multi-model-delegation

# Multi-Model Delegation Protocol for consulting *other* models — kimi, glm, gemini, gpt via the PAL MCP gateway (`chat`, `consensus`) — on design and judgment work, and for acting on what comes back. The core insight, which inverts the naive approach: > **The value is the disagreement, not the union.** Two competent models > briefed identically converge on the obvious 80% — the part you'd have > written anyway. Where they **split** is a precise pointer at the one > decision that is genuinely load-bearing and underdetermined by the prompt. > Resolve the split against **the codebase** — which usually already decided, > and which the models structurally cannot see — never by picking the more > confident model. Treat delegated models as *idea generators*, never as *authorities*. Taking the majority answer, or the more confident one, launders a coin-flip into a decision that merely looks researched. You — the orchestrating Claude session — run the whole consult: you dispatch the PAL MCP calls, collect the replies, and do the judgment steps (diff, adjudicate, synthesize) yourself in the main loop. ## The Tool Prefix Is Derived, Not Fixed PAL's tools are reachable as **`mcp__<server-name>__<tool>`**, where `<server-name>` is the key the server is **registered under** — not the product name, and not the binary. That key can live in any registration scope: a project `.mcp.json`, a **user**-scope entry in `~/.claude.json` (`claude mcp add -s user …`, which has no `.mcp.json` at ...