AIWG 2026.10.12 - Live session follow, safer approvals and checked worker results

AIWG 2026.10.12 lets you watch a live AI session as it happens and import more of your session history automatically. It also gives your project a safe way to stop asking for approval on routine work, and makes it harder for a helper worker's "all done" to slip through unchecked.

Follow a session live

`aiwg sessions tail --provider claude --latest --follow` shows new events from a running session as they arrive, usually within a couple of seconds. It only reads; nothing is saved to your catalog unless you add `--import`. Providers that can't be followed live say so plainly.

More of your history, imported automatically

Grok Build and Hermes sessions can now be found and imported directly. Point AIWG at them with `--grok-root` or `--hermes-root`. Hermes is read through a safe copy, so its live database is never touched. If you download Claude admin exports, `aiwg sessions import --from-dir <folder>` imports a whole folder at once and skips files it has already seen.

Fewer approval prompts for routine work

A project can now pre-approve bounded, reversible work, for example routine edits in a testnet repository, with an `authorization.hitl` profile in `.aiwg/aiwg.config`. Run `aiwg config authorization preview` to see exactly what would be allowed and why. Secrets, destructive changes, production and policy overrides can never be pre-approved, and text inside an issue never counts as approval. Workflows will start using these profiles in a later release; for now the preview lets you set one up and check it.

Worker results that have to earn acceptance

When AIWG hands work to helper CLIs, a task can now spell out what must not break and which checks decide that it's done. Final acceptance requires a review of the exact final version from a different model family, and findings are tracked across review rounds so the loop has a clear finish line.

Smaller fixes you may notice

• Project-local bundles can bind a rule to an agent in the same bundle (thanks to the external report on GitHub #373).
• Deploying to Grok Bot at project scope works.
• `aiwg doctor` checks every configured provider, and a failed worker qualification now exits with an error.
• `aiwg models resolve` accepts tighter per-task limits such as `--max-attempts` and `--max-usd`.
• The model catalog lists current Claude and Codex models. The top "flagship" tier is used only when you choose it.

If you are upgrading

• Run `aiwg use all` to pick up the refreshed model defaults and skills.
• Requalify worker approvals: `aiwg doctor --probe-workers --qualify`.

Full technical detail is in the CHANGELOG entry.