Auto-resume message after weekly-limit reset reads as user authorization to keep burning the new week's quota

Context

I asked Claude to intentionally use a non-default model (Fable) with heavy multi-agent orchestration (Workflow tool, "ultracode" mode) to make productive use of ~78% of that week's remaining usage before the weekly reset. That was a deliberate, one-time instruction. In the same breath I gave a standing rule for what should happen at the reset: "once we hit 100% usage, don't burn into the new week — switch to a cheaper/downgraded model."

What happened

Mid-task, a background sub-agent call failed with You've hit your session limit · resets HH:MM (timezone). After the reset time passed, the harness auto-injected a resumption message into the transcript, formatted as if it came from my (the user's) turn:

"Alcancé mi límite de uso mientras trabajabas, pero ya se restableció. Continúa donde lo dejaste."

I never typed this — it's the CLI's own auto-continuation nudge. Claude then relaunched the same heavy Fable-based workflow without pausing to ask, and even described it in its own follow-up text as "sigo con Fable como pides ahora" ("I'll keep using Fable as you're asking now") — attributing live authorization to what was in fact an automated system message, directly contradicting the standing rule I'd given for exactly this moment. By the time I noticed and manually ran /model to switch, roughly 20% of the freshly-reset weekly quota had been consumed on the expensive model.

Why this is easy to get wrong

The auto-resume text is styled and positioned exactly like a user turn, so it's easy — for the agent, and honestly for a human skimming quickly — to read it as live consent to keep going, especially when a model choice was explicitly time-boxed ("use up leftover quota this week only"). There's no visible marker distinguishing "the harness nudging the agent to continue" from "the user actively re-authorizing continued spend into a new billing period."

Suggested improvements

1. Make the auto-continuation message unambiguously non-user (e.g. wrap it as a clear system/tool event rather than user-turn-styled text), so it can't be quoted or reasoned about as if the user said it.
2. When a session was running a non-default, explicitly costly model and hits the weekly limit, consider resetting to the default/session model on the next turn rather than silently carrying the same model forward — require an explicit re-selection.
3. Some visible confirmation step when a session's activity crosses into a new weekly quota, so carryover spend isn't a surprise.

Impact: ~20% of a freshly-reset weekly quota consumed in one session before manual intervention, on an instruction that had explicitly told the agent to downgrade at that exact point.

I understand this repo's issues aren't a billing/refund channel — I'm not asking for a credit here, just flagging the product gap that made the mistake easy to make.