conductor-skills
Conductor Skills
Teach your AI coding agent to create, run, monitor, and manage Conductor workflow orchestrations.
What You Get
Once installed, your AI agent can:
- Create workflow definitions with any task type (HTTP, SWITCH, FORK, WAIT, AI/LLM, MCP, etc.)
- Run workflows synchronously or asynchronously
- Monitor executions and search by status, time, or correlation ID
- Manage workflows — pause, resume, terminate, retry, restart, rerun, skip-task, jump
- Signal WAIT and HUMAN tasks for human-in-the-loop patterns
- Schedule workflows on cron — schedules are part of OSS
- Write workers in Python, JavaScript / TypeScript, Java, or Go (built-in patterns), or C#, Ruby, and Rust from each official SDK’s README
- Build agents — scaffold, test, deploy, and invoke Conductor Agents with the SDK (Python, TypeScript, Java, C#) or bring a LangChain / LangGraph / OpenAI Agents / Google ADK / Vercel AI / Claude Agent SDK agent; call them from workflows with the
AGENTtask, with human-in-the-loop, multi-agent, scheduling, and evals; call remote A2A agents - Visualize workflows as Mermaid diagrams
- Review & optimize existing workflows and agents against a checklist of 40 structure / reliability / performance / security / wrong-tool / agent rules
- Tell OSS from Orkes — a server that requires auth is Orkes (Enterprise), one that doesn’t is OSS; the skill checks before anything that differs between them
- Manage secrets and webhooks (Orkes)
- Configure AI providers on Orkes — create LLM / vector-DB integrations, register every model the provider key can use (discovered live from the provider), and grant access, so
LLM_*tasks work on an Orkes cluster - Manage prompt templates on Orkes — LLM prompts are saved as named templates, associated with the models that may use them, and referenced by name from tasks (inline prompts need
allowRawPrompts)
On Claude Code, five slash commands give you direct entry points:
| Command | Purpose |
|---|---|
/conductor |
Menu — lists subcommands and natural-language examples |
/conductor-setup |
First-time setup (CLI, server, auth) |
/conductor-optimize |
Review a workflow against the optimization checklist |
/conductor-scaffold-worker |
Generate a worker stub in your language |
/conductor-scaffold-agent |
Scaffold, test, deploy, and wire up a Conductor Agent (native SDK or a framework agent) |
Everything else (run, status, schedule, retry, signal, visualize, create, run/observe/approve/cancel agents) works through plain English — no command needed.
Quick Install
Claude Code Plugin (recommended for Claude users)
Install as a Claude Code plugin from the marketplace:
/plugin marketplace add conductor-oss/conductor-skills
/plugin install conductor@conductor-skills
Or test against a local checkout during development:
/plugin marketplace add /path/to/conductor-skills
/plugin install conductor@conductor-skills
Install via npm
If you have Node.js, you can install through npm — useful in CI and dev containers:
# One-off, no global install
npx @conductor-oss/conductor-skills --agent claude
npx @conductor-oss/conductor-skills --all
# Or globally
npm install -g @conductor-oss/conductor-skills
conductor-skills --agent cursor
conductor-skills --all --upgrade
conductor-skills --agent cline --uninstall
The npm package bundles all skill content, so it works offline once installed. For Claude Code, just restart your session after install to pick up the plugin.
Install for all detected agents
One command to auto-detect every supported agent on your system and install globally where possible. Re-run anytime — it only installs for newly detected agents.
macOS / Linux
curl -sSL https://conductor-oss.github.io/conductor-skills/install.sh | bash -s -- --all
Windows (PowerShell)
irm https://conductor-oss.github.io/conductor-skills/install.ps1 -OutFile install.ps1; .\install.ps1 -All
Windows (cmd)
powershell -c "irm https://conductor-oss.github.io/conductor-skills/install.ps1 -OutFile install.ps1; .\install.ps1 -All"
Install for a specific agent
These agents don’t support global install — run the command from your project directory.
macOS / Linux
| Agent | Command |
|---|---|
| Amazon Q | curl -sSL https://conductor-oss.github.io/conductor-skills/install.sh \| bash -s -- --agent amazonq |
Windows (PowerShell)
| Agent | Command |
|---|---|
| Amazon Q | irm https://conductor-oss.github.io/conductor-skills/install.ps1 -OutFile install.ps1; .\install.ps1 -Agent amazonq |
Windows (cmd)
| Agent | Command |
|---|---|
| Amazon Q | powershell -c "irm https://conductor-oss.github.io/conductor-skills/install.ps1 -OutFile install.ps1; .\install.ps1 -Agent amazonq" |
All other agents are installed globally via the install all command above. You can also use
--agent <name>for any agent to do a project-level install. GitHub Copilot and Amp share the same.agents/skills/directory as Codex/Gemini/Cursor/OpenCode — installing any one of them writes it.
