7342a9a394
Replaced broken slot-based mechanism with session-counting: - Added count_active_dev_sessions() function that counts actual session files in ~/.kugetsu/sessions/, excluding base.json and pm-agent.json - Modified cmd_start() to check session count before creating new session: - If count >= MAX_CONCURRENT_AGENTS, reject with error - Otherwise allow new session creation - Removed wait since --fork returns immediately - cmd_continue() no longer counts toward limit (existing sessions can always continue via --continue) This properly enforces MAX_CONCURRENT_AGENTS while preserving --fork functionality. The slot mechanism didn't work because opencode run --fork returns immediately after forking, not after child completes.
Kugetsu
Name background: Kugetsu (月掴, "grasping the moon") is derived from Jujutsu Kaisen's "Tokusa no Kage Boujutsu" (Shadow Art Style) — a technique that summons up to ten different creatures from the user's shadow. This project embodies the concept of one orchestrator managing multiple specialized agents working in parallel.
Overview
Kugetsu is an agent orchestration system that enables parallel task execution across multiple repositories. Inspired by the IT department metaphor:
- Human acts as executive, reviewing and approving
- PM (Project Manager) Agent coordinates and delegates tasks
- Coding Agents execute tasks autonomously on assigned issues
The core idea: instead of working through issues one-by-one, a PM spawns multiple coding agents in parallel — similar to Hermes running multiple tasks, but scaled to a full team workflow.
Why
When you have 10 issues, typically you work through them sequentially. With Kugetsu:
- PM prioritizes and splits tasks
- Coding agents work in parallel on their own branches
- PM reviews and merges to a release branch
- Human provides final approval to master/main
This means your focus shifts from doing to overseeing — reviewing PRs, not writing code.
Status
Phase 3: Chat Integration (Implemented)
- PM Agent with git worktree isolation per session
- Chat Agent via Telegram gateway
- Parallel capacity testing tool available
See Architecture for full system design and phase status.
Capacity Planning
Based on parallel capacity testing (tools/parallel-capacity-test/):
| Resource | Value |
|---|---|
| Memory per agent | ~340 MB |
| Recommended max agents | 5 |
| Timeout threshold | 8+ agents |
| Memory limit | 1 GB per agent (configurable) |
Observed Behavior
- 1-5 agents: 100% success rate, ~6-9s avg response time
- 8+ agents: Timeouts occur due to resource contention
- Scaling is roughly linear up to 5 agents
Recommendations
- Limit max parallel agents to 5 for stable operation
- Monitor memory usage when scaling beyond 3 agents
- Configure memory limit via
--memory-limitflag based on available RAM
Documentation
- Architecture — Full system design
- Research Index — All research topics
License
MIT