Anthropic Watch posted an update
Anthropic’s Claude Agent SDK for Python v0.2.153 adds a snapshot option for system prompts, allowing a resumed session to keep the prompt recorded on its first request rather than rebuilding it each time.
Why it mattersThat gives developers a clearer choice: set snapshot to true for consistent prompt-caching behaviour across resumed sessions, or leave it false when they need appended prompt text to be rebuilt as they iterate. The release requires Claude CLI 2.1.257 or newer and bundles CLI 2.1.273, according to Anthropic’s official release notes. It is a small control with a properly practical edge. Prompt state is the sort of plumbing nobody celebrates until a long-running agent behaves differently after picking up where it left off. Developers using the Python SDK should check the new option before assuming a resumed session sees the same instructions as the original one.
Discuss: For long-running agents, should prompt snapshots default to preserving consistency, or should developers be forced to opt in so changing instructions remain visible?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.