That’s it — ask your agent to connect to your server (see Try It below).
Try It
After installing, try these prompts with your agent. Or run the slash command shown in brackets where one exists.
Setup & configure (or /conductor-setup)
- “Set up Conductor — I want to start a local server”
- “Connect to my Conductor server at https://developer.orkescloud.com/api”
- “Save my Conductor server config as a profile called production”
- “Switch to my staging Conductor profile”
- “How many workflows in dev vs prod?”
- “Is my server OSS or Orkes?”
Create & run
- “Create a workflow that calls the GitHub API to get open issues and sends a Slack notification”
- “Create a FORK_JOIN workflow that fetches inventory and pricing in parallel”
- “Run the my-workflow workflow with input {"userId": 123}”
- “Run order-processing synchronously and wait until the approval step”
- “Run cleanup-workflow with correlation id daily-2026-05-11”
Monitor & diagnose
- “Show me all failed workflow executions from the last hour”
- “What’s the status of execution abc-123?”
- “Why did wf-789 fail?”
- “Show me the running executions of order-processing”
Manage
- “Retry all failed executions of my-workflow from today”
- “Pause the running execution xyz-456”
- “Terminate wf-123, customer cancelled”
- “Restart wf-456 on the latest workflow version”
- “Skip the email step in wf-789”
- “Jump wf-789 to fulfill_order”
Human-in-the-loop
- “Signal the wait task in execution abc-123 with approval: true”
- “Reject the wait task in wf-456 — don’t retry”
- “What’s blocking wf-789?”
Schedule (part of OSS)
- “Schedule cleanup-workflow to run daily at 2am”
- “Schedule order-report every Monday at 9:30 weekdays only”
- “Pause the cleanup schedule”
- “Show the last 50 scheduled cleanup runs”
Modify
- “Add an error-handling branch to the order-processing workflow”
- “Add a WAIT task before the payment step in my checkout workflow”
- “Convert the inline JS in compute_pricing to a worker”
- “Extract the fulfillment chunk into a sub-workflow”
Review & optimize (or /conductor-optimize)
- “Review the order-processing workflow”
- “Optimize this workflow JSON — what should I fix?” (attach a file)
- “Audit weather-notification for missing timeouts and retries”
- “Make this workflow simpler”
- “Are there any secrets being passed through workflow input that shouldn’t be?”
Workers (or /conductor-scaffold-worker)
- “Write a Python worker that processes image thumbnails”
- “Write a TypeScript worker that validates email addresses”
- “Generate a Go worker that fetches data from a REST API and transforms the response”
- “Scaffold a Java worker for the charge_card task”
AI / LLM workflows
- “Create a workflow that summarizes text with Claude”
- “Build me an AI agent that lists MCP tools, picks one, and calls it” — the canonical first-agent flow (example)
- “Create a RAG workflow that searches my Pinecone index and answers questions with sources” (example)
- “Build an autonomous agent loop that runs up to 10 think/act/observe iterations” (example)
- “Add a HUMAN approval step before the agent calls any tool”
- “Create a workflow that classifies support tickets with an OpenAI model and routes to the right queue”
- “Index this document into the knowledge base” — using
LLM_INDEX_TEXT - “Generate an image from a prompt with OpenAI’s image model” — using
GENERATE_IMAGE
Conductor Agents (or /conductor-scaffold-agent)
- “Build a Python agent that looks up orders and issues refunds only after a human approves”
- “Scaffold a TypeScript support agent with a lookup_order tool and tests”
- “Port my LangGraph graph to run durably on Conductor” — bridges: OpenAI Agents, Google ADK, LangChain, LangGraph, Vercel AI, Claude Agent SDK, LangChain4j, LangGraph4j, Semantic Kernel
- “Write mock_run tests for support_agent that prove it never calls issue_refund without lookup_order”
- “Show me what support_agent compiles to before I deploy it”
- “Deploy support_agent and start serving its tools”
- “Add an AGENT task to order-processing that calls support_agent and pauses for a human if it asks a question”
- “Run planner and researcher in parallel from one workflow”
Agent operations (plain English)
- “Run support_agent with ‘where is order 123’ and stream it”
- “Show failed agent runs from the last hour”
- “What is agent execution abc-123 waiting on?” → “Approve it” / “Deny it — amount too high”
- “Cancel agent run abc-123, customer withdrew”
- “Schedule nightly_digest to run at 2am UTC with prompt ‘Summarize yesterday’”
- “Review support_agent for production readiness” (or
/conductor-optimize support_agent) - “Call the currency agent at https://agents.example.com from my workflow” — remote A2A
Visualize
- “Show me a diagram of the order-processing workflow”
- “Render the FORK_JOIN flow as a Mermaid chart”
Orkes only — secrets, webhooks, AI integrations (Orkes Conductor required)
- “Save STRIPE_KEY as a secret”
- “Create an OpenAI integration on our Orkes cluster with the key in OPENAI_API_KEY and register all its models”
- “Which AI integrations and models does our cluster have?”
