escalation-and-recovery

name	escalation-and-recovery
description	Recover from blocked, failed, low-confidence, or contradictory work without losing completed value. Use when an agent cannot access a dependency or tool.
license	MIT
metadata	{"category":"agent-orchestration","origin":"custom-synthesis","revision":"1.1.0"}

Escalation And Recovery

Trigger

• Validation fails repeatedly.
• Sources conflict or a specialist returns low confidence.
• A delegated task exceeds its boundaries.

Inputs

• Failure evidence
• Task graph
• Completed outputs
• Fallback skills and tools

Procedure

1. Classify the failure: input, access, capability, dependency, implementation, validation, or policy.
2. Preserve reproducible evidence and the last known-good artifact.
3. Retry only when a changed condition makes success plausible.
4. Reduce scope to the smallest failing unit.
5. Reassign to a more suitable skill or use a bounded fallback.
6. Invalidate downstream work only where dependency impact exists.
7. Return a partial but usable result with explicit gaps when recovery is not possible.

Output contract

• Recovery decision
• Revised task graph
• Preserved artifacts
• Failure and limitation record

Quality gates

• No blind retry loops.
• Recovery does not weaken hard constraints.
• Partial results are clearly labeled and still validated.
• Security or policy blocks are never bypassed.

Handoffs and dependencies

• work-delegator
• dependency-aware-scheduler
• result-integrator

Boundaries

• Do not ask the user to repeat known information.
• Do not invent success after an unavailable check.
• Do not discard useful completed work.