[rmkr-dev]

Adds a Skillbook entry and a matching Promptbook entry for production routing and fallback on an LLM gateway.

Skill: llm-gateway-routing-fallback-guardrails

• Alias to ordered targets (provider, model or deployment, region), with priority or weighted selection and sticky versus per-request
• Health-aware skip, then fall through on timeout, 429, 5xx, connection errors, and auth or deployment-config failures
• 400, 413, and 422 stop the walk unless that target sets fallbackOnBadRequest (default off), and those 4xx responses do not open the breaker
• When every target fails, return the last meaningful upstream status. A caller 4xx stays that 4xx rather than a synthetic 502
• No provider switch after the first streamed token; cancel the upstream and emit stream_interrupted
• One adapter per provider family, secrets by reference, per-attempt metrics under one correlation id
• Two-stub tests for fallback order, no fallback on 400, last 4xx preserved, and no mid-stream switch

This skill owns outbound model routing and fallback. Cost and latency budgets, prompt injection, eval gates, HTTP client pools, circuit-breaker primitives, inbound rate limits, and Idempotency-Key stay with the existing skills.

Prompt: add-llm-gateway-routing-fallback

• Placeholders for the gateway name, providers, aliases, streaming, SLO, and observability stack
• Numbered requirements from inventory through routing, failure classes, final status, streaming, adapters, tests, and rollout notes
• Deliverable is the inventory, the config and code diff, the tests, and rollout notes

Catalogs go to 53 skills / 53 prompts. node scripts/site-check.mjs passes locally.