What
A sandbox is a restricted space that limits what a command can reach, such as your files or the network. Claude Code tells the model about these limits through a note attached to the result of a tool call (an action the model takes, like running a command).
- A new "sandboxed" note replaces the old
sandboxed_failurenote. It now appears on sandboxed commands in general, not only when they fail. - The note says whether the command was limited for both files and network (
files_and_network) or for network only (network_only). - The first time in a conversation, the note is shown in a slightly different wording.
- When a conversation is compacted (shortened by summarising it), the flag that records the note was already shown,
toldOfSandbox, is reset, so the note can appear again. - The record kept for a compaction turn now also includes
responseBeforeCompaction.
Why
The model is told about sandbox limits more consistently, including after compaction, so it is less likely to be surprised when a command cannot reach a file or the network.