$ the-wire · showcase
Verbatim prompt mode closes an @-path file read in the Python SDK
By RepoJournal · Filed · About Anthropic · Composed from the cited sources · methodology
The Claude Agent SDK for Python gains a verbatim_prompts option so that in headless/SDK mode an `@/absolute/path` token in a user message is no longer silently expanded into the contents of that file before the model runs.
Add verbatim_prompts option to deliver prompts as written anthropics/claude-agent-sdk-python
The expansion happens before the model runs, outside the working directory, and without a tool call, so it is not stopped by tools=[], allowed_tools, or disallowed_tools=["Read"]. Any app that inlines text its end user did not type (earlier turns, tool output, third-party content) into a prompt can have local files read and echoed back, which makes verbatim_prompts the setting to adopt in servi...
Add verbatim_prompts option to deliver prompts as written (#1269) anthropics/claude-agent-sdk-python
Landed on main as commit f7547d7233527739ece8b12ed28c57be96c966b5, delivering prompts as written rather than expanding @/absolute/path tokens.
Action items
- → Adopt the verbatim_prompts option in claude-agent-sdk-python apps that inline earlier turns, tool output, or third-pa... anthropics/claude-agent-sdk-python [plan]