v2.0.0-beta.1

Changelog

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

[2.0.0-beta.1] - 2026-10-21

Highlights

CatchTwo 2.0 brings the AI catcher to Android and makes the send path more reliable.

• AI catching on Android: the AI now runs through TensorFlow.js without native binaries, so it works on Android with Termux.
• A better model: the new 1,279-class model reaches 99.451% validation top-1.
• Lower memory: TensorFlow.js loads only when AI is enabled, so AI-disabled runs use about half the memory.
• Catches come first: catches, hints, commands, and spam share one priority queue. A full route queue is retried instead of silently dropping the send.
• More commands: ping and support are new, and predict, stats, version, and help were reworked.

Added

• Local AI catch model (sepconv-1279-real-ft) with 1,279 classes, 1,621,951 parameters, and 64x64 RGB input. Validation top-1 is 99.451% and top-5 is 99.870%.
• Center-crop preprocessing for live spawn images. The center 0.8x of the shorter side is cropped before the 64x64 resize, and images at or below 128 px are left as-is.
• CatchTwoRestManager, a priority send queue for Pokétwo traffic. Catches outrank hints, hints outrank commands, and spam runs last.
• Startup config validation and idempotent config migration.
• ping and support commands, plus reworked predict, version, help, and stats commands.
• Cross-platform setup for Windows, Linux, macOS, and Android (Termux).
• TypeScript checkJs coverage for opt-in modules, Prettier, ESLint, and a Node 22/24/26 CI matrix.
• 64 node:test tests covering services, commands, queueing, and lifecycle behavior.

Changed

• Breaking: Migrated from discord.js-selfbot-v13 to discord-self-lite@0.2.0. Clients use the library's restClass injection, and the queue uses RestManager.destroy(), RouteQueueFullError, and the gateway channel cache.
• Breaking: Stats and account state are keyed by user ID instead of username. Existing username-keyed stats do not carry over.
• Breaking: spamming.SpamSpeed is a number (default 2000, valid range 1000 to 120000). logging.LogCatches is split into Pokemon, Quests, LowIV, HighIV, and SpecialCatches, with LowIVThreshold and HighIVThreshold.
• Breaking: incense.IncenseMode now disables spam while enabled. When disabled, incense spawns are ignored. AutoIncenseBuy requires IncenseMode.
• Breaking: The model bundle and labels.json must be copied as a set. A model/label mismatch fails before inference.
• Switched from @tensorflow/tfjs-node to @tensorflow/tfjs, loaded lazily so AI-disabled runs use about half the memory.
• Replaced the old model weights with the 1,279-class model.json bundle and its two shards.
• Outbound Pokétwo traffic now uses a bounded priority queue with per-class deadlines, bucket-aware pacing, and catch protection. Spam yields to catches and is dropped when paused.
• A null REST response now means only 204 No Content. Full route queues reject with RouteQueueFullError and are retried before the item's deadline.
• The confidence floor is 0.5. Known weak classes always go to the hint path instead of being caught.
• Install and update now use main. start.bat and the package scripts were updated.
• package-lock.json is tracked and CI runs npm ci.
• The supported Node baseline is >=22.13.0.

Deprecated

• incense.IncenseChannel is still required for AutoIncenseBuy, but the setting is deprecated.

Removed

• behavior.Sleeping.
• The assisted-catch path from the 1.3.x main branch. The local AI model replaces it.
• discord.js-selfbot-v13, @tensorflow/tfjs-node, the old weights.bin, src/data/ai.json, src/data/levelup.json, and the old logo assets.
• The per-message channel cache workaround in listenEvents.js.
• Temporary spawn calibration logging and analysis scripts.

Fixed

• Full-scene spawn images are center-cropped to the training framing instead of being squashed, which restores confidence on live 800x500 attachments.
• Queued Pokétwo sends are retried on RouteQueueFullError instead of treating a null response as a dropped send.
• REST shutdown aborts in-flight requests and rejects queued sends and pending sleeps.
• The 50013 permission check reads DiscordAPIError.code instead of fullError.code.
• Event handlers no longer fetch and cache each message channel themselves.
• npm run update and npm run force-update preserve local config.js and tokens.txt changes.

Model

• Production run: 20260723T122145Z-sepconv-1279-real-ft from the kyan0045/poketwo-ai repository. Checkpoint SHA-256 starts with b49e33c3.
• Live validation at the 0.5 floor: 184 spawns, 182 caught (98.9%). On caught spawns, 180/182 were top-1 (98.9%) and 174/176 were accepted (98.9%).
• Known weak classes always use hints: Type Null, Type, Galarian Articuno, Galarian Zapdos, MissingNo, Regieleki, Scolipede, and Xurkitree.

Full Changelog: v1.4-beta.3...v2.0.0-beta.1