Repository navigation
Support cancellation during first-turn sandbox initialization #433
Description
Activity
@thesujai Please take a look at the issue wording and modify as per requirement
I reproduced this on
7dc87574and prepared a candidate patch that forwards the turn'sAbortSignalintoSandboxand stops the turn from waiting when sandbox creation or initialization is cancelled.The patch includes 15 sandbox-level regression tests and two route-level tests, including a real loopback HTTP request to the session cancellation endpoint. Coverage includes cancellation during creation and post-create initialization, normal completion, provider errors, listener cleanup, and late completion after cancellation.
One boundary needs a maintainer decision: this stops the cancelled turn from waiting; it does not stop an already-running provider operation. A late create can leave a provider-side sandbox, and late initialization can still issue commands against an identity subsequently reused by another turn. The shared-identity test observes those calls; it does not establish that their filesystem or credential effects are harmless.
Would you approve opening a PR with this bounded scope, or should provider cancellation/shared-identity coordination be included before submission? I have not opened a PR, in accordance with the contribution guide. This is separate from #407's exec-cancellation work.
Description
When a user cancels the first turn while the sandbox is still initializing, the turn does not cancel promptly because the cancellation signal is not propagated to the sandbox initialization methods.
This is separate from sandbox
execcancellation addressed in #407.Steps to reproduce
Expected behavior
The cancellation signal should be propagated through the sandbox initialization flow so that cancelling the first turn stops initialization promptly.
Actual behavior
Sandbox initialization does not currently receive the cancellation signal, causing cancellation to be delayed.
Acceptance criteria