AIWG 2026.10.1 - Turnkey Jev decision offload

AIWG 2026.10.1 makes Jev decision offload work from a token. Agents can send bounded decisions (yes/no gates, picking one of N options, 1-N scores) to the Jev decision classifier instead of reasoning them out in the frontier model.

What changed

• `aiwg decision setup jev`: pipe your Jev token in once and it is stored

in a user-level 0600 credential file. `--verify` checks it with one live call.

• `aiwg decision ask`: one command sends one bounded decision to Jev and

returns a typed answer with confidence. If Jev is not set up, abstains, errors, times out, or answers below `--threshold` (default 0.8), the result is `fallback: "llm"`, and the agent decides as it would have anyway.

• Agents use it automatically: the decision-engine addon now ships a

`decision-offload` rule and skill that tell agents to ask Jev first for bounded decisions. MCP clients get a `decision-ask` tool.

• The steward can set it up: ask it to "help me set up jev" or "use my jev

token" and it walks you through installation, token setup, and verification.

• Fixes: valid Jev Score answers are no longer rejected for two-decimal

rounding, and `aiwg use decision-engine` now keeps the addon's rule after deployment.

Try it

npm install -g [email protected]
aiwg use decision-engine
printf '%s' "$JEV_TOKEN" | aiwg decision setup jev --token-stdin --verify
aiwg decision ask --question "Is this test failure a timeout?" --yes-no --context-file out.log

Savings depend on your workload: the more small bounded decisions an agent makes, the more it offloads. The Jev quickstart covers fallback behavior, privacy (questions and context are sent to Jev), and troubleshooting.

The full change history is in the changelog.