CLAUDE_CODE_WORKFLOW_SUBAGENT_MODEL: model choice vs Sume spend gates

CLAUDE_CODE_WORKFLOW_SUBAGENT_MODEL, added in Claude Code 2.1.296 on October 9, 2026, runs every workflow agent on one model while other subagents keep theirs (changelog read 2026-10-13). If those workflow agents call Sume, switching to a cheaper model changes your token bill, not Sume's rules: paid tools still need an idempotency_key, and dry_run and max_spend_credits work the same on any model.

What does the variable do?

The changelog entry is one line: it runs every workflow agent on one model while other subagents keep theirs. The same release also priced Sonnet 5.5 cache reads at $0.10 per million tokens in /cost, the status line and --max-budget-usd (was $0.20), and Claude Code 2.1.293 added Claude Haiku 5.5 as the default Haiku at $0.10 and $0.50 per million tokens, with $0.50 and $2.50 above 100K-token prompts (all read 2026-10-13). I did not find the variable on the environment variables page I fetched, so confirm the accepted values in your install.

Should workflow agents use a small model for Sume work?

A workflow agent that only submits jobs, polls jobs_status and writes ids down is a good fit for a small, cheap model, because the expensive work happens in Sume's render, not in the agent's tokens. The risk is not capability but fan-out: many workflow agents can submit at once, and they all share your workspace's generation capacity.

Keep the paid step narrow. Give only one agent a Sume server with write access, or use a read-only OAuth scope for the others; with mcp:read only, write and paid calls return insufficient_scope.

What to set before turning workflows on

• Have a parent generate the idempotency_key per intended job, so a retried agent reuses it.
• Run dry_run=true first and read usage.credits.
• Pass max_spend_credits on every paid call.
• Use generation_admission_preview before an expensive burst, as the MCP tools page recommends.
• Treat --max-budget-usd as a limit on Claude Code's cost only; it is not a Sume limit.

What this does not cover

Sume does not know which model called it, so it cannot enforce a model policy. If you need that, enforce it in Claude Code settings. I also have not measured any workflow run here; the figures above are only the prices and defaults the changelog lists.