ZeroScript Free Files

ZeroScript now runs on ChatGPT as an eighth provider, alongside DeepSeek, Gemini, GLM, Kimi, Qwen, LMArena and Meta AI.

This release is mostly about one thing: what ZeroScript reads off the page. Since 1.5.0, several sites moved their code blocks into editors and viewers that do not put the real text in the page — they render an approximation of it. A command could look perfect on screen and still reach Roblox glued onto one line, cut in half, or with its biggest argument replaced by a placeholder. Every fix below is a different shape of that same problem.

Added

• ChatGPT support (chatgpt.com). Image input is deliberately off there: ChatGPT's free tier caps files and images on a quota separate from messages, so vision would only work part of the day. The model picker and reasoning mode stay entirely your choice.

Fixed

• ChatGPT: most tool calls failed outright. ChatGPT renders code blocks with CodeMirror, which emits one element per line and no newline characters at all, so a valid command was read as a single giant line and came back as "your code block was empty". Replies are now read with their real line structure.
• ChatGPT: long commands ran truncated. Past roughly 2000–4000 characters CodeMirror renders only part of a long line and then stops updating — the tool chip's token counter would climb, fall back to about 500, and freeze while the command executed cut off. Measured live: a 21273-character command of which the page exposed 4049. ZeroScript now reads the editor's true document through a MAIN-world tap. A 5.3k-token multi_edit applies whole.
• ChatGPT: the raw command text stayed visible when the model wrote it outside a code fence.
• Meta AI: big commands failed with "bad JSON". Meta shows a JSON block in an interactive viewer whose default Tree view abridges large values: a 19103-character multi_edit was present in the page as 223 characters ending in "edits":[1 item]. That truncated object is what reached the parser. It is also why the token counter collapsed to ~44 tokens the moment the block finished rendering — the counter was faithfully reporting what could be read. Commands are now taken from the viewer's Raw tab, in full.
• The agent no longer dies silently when a model calls a tool the function-calling way. A reply like {"toolName": "get_studio_state", …} names a real tool with the wrong key, so nothing recognised it as a command and the loop simply ended, leaving the agent looking frozen. ZeroScript now spots it and asks for the proper envelope.
• Clearer message when Roblox refuses to parse your Luau. ZeroScript used to always blame an empty code block or a wrong ###LUA### marker — even when a full script had been sent, sending the model to fix a problem that did not exist, so it retried the same payload and failed again. It now reports how many characters were sent and names the real causes: invalid syntax, or code too large for the parser.
• Fixed a stylesheet error that silently disabled command hiding on GLM, Kimi, Qwen, LMArena and Meta AI.