buddyx

BuddyX (buddyx)

A cross-harness Agent Skill that turns your AI coding assistant into BuddyX, a master orchestrator. Every command runs a disciplined loop — plan → route → delegate → verify → report: it maps your /buddyx request to an action, routes it to the best-matching specialized sub-agent in your project (invoked as a real sub-agent when the harness supports it), fans independent work out in parallel, gates the result through a review/verify step, and closes with a consistent report. It creates a new sub-agent — with a proper frontmatter contract and least-privilege tools — when none fits. Distributed as an npm CLI that installs the skill into Claude Code, opencode, and other assistants.

v2 is a ground-up rebuild: real sub-agent delegation (not inline role-play), parallel fan-out, verification gates, planning + status, a security posture, and a thin SKILL.md router over an on-demand references/ corpus. See What's inside.

Install method mirrors ganesh-gyan-cli: a small CLI (buddyx) copies the bundled skill into each assistant's skills folder.

Install

npm install -g buddyx

Then add the skill to your assistant (run inside a project for project scope, or add --global for all projects):

buddyx init --ai claude          # Claude Code   → .claude/skills/buddyx/
buddyx init --ai opencode        # opencode      → .opencode/skills/buddyx/
buddyx init --ai claude --global # → ~/.claude/skills/buddyx/ (all projects)
buddyx init --ai all             # every supported assistant
buddyx init                      # auto-detect + interactive picker

Or without a global install:

npx buddyx init --ai claude

Restart the assistant, then invoke it by typing /buddyx or just asking, e.g. "buddyx create payment-gateway module". The skill's description also lets the model trigger it automatically on matching orchestration requests.

New project? Run /buddyx bootstrap first — BuddyX scans your stack, proposes a tailored set of specialized sub-agents (backend, frontend, db, testing, reviewer…), and on your approval scaffolds them, so every later command has specialists to route to. (This is the in-assistant command; don't confuse it with the terminal buddyx init, which installs the skill.)

Supported assistants

claude (Claude Code) · opencode · cursor · windsurf · codex · gemini · antigravity — or all.

Each install is self-contained: SKILL.md plus the full references/ corpus is copied into <assistant-dir>/skills/buddyx/.

What BuddyX does

BuddyX is an orchestrator, not a stack. It ships no sub-agents of its own — instead it discovers the specialized agents already in your project (.claude/agents/, .opencode/agent/, etc.), routes to the best match on their description, and creates a new agent (to a frontmatter contract) when none fits.

Every command runs the same loop:

PARSE → PLAN → ROUTE → DELEGATE → VERIFY → REPORT

• Real delegation — invokes matching sub-agents via the harness's Agent tool (subagent_type); falls back to inline role-play only when the harness can't spawn them.
• Parallel fan-out — independent subtasks run concurrently; dependent ones are sequenced.
• Verify gate — non-trivial output is routed through a read-only reviewer/tester before it's called done (adversarial multi-reviewer for risky changes).
• Planning + status — multi-step work gets an explicit plan; status reads from it.
• Safety posture — confirmation gates for destructive/outward-facing ops, least-privilege tools on created agents, faithful reporting.

Commands

/buddyx <action> [target] [options]

Command	Description
bootstrap	Scan the project and scaffold a tailored agent fleet (propose → approve → generate)
create <target> <type>	Create a module, component, or agent
analyze <target>	Analyze code, structure, or requirements
refactor <target>	Refactor existing code
test <target>	Create or run tests
docs <target>	Generate documentation
review <target>	Review code for issues
agent create <name>	Create a new specialized agent
agent list	List all available agents
status	Check project status
help	Show available commands

What's inside the skill

A thin SKILL.md router over an on-demand references/ corpus (loaded only when the task needs it):

• delegation-protocol.md — real sub-agent invocation, parallel vs sequential, recovery, degradation
• routing.md — capability index, task→agent matching, tie-breaks, fallback ladder
• agent-spec.md — the frontmatter contract, agent create procedure, quality bar
• report-format.md — the standard result block
• safety.md — confirmation gates, least-privilege, secrets, faithful reporting
• workflows/ — per-command playbooks (bootstrap · create · refactor · review · analyze · test · docs)
• checklists/done-criteria.md — acceptance criteria that define "done"

Update

npm install -g buddyx@latest
buddyx update --ai claude       # re-copy the latest bundled skill over an existing install

Uninstall

buddyx uninstall --ai claude
buddyx uninstall --global
buddyx uninstall               # detect installs in the current project and confirm

CLI reference

Command	Description
buddyx init [--ai <type>] [--global] [--force]	Install the skill
buddyx update [--ai <type>] [--global]	Re-copy the bundled skill over an existing install
buddyx uninstall [--ai <type>] [--global]	Remove the skill
buddyx versions	Print the CLI version

License

MIT Chintan Bagdawala