claude --fallback-model when overloaded: Sume paid calls

--fallback-model makes Claude Code switch to a named model when the primary is overloaded. A fallback changes who writes the next turn, not what Sume does: max_spend_usd, idempotency_key and the job you already started are unchanged, so a mid-run switch cannot double-bill a render if the agent reuses the idempotency_key or reads the job it already started.

What does the flag do?

The CLI reference lists --fallback-model as automatic fallback to the specified model or models when the primary is overloaded. It says nothing about what the second model remembers, so assume it reads the same conversation and nothing more.

What can go wrong with Sume jobs?

The risk is a model that re-runs a create it did not see finish. Two guards cover it on Sume:

• Every paid or write MCP call needs idempotency_key. Use a key derived from the task, not from the turn.
• Print the job id as soon as a create returns, so the next model waits on it with jobs_wait instead of submitting again.

Does the fallback model need the same tools?

The tools come from the MCP config, not the model. Whether a given Sume model is available to your agent is a separate question: the catalog is read with catalog_list. Do not assume a model name from a changelog is in it.

Which cap protects you?

Claude Code's budget flag is a client-side estimate. The Sume cap is max_spend_usd on the call. Use both; see MCP tools and gates.