Inspiration
It happens a lot on software projects: a decision gets made in a Jira comment, everyone agrees, and three months later someone opens a ticket that does the opposite. Nobody did it on purpose. The decision was there, just buried in a thread nobody reads anymore.
The Codegeist theme is about context, and I think past decisions are one of the most useful kinds of context a team has. So I tried to build something that remembers them.
Rovo already has an agent to help you make a decision. This one is for remembering the ones you already made.
What it does
There is a decision registry, and you can reach it from four places.
In Rovo, you can ask the Decision Keeper agent things like "why did we choose PostgreSQL?". It looks in the registry, in Jira comments and in Confluence, and gives you the answer with the source. You can also describe what you're about to do ("I want to move the cloud data layer to MongoDB") and it tells you if that contradicts something already decided. If the team really changes its mind, the old decision is marked as superseded, not deleted.
In Jira, every issue can show a Decisions panel. It spots comments that look like decisions ("we decided", "agreed"...) and lets you save them in one click. It also shows decisions from other issues that seem related, so you see them before you start.
The same actions are exposed as an MCP server, so you can use them from Claude, Cursor or VS Code.
And decisions are pushed to the Teamwork Graph, linked to their Jira issue, so Rovo search and other agents can find them too.
How I built it
All on Forge: a rovo:agent with 4 actions, a jira:issuePanel in UI Kit, the rovo:mcp module, and a graph:connector using @forge/teamwork-graph. Decisions are stored in Forge KVS. Jira and Confluence are searched with JQL and CQL through api.asUser(), so people only see decisions from content they can already access.
The registry code is in one file shared by everything, so the agent, the panel and MCP behave the same way. Whenever a decision is saved, wherever it comes from, it's also sent to the graph.
Challenges
Finding decisions in comments without flagging every comment was harder than I expected. I ended up combining a few decision words with keyword matching on the issue, and I had to stop the panel from suggesting a decision that was already saved from Rovo.
For contradictions, keyword matching isn't enough, it needs some reasoning. So I let the Rovo agent do that part and just give it precise data (ids, dates, sources) so it doesn't make things up.
rovo:mcp and the graph connector are new and there aren't many examples yet. I figured out the document format by reading the SDK type definitions. Setup took a while too: Forge CLI 15 wanted a newer Node.js than mine, and the MCP and connector settings are hidden in Atlassian Administration.
What I'm happy with
It works end to end on a real site: the agent answers with sources and catches the MongoDB case, the panel saves decisions from comments, the MCP Inspector connects and sees the tools, and the connector sent the decisions to the graph.
What I learned
Rovo actions are very reusable. The same four functions run the agent, the MCP tools and most of the panel.
What's next
Showing decisions directly in Confluence pages, suggesting a decision automatically when someone writes "we decided", and reminding people to review old decisions after a while.
Built With
- atlassian-forge
- confluence
- forge
- forge-kvs
- javascript
- jira
- mcp
- node.js
- react
- rovo
- rovo-agent
- teamwork-graph
- ui-kit