You can also browse, search and filter all release notes directly in the application → Open release notes in TAI.

​

New Feature.
​

Open a chat and the foot of the thread says what the session consumed; open a report and the same figures sit in the run details, beside who ran it and when. Credits are the unit your plan is measured in, so this is the figure the Billing screen totals up — brought to the work that produced it.
​

Always on the line: the session's credits, and how many connector calls it made. On hover: the detail behind both — credits split into AI compute and connectors, and the calls split by connector, so you can see which system an agent leaned on. A report lays the same detail out in its Consumption section, with nothing to hover.
​

Consumption is counted per chat session rather than per message, because that is how it is billed. A figure still being counted — a chat you are in the middle of, a report that ran today — carries a ≈ until the day's usage is finalised. In a chat it updates as the agent works, so a long answer's cost accrues in front of you rather than arriving at the end.
​

These figures are a Workbench view. The chat SDK embedded in TripBuilder or any other host application shows no consumption.
​

The line under a chat always shows the session's credits and connector calls; hovering it breaks both figures down.

​

New Feature.
​

Once a month you'll get a short email listing the scheduled reports running under your name, with how often each runs, how many times it has run so far, and when it last ran. Each one links straight to its schedule, so anything you no longer need is quick to pause or delete — which also stops it consuming usage credits. If everything on the list is still useful, there's nothing to do. Owners with no active schedules don't receive the email at all.
​

The monthly email: every active schedule you own, how often it runs, how many times it has run, and when it last ran.

​

New Feature. Only Company Admins.
​

Nezasa staff administering your company can manage its configuration but cannot run agents or read generated content. Settings → Data access is how you open that up: name a Nezasa user, confirm what it exposes, and they can run agents in your company until you revoke it. Every grant and revoke emails your company admins and the named user, and lands in the audit log.

​

New Feature.
​

Agents can now filter and group itineraries by an external reference code — both the codes on the booked products themselves (activities, accommodation, transfers and other inventory) and the code on the template an itinerary was built from. That makes it easy to pull together, for example, every departure of a group tour that shares a transit activity's reference, or to reconcile itineraries against an external system's identifiers. Ask a data or operations agent something like "show me all itineraries with an activity that has external reference ABC."

​

New Feature.
​

A report you no longer want no longer has to run to the end. Cancel run sits next to the live status while a report generates, and on the run's own page it stops the run from anywhere — a different tab, or after you navigated away and came back. Queued scheduled runs can be stopped the same way, before they produce or email anything.
​

A stopped run now says so: it lands in the library as Cancelled rather than Failed, keeps whatever it managed to produce, and offers Re-run so you can start again with a narrower request. Closing the browser tab is no longer the only way to end a run.

​

Improvement.
​

Run details and input parameters flow as a single line of label-value pairs instead of a grid of tiles, and consumption reads as summary lines — the credit total followed by what it is made of, the connector calls followed by which connectors made them. Run details also gains Duration, measured from when the run actually started rather than from when it joined the queue. A figure we could not work out now says so instead of going quiet. Every figure is the one you saw before, with far less to scan.

​

Improvement. Only Company Admins.
​

The Invocations tab on a connector lists each call it made — who triggered it, how long it took, how much data it moved. That list now reaches back one year rather than indefinitely. Nothing else changes: the tab is unchanged for any call inside that window, and no other connector data is affected.

​

Improvement.
​

Every report run — scheduled or started by hand — now explicitly asks the agent for the rendered report in the channel's format: the full HTML document for an HTML report, a Markdown document for a Markdown one. Reports that occasionally arrived as a short summary instead of the full document should now arrive complete. Producing the report is still the agent's own step, so this makes the full document far more likely rather than guaranteed — runs that fall short are now recorded, so we can find and fix the ones that do.

​

Improvement. Only Company Admins.
​

When Nezasa staff enter your company to help you configure it, your own Audit Logs now record their arrival and departure alongside the actions they take — each entry showing the person's email address. The Administration scope filter groups the whole history in one click. Previously these entries were filed on Nezasa's side, so that filter never matched anything in your account.

​

Improvement.
​

Ops and analytics agents can now return a booking's external booking reference — the id an external or back-office system holds for it, such as a Midoco order number, a CRS reference, or a Booking-API-supplied id. It is available on both connector paths: live, per-itinerary via the Cockpit itinerary connector (it rides the lean default, so a plain itinerary lookup returns it), and in bulk via the TripBuilder data-warehouse itineraries mart. Ask an agent for a booking's Midoco number, order number, or booking reference and it resolves to this field — so you can cross-reference a booking with Midoco and crisis- or back-office tools without leaving the agent.

​

Improvement. Only Company Admins.
​

The admin-only agent surfaces had shipped in English to all 13 non-English languages: the Agent access matrix and its legend, the per-company access overrides that decide who may run a built-in agent, and the Thinking, Show thinking and Narrator behaviour toggles with their explanatory hints.

​

Improvement.
​

If you use the Workbench in one of the 13 non-English languages, whole panels used to arrive in English next to translated ones. Now translated: Analytics charts, summaries and filters; the Knowledge, Skills and Connectors showcase panels; the Manage all translations dialog, which had been explaining how translations work while itself untranslated; and shared labels you meet everywhere — Overview, From, To, Details, Source, Profile.
​

Wording follows each language's existing terminology, so a term keeps one name across the pages that use it. Brand names, code examples in input placeholders, and acronyms like JSON and MCP deliberately stay as they are.

