v0.0.32 — Model Refresh (May 2026)

⚠️ Superseded by v0.0.33. The v0.0.32 npm publish was blocked by an expired token. v0.0.33 ships the same content, successfully published to npm + PyPI. Use v0.0.33.

---

Agent Model Refresh (May 2026) (#44)

All six agent registries refreshed to the latest models providers serve as of May 2026. Two of these were forced removals — providers shut them down — so this is partially breaking.

New models

Agent	Added
claude	opus → claude-opus-4-7, sonnet → claude-sonnet-4-6
codex	gpt-5.5, gpt-5.4-mini, gpt-5.3-codex
gemini	gemini-3.1-flash-lite-preview
qwen	qwen3-coder-next (recommended), qwen3-coder-flash
kimi	moonshot/kimi-k2.6
opencode	All of the above mirrored as openrouter/*

Default changes

Agent	Old default	New default
codex	gpt-5.2	gpt-5.4
gemini	gemini-3-flash-preview	gemini-3.1-pro-preview
kimi	moonshot/kimi-k2.5	moonshot/kimi-k2.6

Note for Gemini users: the default moved from Flash to Pro. If you call withAgent({ type: "gemini" }) without an explicit model, you'll get gemini-3.1-pro-preview — substantially more capable, also substantially more expensive than Flash. Pin Flash explicitly if cost matters.

Breaking — removed models

Removed	Reason	Use instead
gemini-3-pro-preview	Provider shutdown 2026-03-09	gemini-3.1-pro-preview
moonshot/kimi-k2-turbo-preview	Kimi K2 series discontinued 2026-05-25	moonshot/kimi-k2.6
gpt-5.2-codex	Superseded	gpt-5.3-codex
gpt-5.1-codex-max	Deprecated	(no direct replacement)
gpt-5.1-mini	Deprecated	gpt-5.4-mini

The same openrouter/* mirrors were also removed.

Examples

TypeScript:

import { Evolve } from "@evolvingmachines/sdk";

const evolve = new Evolve()
    .withAgent({ type: "codex", model: "gpt-5.5" });

const evolve = new Evolve()
    .withAgent({ type: "kimi", model: "moonshot/kimi-k2.6" });

Python:

from evolve import Evolve, AgentConfig

evolve = Evolve(
    config=AgentConfig(type='codex', model='gpt-5.3-codex'),
)

Note: v0.0.31 shipped initial gpt-5.4 support via #40. v0.0.32 expands and corrects the entire matrix.

---

Cleanup

• Removed obsolete --experimental flags for Gemini and Qwen (#37) — the upstream CLIs no longer require them.
• Removed docs-site/ from the repo (Nextra build now happens out-of-tree).
• Removed agent-builder cookbook and stale plans/ + local changelog files from version control.

---

Migration checklist

If you're upgrading from v0.0.31:

• Replace any hard-coded gemini-3-pro-preview → gemini-3.1-pro-preview (provider shutdown)
• Replace any hard-coded moonshot/kimi-k2-turbo-preview → moonshot/kimi-k2.6 (provider shutdown 2026-05-25)
• Replace any gpt-5.2-codex → gpt-5.3-codex
• Replace any gpt-5.1-codex-max / gpt-5.1-mini → see table above
• If you rely on the Gemini default, decide whether you want Pro (new default) or Flash (set explicitly)