Opinionated multi-model delegation for very large repos.
> Big work in big repos is won by decomposition + heterogeneity, not by one smart model.
thunderkit is an *opinionated* skill pack. It exists to take a large change in a large repository and:
Claude Opus, Codex Sol — via the claude, codex, and opencode/hermes harnesses).
finished. tk-plan emits lanes or it fails.
tk-review fans work to ≥2 families (e.g. Opus + Sol) and consolidates.
no "done" — the lane is *blocked*.
actually installed and authenticated, and *name* what's missing instead of failing dark.
outcome (the critical path, the final review), thunderkit offers the top models and asks — it does not silently default.
.thunderkit/ (NORTH_STAR.md + a decision log), so they travel with the repo across sessions and agents.
turning big-repo changes into parallel, cross-reviewed, evidence-gated work.
SKILL.md files that anyAgent-Skills-compatible agent can read.
works on anyone's machine.
Different work has different cost/quality curves. Wide exploration and cheap retrieval want a fast, login-free model (Fable 5.1). The critical implementation path wants the strongest coder (Opus). Review wants a *different* family than the author, so blind spots don't survive. Sol (Codex) brings a genuinely different lineage for that second opinion. One model can't be all of these at once; a fleet can — if the work is decomposed to feed it. That decomposition is the whole game, and it's what thunderkit is opinionated about.