Changing the model re-prices the Optimizer's answer. The result panel
held whatever the last press of Optimise computed, so flipping the model
selector left a report priced for another model on screen, labelled with
the old name — a reader comparing models saw "nothing changes" (the token
reduction should not change: the rules are deterministic text transforms
and the counter is one heuristic, deliberately not a per-vendor
tokenizer). Now, once a result exists, a change to the scenario — model,
calls, output tokens, cache hit rate, batch, level, reorder — re-runs the
free deterministic pass, debounced, so the dollars and the advisories
follow the selector. Two refusals hold the shape: the auto-run never fires
before the reader's first Optimise, and never while the LLM or suggestion
pass is enabled — a dropdown change must never spend a provider call
unasked. Automatic runs also stay out of the reader's history.
optimizer-reprice.test.mjs pins all four properties.
The humanizer skill, vendored. From
blader/humanizer at e2e92e7, MIT:
thirty-five patterns of AI-sounding prose and a two-pass rewrite
discipline, covering the one lane no vendored skill covered — the
human-facing English this repository writes constantly. Vendored as
SKILL.md only, into the tracked.claude/skills/ tree with a new
VENDORED.md registry, because the earlier vendoring round's registry
lived under the gitignored .agents/ and survived nothing. The registry
carries the style caveat (the skill's em-dash and bold-label rules lose to
this repository's voice sample) and records one idea adopted without
vendoring, from addyosmani/agent-skills: reviews get the artifact and the
contract, never the conclusion.