agentsop-llm-tool-idempotency

What problem does it solve?

This Skill prevents side-effectful LLM agent tools from executing twice when retries happen due to timeouts, framework resume, or duplicate tool-call emissions, avoiding outcomes like duplicate emails, double charges, and duplicate records.

Core Features & Use Cases

• Decision protocol for at-most-once effects: Implements the rule that the LLM is at-least-once while the tool must engineer at-most-once execution using a caller-supplied idempotency key and dedup at the execution boundary.
• Framework-agnostic operating models (OP-1..OP-8): Covers native idempotency APIs (e.g., Stripe-style headers), internal dedup tables, content-addressed writes, database upsert/insert patterns, saga compensation, and orchestrator-specific guidance for LangGraph interrupts and MCP tool schemas.
• Production guardrails and failure handling: Includes dilemma cases, anti-patterns, conflict handling rules (e.g., same key different args), and cross-framework comparisons to ensure duplicates are blocked rather than silently re-triggered.

Quick Start

Ask an engineer to implement idempotency keys for a side-effect tool so that retries become safe no-ops instead of double-send behavior.