Skip to main content

StackJack 0.1.0.10783+5359cee

StackJack 0.1.0.10783+5359cee

This release lets your AI assistant build automations for you through a downloadable builder skill. It adds a way to stop a run cleanly when there is nothing to do, and lets decision steps judge whole lists and show exactly what they were asked about. It also speeds up the Endpoints, Team and Connectors pages for organizations with many connections, and fixes several automation builder issues.

New features

  • Build automations with your AI assistant. A new "Build with your AI assistant" card on the Automations page links to a downloadable builder skill and a new help page. With the skill installed, your connected assistant walks you through the whole process:
  • It interviews you one question at a time.
  • It looks at your systems in read-only mode to decide what to automate.
  • It designs and builds the automation.
  • It tests the automation in test mode.
  • It helps you improve the automation over time.
  • You stay in control of going live. An automation your assistant creates always starts in test mode. Consent, going live, and switching on a schedule or webhook trigger are each asked as separate, plain questions. Each question explains what the step means for cost and for your customers' systems.
  • Installing the skill. The help page explains how to install the skill in Claude (web, desktop and Claude Code) and how to use it with other assistants. It also includes example requests that span several systems.
  • Stop a run when there is nothing to do. A step that reads data can now carry a "not empty" condition. For example, a list must have at least one item, or a count must be at least a given number. When the condition is not met, the run stops as Skipped instead of Failed. No failure email is sent and no AI usage is spent. You can set this condition in the Portal builder, the chat builder, and through your assistant.
  • Decision steps can judge lists and objects. A decision step's input can now be a whole list or object returned by an earlier step, not just a single value.
  • The limit is 64 KiB for everything sent to one decision.
  • If the input is over the limit, the step is refused and names the largest input and its size. The input is never silently shortened.
  • See what a decision was asked. The Decisions card on a run now has a "What was sent" section showing the exact data each decision call received. This makes it easier to see why a decision came out the way it did. This data is cleared when you delete the run's transcript.

Improvements

  • Any automation can now be offered as a tool to your AI assistant, whatever its trigger type. Previously only on-demand automations could be offered this way. The tool descriptions your assistant sees now state this correctly.
  • When a step's input cannot be resolved, the error message now says why:
  • the earlier step is missing,
  • the earlier step returned a result with no value at that location, or
  • the value is a list where a single value was expected.
  • Builder help text now makes clear that a record count is only returned by HaloPSA list tools. For other connectors, it explains how to pass the short list itself or a single field instead.

Optimizations

  • The Endpoints and Team pages load much faster for organizations with many MCP connections. Each connection's tool list is now read only when you open its tool editor, not on every page load or refresh.
  • The personal sign-ins section of the Connectors page makes fewer lookups, so it loads faster.

Bug fixes

  • Editing an unrelated part of an automation no longer rounds a very precise number in a condition. For example, a threshold of 1.00000000001 is no longer turned into 1. Previously, this rounding could change which runs the condition stopped.
  • If you type something that is not a number into a condition's number box, the builder now tells you what you typed instead of reporting a generic missing value.
  • An answer number too large to compare is now reported as such. It no longer silently passes or fails the condition.
  • A step with a condition that has no credential available no longer skips its condition. Under "abort", the run now ends as Failed before any AI session starts. Under "continue", the step is recorded as unchecked.
  • After a run was retried, the Decisions view could show the data sent by an earlier attempt next to a later attempt's answer. It now shows sent data only when it can be confirmed to belong to the answer shown.
  • A decision step with many inputs now still shows its sent data and result correctly in the run's records.
  • When a decision step was missing its input, later steps that used that decision were also wrongly reported as having no earlier decision step. Now only the missing-input error is shown.

UI & UX changes

  • On the Endpoints page, the Tools column now shows "Custom list" or "All tools" for each connection. The number of selected tools is shown in the tool editor.
  • If a connection's tool list cannot be read, or the connection no longer exists, the tool editor now refuses to open. This prevents it from opening with every tool ticked for a restricted connection.

Breaking changes

  • A decision step with no input can no longer be saved, whether in the Portal builder, the chat builder or through your assistant. The error message says what to add. Existing automations keep running unchanged until you next edit their steps.