Skip to main content

Completion checks and privacy

For both cloud-model and local-model sessions, AdaL can check whether a final reply stops before doing the work you requested—for example, saying “I’ll run the command now” without running it. This check is enabled by default in versions that support it and requires the completion-check service to be available.

AdaL uses Jev, provided by TypeSafe AI, for this check. This is a separate third-party service from the model you selected. The check also applies to subagents, including those using local models. Even when your selected model runs locally, completion checking sends the evidence described below to this external service and requires an authenticated AdaL session and internet access.

What is sent​

Each check sends:

  • User messages from the current turn.
  • The proposed final answer.
  • Names and success/error statuses of up to 100 recent actions in that turn.

It does not send earlier conversation turns, hidden reasoning, system instructions, or raw action arguments and outputs. However, information you put in your message—or that the assistant quotes in its proposed answer—is included. Some recognizable credential formats are redacted, but this does not remove all possible secrets or personal information.

Checks are skipped when the task or answer exceeds AdaL's size limits, or platform authentication is unavailable. These are character limits, not a guarantee that the request fits Jev's token context limit. If Jev rejects a request, AdaL continues without verification. Switching between local and cloud models does not disable checking; use the explicit opt-out below.

What happens after a check​

If Jev detects an unfinished reply, AdaL asks the selected model to continue the missing work. There is no separate limit on these continuation attempts. Your permission settings, cancellation controls, and existing overall execution limits still apply. Repeated attempts can consume additional model credits.

The check does not verify factual accuracy or guarantee task completion. If the service is unavailable, AdaL returns the proposed answer without verification rather than blocking the session.

Opt out​

Set ADAL_COMPLETION_JUDGE=off before starting AdaL. This disables completion checks for that process and its subagents; it does not disable ordinary requests to your selected model provider.

macOS, Linux, or WSL​

ADAL_COMPLETION_JUDGE=off adal

To disable checks for all AdaL sessions launched from a terminal, add export ADAL_COMPLETION_JUDGE=off to your shell configuration, then open a new terminal.

Windows PowerShell​

$env:ADAL_COMPLETION_JUDGE = "off"
adal

Desktop and hosted sessions​

For a locally launched desktop app or SDK process, set the variable in the environment that starts the app, then restart it. An app already running will not inherit a later terminal change. For a hosted session, the host operator must set the variable in the agent process environment.