passepartout/docs/PRD_LLM_CASCADE.org

PRD: LLM Failover Cascade & Multi-Provider Support

• 1. Purpose
• 2. Functional Requirements
• 3. Success Criteria

1. Purpose

Ensure 100% availability of System 1 (Neural) reasoning via an ordered list of providers and automatic fallbacks.

2. Functional Requirements

• Backend Registry: The core MUST allow Skills to register new AI providers (Gemini, OpenAI, Claude, etc.).
• Automatic Failover: The `ask-neuro` function MUST iterate through a `*provider-cascade*` list, automatically trying the next provider if the previous one fails.
• Interactive Re-Ordering: The system MUST allow the user to update the cascade order in real-time from Emacs.

3. Success Criteria

• DONE Daemon correctly falls back from a failing API key to a working one.
• DONE `M-x org-agent-set-model-cascade` successfully updates the live Lisp hierarchy.
• DONE Support for 5+ providers (Gemini, OpenAI, Anthropic, OpenRouter, Ollama).