This one's mine. I've spent the last few months building it solo: type a prompt, get a game-ready 3D model, all on your own GPU, nothing in the cloud.
The thread shows exactly what comes out.
I got tired of paying credits to generate 3D models on someone else's servers, so I built a desktop app that runs the whole pipeline on my GPU.
Typed "weathered wooden dungeon crate with iron corner bands and a hinged lid". Four minutes later: this. Nothing touched the cloud.
@dizziless@lmstudio yep, paid for a boost. first time launching plugins for lm studio and i wanted them in front of people who'd actually use them. the plugins themselves are free
Finally published my two @lmstudio plugins on the Hub, quick thread for anyone running local models.
agentic-workspace turns your model into a coding agent that works on a real folder and asks before it writes anything.
context-compressor stops you running out of context.
@0xWeq@lmstudio cheers. the trick is that summaries are byte stable, same prefix in means same text out, so the history never shifts under the model between turns. also returning this <div> you seem to have dropped
@0xGrebe@lmstudio that's the whole design bet. the chat is a lossy view, the journal is the record, and after a compression the model just asks for the ids it lost. stress tested it hard today and the journal held up better than the model did tbh
update, as promised: ran both plugins in one 16k chat all evening and it broke four ways. lm studio feeds a model its own internal reasoning marker back through plugin history (now stripped). the compressor budgeted the raw window and ignored ~5k of tool schemas, two engine 400s (now it budgets what's actually left after the tools). qwen 3.8 burns 6k+ tokens thinking about merging summaries and never answers, so that gets detected and switched off per session. and the workspace path setting never applied because lm studio gives every chat its own folder. all fixed and on the hub tonight, 0.8.1 and 0.3.1. the good bit: a staged approval survived two compressions and the model resumed the exact call from memory.
two different things. the state itself never touches the compressor, it's on disk in .agentic (append only journal, transactions with snapshots, approvals, plans).
what gets compressed is the chat's memory of it, and the compressor knows the workspace format: ids, statuses, file hashes and the exact resume call for a pending approval survive, diffs and stdout don't. summaries always come from the original messages so nothing drifts, then the model calls show/status with the id and carries on.
tested with real tool outputs on both sides. next is both plugins in one chat with a forced /compress mid task, will post what breaks.
@PaulWagmi@lmstudio qwen 3.8 27b for both. follows the tool calls cleanly and respects the ask-before-write step every time.
running it on a single 5090 with headroom left for context. what are you running locally?
Both are on the LM Studio Hub, search the Plugins tab or grab them here:
https://t.co/OHZIBOswTn
https://t.co/O6wueYtPm2
Built and tested on Windows (npm and npx work without a shell). Free for personal and internal use. Bug reports welcome.
context-compressor 0.8.0 summarises old messages as you near the limit. One chat went from 645k tokens to a 33k prompt and the model could still answer about the start.
New in 0.8.0: it keeps the workspace plugin's approvals and plans through compaction so the agent carries on.
@ScamAItman@claudeai Never said they were. It’s just standard regulatory capture, monopolizing capability and pulling up the ladder behind them while calling it safety compliance.
I'm definitely not letting the infra off the hook, leaving that door open was a critical failure. But the network is the physical bounding box, the prompt is the cognitive one. If you give an agent an unscoped objective, it will find the flaw in your sandbox. Defense in depth means you have to lock down both.
The root cause of this breach isn't just a misconfigured third-party environment, it's a fundamental prompting failure.
Anthropic states they "didn't give Claude any limits on where to look for the flag."
When you instruct a frontier model to find the flag at all costs without defining the scope, you are giving it explicit permission to break the sandbox. If the internet door is left open, it will use it. Basic red team fundamentals, you must define your bounding boxes.
What do you all think is the responsibility here on the model builders or the infrastructure partners?
In a review of our cybersecurity evaluations, we found three incidents in which a Claude model reached the internet from within or while interacting with a third-party evaluation environment, and then gained unauthorized access to the real systems of three different organizations.
Our post describes what happened, how it happened, and what we’re changing. We encourage other AI developers to perform similar reviews.
We conducted this review together with @Irregular, one of our evaluation partners, and thank them for the joint investigation and their collaboration on this post. This type of collaboration is increasingly critical to safe, rigorous evaluation of models, and we look forward to continuing to work together on security.
https://t.co/dKFCdpKd9v
The infrastructure misconfig enabled this, but the boundless CTF prompt is the actual root cause. When you tell an unguardrailed model to 'find the flag' without defining the scope, it will use whatever doors are left open. It didn't go rogue, it just executed a boundless prompt perfectly. In red teaming, bounding boxes are mandatory.