Inspiration (Backed by Our 101-Task Empirical Study)
Coding agents excel when given developer pull-request specifications, yet real customer bug reports in Jira Service Management (JSM) and support channels describe only symptoms. In our 101-task empirical benchmark study (Lingua-SWE-129, CC BY 4.0 across FastAPI and Rich), rewriting developer specifications into symptom-only user reports cut agent solve rates by 31% on 31B models (36.0% to 25.0%) and 62% on 12B models (26.0% to 10.0%), with 72–89% of generated patches failing hidden tests. Yet coding agents still located the right file 83–85% of the time—what was missing was the exact function/class symbol to edit and the intended behavioral contract. We built Agent-Ready Issues for Jira & Rovo MCP to close that exact gap natively inside Atlassian Forge.
What it does
- 5-Dimension Agent Readiness Score (0–100): Evaluates Jira/JSM issues across Reproduction Steps (25), Behavioral Contract (25), Code-Map Symbol Localization (20), Acceptance Criteria (15), and Environment/Trace Context (15), writing
agentReadyScoreandagentReadyStatusto Jira Entity Properties for JQL (agentReadyScore >= 75 AND agentReadyStatus = "READY"). - Zero-Egress Code Map Localization: Indexes only function signatures, API routes, and exception names in Forge KVS (
@forge/kvs), pinning targets likefastapi/dependencies/utils.py :: analyze_param (98%)without source-code bodies ever leaving Atlassian. - Bilingual Reporter Clarification Loop (EN / ID): Drafts up to 3 ranked clarifying questions in English and code-mixed Indonesian for under-specified tickets and posts them to the reporter in one click.
- Closed-Loop Rovo MCP Server (
rovo:mcp) + Rovo Agent & Skill: Exposes 6 typed tools (get-agent-brief,get-readiness,list-agent-ready-issues,post-reporter-questions,record-code-location,record-fix-outcome) alongside the Bug Brief Engineer Rovo agent (agent-ready-briefRovo skill) and an open-sourceclient-skill/fix-from-agent-brief/SKILL.mdfor Claude Code, Cursor, and Rovo Dev. - Runs on Atlassian Eligible: Built with
@forge/llm(claude-sonnet-5) inside an async Forge Events consumer (timeoutSeconds: 600) with zero external egress and a deterministic local code-map fallback.
Challenges we ran into
- Keeping issue panels dense with structured evidence (readiness breakdowns, localized symbol badges, Teamwork Graph precedents) while avoiding horizontal badge overflow inside Jira Cloud's narrow right/main columns (
Inlinewrapping withshouldWrapandrowSpace). - Supporting real code-mixed Indonesian and English customer bug reports without leaking repository code outside Atlassian Cloud.
- Designing 6 validated
rovo:mcptools plus graceful deterministic fallbacks when optional Teamwork Graph scopes or LLM quotas vary across judge sites.
Accomplishments that we're proud of
- Verified Eligible for the Runs on Atlassian program, deployed to production (
v2.6.0), and live on Jira Cloud (PLAT-33510,PLAT-33511,PLAT-33512). - Built end-to-end automated lifecycle coverage (
4/4tests passing vianode --test) verifying MCP brief retrieval, reporter clarification, symbol recording, andSOLVED_BY_AGENTfix-outcome updates.
What we learned
- Autonomous coding agents fail mainly from missing target function symbols and behavioral contracts rather than file discovery; making readiness explicit in Jira Entity Properties gives both triage engineers and MCP agents a reliable quality gate.
What's next for Agent-Ready Issues for Jira & Rovo MCP
- Expanding CI webhooks for multi-language AST signature refresh and Automated JSM customer portal clarification triggers.
Built With
- atlassian-forge
- atlassian-forge-llm
- atlassian-teamwork-graph
- javascript
- jira
- jira-service-management
- mcp
- node.js
- react
- rovo