Ultracode mode can silently consume an entire 5-hour usage window on subagent fleets

What happened

On a Max ($200) plan, a Claude Code session with ultracode mode enabled ran three multi-agent workflows as part of otherwise normal feature work on my repo:

• a 22-agent "assessment" workflow (~1.7M subagent tokens)
• a 12-agent adversarial review (~0.9M subagent tokens)
• a 20-agent adversarial review (~1.5M subagent tokens)

Roughly 4.4M tokens went to subagents — the large majority of my 5-hour usage window — while the visible main-conversation work modified a few dozen files. From my side the session looked like ordinary coding; the burn happened in background fleets I never individually approved.

Why this is a problem

Ultracode's instructions tell the model that "token cost is not a constraint" and to use multi-agent workflows on every substantive task. But plan users have hard usage windows. The combination means a single afternoon session can exhaust a Max plan's window with no warning, no estimate, and no per-workflow confirmation. The model is following its instructions; the instructions don't account for metered users.

Suggestions

1. On metered plans, require a confirmation with an agent-count / rough token estimate before each workflow launches.
2. Give workflows a default token budget unless the user explicitly raises it (the budget mechanism already exists for +500k-style directives — make a conservative default the baseline).
3. Surface subagent share of usage in /status while the session runs, so the burn is visible when it's happening rather than after.
4. Make ultracode opt-in per task instead of session-sticky, or auto-disable it past a usage threshold.

Environment: Claude Code desktop (macOS), Max $200 plan, model claude-fable-5.