How do I hold Claude Code at a read-only Montage checkpoint?
Coding assistants can move from investigation into implementation within one conversation. Establish an explicit Montage checkpoint so that a request to understand the connected workspace cannot be interpreted as approval to save research or change workspace state. Use a named signals_status request to separate remote readiness from implementation. An assistant authorized to edit repository files still needs a distinct instruction before changing retained Montage research records.
Help me hold Claude Code at a read-only Montage checkpoint. Tell Claude Code that this stage concerns only the current Montage workspace and available coverage. Conclude with the verified workspace and unresolved limits, then obtain a new instruction before writes. No writes.Restate the requested boundary
Tell Claude Code that this stage concerns only the current Montage workspace and available coverage. File-edit permission in the repository is not permission to change remote research data. A clear bounded brief says “status only” before describing research goals, rather than relying on a general request to be careful.
Select the status tool explicitly
Ask for signals_status by name rather than an open-ended instruction to test everything. Review the advertised tool and do not introduce undocumented arguments to make a broader request fit. Use the status tool’s advertised input instead of importing parameters from a proposed research operation or another MCP server.
Review any proposed expansion
If the assistant proposes a watchlist or audience operation, stop at the proposal. A useful next step can be discussed without executing it as part of a connectivity check. Ask the assistant to leave a proposed mutation as plain text, naming its intended effect without issuing the call.
Read the execution evidence
Inspect the actual status response in the coding conversation and distinguish server errors from the assistant’s explanation. A narrative saying the workspace looks ready is not sufficient evidence. If the response is unavailable, describe that observed failure instead of allowing the assistant to supply an expected workspace from memory.
Make later approval separate
Conclude with the verified workspace and unresolved limits, then obtain a new instruction before writes. Stopping the assistant does not reverse any remote action that already completed. A later write request should name the desired effect; earlier permission to inspect code does not supply remote-workspace approval.
Keep the boundaries clear.
- Ask the assistant to leave a proposed mutation as plain text, naming its intended effect without issuing the call.
- A later write request should name the desired effect; earlier permission to inspect code does not supply remote-workspace approval.
Start with a read-only check: ask your agent to report actual source coverage before collecting records or exporting an audience.
Primary references
Documentation reviewed 2026-11-15. Product capabilities, client availability and plans can change. Third-party marks identify their products, not an affiliation.