mini-openclaw/COORDINATION_AGENT_GUIDE.md

Review Coordination Agent

Agent 3 is available at /review-workbench. The local preview runs at http://127.0.0.1:3012/review-workbench with the API on loopback port 8012. The application reuses Agent 1 risk snapshots, identity configuration, SQLite storage, audit trail, and mini-openclaw's scoped AgentManager execution.

Workflow

1. Create a coordination project from an existing risk run. Select its industry and complexity (1-5). Only one active coordination record per project is allowed.
2. Configure reviewers with industries, capacity, outside workload, availability, and project exclusions. The recommendation ranks eligible people using industry matching (40 points) and remaining capacity after assignment (up to 60 points). Local workload is the sum of assigned active project complexity. Capacity and outside workload use the same units. These are preview heuristics, not approved Xingbang assignment policy. An administrator confirms the local assignment.
3. Convert findings into follow-up tasks with an owner and due date, or record a reviewer's evidence-based disposition. Material submission does not close a task. Inspection must precede reviewer confirmation. A reviewer can reopen it.
4. Record the proposal: CNY ten-thousands, months, annual percentage rate, deposit percentage, and guarantee conditions. Each actual change stores old/new values and creates affected review items. Decimal-equivalent inputs are not changes.
5. Confirm pending risks, tasks and change reviews before department review, committee preparation, or completion. A proposal must also be recorded.
6. Inspect the current briefing, export text, or request an AI briefing.

Proposal changes during department review or committee preparation return the project to financial analysis. Updating to a later risk snapshot for the same project preserves previous snapshots, invalidates applicability of old risk confirmations, and creates data/report review tasks. Open tasks remain open.

Human Authority

• Administrator: people configuration, local assignment, task preparation, inspections and proposal recording.
• Reviewer: risk disposition, task resolution/reopening, change confirmation, and progression through final review stages; also prepares tasks and proposals.
• Customer manager: sees authorized project records and submits material references.
• Loopback-only preview: all operations enabled for demonstration.

The current roles reuse the existing application roles. Secretary, assigned main reviewer and department-head identity mapping still needs formal organization integration. The workspace records local responsibilities; it does not assign or approve projects in ERP. Configure authentication using RISK_AGENT_GUIDE.md.

Agent Execution

The optional AI briefing uses risk.local_model -> invoke_risk_task -> AgentManager.invoke_scoped with only authorized snapshot-reading tools and the review_coordination/SKILL.md native skill. It cannot assign, close tasks, execute SQL, or change reports. Missing model configuration returns a clear error without changing the project. A briefing is visibly stale after another project update. Tests run the real agent/tool loop with a deterministic test model; a production local model has not been exercised in this environment.

Current Integration Boundaries

• Reviewer information and workload are manually configured, not live ERP data.
• Material submissions record ERP identifiers or file locations. There is no file upload, attachment retrieval or content inspection in this version.
• Change impacts are conservative deterministic review checklists. They do not perform financial recalculation or automatically rewrite Agent 2 reports.
• Source updates select a later existing Agent 1 run; they do not poll ERP.
• The briefing export is text, not the formal committee document template.
• This does not implement experience retrieval from Agent 2, reminder delivery, automatic assignment, or formal approval writeback.
• Sample people and tasks created by browser verification are explicitly labeled as examples. They are stored locally and can be used to try the workflow.

Verification

python -m unittest discover -s backend -p "test_coordination_agent.py"
python -m unittest discover -s backend -p "test_risk_agent*.py"
python -m unittest discover -s backend -p "test_report_agent.py"
python -m unittest discover -s backend -p "test_risk_core.py"
npx tsc --noEmit
node verify-coordination-agent.cjs

The browser script uses the running preview servers and creates labeled demo records. It checks project intake, people, assignment, material submission, plan changes and impacts, briefing download, and layouts at 1440, 1024 and 390 pixels.