v0.87.3
September 10, 2026
**Work queues that lose nothing, confidential data handled with warnings instead of blocks, and an AI call log:** an event reaching a queued case is applied when the case leaves the queue, flow designers see right away where confidential data goes out to automations and models, and every AI call can be recorded and reviewed by administrators.
-
Feature
WIP queue: events are no longer lost
When a phase is at its work-in-progress limit and a case waits in the queue, an event asking to move it is no longer lost: it is set aside and applied, in arrival order, the moment the case leaves the queue. The case detail shows how many events are pending, the Events page recognises them with the "Deferred" outcome and a dedicated card, and a queued case can still be moved manually to another phase. Force admit beyond the limit is now reserved to administrators.
-
Improvement
Confidential data in automations: warning, not block
An Event task that uses a confidential field is no longer stopped: the event goes out with the real values and stays marked "Confidential data" on the Events page, with the list of fields involved. In the designer, on save, a non-blocking warning points to the tasks reaching confidential fields — Event, Webhook (body, address, headers), automatic scripts and AI prompts on a provider without shielding — so designers decide knowingly, with no extra consents at every step.
-
Feature
AI call log
For each AI provider you can enable a call log: metadata only (date, channel, model, tokens, latency, outcome) or also the messages sent to the model, exactly as they left, and the response received. The log is reserved to administrators, under Settings → AI call log: filters by provider, channel, outcome and shielding, the detail of every call, CSV or JSON export and retention configurable in days. With shielding on, the recorded messages contain pseudonyms, never the real values; with shielding off, the provider form warns you.
-
Improvement
Documents and designer: polish
Text extraction also recognises formats declared under alternative names (a PDF uploaded as `x-pdf` is read, an `image/jpg` is treated as `image/jpeg`), with a tool to reprocess documents skipped in the past. In the expression picker, a Library document's text appears only for assets marked "agent context". An upload with an unexpected transport type is rejected right away with a clear message. The "connector not validated" warning lists every phase and task involved, with a direct link to each. The Documents guide clarifies what stays available in templates when agent context is off.