Orchestrator Plugins

A private Claude Code plugin marketplace for the orchestrator multi-agent development ecosystem.

Plugins

Plugin	Description	Requires
orchestrator-core	6 agents, 2 skills, dev-tools MCP server (write_findings pipeline contract), full hook suite (SessionStart, SubagentStop JSON-validated blocking, PostToolUse auto-context, PreCompact snapshot, SessionEnd audit)	uv
ty-lsp	Python LSP via Astral ty	uv tool install ty
vtsls-lsp	TypeScript/JavaScript LSP via vtsls	npm install -g @vtsls/language-server

---

Requirements

MCP dev-tools server (orchestrator-core)

The dev-tools MCP server is a Python process launched automatically when the plugin is active. Requires either:

Runtime	Notes
uv	Preferred — dependencies resolved automatically
Python 3.11+	Fallback — install manually: pip install fastmcp

---

Installation

1. Add this marketplace to Claude Code

/plugin marketplace add rafa1off/orchestrator

Or manually in your project's .claude/settings.json or user settings (~/.claude/settings.json):

{
  "extraKnownMarketplaces": {
    "orchestrator": {
      "source": {
        "source": "github",
        "repo": "rafa1off/orchestrator"
      }
    }
  }
}

2. Install plugins

# Core orchestrator (required)
claude plugin install orchestrator-core@orchestrator

# Python LSP — install for Python projects
claude plugin install ty-lsp@orchestrator

# TypeScript/JavaScript LSP — install for TS/JS projects
claude plugin install vtsls-lsp@orchestrator

Other language servers

For Go, Rust, Java, and other languages, check the official Claude Code plugin marketplace:

/plugin marketplace browse

Official LSP plugins (gopls, rust-analyzer, etc.) are maintained there and install without any additional configuration.

3. Activate in your project

Add one line to your project's CLAUDE.md:

Skill("/orchestrator-core:orchestrator")

That's it. No other setup required.

---

orchestrator-core

A complete multi-agent development harness for Claude Code. The orchestrator session acts as a coordinator, calling specialized subagents on demand.

Agents

Agent	Model	Type	Role
orchestrator-core:reader	haiku	readonly	Maps code paths, returns structured context snapshots
orchestrator-core:researcher	sonnet	readonly	Finds external patterns, library APIs, prior project decisions
orchestrator-core:thinker	sonnet	readonly	Deep reasoning, tradeoff analysis, brainstorming; isolates verbose analysis from main context
orchestrator-core:writer	sonnet	read+write	Produces minimal, focused code changes from a context block
orchestrator-core:verify	sonnet	readonly	Runs lint + typecheck + diff review in one pass; writes verify-findings.json; background: true
orchestrator-core:tester	sonnet	read+write	Identifies missing tests, writes them, runs the suite

Skills

Skill	When to use
/orchestrator-core:orchestrator	Load at every session start — the agent routing guide; L1/L2/L3 dispatch levels, 3 invariants, verify loop
/orchestrator-core:orchestrator-plan	Before any multi-step task — writes a plan to .claude/plans/, then dispatches directly after approval

Hooks

Event	Trigger	Behavior
SessionStart	Session begins or resumes	Clears stale verify-findings.json and snapshot from .claude/pipeline/
SubagentStop (verify)	verify agent finishes	Blocks (exit 2) if the agent stopped without a valid verify-findings.json — validates with jq when available; forces re-invocation of write_findings
PostToolUse (write_findings)	Findings file written	Reads the file and injects its content as additionalContext — orchestrator receives results automatically
PreCompact	Context compaction begins	Snapshots verify-findings.json and progress.md to .claude/pipeline/pre-compact-snapshot.md so state survives compaction
SessionEnd	Session terminates	Appends an entry to .claude/pipeline/session-log.txt and removes stale findings

Dev-tools MCP server

Exposes a single tool used by verify to write structured findings to the pipeline:

Tool	Description
write_findings(source, status, pipeline?, errors?, issues?)	Writes verify-findings.json to .claude/pipeline/ (or a per-track subdirectory for parallel runs)

Verify runs lint, typecheck, and diff review directly via Bash, reading the project's commands from CLAUDE.md first and falling back to marker-file detection (uv.lock → ruff/mypy, package.json → eslint/tsc, etc.).

The 3 Invariants

The orchestrator enforces these rules regardless of task or scale:

1. Read before write — invoke reader (or read files directly for trivial changes) before calling writer
2. Verify after write, max 2 rounds — run verify + tester after a meaningful write phase, always together in the same message turn; escalate after 2 rounds with remaining findings
3. One writer per overlapping file set — serialize writers sharing files; disjoint sets may run in parallel

---

ty-lsp

Python LSP via Astral ty. Provides go-to-definition, find-references, hover, and document-symbol on .py files.

Prerequisite: uv tool install ty

---

vtsls-lsp

TypeScript and JavaScript LSP via vtsls. Provides go-to-definition, find-references, hover, and document-symbol on .ts, .tsx, .js, and .jsx files.

Prerequisite: npm install -g @vtsls/language-server

---

Contributing

Each plugin lives in plugins/<name>/. The marketplace manifest is at .claude-plugin/marketplace.json.

To add a new plugin:

1. Create plugins/<name>/ with a .claude-plugin/plugin.json manifest
2. Add the entry to .claude-plugin/marketplace.json
3. Bump the version in both files

Versioning follows semver. MAJOR for breaking protocol changes, MINOR for new agents/skills/tools, PATCH for fixes and instruction improvements.

---

License

Apache 2.0 — see LICENSE.