A five-role fleet for long-running coding work in Cursor: planner, designer, tech lead, developer,
tester. The orchestrator is a Cursor skill; role work runs through subagents in .cursor/agents/.
Durable state lives in fleet/, so a run survives across sessions.
| Role | Invocation | Writes |
|---|---|---|
| Orchestrator | skill @fleet-orchestrator |
fleet/STATUS.md |
| Planner | agent /fleet-planner |
fleet/PLAN.md, draft fleet/TASKS.md |
| Designer | agent /fleet-designer |
fleet/DESIGN.md |
| Tech lead | agent /fleet-tech-lead |
fleet/ARCHITECTURE.md, refined fleet/TASKS.md |
| Developer | agent /fleet-developer |
code, checks off fleet/TASKS.md |
| Tester | agent /fleet-tester |
fleet/TEST_REPORT.md, follow-up tasks |
Phase order is plan → design → architecture → developer → test → done. Planner, designer,
and tech lead stop at an approval gate before the fleet advances.
Open this repo in Cursor, then in an Agent chat:
@fleet-orchestrator Start a new epic: <what you want built>
The orchestrator captures the request in fleet/REQUEST.md, sets fleet/STATUS.md to phase: plan,
and runs the roles in order. Answer each gate with approve or revise: <note>.
In a new chat, nothing else is needed:
@fleet-orchestrator Continue from fleet/STATUS.md
The orchestrator reads the status board and picks up at the recorded task. This is the whole point of the artifact folder: chat memory ends, files do not.
When you want manual control over one phase:
/fleet-developer Implement TASK-003 only
/fleet-tester Run a test pass on the completed tasks
Copy .cursor/skills/fleet-orchestrator/, .cursor/agents/fleet-*.md (five role agents), and the
fleet/ folder into the target repo. For global use, copy the skill to ~/.cursor/skills/ and the
agent files to ~/.cursor/agents/. Keep paths identical; everything references fleet/ by path.
The fleet was dry-run on a small end-to-end epic in this repo: a stdlib Python CLI that validates
fleet/STATUS.md and prints the next phase. It is small enough that every role produced real output,
which is what exposed the handoff gaps now fixed in the agents. See fleet/ for the artifacts from
that run.