StackJack 0.1.0.7119+9317985

This release turns on thinking and write steps in the agent builder, makes the compact tool-catalog address work with no setup, and corrects a series of billing and error-message defects around runs, requeues, and organizations that bring their own Anthropic key.

New features

• Thinking and write steps in agent bookends. Automations can now declare a model-driven "think" step and write steps in their before/after-run phases. Think steps are metered into the run's credit total at the think model's rate, or run on your own Anthropic key if your organization brings one. Write steps are only executed after you approve them in the builder (see UI & UX changes).
• HaloPSA recurring invoice schedule tool. A new tool updates a recurring invoice's schedule and reads the record back to confirm the change. The create-recurring-invoice tool now states clearly that the start date on the header is ignored, so agents use the schedule tool instead.
• Microsoft 365 Copilot is recognized automatically. Connections made from Microsoft 365 Copilot now receive the compact tool list without any self-identification or configuration.

Improvements

• The compact endpoint address needs no setup. The compact address always serves the compact catalog, regardless of your organization's catalog-mode setting, and ignores any ?tools= parameter. The catalog-mode switch now governs only the standard endpoint. Previously a Copilot organization that had never touched the setting could be served the full tool list.
• Runs are priced on the model they actually ran on. If you change an automation's model while a run is paused, stranded or in flight, the run continues to be billed at the rate of the model its session was created with. The run page's cost breakdown uses the same model, so it always matches the billed total.
• Documentation covers using StackJack from a Microsoft 365 Copilot chat, including example prompts and how to confirm a tool was actually called.
• Microsoft Graph bring-your-own-app documentation now links the published setup script and lists the permissions each tier requires.

Optimizations

• The connectors page and connector usage counters load noticeably faster for organizations with several idle connectors; usage totals are now fetched in a single query instead of one per connector.

Bug fixes

• Own-key problems no longer show as "buy more credits" or a server error. When your organization's own Anthropic key is rejected, cannot be read, or your base subscription has lapsed, the builder banner, the AI-assist button, the create/deploy step, and Run now each show the correct cause and repair guidance. Previously an organization with ample credits could be told to purchase more, and an organization that was simply out of credits could be told its key was rejected.
• A key revoked mid-run now ends the run with a clear key-repair outcome instead of "Anthropic API unreachable", and triggers receive a non-retryable refusal rather than a retryable server error.
• Run page said "not billed" on a run that was charged. Runs settled by recovery paths (for example cancelling a paused run) now record the charged amount on the run, so the page no longer claims the run was free or attributes it to a bring-your-own-key or staff-initiated run.
• Run page cost after a requeue now reflects the current attempt's model rather than the previous attempt's, without needing a page reload.
• Requeued runs use the right funding. A requeued attempt is funded the same way as its first attempt; a platform-funded run no longer switches to your own key (or vice versa) partway through.
• A stranded run's session is stopped before the run is settled, so a live session no longer keeps consuming budget after the run has been marked failed.
• Builder assistant answers survive interruptions. Closing the tab mid-turn or a storage fault during finalization no longer loses an answer that has already been generated and billed.
• Destructive-step approval is bound to the list you approved. If the write-step list changes and you reload the page, the builder re-asks for approval instead of carrying a stale tick forward.
• HaloPSA schedule dates. A schedule start date must include a full calendar date; a time-only value is refused up front instead of silently being applied to today's date. A start date with sub-millisecond precision is no longer written correctly and then falsely reported as unverified. Refusal messages now point to an invoice-scoped read tool that the invoicing permission can actually reach.
• Refused stored credentials now surface their specific guidance instead of "unknown error"; WatchGuard token refusals reach their own guidance.
• Microsoft Graph setup script no longer removes redirect URLs on the app registration that it did not create.
• Requeue refusal wording no longer claims a run was started with your own key when that cannot be known; it now simply states no key is on file for the run.

UI & UX changes

• The compact endpoint card in the Portal is always marked Ready and its Copy button is always enabled. The card no longer warns that catalog modes must be enabled first, and the inline enable action is described as applying to the standard endpoint.
• The builder's approval box now lists each declared write step by name and re-asks for approval whenever the list changes.

Breaking changes

• Builder assistant turns bill your own Anthropic account if your organization brings its own key. Previously these turns ran at no charge on StackJack's key. If your key cannot be read, the turn is refused with repair guidance rather than falling back to StackJack's key.
• StackJack no longer absorbs AI usage on runs that end abnormally. Runs that fail before starting, are cancelled while paused, hang and are reaped, or are requeued after a restart are now billed for the model usage they actually consumed (including think steps), capped at the automation's per-run limit. Only usage that never occurred is refunded. Organizations using their own key are unaffected — those runs continue to bill zero credits.