Hivemind for Jira and Rovo

Your team's AI coding agents keep repeating mistakes someone already corrected. Hivemind turns those corrections into approved team rules, now inside Jira and Rovo.

Comment

Inspiration

We code with AI agents at every hackathon, each of us with our own. When one of us corrected an agent for breaking a project rule, the other's agent made the same mistake an hour later. Every team running more than one coding agent has this problem: a correction lives in one chat and disappears. We built Hivemind so a correction becomes a rule the whole team's agents follow, but only after a lead approves it.

Most of a software team's context already lives in Jira, so for Codegeist we brought Hivemind's rules to where the work is planned and reviewed.

What it does

Hivemind for Jira and Rovo is a Forge app with two parts:

  • A Jira work item panel, "Hivemind team rules." It reads the work item's summary, labels and components and shows the team's approved rules that apply, with locked rules (set by an admin, never overridden) first. If someone corrected an agent while working on the item, they can propose a lesson right from the panel.
  • A Rovo agent, "Hivemind." Ask it "which locked rules apply before I push?" and it answers only from the team's approved rules, by name. Tell it "remember that we always run the tests before we push" and it proposes that as a lesson. It never approves anything.

Every proposal from Jira or Rovo arrives in Hivemind as pending. A lead approves it in Hivemind, and from then on every connected coding agent loads it at the start of each session.

How we built it

  • Forge modules: a rovo:agent with two action modules (get-team-rules, propose-lesson), and a jira:issuePanel built in UI Kit with a resolver.
  • Jira context: the resolver and the Rovo actions read the work item through the Jira REST API (read:jira-work) and rank the team's rules by overlap with its summary, labels and components.
  • Hivemind API: Forge functions call the hosted Hivemind API through a declared egress permission, with a team token stored as an encrypted Forge variable. The token can read rules and propose lessons. It can't approve them.
  • Hivemind itself: Next.js, Postgres and Claude for turning corrections into lessons. Hivemind existed before Codegeist (we built it at a hackathon in October); the Forge app is new for Codegeist.

Challenges we ran into

Keeping humans in charge. An agent that can write its own rules can also learn the wrong ones, so nothing from Jira or Rovo becomes a rule on its own: the token the Forge app uses can only propose.

Accomplishments that we're proud of

In our own test on a mock repo, agents without lessons broke team rules in 22 of 36 runs; with lessons we wrote by hand, none did. Hivemind also won Lemma's track at an LA Tech Week hackathon.

What we learned

Rovo decides when to call an action from its description alone, so "propose, never approve" has to be written into both the agent's prompt and the action.

What's next

A Bitbucket pull request view that flags changes breaking a locked rule, and exposing the actions through rovo:mcp so teams can add them to their own agents.

Built With

Share this project:

Updates