dogukan

A multi-agent swarm coding CLI built on top of Claude Code. You give it a topic, it plans like a senior engineer, you trigger the loop, and a swarm of coding agents builds the plan in parallel — each in its own git worktree — while you talk to any one of them live.

It's the "type and talk" feel of Claude Code, but orchestrating many agents you can each converse with mid-task.

The flow

topic → plan → (review/refine) → /loop → swarm → talk to agents → /merge → done

1. Topic — you say what you want to build.
2. Plan — a planner agent produces a full plan (problem, actors, acceptance criteria, data model, API, sequence flow, architecture, slices, tests, risks) following the senior-engineering rules in src/prompts/senior-coding-agent.md.
3. Review — read it in the planner pane; chat with the planner to refine.
4. /loop — the plan is decomposed into independent slices; one worker agent per slice spawns in its own git worktree (parallel, no file clobbering).
5. Talk to any agent — Tab switches the focused agent (tmux-style sidebar + conversation pane). Type to steer the focused agent; Ctrl-C interrupts it; /broadcast messages every worker.
6. /merge — agent branches merge back into the base branch.

Architecture

• Engine: each agent is a Claude Code session via the Claude Agent SDK (@anthropic-ai/claude-agent-sdk), running locally with cwd set to its worktree. Streaming-input mode + interrupt() are what let you steer a running agent — that's the load-bearing capability the whole tool rests on.
• TUI: Ink (React for the terminal) — focus-switch layout.
• Isolation: one git worktree + branch per worker; sequential merge back.

Key modules: src/core/agent.ts (steerable session wrapper), orchestrator.ts (state machine + scheduler + bus), planner.ts / decompose.ts / worktrees.ts / store.ts, and src/ui/*.

Requirements

• Node 20+ and git.
• Claude Code authentication: an existing claude login or ANTHROPIC_API_KEY.

Install & run

npm install
npm run build
node dist/cli.js            # or: npm link  →  dogukan
# work in another repo:
node dist/cli.js --repo /path/to/your/project

Dev (no build step):

npm run dev

Keys & commands

• Tab / Shift-Tab — switch focused agent
• Ctrl-C — interrupt the focused agent (press twice to quit)
• /loop — decompose plan and start the swarm
• /merge — merge agent branches back
• /focus N, /broadcast <msg>, /pause [N], /resume [N], /agents, /help, /quit

Configuration

Optional .dogukan/config.json in the target repo (all fields have defaults):

{
  "model": "claude-opus-4-8",
  "workerModel": "claude-opus-4-8",
  "maxConcurrentAgents": 4,
  "permissionMode": "acceptEdits",
  "worktreeBase": ".dogukan/worktrees"
}

To cut cost while iterating, set "workerModel": "claude-sonnet-4-6" and "maxConcurrentAgents": 2.

Tests

npm test          # vitest — uses a mocked SDK (no tokens) + a temp git repo
npm run typecheck

Status / known simplifications

• A worker reaching a turn boundary counts as its slice "complete" for scheduling (frees a concurrency slot and unblocks dependents). You can keep steering it afterward; acceptance criteria aren't auto-verified yet.
• Merge conflicts are surfaced for manual resolution (no auto-resolver agent yet).
• Off-focus agents update status live; their full transcript is read when focused.