StackJack 0.1.0.8599+b93b6ac

This release lets automations run on your own Azure AI Foundry, including selected models other than Claude. It also adds thinking-step and write-step editing to the automation builder, usage alerts for connector allowances, clearer ConnectWise setup, and fixes for the chat builder and supervised runs.

New features

• Run automations on your own Azure AI Foundry. A new Foundry tab on the Automations page connects your Azure AI Foundry, either by pasting a key or through guided setup. Guided setup requires Owner rights on the Azure subscription and shows a terms screen that is never pre-filled.
• Each automation has a "Run on" choice in the builder, the chat builder and the AI assistant tools.
• Model usage on Foundry is billed to your Azure subscription, and StackJack charges no credits for it.
• You can set a spend guard for each automation. It is estimated at Azure list prices, and the builder tells you when it cannot estimate.
• Models other than Claude on Foundry. You can now use selected non-Claude models deployed in your Azure AI Foundry, including checked OpenAI models, or declare another Azure model by name.
• StackJack checks each deployment before an automation can run on it. The connection page shows each deployment's check result.
• The "Run on" model picker lists every mapped model. A model that is still being checked is shown but cannot be selected yet.
• Daily deployment check on Foundry. Once a day, StackJack re-checks each non-Claude deployment to confirm it still works.
• A check is normally two small test calls on your Azure bill, typically well under a cent. An occasional full-size repeat can cost up to a few dollars.
• A deployment that fails is switched off, and you get one e-mail per connection that night listing what was switched off.
• Web search on checked OpenAI models. Web search is available for automations on checked OpenAI models through Foundry. Before those automations run, someone in your organization must confirm a notice about how web search works on that model.
• Thinking steps in the builder. You can now create and edit AI thinking steps directly in the fixed-steps section of the automation builder, including provider, model, instructions, input, output and time limits. Previously these could only be viewed.
• Steps that change something. The builder now lets you add steps that make changes in your tools. These steps are clearly labelled and use the existing approval flow.
• Usage alerts for connector allowances. The account owner now gets an e-mail when 80% of a connector's usage allowance is used, and another when all of it is used.

Improvements

• ConnectWise:
• New ConnectWise connections use StackJack's client ID by default.
• Existing connections that have no stored client ID are offered StackJack's client ID.
• ConnectWise credentials for automations now include the public key, the client ID choice and the x360Sync company ID.
• Error messages for rejected credentials:
• When a connector rejects stored credentials, the error now names that vendor's own credentials.
• Connectors that use sign-in now tell you to sign in again.
• For ConnectWise connections using StackJack's client ID, authorization errors no longer tell you to regenerate your keys.
• Axcient: The connection drawer lists Axcient's four products and says which ones Test Connection does not check.
• SuperOps, Ingram Micro and Backup Radar: Requests are now paced below these vendors' published rate limits, so you should see fewer rate-limit errors.
• Avanan: The authenticated-threats report now defaults to a check it can resolve.
• ImmyBot: Script-creation tools now describe the correct fields. Identity tools now also cover personal access token connections.
• ConnectWise dashboard tool: The board health dashboard now points to the correct tool for discovering boards.
• Foundry deployment checks: After you press Test connection, the Foundry tab keeps refreshing until the checks finish. You no longer need to reload the page or run the test again.
• Foundry guided setup: A setup that has been stuck for 30 minutes now offers "Retry setup". Picking more models than one setup can hold is refused with a plain message before anything is created.
• Staging copies: Deploying a staging copy whose "Run on" setting differs from production is now refused with a clear message, so production never silently changes where it runs. On Foundry, the spend guard now carries over when you deploy a staging copy.
• Reaped runs: When a run is stopped by the automatic recovery process, its verify steps are now marked "Not verified" instead of being left blank.
• Pending invitations: Re-inviting someone who already has a pending invitation adds the selected roles again, as the Team page describes.

Optimizations

• The automation builder's tool configuration and cost panel now recalculate only when their inputs change, so editing is more responsive.

Bug fixes

• Supervised runs:
• A supervised run whose approval did not go through now waits for approval again instead of completing without it.
• A supervised run that was restarted now stops for its own approvals instead of hanging until its time limit.
• Concurrent edits: If two people save the same automation at the same time, the second save now shows a message asking you to reload. This also applies when approving proposed tools. Previously these cases could return an unexpected error.
• Chat builder:
• Your message is no longer lost when you send it while a knowledge source is still being saved.
• A newly attached knowledge source is no longer dropped from the automation by the next turn.
• A question card keeps your typed answer and tool selection while a knowledge change finishes saving.
• Sending a message no longer freezes the page if clearing the message box fails.
• After a reload, the chat builder waits longer for a turn that is still being set up. This avoids offering a Retry that would charge for a second turn.
• A turn that runs into a problem partway through now keeps the answer you were already charged for, instead of discarding it.
• Words like "azure" in your instructions no longer switch where an automation runs.
• Automations page: "Test it now" now opens the run panel on the automation's detail page instead of doing nothing.
• Team page: Revoking or reinstating a member no longer reports a failure when only the page refresh afterward failed.
• Foundry:
• The cost panel now updates when you change only the "Run on" setting.
• A run on a non-Claude model that was interrupted by a platform restart before it started is now restarted, instead of failing.
• A temporary problem reading your Foundry key no longer tells you to reconnect Foundry.
• A run that is restarted keeps its spend guard.
• A daily check can no longer overwrite a newer passing check that you started yourself.
• The daily-check e-mail now covers every deployment it switched off. If there are too many to list, it counts the rest.

UI & UX changes

• For a Foundry automation, the run page, run breakdown and builder cost panel show an Azure cost estimate instead of credits.
• The run page's cost card says credits were absorbed only when the bill is exactly at the automation's credit limit.
• Foundry automations explain why they cannot include a thinking step.
• You are now asked to confirm before removing a personal sign-in that is set up but currently failing.
• Connection forms lock their inputs while saving and show more helpful hints when you add a connection.

Breaking changes

• The maximum credits per run setting now stops a run when it is reached. A run that reaches the limit is charged for the usage it actually consumed, including the turn that crossed the limit.
• A run the platform restarts can go past its limit before the limit check resumes, and is charged for what it used.
• A restarted run's earlier attempt now counts toward its credit limit.