atlassian-rovo（阿特拉斯罗沃）

Atlassian ACLI + REST API Automation

Configuration Discovery

Before starting any workflow, gather these from the user or discover via CLI:

Always ask the user which project/space to use. Never assume defaults. Use --limit 10 for all Jira list operations and limit=10 for Confluence REST API calls.

Setup & Authentication

For full setup, API token generation, and troubleshooting, see setup.md.

Quick config — add a .env file to your project root:

ATLASSIAN_EMAIL="you@example.com"
ATLASSIAN_API_TOKEN="your-api-token-here"
ATLASSIAN_SITE="https://yoursite.atlassian.net"

Install ACLI and authenticate:

brew tap atlassian/homebrew-acli && brew install acli
set -a; source .env; set +a
# ACLI expects the bare domain, not the full URL — strip the https:// prefix
ACLI_SITE="${ATLASSIAN_SITE#https://}"
acli jira auth login \
  --site "$ACLI_SITE" \
  --email "$ATLASSIAN_EMAIL" \
  --token < <(echo "$ATLASSIAN_API_TOKEN")

For Confluence, use the REST API via curl with basic auth:

curl -s -u "$ATLASSIAN_EMAIL:$ATLASSIAN_API_TOKEN" \
  "$ATLASSIAN_SITE/wiki/rest/api/content" ...

See reference/atlassian-helpers.sh for convenience shell functions.

For image upload configuration, see image-upload.md. For git branch integration (GitHub/Bitbucket), see git-integration.md. For Confluence comment operations, see confluence-comments.md.

Agentic Project Management Workflow

Full lifecycle project management using Jira + Confluence.

Execution Modes

This workflow supports two modes:

• Multi-Agent Mode (Claude Code with Agent Teams): Parallel execution with TeamCreate/TaskCreate orchestration. Requires CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 in ~/.claude/settings.json.
• Single-Agent Mode (all agents): Sequential execution — one workstream at a time. Same Jira tracking and Confluence updates, no multi-agent dependencies.

Prerequisites

• ACLI installed and authenticated (Jira)
• API token configured in .env (Confluence REST API)
• For multi-agent mode: Claude Code with CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1

The 4-Phase Cycle

Phase 1: INTAKE   -> User describes request -> Confluence plan + Jira Epic
Phase 2: EXECUTE  -> Agent team works tickets with Jira tracking
Phase 3: RESUME   -> New session resumes from Jira state (fully autonomous)
Phase 4: COMPLETE -> All tickets Done -> Epic closed -> summary delivered

Phase 1: Intake & Planning

Classify the request, create Confluence plan page, Jira Epic with [AI-PM] prefix, child tickets per workstream, link dependencies, then present plan to user for confirmation.

See phase-planning.md for step-by-step details and templates.

Phase 2: Execution

Multi-agent mode: Create agent team, map tickets to tasks with dependency tracking, spawn teammates with standard prompt. Orchestrator monitors progress and updates Confluence plan page.

Single-agent mode: Work each ticket sequentially, respecting dependency order. Track progress via Jira transitions and Confluence updates.

See phase-execution.md for both modes (multi-agent and single-agent workflows). Both modes share common protocols for transitions, branch creation, and publishing — see common-patterns.md.

Phase 3: Resume

Auto-detect open [AI-PM] Epics, read Confluence plan, query incomplete tickets, classify state.

Multi-agent mode: Spin up right-sized team for remaining work. Single-agent mode: Pick up the next incomplete ticket and continue sequentially.

See phase-resume.md for both modes (resume protocol and JQL patterns).

Phase 4: Completion

Update Confluence page status, transition Epic to Done, add final summary comment, report deliverables to user.

See phase-completion.md for completion checklist.

Team Sizing (Multi-Agent Mode)

One ticket per teammate. Lead coordinates only (no workstream tasks). Min team = 2.

Key Constraints

• [AI-PM] prefix on all managed Epics — the resume phase uses JQL summary ~ "\\[AI-PM\\]" to auto-detect managed Epics.
• One writer per resource — Confluence uses optimistic locking with version numbers. Concurrent writes cause version conflicts and data loss. Only one agent should update a given page/issue at a time.
• --limit 10 on all Jira list operations; limit=10 on all Confluence REST calls — larger result sets consume excessive tokens.
• Transition to In Progress when work starts — this signals to other agents (and the resume phase) that a ticket is actively being worked.
• Assign tickets when starting work — use acli jira workitem edit --key {key} --assignee "@me". Unassigned tickets look abandoned during resume.
• Publish findings as child pages of the plan page (using parentId) — keeps all deliverables organized under one parent.
• GET before PUT for Confluence — the REST API PUT replaces the entire body. Always read the current version first, then PUT with incremented version number. Also check for inline/footer comments before updating — users leave review feedback as comments.
• JQL != workaround — ACLI escapes ! in JQL. Use NOT IN (Done) instead of != Done.
• JQL [ / ] escaping — square brackets are JQL range-query operators. Use \[AI-PM\] (escaped) in summary ~ queries, not [AI-PM].
• Confluence search uses v1 API — CQL search for pages must use /wiki/rest/api/content/search, not v2 /search.
• Comment replies use v1 API — creating threaded replies requires POST /wiki/rest/api/content with ancestors array. The v2 children endpoint returns 405. See confluence-comments.md.