Guide
The hands-on path. Concepts page is the what; this is the how.
Getting started
1. Install & first run : Clone, install, run the canary hello-tool example, verify your environment.
2. Your first experience : Build a 2-tool flow from scratch with state.schema declaration. No LLM yet.
3. Run with an LLM : Add an agent node with structured output. AJV schemas. Env-var setup. Default models per provider.
4. The TUI dashboard : Anatomy, status glyphs, activity strings per dispatcher.
Authoring
Hand-writing experience.yaml : The full YAML shape, from state.schema through graph.nodes to edges.
oe ultra — LLM authors for you : One keyword, full experience synthesized into a validated draft directory. Now with a critique→revise quality loop and --run smoke test.
From inside Claude Code : Run /oe-authoring (or invoke the oe-mcp server) without leaving your editor.
Editor support (autocomplete + validation) : Wire the experience.yaml JSON Schema to VS Code / any yaml-language-server editor for autocomplete, hover docs, and inline validation.
Tool stubs in .mjs : The default-export signature, state_delta vs edge_output, error handling.
Prompt files : Markdown prompts with interpolation. v-pre gotchas. System vs user.
Running
Concurrency + 429 retry : runtime.concurrency vs for_each.concurrency. Per-provider 429 retry with backoff.
Resume + cache : How oe resume works. When the content-hash key hits. When it doesn't.
Error policies (on_error) : fail (default) vs skip vs retry. Cascade behavior. When to use which.
Self-hosted LLMs (vLLM, Ollama, LM Studio) : Point OPENAI_BASE_URL at your local server. The default-model override pattern.
Visualize & report : Render the DAG with oe graph (Mermaid, paste into READMEs) and produce an HTML run report with oe inspect --html.
Evolving
The advisor : What EvolutionAdvisor.analyze() reads, what it proposes, confidence levels. Plus oe evolve --runs a,b,c cross-run analysis (stable patterns vs one-off blips).
Applying proposals : Extracting the embedded diff. git apply workflow. Handling drift.
Author → run → evolve loop : The full lifecycle with one LLM provider driving all three stages.
Integration
MCP server (use OE from Claude Code) : oe-mcp exposes 8 tools so Claude Desktop / Cursor can invoke OE primitives.
cli-agent node (call CLIs from OE) : Outbound: have an OE node spawn a Claude Code / Codex / Gemini session.
Skills + SKILL.md : Package reusable LLM capabilities as redistributable SKILL.md packages.
→ Stuck? Hit FAQ + troubleshooting. → Pattern-shopping? Browse the cookbook. → Want concepts first? Start at Concepts.