tier-offloadlisted

# tier-offload Reads every `SKILL.md` under your `.claude/skills/`, classifies each skill's LLM-bearing roles with a fixed taxonomy, then applies the Switchboard routing rule to decide which (and only which) sub-tasks are safe to run on a cheap **local** model instead of Claude. It emits two artifacts: a human-readable **inventory** (markdown, grouped by routing verdict) and a **switchboard config** (`enabled_call_sites` map) that the `switchboard` `local_judge` client loads directly. This is the generic, contributable **scanner** half of Switchboard offload. It is **discovery + config only** — it identifies and configures the local-safe slices; the actual wiring of each skill to call `local_judge` stays a guided, per-user edit you make afterward. It never edits a skill. ## When to use - You've adopted the `switchboard` local-offload core and want to know which of *your own* skills have a sub-task safe to route to the local model. - You want to regenerate the offload inventory after adding or changing skills (the safe surface drifts as skills evolve). - You want a config file the `local_judge` client can load, without hand-authoring the `enabled_call_sites` map and risking a shape mismatch. ## Background — the routing rule this skill encodes The offload's one governing rule (Switchboard Decision 9 / the 3-tier judge split): **authorship, planning, orchestration, and any final/gating judgment stay on Claude; only a fan-out array of cheap judges/graders goes local; mechan