Inspiration
In export logistics the departure date (ETD) of a shipment moves again and again: the carrier rolls the booking, a transshipment port is congested, a later revision pulls it two days earlier. Each change is typed into a field, and the reason ends up in a comment, an email or nowhere. When a customer asks "why is my cargo late?", someone scrolls the history and rebuilds the story by hand. AI assistants make this worse in one specific way: asked for a reason, they can produce a plausible but unsupported one.
What it does
- Jira issue panel. On a shipment issue it reads the change history of the ETD (due date) field and lists every change: when, who, old date, new date and the shift in days, with the baseline and the current date at the top.
- The comment written with each change, linked. For each change it shows the comment the same person wrote around that moment (plain text, whitespace normalized), linked to the original. The match is by author and time only, so it is a potentially relevant comment, not a verified explanation of the cause. If there is no such comment, the row says "Reason not recorded".
- Rovo agent with one read-only action. Ask "Why did the ETD of SHIP-1 move?" and the agent calls the action, gets the same history, and answers with baseline, current date, total shift and every change, quoting the matched comments. Ask it to "draft a short customer update" and it is instructed to use only those comments and to promise to confirm a missing reason rather than supply one.
How we built it
- Atlassian Forge (Node.js runtime):
jira:issuePanel(UI Kit),rovo:agentand oneactionbacked by the same function. - Only the
read:jira-workscope. The app never writes to Jira and stores nothing outside it. - The history logic is a small pure module (pair each ETD change with the nearest non-empty comment by the same author, from one hour before to 24 hours after the change and never past the neighbouring changes; each comment is used once) covered by unit tests, including consecutive changes and changes without a comment.
- The agent is instructed to use only the action output, and the action returns "Reason not recorded" explicitly, so the model has a truthful value for the gap. This reduces unsupported explanations; it cannot guarantee them away.
Challenges we ran into
- Rovo chat is not available on a free developer site, so the demo runs on a
forge site provisiondemo site. - Time zones: Jira returns history in UTC; the panel labels times in UTC to avoid silent off-by-one-day shifts.
- Matching comments to changes without guessing: only the same author, a bounded time window, never crossing the neighbouring change, and each comment used once. Outside that, the row stays "Reason not recorded".
Accomplishments that we're proud of
In the recorded demo the agent quoted the two matched comments and reported the third change as "Reason not recorded"; it did not add a reason of its own, and the customer update it drafted promised to confirm the missing one. An export coordinator can review the whole ETD story in one view before writing to a customer.
What we learned
For an AI agent in operations, an explicit "not recorded" value is a feature: it gives the model a truthful answer for the gap instead of an empty field it may be tempted to fill.
What's next
Support a custom ETD field chosen per project, ETA alongside ETD, and a Confluence report of all shipments whose date moved this week.