session-orchestrator Stop doing the work. Start leading it.
AI writes code one task at a time, while a person still plans, hands out and checks the work. This plugin for Claude Code, Anthropic's AI coding tool, makes one AI session the lead of a team that does it.
Designed and built by Marko Humar
Six jobs you never signed up for
It takes all six. You keep one job: deciding.
AI writes the code. Everything around it still lands on you.
- Planner You did: Split the feature by hand Now it does: Small, reviewable tasks, planned for you
- Dispatcher You did: Brief and chase every session Now it does: A lead session runs the team
- Model picker You did: Guess an AI model for every task Now it does: Strong models where the work is hard, cheaper ones where it isn't
- Reviewer You did: Read every giant change Now it does: Layered reviews before you look
- Memory You did: Re-explain the project, every time Now it does: Knowledge that grows each run
- Expert You did: Know a field you never studied Now it does: An expert view while it plans
A run, end to end
You set the goal, approve the plan and merge. The rest runs on its own.
Each finished task arrives as a pull request: one change, packaged for review.
-
You give a goal
"Add a refund flow."
-
It plans the work
Tasks, a model for each, an expert view
-
You approve the plan
Or ask for changes
-
Workers run in parallel
One AI session per task, tests first
-
One pull request per task
After its own fresh review
-
It reviews every one
Findings with severity and evidence
-
You merge
Grouped by risk
-
Wrap-up
Results, and what it learned
Your part
Layers, not one gate
Each layer catches some problems and misses others. Few get through all six.
Six layers of checks stand between a task and your main branch, the version of your code that goes live.
Tests first
Each piece is written against a test, and the full suite runs after it.
Fresh review of the change
A new reviewer reads it before it is offered to anyone.
The lead's review
Every finding names its severity, file and line, and is double-checked.
Specialist review
Hard tasks get one more review from the strongest model.
You merge, by risk
Risky changes come with the files worth reading.
Tests on the updated base
After merges, before waiting tasks start.
When work fails, it climbs a ladder
Cheap fixes first, the expensive model only when it's needed, and you decide before it goes further.
Reviews stop after three rounds, then it asks you.
-
A way to check the work
No test to run? It gets one, then tries again.
-
More effort
The same model, thinking longer.
-
A stronger model
From a quick model to a stronger one.
-
A second failure comes to you
A stronger model, split it, hand it to the lead, or drop it.
What it never does
It does the work and the bookkeeping. The decisions that matter stay with you.
- Merge a change. Every task ends in a pull request, and you merge it.
- Push to your main branch.
- Touch the files you are working in. Workers use separate copies.
- Send usage data anywhere.
- Overrule your project's own instructions.
- Store secrets.
It learns each project
Knowledge lives in your code and grows with every run.
It counts interruptions and rework, so each run asks less.
First run
It reads the code, asks up to five questions and drafts notes.
Every run after
Plans and reviews use them. Your team reviews each update like code, and your edits win.
Tested like code
With the plugin, every failure scenario scored 1.00. Without it, mostly zero.
Five scenarios where work goes wrong, three runs each, with and without the plugin. The score is the share of graded checks passed: 1.00 means all of them.
-
More effort after a failure
orch-effort-escalation -
A stronger model when you choose it
orch-retier -
Strongest model fails, so it thinks longer
orch-top-tier-escalation -
A worker passes a stuck piece to a stronger model
worker-native-escalate -
Reviews stop at three rounds
orch-review-round-cap
- Mean gain
- +0.68
- Behaviour test cases
- 63
- Test files
- 21
Try it
For the Claude Code desktop app or CLI. Needs git, gh and a GitHub remote. The code is open on GitHub.
claude plugin marketplace add donVortex/session-orchestrator
claude plugin install session-orchestrator@session-orchestrator --scope user
/session-orchestrator:orchestrate <goal>
Have an idea that needs building?
I designed and built session-orchestrator, from the first sketch to the tests. If you have an idea or a product to bring to life, I would like to hear about it.