Inspiration
The spec lives in Confluence and the work lives in Jira, and nobody checks one against the other. A PRD promises guest checkout, but a story says "redirect to sign-up". The spec says "email within 60 seconds", but the ticket says "batched every 15 minutes". Nobody creates the accessibility story at all. These gaps usually show up in QA or with customers, when they are expensive to fix. We wanted the gap to show up while the work is still being planned, inside the tools the team already uses.
What it does
Driftwise reads the Confluence spec linked to a Jira epic, along with the epic and all its child work items. It then judges every requirement as covered, partly covered, not planned or contradicted. It also lists the work that is not in the spec (scope creep) and the open questions.
- Jira work item panel. Shows a spec-confidence score, then each finding with "Spec says" and "MOBL-68 says" quotes side by side. Each quote carries a verified quote badge. Every finding has a one-click fix: add the suggested, testable acceptance criterion to the story; create the missing work item under the epic; or dismiss the finding as intended. If the spec page or the work changes after a check, the panel says so and offers a re-check.
- Jira project page, "Ship with confidence". Lists every epic in the space with its spec, confidence score and drift counts, worst first, so a lead can see release risk at a glance.
- Confluence byline on the spec page. Shows which Jira work follows this page, and how closely.
- Rovo agent "Driftwise". Ask "Is this epic ready to ship against its spec?" in Rovo chat. The agent starts or reads the check, answers questions about the spec with quotes, and with your agreement adds acceptance criteria or creates work items. Its four actions (get drift report, read spec and work, add criterion, create work item) are ordinary Rovo actions, so other agents can use them as tools.
- Spec discovery. If no spec is linked, Driftwise suggests the likely spec. It uses Teamwork Graph to find pages in the Confluence spaces linked to the Jira space, falls back to a title search, and links your choice with one click.
In the demo, Driftwise scored the "Checkout v2" epic at 30%. It found 3 contradictions (guest checkout, stacked promo codes, batched email), 4 requirements nobody planned (including inventory hold and WCAG 2.2 AA), and 2 out-of-scope features being built (PayPal, loyalty points). After two one-click fixes, the Rovo agent reported 44%.
How we built it
Driftwise is a single Forge app. It is installed on Jira and connects to Confluence, and it uses no external servers.
- UI Kit (
@forge/react) for the issue panel, project page and Confluence byline. All three share one resolver. - Async events queue with a 300-second consumer, so a check over a long spec and many stories never blocks the UI. The UI polls the job state.
- Forge LLMs (Claude) with a strict reviewer prompt that returns JSON only. The model must quote both the spec and the Jira text.
- Quote verification. Each quote is matched against the source text (an exact match, or at least 85% word 4-gram overlap for long quotes). Unverified quotes are labelled "check wording", so the team never acts on an invented sentence.
- Forge KVS stores each result with a fingerprint of the issue update times and page versions, which is how stale results are detected. It also stores a project index and a page→issue map for the project page and the byline.
- Rovo
rovo:agentplus fouractionmodules backed by Forge functions. - Teamwork Graph through
requestTeamworkGraph, using a registered persisted Cypher query (JiraSpace ← linked ConfluenceSpace → ConfluencePage) with pagination. - Jira REST v3 and Confluence REST v2. Fixes run as the user, so Jira permissions are respected.
- Cost guard. AI spend is tracked per installation and capped each month; the app stops calling the model when the cap is reached.
Challenges we ran into
- Hallucinated quotes were the biggest trust problem with LLM review, which is why we built verification into the core.
- Teamwork Graph only runs registered (persisted) Cypher queries, needs granular scopes, and limits page sizes, so we built pagination and a search fallback.
- Localized Jira sites name issue types in other languages ("史詩", "故事"), so Driftwise picks types by hierarchy level, not by name.
- Long checks do not fit in a resolver call, so the work moved to a queue consumer.
Accomplishments that we're proud of
- Every finding is grounded: you can see the exact sentence in the spec and in Jira.
- One check feeds four surfaces (issue panel, project page, Confluence byline and Rovo chat), and the agent can act, not just talk.
- It runs entirely on Atlassian: Forge functions, Forge storage and Forge LLMs.
What we learned
AI review is only useful when people can trust and check it. Quotes, verification badges and "Not an issue" dismissals matter more than clever prompts. We also learned how far Forge has come: LLMs, Rovo actions and Teamwork Graph all live inside one app.
What's next for Driftwise
- Checking specs during sprint planning and in pull requests, through Bitbucket and GitHub links.
- Watching spec pages and alerting the epic owner when a requirement changes after work has started.
- A Marketplace listing.
Built With
- claude
- confluence
- forge
- forge-events
- forge-kvs
- forge-llms
- javascript
- jira
- react
- rovo
- teamwork-graph
- ui-kit