- “Give the data-science group access to the openai-prod integration”
- “My LLM task works locally but fails on Orkes with ‘integration not found’ — fix it”
- “Save this system prompt as a template for openai-prod:gpt-4o-mini and use it in ticket-summary”
- “Why does Orkes say my prompt is not associated with integration openai-prod and model gpt-4o-mini?”
- “Create a GitHub webhook that triggers github_pr_handler”
- “List my webhooks”
Examples
Operational:
| Example | Description |
|---|---|
| Create and Run a Workflow | Define a workflow, register it, and execute it end-to-end |
| Monitor and Retry | Search executions, diagnose failures, and batch-retry |
| Signal a Wait Task | Human-in-the-loop with WAIT tasks and external signals |
| Review and Optimize | Apply the optimization checklist to an existing workflow |
Design patterns:
| Example | Description |
|---|---|
| FORK_JOIN Parallel Branches | Run independent tasks in parallel and converge with JOIN |
| DO_WHILE Loop | Iteration counter via the self-reference pattern |
| SUB_WORKFLOW Composition | Compose reusable child workflows under a parent |
AI / LLM patterns:
| Example | Description |
|---|---|
| Minimum LLM Workflow | Single LLM_CHAT_COMPLETE — building block |
| Multi-turn Chaining | OpenAI / Azure Responses API turns linked with previousResponseId — no message-history resend |
| AI Agent with MCP Tools | List tools → plan → call → summarize (the canonical first-AI-agent tutorial) |
| Autonomous Agent Loop | ReAct-pattern DO_WHILE loop until the LLM decides it’s done |
| RAG — Retrieval Augmented Generation | Vector search + grounded LLM answer with sources |
Conductor Agents:
| Example | Description |
|---|---|
| Build, deploy, and invoke an agent | SDK Agent + tools → run → deploy + serve → AGENT task → HITL → observe → schedule → cancel |
| Call a remote A2A agent | GET_AGENT_CARD → AGENT a2a → multi-turn resume; expose a workflow as an A2A server |
| Agent code samples | Native scaffolds and framework bridges in Python, TypeScript, Java, C# |
Raw JSON workflow definitions live in skills/conductor/examples/workflows/ — pass any directly to conductor workflow create.
References
| Reference | Description |
|---|---|
| Setup | Install the CLI, choose a server, configure auth, named profiles |
| CLI Index | Verb → CLI command lookup, grouped by lifecycle / intervention / tasks |
| Workflow Definition Schema | Full JSON schema, every task type, expression syntax |
| GraalJS Gotchas | JS-evaluated task pitfalls — Java-Map proxies, $.varName rule, scope rules, IIFE convention for DO_WHILE |
| Template Resolution | ${...} resolution pitfalls — missing-field-returns-parent, object → string toString, iteration paths |
| Writing Workers | Worker patterns for Python, JavaScript / TypeScript, Java, Go; official SDK repos for C#, Ruby, Rust |
| Conductor Agents | Decision table, lifecycle verbs, Agent fields, tools, HITL contract, operations, hosted agents, testing, gotchas |
| Agent SDKs | Per-language scaffolds and the framework-bridge matrix (LangChain, LangGraph, OpenAI Agents, ADK, Vercel AI, Claude Agent SDK, LangChain4j, Semantic Kernel) |
| API Reference | REST endpoints for direct API access |
| Visualization | Mermaid mappings for every Conductor construct + UI link |
| Schedules | Cron schedules (OSS) — schema, format, idempotency patterns |
| Optimization Checklist | 40 review rules across structure, reliability (incl. LLM-specific gotchas), performance, security, wrong-tool, and agents |
| Troubleshooting | Common errors, diagnosis flow, stuck-workflow recovery |
| Orkes Enterprise | Secrets, webhooks (Orkes Conductor only) |
| Orkes integrations | AI provider / vector DB integrations, models, access, prompt templates (Orkes Conductor only) |
| Fallback CLI | Bundled Python REST script — CLI equivalents when the CLI isn’t available, plus agent and Orkes-integration verbs the CLI lacks |
Evaluations
The evaluations/ directory holds 86 scenario evals: an LLM gets the skill as context, answers a user request, and a judge model scores the answer against each eval’s success criteria (an eval passes at ≥ 80% of its criteria). See evaluations/README.md for details.
