Delegation Failure Recovery Planner

Delegated workflows often fail in ways that manual retries cannot fix. When a tool timeouts, an API returns a 500, or a script is interrupted, you are left with an ambiguous state where side effects may have partially occurred. This skill provides a systematic framework to reconstruct what happened and execute a safe recovery without creating duplicates or corrupting production data.

What it does

• State reconstruction classifies failed actions into precise states like PARTIAL_SIDE_EFFECT or OUTCOME_AMBIGUOUS based on available logs.
• Risk assessment evaluates idempotency and duplicate risks before any retry is attempted.
• Blast radius mapping identifies downstream impacts across financial, customer-facing, and system-level artifacts.
• Recovery planning selects the optimal strategy between resume, reconcile, compensate, rollback, or human escalation.
• Evidence verification defines observable criteria that must be met before a workflow is considered recovered.

How it works

1. Freeze and Fact-Find by capturing the task objective, observed symptoms, and stopping all automated retries.
2. Build the FSCRL Ledger to map every failed step to its specific execution outcome, containment state, and reversibility class.
3. Establish Recovery Authority by verifying if the current principal has explicit permission to perform recovery actions like voiding or deleting.
4. Generate the Control Pack which provides a deterministic sequence of containment, reconciliation, and verification steps.

Frameworks & tools

This skill is platform-agnostic and works with any delegated agent environment including Claude Code, Cursor, AutoGPT, and custom LangChain or CrewAI implementations. It handles REST APIs, database transactions, and cloud provider operations.

Why this beats prompting it yourself

Generic prompts often lead agents to blindly retry failing actions, which causes duplicate payments, double-posts, or data corruption. This skill enforces strict evidence discipline and human-in-the-loop gates for all irreversible actions.

Use cases

• API Timeout Recovery where a payment or message may have been sent but no receipt was recorded.
• Interrupted Deployment where a CI/CD pipeline fails mid-run leaving infra in an inconsistent state.
• Approval Drift when a workflow payload changes after a human has already signed off on an earlier version.
• Data Pipeline Failures where partial ETL loads require precise reconciliation before resuming.

Known limitations

This skill designs recovery procedures but cannot execute them without direct tool access provided by the host environment. It relies on the accuracy of provided logs and system receipts.