Declarative semantic failure recovery layer for multi-step AI agent workflows that routes on failure type instead of blind restart

When a multi-step AI agent workflow fails, the standard recovery is to restart the run. But restarting after garbage output produces identical garbage: same model, same input, same failure. Production engineers building AI agents have learned from distributed systems that you need supervision trees with different recovery strategies per failure type. For AI agents the equivalent is semantic failure routing: a malformed tool call retries with an auto-rephrased prompt; context rot triggers a session reset and context re-injection; a consistently wrong model escalates to a stronger one; an ambiguous task pauses for human clarification. No tool provides this as a declarative config layer on top of existing coding agents. The production concurrency article on HN (149 pts, 87 comments, Feb 2026) explicitly named this gap: restarting an agent that received garbage output yields identical problems, requiring custom semantic supervision trees built from scratch each time. Forge (687 pts, May 2026) proves the demand for agent reliability primitives but only addresses structural tool-call failures, not semantic failure taxonomy or model escalation.