export ANTHROPIC_API_KEY=... # agent + judge (OpenAI / Gemini models also supported)
python3 scripts/run_evals.py # all evals, 4 in parallel
python3 scripts/run_evals.py evaluations/setup-flow.json -v # one eval, verbose
python3 scripts/run_evals.py --retries 0 # strict: no re-sample of a failed eval
A failed eval is re-sampled once by default; evals that pass only on the retry are listed as flaky. CI runs the suite in 5 shards (.github/workflows/evals.yml) on pushes to main, weekly, on demand, and on PRs labeled run-evals.
Upgrade
Check for a newer version and upgrade all installed agents:
macOS / Linux
curl -sSL https://conductor-oss.github.io/conductor-skills/install.sh | bash -s -- --all --upgrade
Windows
irm https://conductor-oss.github.io/conductor-skills/install.ps1 -OutFile install.ps1; .\install.ps1 -All -Upgrade
Or upgrade a single agent: --agent <name> --upgrade
Supported Agents
| Agent | Flag | Global install | Project install |
|---|---|---|---|
| Claude Code | claude |
~/.claude/skills/conductor/ + plugin (via /plugin marketplace add + /plugin install) — also adds /conductor* slash commands |
— |
| Codex CLI | codex |
~/.agents/skills/conductor/ |
.agents/skills/conductor/ |
| Gemini CLI | gemini |
~/.agents/skills/conductor/ |
.agents/skills/conductor/ |
| Cursor | cursor |
~/.agents/skills/conductor/ |
.agents/skills/conductor/ |
| Windsurf | windsurf |
~/.codeium/windsurf/memories/global_rules.md |
.windsurf/skills/conductor/ |
| Cline | cline |
~/.cline/skills/conductor/ |
.cline/skills/conductor/ |
| GitHub Copilot | copilot |
~/.agents/skills/conductor/ |
.agents/skills/conductor/ |
| Aider | aider |
~/.conductor-skills/ + ~/.aider.conf.yml |
.conductor-skills/ + .aider.conf.yml |
| Amazon Q | amazonq |
— | .amazonq/rules/conductor.md |
| Roo Code | roo |
~/.roo/rules/conductor.md |
.roo/rules/conductor.md |
| Amp | amp |
~/.agents/skills/conductor/ |
.agents/skills/conductor/ |
| OpenCode | opencode |
~/.agents/skills/conductor/ |
.agents/skills/conductor/ |
Codex, Gemini, Cursor, OpenCode, GitHub Copilot, and Amp share one .agents/skills/ install — installing any one of them writes the shared directory; installing another just points at the same copy. Uninstalling only removes the shared directory once every agent using it has been uninstalled.
Uninstall
macOS / Linux
# Remove a global install
curl -sSL https://conductor-oss.github.io/conductor-skills/install.sh | bash -s -- --agent <name> --global --uninstall
# Remove a project-level install
curl -sSL https://conductor-oss.github.io/conductor-skills/install.sh | bash -s -- --agent <name> --uninstall
Windows (PowerShell)
# Remove a global install
irm https://conductor-oss.github.io/conductor-skills/install.ps1 -OutFile install.ps1; .\install.ps1 -Agent <name> -Global -Uninstall
# Remove a project-level install
irm https://conductor-oss.github.io/conductor-skills/install.ps1 -OutFile install.ps1; .\install.ps1 -Agent <name> -Uninstall
Windows (cmd)
:: Remove a global install
powershell -c "irm https://conductor-oss.github.io/conductor-skills/install.ps1 -OutFile install.ps1; .\install.ps1 -Agent <name> -Global -Uninstall"
:: Remove a project-level install
powershell -c "irm https://conductor-oss.github.io/conductor-skills/install.ps1 -OutFile install.ps1; .\install.ps1 -Agent <name> -Uninstall"
License
Apache 2.0 — see LICENSE.txt.
Built for Conductor OSS. Enterprise features powered by Orkes.