Course contentModule 1 · Lesson 1
What the Session Orchestrator is for and what you need
You see how the Session Orchestrator structures work in Claude Code, where the plugin stops, and what you need to get started.
MODULE 1 / GETTING STARTED
Note: Version and verification basis
Version: session-orchestrator 5.20.0 from November 14, 2026. The workflows were checked against this version's source code, not tried out in a fresh installation. If your version differs, the project's README and User Guide take precedence (repository with README and changelog).
A long conversation with Claude Code drifts
If you've worked on real tasks with Claude Code, you know how it goes. You describe a feature, the agent gets going, and after many messages the work is far from your first sentence. Usually one of these things is missing:
- A plan. Which files belong to the task, and what counts as done? At best that's scattered across the chat history.
- Checks between the steps. Tests, typecheck and lint run at the end or not at all. A mistake from step two only shows up in step eight.
- A state that stays. Close the window and the state is gone, and the next conversation starts from zero.
This is where the session-orchestrator plugin comes in. It's a free, open-source plugin for Claude Code under the MIT license. Claude Code stays your tool, and the plugin wraps a fixed process around your work.
The flow in five stations
You type three commands, and the plugin works in between. The course calls one run from /session to /close a session.
You need three more terms from here on. A wave is a round in which several agents work at the same time, each with its own list of files. A gate is a check decided solely by the exit code of a command, which is the number a program returns when it finishes. 0 means passed, any other value means failed. The Full Gate is the big check made up of typecheck, tests, lint and a search for leftover debug code.
The stations in detail:
/sessionreads the state of your repo and suggests a focus. Without your confirmation, no plan gets created.- The plan lists, per wave, the agents, their files and the acceptance criteria. It ends with the prompt to type
/go. /goruns the plan. It doesn't replan.- After the implementation waves, tests and typecheck run as a gate. After the Quality wave, the Full Gate runs.
/closeis the wrap-up. The plugin compares the result with the plan, runs the Full Gate once more, commits, and writes a final report with the decisions that are yours.
Where the plugin stops
- It needs no account and no server. The work runs on your machine.
- According to the README, it doesn't change code without your approval.
- It isn't a sandbox. The project docs describe the built-in safeguards, such as the per-wave list of allowed paths, as "an accident guard for a cooperating agent, not a containment boundary". That's protection against slips by an agent that plays along, and no security boundary. Treat your repo and your credentials as if this protection didn't exist.
- It's independent of Anthropic. The project describes its upkeep as "best-effort maintenance and no SLA", meaning it's maintained as time allows, with no promised availability.
- On the load it puts on your Claude plan, the docs say nothing. The plugin measures token totals in a log file, more on that in Lesson 5.
The plugin's skills also run in Codex, Cursor and Pi. There the hooks differ, meaning the scripts that run automatically on certain actions. The course sticks with Claude Code.
What you need to bring
- Node.js version 24 or higher.
- Git. You create the repository for the course in Lesson 2.
- Claude Code. The docs name no minimum version.
- Optional:
glaborgh, the command-line tools for GitLab and GitHub. The plugin uses them for issues and merge requests on the respective platform. The course doesn't set them up. Lesson 3 shows what happens without them.
An update check asks npm for new versions, and you can switch it off. According to the docs, telemetry is off until you agree. When you get asked is covered in Lesson 3.
What the course adds to the project docs
The project docs are an English reference. The course adds to them in four places:
- The docs explain commands and options. The course gives you a practice task with a result you can verify.
- The docs describe what each wave does. The course shows typical failures and what you do about them.
- The docs list every config key. The course names the three settings that matter on day one.
- The docs are organized as a reference. The course builds lesson by lesson, each on the one before.
In a few places the project's documents contradict each other, for example on the default session type. The course doesn't take a side there. That's why you always state the session type explicitly throughout the course, so nothing depends on a default value.
If a command answers differently than described here, you're probably on a newer version. The project's changelog lists what changed.
Check yourself
You can move on to Lesson 2 once all of this is true:
node --versionshowsv24or a higher major version.git --versionandclaude --versioneach answer with a version number.- You write down the five stations from the diagram from memory and mark the points where you decide. Then compare with the text above: you confirm the focus at
/session, you release the plan with/go, and you read the open decisions in the final report of/close.
Knowledge Check