# Decide workspace commit semantics for failed runs

Type: grilling
Status: open

## Question

When an agent run fails, is canceled, times out, loses the database, or is
killed mid-tool, should its workspace changes remain as partial collaborative
work, be quarantined for review, or roll back atomically? Decide the contract,
including the relationship among provider cursor, AgentMessage,
AgentFileChange, git/filesystem state, and the next run, then create the
smallest implementation tickets and crash tests that enforce it.