Clarify that model routing is not process overhead

[lorenzstorm1]

Problem

When a plan already contains complete implementation code (e.g. 500 lines across two files), ponytail's "skip subagents — that's over-process" heuristic prevents dispatching to a cheaper model. The result: the expensive frontier model re-generates pre-written code verbatim, burning ~5x more tokens for zero additional value.

Real-world conflict

The superpowers plugin includes model-routing configuration that assigns plan tasks a tier (mechanical, standard, frontier) and dispatches subagents to the appropriate model. Tasks with complete code in the plan steps are mechanical → routed to Haiku.

Ponytail intercepts this flow: "The plan already has the code, dispatching subagents to copy-paste it would be over-process — just write it directly." This sounds lazy but isn't: Opus outputting 500 lines of pre-written code costs ~5x what Haiku would.

Fix

New bullet in Rules clarifying that model routing ≠ ceremony. The ladder governs what the solution looks like, not which model tier writes the files. When the code is already designed and written in the plan, dispatching to a mechanical-tier model is the lazier (cheaper) path.

Example

Plan contains two complete scripts (~500 lines total), tier: mechanical. Without this fix:

• Ponytail: "just write it yourself, subagents are over-process" → Opus outputs 500 lines → expensive

With this fix:

• Ponytail recognizes dispatch-to-Haiku as the lazy choice → Haiku outputs 500 lines → 5x cheaper, same result

🤖 Generated with Claude Code