see Runtimes.
Configuration
The cooldown is at least a minute. Every field is in the Configuration Reference.