continue-on-fallbacklisted

# continue-on-fallback — keep the work moving past a dead lane You are mid-work and the lane you were running on just stopped: the cloud budget is spent, the local MLX endpoint died, or the context is too large to continue. The goal is to **keep producing** on whatever lane is actually available — local first, then an explicitly-approved remote free allowance — and, if you must stop, to leave a handoff that another agent can resume exactly where you stopped. ## Decide the routing mode from the situation, not from habit Pick the mode that matches what actually stopped. Local is free and private; remote is the emergency, and it is your data leaving the machine. - **Budget spent, local MLX alive** → `local-first-free`. Local does the work for $0; a remote free tier is the backup only if local fails for a clean reason. - **Budget spent, local MLX also dead** → `remote-free-only` (a remote free allowance is the only thing left that isn't spending a dollar). - **You will genuinely stop and resume later** → build a handoff (below) *before* the context is too large to carry the state. Default to the least-permissive mode that still makes progress. Do not reach for `include-trials` or `research-emergency` on a routine task — those touch non-renewing or unpredictable allowances and are for a research box whose only goal is to keep output flowing. ## Remote is a privacy event — make it visible, never silent The router and dispatch print a loud `⚠ REMOTE provider=… model=… k