Ralph Router — Automatic level selection

Before you act, pick the right level. Never start a task without loading the correct skill.

Decision tree

I receive a task.
│
├─ How many steps do I expect?
│    │
│    ├─ 1–3 steps / single operation / lookup / quick edit
│    │    └─ → Load ralph-small
│    │         Path: /home/ubuntu/.openclaw/skills/ralph-small/SKILL.md
│    │
│    ├─ 4–10 steps / sequential workflow / moderate dependencies
│    │    └─ → Load ralph-medium
│    │         Path: /home/ubuntu/.openclaw/skills/ralph-medium/SKILL.md
│    │
│    └─ 10+ steps / debugging / refactor / architecture / high complexity
│         └─ → Load ralph-huge
│              Path: /home/ubuntu/.openclaw/skills/ralph-huge/SKILL.md
│
├─ Unsure about complexity?
│    └─ → Err on the side of more discipline. Prefer ralph-medium on a simple task
│         over ralph-small on a task that turns out complex.
│
└─ Already using ralph-small or ralph-medium and exhausted attempts?
     └─ → Escalate to the next level immediately.
          Do not wait. Do not retry the same tier. Escalate.

Automatic escalation rule

ralph-small  →  2 failed attempts  →  ralph-medium
ralph-medium →  3 failed attempts  →  ralph-huge
ralph-huge   →  4 failed attempts  →  Circuit breaker + wait for user

There is no escalation beyond ralph-huge. After four failed unique attempts, stop, emit the diagnostic, and wait for human input.

After success (self-improvement)

When the task was run under ralph-small, ralph-medium, or ralph-huge and you have verified success (evidence in the same turn):

• Follow the Closure: self-improvement section at the end of whichever Ralph level you used — it links to self-improvement and .learnings/ without blocking the user reply.
• Skill path (OpenClaw): /home/ubuntu/.openclaw/skills/self-improvement/SKILL.md

This is optional but high-value for recurring patterns, non-obvious fixes, and tool/gateway gotchas.

Note for the agent

Never tell the user “I’m using ralph-[level]” — that is internal operations detail. Select, load, act. The user wants outcomes, not narration of your selection process.