Escalate — rare, scoped, cost-capped

The default is $0 local grinding. Escalation to a cheap cloud frontier model is the exception, not the habit. mcp-escalation is OFF by default and has a hard daily USD cap enforced in the server.

Escalate ONLY when ALL hold

• The subproblem is genuinely hard for the local model (subtle algorithm, dense cross-file reasoning, a tricky proof-like constraint) — not routine work.
• It is well-scoped: you can state the problem and the success criterion in a self-contained prompt, with the relevant code/context attached.
• You have already tried locally and have a clear record of what failed (you've effectively passed the workflow-stuck bar on reasoning, not on a missing decision).

Do NOT escalate

• Routine edits, boilerplate, formatting, test scaffolding.
• Anything blocked on a human decision (use workflow-stuck → ping instead).
• Repeatedly, to brute-force a vague problem — that just burns the cap.

Repeated stuck mid-run → rises to the CONDUCTOR

The conductor planned the task; the escalation ladder rises back to it. When you are stuck or an issue repeats mid-run — a step failed twice, you hit an architectural decision PLAN.md didn't cover, or you're spinning — do not keep grinding locally and do not silently lower the bar. Rise to the conductor:

• A reasoning gap (you can't resolve a design/algorithm question in budget, a step is complexity: HIGH, or verify failed twice) → call reasoning_escalation — the SAME conductor/frontier reasoner that wrote your plan. Act on its ## Frontier guidance.
• A plan step is impossible as written (a referenced API doesn't exist, the approach can't work, verify failed 3+ times on one step) → call review_and_adapt; the conductor REVISES PLAN.md from that step. Do not attempt impossible implementations.

Both are run-capped (see workflow-plan). This is the rule: a repeated local failure escalates to the conductor, it does not get brute-forced or quietly dropped.

How (scoped one-off escalation)

1. Assemble a self-contained prompt: problem statement, the minimal relevant code, constraints, and the exact success criterion.
2. Call escalate(task, tier) (mcp-escalation). If it returns disabled or cap-reached, respect that — fall back to workflow-stuck and ping the operator. Never try to route around the cap.
3. Treat the cloud result as a proposal: integrate it, then run the verify gate (workflow-task-finish) before accepting it.

Never

Tier-3 (Opus / Claude Code) is not wired here and must not be — it lives on the laptop's separate Claude Code to avoid auth collisions.