Inspiration

Coding agents will start on anything. Give one an issue that says "Implement live call monitoring" and nothing else, and it does not push back. It guesses, and the cost shows up later as a pull request that solves a different problem.

Teams already know what a workable issue needs: a description, acceptance criteria, steps to reproduce for a bug, no open questions, no unfinished blocker. That knowledge lives in people's heads and in review comments. Nothing checks it at the moment work is handed to an agent.

What it does

Handoff is a Forge app for Jira that checks whether an issue is ready for a coding agent and gives the same answer in three places:

  • Issue panel. A verdict (Ready for an agent, Needs clarification, Blocked), each check with a tick or a cross, and the reason for every failed check. It updates as the issue is edited.
  • MCP tools for coding agents. get-work-brief gives any MCP client a compact brief of the issue (description as plain text, links, subtasks, recent comments) with the verdict, what is missing, and whether to start. ask-for-clarification lets the agent post its open questions on the issue as one comment instead of guessing.
  • Rovo agent "Handoff Coach". Inside Jira it reports what is missing, asks the questions needed to fill the gaps, drafts a description to paste, and can post open questions for the reporter after the user confirms.

The brief and the panel also carry connected work from Teamwork Graph: the goals the issue contributes to, linked pull requests and linked designs. An agent is told when a pull request is already open or merged, so it does not redo work in progress.

The checks are deterministic and use no LLM. The same issue always gets the same verdict, and every failed check points at something a person can fix.

How we built it

Everything runs on Forge, with no external egress and no external storage.

  • rovo:mcp exposes two action modules as MCP tools at mcp.atlassian.com/forge/<app-id>.
  • rovo:agent uses the same two actions. Its prompt forbids it from judging readiness itself: the verdict comes only from the action.
  • jira:issuePanel (UI Kit) calls the same readiness code and refreshes on issue change events.
  • Connected work comes from three curated Teamwork Graph queries sent in one request as the current user. It is context only: it never changes the verdict, and a failed lookup is reported inside the brief.
  • The rules are plain JavaScript modules (ADF to text, readiness, brief, graph) covered by 22 node:test tests that run in CI without Forge.
  • Output is bounded: description capped at 6,000 characters, 5 comments, 20 links, 20 subtasks, 10 questions per comment.

How it meets the technology requirement

Handoff is built on Forge and integrates with Jira (issue panel, Jira REST API through @forge/api), Rovo (rovo:agent, rovo:mcp, action) and the Teamwork Graph API. The context it uses is the issue itself, its links, blockers, subtasks and comments, and the goals, pull requests and designs Teamwork Graph connects to it.

Challenges we ran into

External MCP calls do not always arrive with a user session. On our development site asUser() threw Authentication Required on the first calls from an external client, and worked once the same user had used the app inside Jira and re-authorised the client. Forge passes the caller's account ID either way, so when there is no session Handoff acts as the app only after asking Jira whether that account may browse (and, for comments, comment on) that exact issue. A caller without access gets the same answer as for an issue that does not exist. We verified this with a second user and an issue in a private space.

Accomplishments that we're proud of

One set of rules, three surfaces, one verdict. The panel, the Rovo agent and an external coding agent were checked against the same issue and agreed, on the production app.

What we learned

Keeping the rules as plain modules with no Forge imports paid off: the verdict logic is tested in milliseconds without a tunnel or a site, and the panel, the agent and the MCP tools reuse it unchanged.

What's next

  • Per-project rules, so teams can set their own definition of ready.
  • Marketplace listing.

Built With

Share this project:

Updates

Submission history