Repository navigation
Replies: 2 comments
|
Not a promise about side effects the checkpoint is written from the UI layer only for calls sitting in const restorableToolCalls = toolCalls.filter(
(toolCall) =>
EDIT_TOOL_NAMES.has(toolCall.request.name) &&
toolCall.status === CoreToolCallStatus.AwaitingApproval,
);and that set is two tools export const EDIT_TOOL_NAMES = new Set([EDIT_TOOL_NAME, WRITE_FILE_TOOL_NAME]);second gate
the storage layer is not the limit though so a shell command that writes inside the project root would be captured and rolled back fine two edges to keep in mind if it ever gets extended |
|
The gap is real in practice, not just on paper — I lost work to exactly this shape twice, 1. A shell-driven build overwrote a hand-maintained file. A worker ran a generator whose 2. A backup taken by a shell command became the incident. On the boundary question itself, I'd separate two things that the current design conflates:
If the guarantee is what's blocking coverage, the honest middle is to snapshot before shell One caveat from the same incidents, in case it informs where the hook goes: the approval layer |
Uh oh!
There was an error while loading. Please reload this page.
The checkpointing doc says a snapshot lands before approved file-modifying tools such as write_file and replace, bundling the shadow commit, the conversation, and the pending call. run_shell_command can modify files too, and I don't see a checkpoint preceding it, so the restore list has gaps exactly where changes are least predictable. This is a boundary question rather than a feature request. Was excluding shell commands a deliberate call, on the grounds that a snapshot can't promise to contain effects that may leave the project directory anyway? Or is it mainly that the hook predates the shell tool and hasn't been extended yet?
All reactions