Inspiration

A support agent needs to answer a Jira issue using the correct Confluence procedure. An AI assistant's proposed response may lack a source or cite a page version that has since changed, and customer-support writers already warn about AI answers that are "presented confidently as fact" while wrong (Help Scout, 2025) and recommend that "human help should always be easy to find" (Help Scout, 2025). Before authorizing publication, they need to inspect the destination, comment and exact quote. Evidence Review brings the cited draft into a Jira review panel and requires human approval before a separate execution attempt rechecks the evidence. The development-site acceptance record does not measure customer demand, time savings or production outcomes.

What it does

In the synthetic development-site run on 13 November, Rovo retrieved Confluence page 6684673, version 1, and prepared a cited draft for DESIGN-10 without publishing: the issue still had 0 comments. The Jira panel displayed the destination, comment and citation before approval for one attempt. History showed the approval as pending. A separate execution returned succeeded with comment ID 84934; Jira then contained 1 comment with its source line and exact quote, and history showed the approval as consumed. [development version 3, 13 Nov 2026]

The included Evidence Review Assistant, opened from Jira, listed consumed, expired and revoked approvals and repeated find_related_pages for DESIGN-8, returning synthetic Confluence page 5767169. Suggested pages are discovery results; collect_evidence must retrieve their version and excerpt before use as evidence. [development version 3, 13 Nov 2026]

Earlier runs established draft preparation, panel approval and corrected citation rendering. [2.5.0–2.6.0, 29 Oct 2026] Subsequent runs checked agent history, ISO dates, approval states and related-page discovery from Jira. [2.10.0–2.12.0 and 3.3.0, 7 Nov 2026]

Preparation, review, authorization, execution and inspection remain distinct stages. A matching quotation does not establish that the proposed answer is correct; human judgment remains necessary.

How it differs from built-in options. Atlassian documents that Rovo agents ask for confirmation before consequential tools, and that Jira Service Management AI answers reply automatically in connected channels once activated (Atlassian support pages, read 11 Nov 2026). Evidence Review binds a proposed Jira comment to exact Confluence page versions and quotes. Human approval is tied to the user, installation, destination and draft hash, expires after 15 minutes and authorizes one execution attempt; before an attempt the app rechecks the cited evidence, and its history links approvals to drafts, destinations and recorded execution outcomes. We have not tested the alternatives, so we do not claim exclusivity.

How we built it

The project combines JavaScript, Forge functions, a UI Kit issue panel, entity storage, and product integrations. Its manifest declares the MCP tools, the panel, and rovo:agent (evidence-review-agent). The included agent exposes three reusable actions: collect_evidence, prepare_comment, and get_action_history, excluding execute_approved_comment. A fresh Confluence installation initialized its own store and returned empty history. [2.10.0, 2026-11-07] History supplies ISO time fields [2.11.0, 2026-11-07] and per-approval effectiveStatus. [2.12.0, 2026-11-07] The intended authorization boundary is deliberate: a person approves through Jira; an MCP tool never grants approval.

Development-site collection worked with trusted user and installation context, while preparation and execution used issue-specific authorization. [2.5.0–2.7.0, 29 Oct]

We used synthetic page 3801089 and issues DESIGN-8 and DESIGN-9 to distinguish source retrieval, draft preparation, and actual comment outcomes in the acceptance record. [2.5.0–2.7.0, 29 Oct]

Challenges we ran into

The initial successful comment omitted its citation, and reusing an approval surfaced an HTTP 500 error. [2.5.0, 29 Oct]

The next documented version rendered the citation and returned a structured APPROVAL_USED rejection. [2.6.0, 29 Oct]

An object-valued proposalRequest still failed; Rovo succeeded with a string containing serialized JSON. [2.6.0, 29 Oct]

Conditional-write errors also exposed missing classifier wiring during preparation; the correction was deployed after diagnosis. [2.7.0–2.8.0, 29 Oct]

These findings made deployment evidence essential: passing local checks alone could not establish the observed product behavior.

Accomplishments that we're proud of

The latest synthetic development-site run demonstrated the complete sequence on Confluence page 6684673, version 1, and Jira issue DESIGN-10: a draft with 0 comments; panel review of destination, comment and citation; approval for one attempt; pending approval history; execution returning succeeded with comment 84934; then 1 comment containing the source line and exact quote, and history showing consumed. [development version 3, 13 Nov 2026]

The included agent, opened from Jira, listed consumed, expired and revoked approval records and repeated find_related_pages for DESIGN-8. [development version 3, 13 Nov 2026]

Earlier runs recorded identical replay without another comment and rejection of cross-issue approvals, expired approvals and changed source versions. [2.5.0–2.7.0, 29 Oct 2026] Agent history and related-page discovery were subsequently checked from Jira. [2.10.0–2.12.0 and 3.3.0, 7 Nov 2026]

These are bounded observations on synthetic development-site data. Customer adoption, time savings and production outcomes have not been measured.

What we learned

A retrieved quotation does not establish that a proposed conclusion follows from it. Human review remains central. Source versions, issue destinations, approval state, and execution results deserve separate attention.

We also learned to keep historical versions visible in demonstrations. A correction deployed later cannot retroactively change what an earlier run proved. Synthetic development-site evidence should be presented as such.

What's next

Complete real-site checks for other identities and installations, denied permissions, changed draft hashes, and durable concurrent comment execution. Preparation concurrency (case E) does not establish concurrent execution.

Confirm administrative MCP enablement on a judge's test site. Validate the support problem with users before making impact claims.

Built With

Share this project:

Updates

Submission history