v2026.11.6 — Use Local Models and Check Work Before Delivery
Run workers with self-hosted models, check finished files before delivery, keep local-only files out of cloud work, and adapt AIWG with documented commands and bundles.
AIWG 2026.11.6 - Use local models and check work before delivery
Use the model running on my own computer for this task. Check that my worker tool connects to it, keep my local-only files away from cloud workers, and show me the checks before delivering the result.
AIWG 2026.11.6 gives you more choice over where AI work happens and stronger checks before a worker's result becomes your finished file. A worker is a separate AI session your assistant gives a task to.
Use models on your own computer or server
Set up a worker with my existing Ollama, llama.cpp or vLLM server. Show me which model is available and ask before running a task.
Direct workers can now use self-hosted open-weight models through an OpenAI-compatible endpoint: a server address that supported AI tools can connect to. Your assistant registers the model, checks its availability and completes worker setup. It must also check that the worker tool actually sends tasks to that server; finding a model alone does not prove where work runs.
OpenCode installed from npm now qualifies for worker use alongside the separate official-installer release line. See the local model guide for setup details.
Check the actual result before delivering it
Give a worker this task. Check its finished file against my requirements, including cases that should fail, before publishing it. If a rule blocks the task before any worker starts, deliver a clear explanation.
The protected delivery route requires an explicitly named file written by the worker before it exits. Verification rejects checks that only test whether a file exists or inspect its metadata, and requires a check aimed at its contents. Checks should come from your requirements; passing them still depends on choosing useful checks.
On supported Linux hosts, the assistant coordinating the work cannot write directly into the protected delivery folder. Accepted files go through verified delivery. A blocked task can also produce a structured refusal explaining which rule stopped it, even when no worker ran. Ask your assistant to check host support using the delivery guide.
Keep local-only files out of cloud work
Mark these files as local-only. Before handing work to a cloud worker, confirm its working copy excludes them.
Files declared local-only are removed from prepared worker copies, hidden from the protected coordinating session, and checked again before worker launch. A cloud worker is refused if its working copy still exposes a protected path. Qualified local workers can use the declared local-processing route.
Make AIWG fit your own workflow
Show me how to package our team's instructions and commands as an AIWG bundle, with an example I can adapt.
A new architecture and extension guide explains how commands, bundles and discovery fit together, with diagrams and a runnable example. Teams can find cached plugins and discover workflows supplied by their own addons too.
After updating, ask your assistant to refresh AIWG for your tools and check worker setup. See the full CHANGELOG for every change.