​

Improvement. Only Company Admins.
​

API callers: DELETE /reports/{id}, DELETE /report-runs/{id} and DELETE /report-schedules/{id} now respond with { "id": …, "status": "DELETED" } instead of the full record. A delete has no business echoing the content it just removed — for a schedule that included its inputs and recipient list. This is a breaking response-shape change — update any integration that reads fields off the delete response. The Workbench needs no change.

​

Improvement. Only Company Admins.
​

Two machine-side principals lose the Reports schedule surface — create, run-now, and the lifecycle actions (edit, pause, resume, delete):

Neither can now leave behind a recurring schedule it has no way to pause or delete. Ad-hoc report runs and chat are unaffected for both, and managing schedules in the Workbench is unchanged.

​

Improvement.
​

Adding a skill or connector to an agent no longer means choosing from a list of bare names. Each row now carries its description, its maturity, and its reference, so you can tell what you are attaching before you attach it.
​

Search widened to match too: typing a word from a description — or from what a skill is for — finds it, not just an exact name.
​

Skills and connectors reserved for their own built-in agent stay listed but unselectable, and now say so with a Reserved badge instead of only on hover. Where nothing in the list can be picked, the picker explains why rather than looking broken.
​

Each skill now says what it does, and how mature it is, before you add it.

​

Improvement. Only Company Admins.
​

Agents ask for Fast, Balanced or Most capable and we pick the model — the vendor and the exact model id moved under Advanced for the cases that still need them.

​

Fix.
​

A report whose CSV, spreadsheet or PDF could not be delivered still finished green, with its summary claiming the file was attached and nothing to say otherwise. Every surface now names the file and why: the report, the live run, the run's own page when it later fails, the exported PDF and HTML, and the scheduled email.

​

Fix.
​

conversation.md is now written after the files are packed, so a file that could not be included is named with the reason instead of listed as delivered.

​

Fix.
​

Credit and token figures were formatted by your browser's language rather than the one you chose in TripBuilder AI, so 12.5 credits could read as twelve thousand to a German or French reader. Report run details and the Billing Center now both follow your chosen language.

​

Fix.
​

A call we can no longer match to a connector used to appear as a bare id. It now reads Unknown connector, with the id still available on hover, and the counts add up as before.

​

Fix.
​

An agent that wrapped its report in a code block had that wrapper stored with the report, so the whole document showed as raw source instead of formatted text. The wrapper is now removed when the report is stored, and reports saved before this fix render correctly too — in the app and in the HTML and PDF exports.

​

Fix.
​

​

Fix.
​

The Traveller Locator's detail table listed an Agency Email column that read "Not captured" on every booking, suggesting the contact details had never been recorded. They had — the agent simply wasn't asking TripBuilder for them. The column now shows the selling agency's real email, with its phone number alongside, and "Not captured" appears only where TripBuilder genuinely holds no value. During an incident that agency is often the fastest route to a traveller who can't be reached directly.

​

Fix.
​

Editing a scheduled report now saves only the fields you actually changed. A rename or a cadence change no longer fails over recipients a later delivery-policy change blocked, or over inputs its agent has since stopped accepting. Recipients and inputs you do edit are still checked as before.

​

Fix.
​

Tables in agent replies briefly failed to load, showing a "Couldn't load table" message that a retry could not clear. Sorting, copy, CSV export and fullscreen expand all work as before.

​

Fix.
​

Sorting a chat table column mixing numbers and text could reorder rows unpredictably, and blank cells sorted as if they were zero. Both are fixed.

​

Fix.
​

Maturity (POC / Preview / Beta / GA) says how far along a Nezasa-shipped catalogue entry is. Entries you create are yours from day one, so their maturity slot now shows a muted dash instead of a POC badge, and they no longer turn up under the POC maturity filter. Agents already behaved this way.

​

Fix.
​

When an agent produced a file the server could not store, the file was dropped silently while the answer still said it was attached — leaving no download and no explanation. Chat now names the file and the reason on the same turn, in the reader's language, and keeps showing it when the conversation is reopened. A report run whose report file is refused now fails naming that refusal, instead of reporting that the agent produced no content or saving a report whose body was the chat summary.

​

Fix.
​

Typing a starter prompt and pressing Enter used to save the agent mid-edit. Enter now opens the next prompt below the one you're on and puts the cursor in it, so you can type a list straight through.
​

Cmd+Enter (Ctrl+Enter on Windows) still saves, from anywhere in the form. The same fix applies to starter prompts on skills and connectors.

​

Fix.
​

When an agent fetched a PDF or image from the web, the file reached it as an unusable text dump. It now arrives as the document itself.

​

Fix.
​

Typing a filter and pressing Enter used to save the whole form. Every skill, connector and knowledge picker now claims Enter while its list is open; Escape closes the list and hands Enter back to the form.

​

Fix.
​

A past run's uploaded file isn't kept, so a report agent that requires one could never be re-run in a single click — it launched without the file and bounced back with a validation error. Re-run is now disabled for those reports and says why; Edit & re-run takes you to the form to choose the file.

​

Fix.
​

Uploads are judged by the file's extension, not by what the browser calls it — so a .csv your computer mislabels as a spreadsheet now goes through. The same rule means a file saved with no extension at all (export rather than export.csv) is turned away, even when its contents are fine: rename it and upload again.