agent-orchestration
Claude Code plugins: session-hub (multi-repo master/child sessions) and model-router (model and effort routing with a quality loop)
Safety scan
A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.
Badge for your README
Markdown
[](/listings/mrbogomips--agent-orchestration)HTML
<a href="/listings/mrbogomips--agent-orchestration"><img src="/api/listings/mrbogomips--agent-orchestration/badge.svg" alt="findsafeskills safety rating" /></a>Install
Install instructions from the README
As written by the repo's author. findsafeskills has not verified these steps.
Both plugins are listed in the `mrbogomips` marketplace:
/plugin marketplace add MrBogomips/claude-plugins /plugin install session-hub@mrbogomips /plugin install model-router@mrbogomips
Requirements: `node` on `PATH`, and for `session-hub` Claude Code 2.1.224 or later (cross-session messages).
The skills and the classic hooks work with mods turned off. The mods need Claude Code 2.1.329 or later with mods turned on: in `session-hub` the send guard, the shell check, message marking and the children panel; in `model-router` routing, measurement, the per-Agent lines and the usage view.
What was scanned
The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.
README.md · README · 119 lines
- # agent-orchestration
- **Run many Claude Code sessions across many repositories, and run each piece of work on the right model.**
- Two plugins for teams that have outgrown one session in one repository:
- - **[`session-hub`](session-hub/)**: one master session coordinates child sessions, one per repository, each
- limited to its own scope.
- - **[`model-router`](model-router/)**: every subagent gets the model and effort its task needs, and you can see
- on your own machine whether those choices worked.
- Each plugin works on its own and does more alongside the other.
- 🌐 **[Documentation site](https://mrbogomips.github.io/agent-orchestration/)**, in English and Italian: concept, scenarios,
- quickstart and architecture.
- 📘 **[Read the user manual (PDF)](docs/manual/user-manual.pdf)**: every feature, with examples and screenshots.
- ---
- ## session-hub: one master, many repositories, clear limits
- Your product spans a backend, a frontend, a design system and the docs. A single agent working across all of them
- skips each repository's own instructions, skills and agents. Several agents working side by side step on each
- other's files.
- `session-hub` turns a folder of repositories into a **hub**. A master session starts one child session per
- repository, and each child works inside that repository with its rules loaded. Every child knows its task, what
- it may change, which model it uses, and that it reports only to the master.
- 
- **What you get**
- - **Children with a real scope.** A hook blocks file edits outside the child's write root (its repository or
- worktree), resolving links and `..` first. When the guard cannot decide, it blocks and says why: it fails
- closed.
- - **Requests go up, not sideways.** A child that needs a change elsewhere sends an `out-of-scope` message to the
- master instead of editing another repository.
- - **A record and a panel.** The master keeps a record of every child, stored outside the hub, and a live panel
- shows each child's state, model and last message.
- - **Shell commands are checked too.** With mods on, a best-effort check also blocks redirections and common
- file-writing commands that target a path outside the write root, through `cd` and wrappers.
- - **Messages are not orders.** Messages from other sessions, the master's included, are marked as information,
- not authorization, so a child acts on them only within its scope.
- - **Works where you work:** Herdr, tmux, `claude --bg`, plain commands to paste, or the desktop app (where the
- scope rests on the handoff alone).
- 
- ## model-router: the right model for each task, and a measure of how well it worked
- Left to itself, an agent runs every subagent on its own model. Simple searches run on the most expensive model,
- while hard reviews sometimes run on a cheap one and come back for rework.
- `model-router` gives the agent **one short grid** for choosing a model and an effort for each task. The agent
- passes both on every subagent start, and you see each choice the moment it is made:
- 
- When a built-in subagent start (general-purpose or Explore) still carries no model, a fast judge model
- classifies the task and sets one. Explicit choices,
- forks and agents with a fixed model are left alone. Any error leaves the request unchanged: the router fails open.
- **Then it measures.** For each kind of task and each model and effort, `model-router` tracks:
- - **First-pass yield**: how often the work was right the first time;
- - **rework**: how often it had to be redone;
- - **cost per success**: what a task that worked actually cost.
- `/model-router` opens a usage view modelled on Claude Code's own `/usage`, with Overview, Models, Classes and
- Decisions tabs, a choice of period, and user or project level:
- 
- One key writes a self-contained quality report with the trend week by week:
- 
- **Your measurements stay on your machine**, under your Claude Code config folder, and nothing is written inside
- your repositories. The judge's one call goes through your session's own client, like any subagent. A
- project can record its own exceptions, such as "always use opus for migrations", in its `AGENTS.md`.
- *Coming in 0.4: the router uses its own quality data to adjust its choices.*
- > The screenshots come from the manual and show demonstration data.
- ---
- ## Install
- Both plugins are listed in the `mrbogomips` marketplace:
- ```text
- /plugin marketplace add MrBogomips/claude-plugins
- /plugin install session-hub@mrbogomips
- /plugin install model-router@mrbogomips
- ```
- Requirements: `node` on `PATH`, and for `session-hub` Claude Code 2.1.224 or later (cross-session messages).
- The skills and the classic hooks work with mods turned off. The mods need Claude Code 2.1.329 or later with mods
- turned on: in `session-hub` the send guard, the shell check, message marking and the children panel; in
- `model-router` routing, measurement, the per-Agent lines and the usage view.
- ## Documentation
- | Document | For |
- |---|---|
- | [Documentation site](https://mrbogomips.github.io/agent-orchestration/) | The plugins explained: concept, ways of working, quickstart, model-router in depth, architecture. Source: `site/`, see `site/README.md`. |
- | [User manual (PDF)](docs/manual/user-manual.pdf) | Using both plugins: features, commands, examples, screenshots. Source: `docs/manual/manual.html`, built with `docs/manual/build.sh`. |
- | [`session-hub` README](session-hub/README.md), [`model-router` README](model-router/README.md) | Each plugin's components, commands and options. |
- | [Architecture](docs/architecture.md) | How both plugins work, as built, and what is planned. |
- | [Design](docs/design/) | The design of both plugins and the tests run before building them. |
- | [Research](docs/research/) | The research behind the design. |
- ## License
- MIT, see [LICENSE](LICENSE).
Ratings and comments
Sign in to rate this skill.
Sign inFree to use. If it’s helped, pass it on.