AIWG 2026.10.16 - Start with a prompt and check room for helpers

AIWG 2026.10.16 puts example prompts first in the guides you use to get started, connect your assistant and do everyday work. It also gives your assistant one resource check before starting helper workers, so it can plan around the machine you have.

Start with what you want done

The getting-started, provider, framework and everyday guides now open tasks with what to ask your assistant and what to expect back. Commands follow under “Command reference,” with their limits and approval boundaries kept in place.

Start with the example closest to your task, then add your inputs, boundaries and the result you want to review. For example: “Help me connect AIWG to the assistant I use. Ask which provider I use, walk me through its setup, and show me how to check it worked.” The getting-started guides give you a place to begin without assembling command options yourself.

Check the machine before adding helpers

Before asking for parallel work, try: “Check this machine's available resources and running AIWG workers before starting any helpers. Tell me how many can fit, and wait if there is no room.” Your assistant can run:

aiwg doctor --resources

The read-only report brings together load, memory, memory pressure, free disk and running workers. It tells you how many helpers can fit right now, and what is limiting that number. If there is no room, your assistant won't start more helpers and tells you why. The worker cap defaults to five; you can configure it for your machine. See the orchestration guide for the workflow.

Less setup clutter, more predictable recovery

When you deploy to Claude Code, `aiwg use` now turns off its automatic AI co-author trailer on commits, PR attribution and session links by default. Your existing attribution choices are kept with a warning. To keep attribution on, add `--no-attribution`, which tells AIWG to leave your Claude attribution settings alone (`aiwg use all --provider claude --no-attribution`); the Claude Code guide explains the setting.

AIWG now generates one root AIWG.md instead of two copies. Existing operator notes survive migration. `aiwg daemon` works again, and its guide now uses supported commands. If a task that writes a file is resumed after an interruption, it skips the write when it already happened.

An approval now covers only the exact edit you approved, and the records that guard evaluation data are signed, so a faked access time is refused.

Research tools that keep your intent

Citation checks now match the full reference code, so a short code can't be mistaken for another paper. Notes on where a paper was found now keep quotes and special characters intact. Rate-limited lookups retry, then report that you should try later rather than calling the paper missing. You can also ask your assistant to review suggested papers: `aiwg corpus review` supports accepting, rejecting or deferring candidates, with changes saved only when requested through `--write`.

If you are upgrading

• Run `aiwg use all` to refresh your deployed guides, skills and settings.
• Run `aiwg regenerate` if doctor reports registration drift, including existing OpenCode projects that list only WORKSPACE.md.
• To keep Claude attribution on, deploy with `--no-attribution` so AIWG leaves those settings alone. If you already applied the defaults, set your preferred attribution in Claude settings.

The CHANGELOG entry has the full technical detail.