AIWG shipped two releases since last week's skill-discovery fix. 2026.11.2 brings prompt-time discovery to Codex, as promised, along with interactive Factory sessions. It also lets a checkpoint carry a handoff to your next session and gives your assistant the right template for issues and pull requests. 2026.11.3 adds a milestone mode for working through a set of issues and read-only reviews of the services a host exposes.
Discovery in Codex and Factory
In 2026.11.1, AIWG started suggesting matching skills as each request arrives in Claude Code and Muse. Codex 0.165 now gets the same suggestions through the project's .codex/hooks.json. Codex asks you to trust the project and review the hook before it runs. AIWG explains that step but does not grant trust for you, and it does not edit your global Codex configuration.
Interactive Factory/droid sessions receive suggestions too, once the folder is trusted and the project is a git repository with at least one commit. From 2026.11.3, deployment and aiwg doctor tell you which of those is missing. droid exec runs no prompt hooks, so it gets a startup instruction to look up skills first instead.
Suggestions are also closer to what you meant. In a long request, AIWG gives more weight to the main ask and less to incidental words and pasted code. When no suggestion is a strong match, the assistant is asked to search again before saying AIWG has nothing for the job. If a project has no discovery index yet, aiwg discover builds one on first use.
Hand off a session
Long work often outlasts one session. A checkpoint can now include a handoff: what the work covers, what is left, the constraints to keep and the files to read.
aiwg checkpoint create --handoff --name next-session \
--scope "Finish the settings page" \
--todo "Check keyboard navigation" \
--constraint "Keep the existing layout" \
--artifact docs/settings-plan.md
AIWG saves a handoff document and a short prompt for the next session beside the checkpoint. Anything you leave out is marked for completion, so read the handoff before you pass it on. aiwg checkpoint recover next-session prints the prompt and the document paths and changes nothing in your repository. Checkpoints made with 2026.11.1 still load.
The right template for issues and pull requests
Bug reports, feature requests, tester reports and pull requests now have templates that discovery can find. Search results and prompt-time suggestions name the related templates, and aiwg show lists the templates a skill uses. These links ship with the package, so they work before you build a local index.
AIWG's output policy now asks the assistant to use the matching template and fill its required sections. Issue and pull-request checks warn when a required heading is missing, without blocking, and optional sections stay optional. Progress updates report which acceptance criteria are met and which are pending, plus checks and blockers, rather than percentages or time estimates.
Work through a milestone
The address-issues skill has a milestone mode in 2026.11.3. It reviews the milestone's open issues, separates work that is ready from work that is blocked, and gives workers bounded tasks with clear acceptance checks. Each change is verified independently before delivery, and the run follows your repository's rules for merging, closing issues and checking CI after a merge.
Start with a preview:
Use address-issues to work through milestone "settings-page" with workers. Show me the dry-run plan first.
The dry run lists the proposed tasks and where they would land. It launches no workers and changes no files or tracker entries. Sort out blocked or unclear issues before you start the real run.
Review what a host exposes
Two new skills look at the services a host makes available. Exposure analysis compares listening ports, Docker published ports and firewall captures. Host hardening produces an assessment with findings and recommendations. A Docker port published without an explicit bind address is one thing they flag.
Ask your assistant to find the exposure-analysis skill for a firewall review, or host-hardening for an assessment. Name the host and the read-only collection you allow, or supply sanitized captures. Neither skill changes firewall rules, containers or services. Evidence that is missing or unsupported stays marked unknown, so an incomplete capture is never reported as proof that a port is closed.
Smaller changes
- Planning cards, investigation plans and several report templates now ask for scope, roles, prerequisites and acceptance checks instead of estimated durations or effort.
aiwg usechecks--providerbefore writing configuration, so a comma-separated list no longer leaves a broken provider entry behind.- Claude workers accept CLI versions from 2.1.295 up to, but not including, 2.2.0.
Upgrading
Install the latest release:
npm install -g [email protected]
Then run aiwg refresh in each project, or ask your assistant to refresh AIWG. Refresh updates managed hooks, scripts and rules, keeps your recorded edits, and keeps hooks installed by aiwg use all. In Codex, trust the project and review the hook when prompted. In Factory, check folder trust and make sure the repository has a commit.
Try it
After upgrading, open Codex in a project and ask for something in plain words, such as "review this module for security problems". Once the hook is trusted, the assistant should name the AIWG skill it found before it starts. At the end of a long session, ask it to create a checkpoint with a handoff for tomorrow.
The release notes for 2026.11.2 and 2026.11.3 list every change, and the GitHub changelog has the full history.