v0.104.0

0.104.0 - 2026-07-17

Added

• moonshot joins the canonical catalog (17 providers · 11 cloud + 5 local + 1 test)
  (ADR-105 · the ADR-104 promotion precedent applied) — model: moonshot/kimi-k3
  runs Kimi K3 first-class: 1M context · thinking model (reasoning spends output
  tokens · budget max_tokens accordingly) · MOONSHOT_API_KEY · endpoint
  api.moonshot.ai/v1/chat/completions · weights announced open 2026-07-27.
  Catalog row refreshed (default kimi-k3 · cheap kimi-k2.6 · real context/output
  limits) · LSP completion + vocab carry the new id (spec-token order preserved:
  token-absent providers trail the pinned set until the spec adds them).
• kimi pricing — the models.dev snapshot refresh now maps moonshot (10 models
  priced · kimi-k3 $3 in / $15 out / $0.30 cache-read per M) so --max-cost-usd
  is a hard ceiling on metered kimi spend, not an UNBOUNDED floor.

Fixed

• nika-lsp semantic_document fixtures spoke the pre-argv grammar
  (command: "x" string form · stale since the 0.103.0 exec tightening) — the
  three tests shipped red at the v0.103.0 tag and CI never ran nika-lsp --lib
  (banked as a CI-coverage finding). Fixtures now speak argv (command: ["x"]);
  the crate is green end-to-end.

Full Changelog: v0.103.0...v0.104.0