StackJack 0.1.0.7553+86f0ae6

This release adds a Settings page for organization-wide preferences, introduces fast decisions in automations (decision steps and smart filters on webhook triggers), lets the chat assistant edit existing automations, and makes knowledge sources editable in place. It also fixes a long list of data-loss and save-consistency bugs in the automation builder and assistant.

New features

• Settings page. A new Settings page (first under "More" in the sidebar) holds your organization-wide preferences: the tool catalog defaults (catalog-mode opt-in, default catalog mode, default pinned tools), the per-connector response-shaping rules, and, in regions that require it, the fast-decisions consent. Catalog defaults previously lived on MCP Setup; the per-client catalog mode stays there. The old response-shaping address forwards to the new section and keeps the connector you were opening.
• Decision steps in automations. An automation author can add a decision step that asks a typed question once per run, records the answer before it is used, and replays it if the run is re-queued. A write step can act on the option the decision picked, and the approval list shows which write step acts on which decision option. The run record says what was decided.
• Smart filters on webhook triggers. A webhook trigger can now judge whether an incoming delivery is worth a run. Each evaluation leaves a receipt in the automation's receipts panel showing the verdict and, when the delivery was not judged, why.
• Fast decisions for your organization. Where the region asks for it, an owner or administrator can consent to the fast-decision lane and store the organization's own key for the decision provider from the Settings page.
• The chat assistant edits existing automations. The assistant can now maintain an automation you already have: attach, edit in place, or remove knowledge sources (removals only when you name the source), set recording, the evaluation requirement, repeat-delivery skipping and test mode, clear the trigger configuration or the alert address, and hand the draft to the advanced builder with your changes intact.
• Knowledge sources can be edited in place. From the knowledge panel you can change a text source's name and text, a URL source's name and address, or a file source's name, without removing and re-adding it. The agent-knowledge MCP tool gains a matching update action.
• Claude Opus 5.5 is available as an automation model.
• Builder tools from your MCP client can maintain live automations, not only draft new ones. Restoring an automation version by its number now works, and installing a template as a draft from your MCP client no longer fails.

Improvements

• Tool search and tool suggestions are reranked by a decision call and fall back to the previous ordering whenever a clean answer is not available.
• Adding, removing, renaming, relabelling or reordering the options of a decision that feeds a write step now withdraws the destructive-action acknowledgment and asks for it again, and the consent card says so. The withdrawal note now states exactly which steps stop running until someone acknowledges again.
• In the tool-grant panel, ticking the destructive-action acknowledgment for an automation that declares write steps is refused and you are pointed to the builder's consent card, which lists those steps.
• A recorded acknowledgment that no longer covers the current write steps is now reported as pending by the MCP consent tool instead of "already on record", so it can be re-acknowledged from there.
• The template list from your MCP client now includes your own private templates.
• The MCP grant confirmation names the tools it is granting.
• Removing every knowledge source of an automation at once is refused and explained; the refused sources are put back in the panel so you can remove them individually.
• The knowledge panel refuses a blank text or address on a staged source as well as a saved one.
• A webhook delivery whose trigger configuration cannot be read now leaves a receipt instead of disappearing silently, and a sender that hangs up mid-delivery can no longer leave a receipt claiming no run started.
• A decision reply from the provider that is incomplete, out of range, or outside the declared options is treated as unavailable and follows your automation's outage policy, instead of being read as a real verdict.

Optimizations

• Tool search, tool suggestions, the tool picker, smart filters and the assistant's typed fields no longer wait on usage bookkeeping after the answer is already in hand; the answer is returned as soon as it arrives.
• The assistant no longer makes an extra typed-field call on early conversation turns where nothing it returns could be used, shortening those turns.
• An automation run resolves its decision context only when its own document declares a decision.

Bug fixes

• Values typed into a decision's number boxes (deadline and condition thresholds), a decision's question boxes, and the assistant's Core Knowledge box are now kept when you save with the cursor still in the box. Previously the save reported success but kept the old value.
• A trailing decimal separator ("0.") is no longer committed as a number, so a probability box cannot accidentally commit a threshold of 0.
• A typo in a decision condition keeps the condition rather than clearing it.
• The trigger editors in the builder now receive the stored configuration rather than an empty form.
• Handing a draft from the assistant to the advanced builder opens the automation being edited, not a new one, and the conversation's knowledge edits and removals are delivered by the builder's save instead of dropped.
• An edit conversation that never mentions knowledge no longer deletes the automation's knowledge sources, and a restored conversation no longer deletes or reverts Core Knowledge, knowledge sources or the response profile that were changed elsewhere in the meantime. A source edited both in the conversation and elsewhere is reported as a conflict instead of overwritten.
• Restoring or editing a conversation in the assistant no longer resumes a paused automation.
• Clearing the trigger configuration or the failure-notification address in chat now reaches the automation, and a cleared notification address is stored as empty rather than as a blank value.
• After a partial save in the assistant, a retry re-attempts only the knowledge sources that failed, and a source that did not attach is shown with its retry material kept. Two uploaded files sharing a name keep their own identities when a turn reorders them.
• A knowledge source is never deleted to make room for a replacement that was not saved.
• Refreshing a URL source, or changing its auto-refresh, no longer discards a pending rename or address change made in the assistant.
• A refresh of a URL source that was moved while its old page was being read no longer stores the old page under the new address. An edit that pushes the prompt over its size limit rolls back only what that edit changed, and if it could not be rolled back, the response says the change was kept. The synthetic Core Knowledge row can no longer be renamed.
• A destructive-action acknowledgment given from your MCP client is checked against the write steps that were actually shown; if a colleague changed them in between, the acknowledgment is refused and a new confirmation phrase is issued.
• A destructive-action acknowledgment withdrawn by a save is now reliably removed from the automation's own client credential, and a stranded stamp is repaired by the next re-sync.
• A decision step's yes/no answer outside 0–1, or a choice outside the declared options, is no longer bound or replayed as a real answer.
• A decision step that timed out after its request was sent, or whose request the provider rejected, is now recorded accurately in the run record.

UI & UX changes

• Every pointer to the organization's catalog defaults (per-client dialog, member card, tool selector, onboarding, connector setup notes) now names the Settings page.
• The Connectors card links directly to the response-shaping section on the Settings page.
• On the Settings page, members see the stored values read-only with a note that only an owner or administrator can change them; if the picker catalog cannot be loaded, an alert is shown instead of a savable form.
• The receipts panel explains why a smart filter did not judge a delivery (not opted in, fast decisions switched off, no usable key, or an unreadable filter/empty payload) instead of a bare "Not judged", and says when a decision was refused because the organization's fast-decision billing is paused.
• The all-source removal refusal reads correctly when there is a single source, and the held-back-removal message is worded neutrally.
• The assistant renders the finished turn when you press Enter to send.

Breaking changes

• The MCP grant-apply action no longer records the destructive-action acknowledgment: passing that flag as true is refused and you are directed to the dedicated consent tool. Record the acknowledgment first, then grant.
• The MCP tool that updates an automation draft now edits live automations with whole-replace semantics and is flagged destructive, so MCP clients that prompt before destructive tools will prompt for it.

Known issues

• Two webhook trigger fields, selected paths and maximum payload size, still commit only when focus leaves the box. Press Tab or click elsewhere before saving to be sure the value is kept.
• Core Knowledge typed in the advanced builder after a hand-off from the assistant, but before that conversation's own save completes, is replaced by the automation's current knowledge; a notice is shown when this happens.