v2.0.0-beta.3

Changelog

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[2.0.0-beta.3] - 2026-11-02

Highlights

Beta.3 updates the catch model, reduces runtime memory and install size, and removes several runtime dependencies.

Added

• New 1,303-class model from run 20261102T141243Z-sepconv-accuracy-ft, with 1,646,551 parameters and 99.8847% top-1 on the cleaned 26,014-image audit split.
• Catch-name mapping for Arceus, Silvally, and Alcremie forms. The model keeps the form classes, but catch commands use the base species.
• CI verification for the float16 model bundle, including exact hashes for model.json, both shards, and labels.json.

Changed

• Model weights now ship as float16 and load as float32 at runtime. The model directory is about 3.3MB.
• The vendored TensorFlow.js bundle now includes only tfjs-core, tfjs-layers, and the CPU backend. It is about 1.14MB, down from 2.07MB.
• The AI model unloads after 15 minutes without a spawn and reloads on demand.
• Catchers start concurrently with a 250ms delay instead of one second per account.
• Runtime dependencies now use native fetch, node:fs, and a local HH:mm formatter. Removed axios, fs-extra, and date-and-time.
• The start scripts set --max-old-space-size=256.
• Finizen, Palafin, and Hero Palafin now fall back to hints instead of attempting a direct catch.
• The beta README installs the dev branch and links to main for the stable release.

Removed

• Runtime dependency on figlet. The startup banner is now pre-rendered at build time.
• The Type and Beta Sentret model classes.

Fixed

• Quest track completions that reward a background now log the correct reward type instead of an unknown badge.

Breaking

• The model class set changed. Use the bundled model.json, weight shards, and labels.json together.

Full Changelog: v2.0.0-beta.2...v2.0.0-beta.3