CATMS

CATMS — Coding Agents Team Management System

A drop-in workflow template for teams building with multiple AI coding agents (Cursor, Claude Code) alongside humans.

**The problem:** AI agents are stateless. Every new session forgets everything. When two agents and a human work on the same codebase, work gets duplicated, context is lost, and nobody knows who did what.

**CATMS solves this** with a lightweight local workflow:

• Local Trackers (tasks.md, sprint.md, decision.md) are the single source of truth for tasks, ownership, state, and design decisions
• CAMS RAG is the cross-agent memory — a local, lightweight semantic memory layer running as an MCP server with pluggable embedding providers (OpenAI, Voyage, Ollama) and flat file storage (memory.ndjson)
• AGENTS.md is the single canonical rules file — read natively by Cursor, Google Antigravity, Codex, and Windsurf; CLAUDE.md imports it for Claude Code
• OpenSpec gives every feature a paper trail before any code is written
• docs/BRANCHING.md ties every branch to a task ID, with a CAMS-backed collision check before you cut one and a depends_on/blocks_on mechanism for real cross-task dependencies — single-stage (feat→main) or two-stage (feat→dev→main), chosen at catms init
• docs/SPRINT-WORKFLOW.md groups multi-task/multi-repo initiatives in sprint.md, with a single-active-sprint WIP limit
• Seven Claude Code skills (task-kickoff, sprint-planning, pr-checks-loop, sprint-close, project-adoption, session-sync, architecture-diagram) turn the docs above into runnable checklists — including a first-run pass that seeds the trackers from an existing codebase, a proactive end-of-session sync so nothing gets forgotten, and a grounded Mermaid architecture diagram saved into the repo

---

Multi-agent support

`AGENTS.md` is the canonical standard — Cursor and Google Antigravity read it live off disk every session, no setup required. Claude Code doesn't re