dkoddkod

Agent Skill

Install the dkod skill to unlock fearless parallel execution across your AI agents — no merge conflicts, no worktrees, no waiting.

Claude Code users: The dkod plugin bundles the skill, MCP tools, agents, and commands into a single install. Use the plugin for the best experience.

Install

One command. Works with Claude Code, Cursor, Codex, Windsurf, Cline, and any skills-compatible agent.

npx skills add dkod-io/skills

Note: The skill is published to the dkod-io/skills GitHub repository. The install command clones the skill files (SKILL.md, references/, evals/) from that repo into your project.

After installation, your agent gains the /dkod command and automatically activates parallel execution when dkod MCP tools are available.

Not a Worktree. Not a Fork. Not a Clone.

This is the dkod Agent Protocol in action.

When agents work in parallel today, the standard approach is to create Git worktrees, forks, or full clones — one per agent. Each agent gets its own copy of the entire repository, works on its own branch, and then someone has to merge everything back together, resolving the inevitable text-level conflicts along the way. It's slow, fragile, and doesn't scale.

dkod replaces all of that. Instead of copying the repo N times for N agents, dkod gives each agent an isolated session overlay — a lightweight, copy-on-write layer on top of a single shared codebase. Agents read and write through their overlay. Changes are invisible to other agents until submitted. When an agent submits, dkod's semantic merge engine analyzes the changes at the AST (Abstract Syntax Tree) level — it understands functions, classes, types, and their relationships. Two agents editing different functions in the same file? That's not a conflict — it's an automatic merge in under 50ms. No branches, no worktrees, no stashing, no multi-step merge ceremony.

The result: one codebase, unlimited agents, zero false conflicts.

What It Does

The dkod skill teaches your AI agent a new default: parallelize everything.

Today, agents serialize work to avoid conflicts. If two tasks touch the same file, one waits for the other. This is because Git sees code as text — two edits to the same file means a merge conflict, even if the changes are completely independent.

With dkod installed, the agent understands that:

  • Multiple sub-agents can edit the same file simultaneously — dkod merges at the AST level, not the text level
  • Session isolation means zero interference — each agent gets a copy-on-write overlay, invisible to others
  • Auto-rebase handles moving targets — if main changes while an agent works, dkod rebases automatically
  • True conflicts are caught and surfaced — with full semantic context, not cryptic diff markers

Usage

Invoke the skill directly with /dkod followed by your task:

/dkod refactor the auth module — rename authenticate_user to
verifyCredentials everywhere, and add rate limiting to the login endpoint

The agent will decompose the task by symbol, launch parallel sub-agents, and let dkod handle the merging.

Or just describe parallel work naturally — the skill activates automatically when it detects independent tasks and dkod MCP tools:

Add input validation to all 12 API endpoints in src/api/.
Do them all in parallel.

What's Safe in Parallel

ScenarioResult
Two agents edit different functions in the same fileAuto-merge
Two agents add different fields to the same structAuto-merge
Two agents add the same importDeduplicated
Two agents modify different sections of a functionAuto-merge
Two agents modify the same function bodyConflict (surfaced with context)
Agent A deletes a function that Agent B callsConflict (caught at merge time)

dkod catches more real conflicts than Git — and eliminates the false positives that waste your time.

How It Works

Under the hood, the skill instructs your agent to:

  1. Decompose by symbol, not file — think functions and modules, not file paths
  2. Launch sub-agents concurrently — each connects its own dkod session via dk_connect
  3. Let each agent work independently — read/write through isolated overlays
  4. Submit and merge — each agent submits via dk_submit, verifies via dk_verify, and merges via dk_merge

The platform's AST-level semantic merging reconciles all changes automatically. Only true semantic conflicts (two agents modifying the same function body in incompatible ways) require intervention.

Prerequisites

The skill requires the dkod MCP server. If you haven't set it up yet:

Cloud (dkod.io)

claude mcp add --transport http dkod https://api.dkod.io/mcp

Self-Hosted

claude mcp add --transport http dkod https://your-instance.example.com/mcp

See the Quickstart for full setup instructions.

Skill Contents

The skill follows the open Agent Skills format:

dkod/
├── SKILL.md              # Behavioral instructions for the agent
├── references/
│   └── mcp-workflow.md   # Full MCP protocol reference (loaded on demand)
└── evals/
    └── evals.json        # Test cases for skill validation
  • SKILL.md — The core behavioral shift: parallelize fearlessly, decompose by symbol, launch concurrent agents
  • references/mcp-workflow.md — Complete protocol reference (connect → context → read/write → submit → verify → merge) loaded only when the agent needs deeper protocol details

Supported Agents

The skill installs on any agent that supports the Agent Skills format:

Next Steps