Skip to content
exit-nodePublic

About

Multi-agent coding loop for omp: scouts, planner, isolated implementers, cross-model review and a model council.

Topics

Resources

Stars

16 stars

Watchers

0 watching

Forks

Latest commit

 

History

29 Commits

Folders and files

Repository files navigation

node-loop

node-loop is a plugin for omp, a terminal coding agent. You give it a goal, and it runs a loop of agents that plan, build and review the change in your repo. Scouts map the repo, and a planner cuts the work into small plan nodes. Each plan node is built in its own isolated checkout and reviewed by a different model until the review passes. When every node has merged, a council of three models reviews the whole branch.

Before you start

  • omp 18.11.1 or newer. Full runs were tested on 18.12.2; 18.11.1 was only checked for loading the extension and reading its settings.
  • git, and a repo with at least one commit and a clean working tree.
  • bash. macOS and Linux work; Windows is not supported.
  • Model access. At least one omp provider you are signed in to. Two providers give cross-provider review; with one, review is same-provider.

Quick start

In omp, run:

/marketplace add exit-node/node-loop
/marketplace install node-loop@exit-node

Restart omp, then run:

exitnode setup
exitnode add magic-link login to the API

exitnode setup picks a model for each role and writes the global settings listed under "What setup changes".

Type What it does
exitnode <goal> Start a run in the current repo
exitnode resume Continue the latest run after a restart
exitnode setup Pick a model for each role and turn on isolation (rerun any time)
exitnode view Open the live forge view in your browser

exitnode only triggers as the first word of a prompt; /exitnode ... works too.

How it works

flowchart LR
  G[goal] --> S[scouts] --> P[planner]
  P --> I[implementer per plan node, isolated] --> F{fence check}
  F -- outside fence --> I
  F --> R[reviewer, other model]
  R -- fail, max 3 kickbacks --> I
  R -- pass --> M[merge to run branch]
  M --> C[council: 3 seats + chair] --> D[report]
  C -- verified blockers --> I
Loading
  • Plan node. One change small enough for one agent to build and one reviewer to pass or fail from the diff.
  • Wave. Plan nodes whose dependencies have all passed; they run in parallel.
  • Fence. The files a plan node may touch. Changes outside it fail automatically, and good ideas outside it go to parking.md.
  • No approval wait. The plan is posted for you to read and the run keeps going; interrupt any time to change it.
  • Fresh agents. A failed review sends the work to a new implementer that reads the findings, not the old conversation.
  • Council. Three models review the branch blind. The chair checks every finding against the code; one verified blocker is enough, with no voting.
  • Isolated work. Each implementer works on its own branch, and nothing is pushed.

Cost and time

These numbers come from three small test runs on one Mac, using Anthropic and OpenAI Codex, with a todo-list goal. They are examples, not promises; your numbers depend on goal size, repo size and providers.

Run Plan nodes Wall time Cached input tokens Output tokens
A 2, no kickback 16 min about 4.1 M about 66 K
B 1, one review kickback 9 min about 3.3 M about 49 K
C 1, one review kickback 10 min about 3.7 M about 48 K

Every agent counts: scouts, planner, implementers, reviewers and the council. Measure your first run before you budget for more.

What a run changes in your repo

  • Your checkout switches to a node-loop/<run-id> branch. Go back with git switch <your-branch>.
  • Each implementer works on an omp/task/* branch. The final report lists the git branch -D commands to clean them up.
  • Run state lives in ~/.omp/node-loop/ (set NODE_LOOP_HOME to move it).
  • Nothing is pushed.

Safety

Agents run shell commands with your permissions, and the acceptance commands the planner writes run unattended. Commit anything you cannot lose before a run. Read plan.md first when the goal touches deployment, credentials or data.

What setup changes

exitnode setup writes to your global omp config, ~/.omp/agent/config.yml:

  • modelRoles: one entry for each of the seven node-loop roles (node-* and council-seat-*).
  • task.isolation.enabled: true, task.isolation.merge: branch, task.isolation.apply: false.

The isolation settings apply to every omp subagent task in every session, not only node-loop runs. To go back to omp's defaults:

omp config reset task.isolation.enabled
omp config reset task.isolation.merge
omp config reset task.isolation.apply

Then delete the node-* and council-seat-* lines under modelRoles in ~/.omp/agent/config.yml.

Models

Each agent uses an omp model role, so you choose the models with exitnode setup. Setup offers only models omp can resolve for you.

Role Used by Example
node-plan planner anthropic/claude-opus-5-5
node-implement implementer anthropic/claude-sonnet-5-5
node-review reviewer openai-codex/gpt-6.1-sol
node-chair council chair anthropic/claude-opus-5-5
council-seat-1..3 council seats one model from each provider you have

The examples are the author's own picks; setup offers an author default only when you have that provider. Put the reviewer on a different provider than the implementer. Setup warns when the reviewer and implementer share a provider, or when council seats share a model.

Uninstall

  1. omp plugin uninstall node-loop (check the name with omp plugin list).
  2. Revert the settings under "What setup changes".
  3. Optional: rm -rf ~/.omp/node-loop to delete run history.

Troubleshooting

Symptom Fix
exitnode does nothing Restart omp. The keyword works only as the first word, and a session that was already open does not load it.
exitnode does nothing under omp -p -p skips the keyword hook. Headless and scripted runs use omp -p '/skill:node-loop <goal>'.
Preflight: model role missing Run exitnode setup.
Preflight: model role unresolved omp does not list that model for you (not signed in, or a typo). Run exitnode setup and pick from the list.
Preflight: task.isolation.* is ... Run exitnode setup. If it still fails, a project config is overriding the global one.
Preflight: git identity missing The target repo has no user.name or user.email. Run git config user.name '<name>' and git config user.email '<email>' in that repo.
Run refuses: dirty working tree Commit or stash, including untracked files.
Run refuses: detached HEAD git switch <branch> first.
Forge says no free port Ports 4747 to 4767 are all taken. Close the other forge.
Forge page is empty No run exists yet. Start one with exitnode <goal>.
Forge stops The forge runs inside omp. Closing omp stops it.

Reporting issues

Open an issue at https://github.com/exit-node/node-loop/issues with the output of omp --version, the output of scripts/preflight, and the last 20 lines of the run's ledger.md. Remove any secrets first.

For contributors and agents

docs/CONTRACT.md is the source of truth for names, file layout, agent schemas, the ledger grammar and the loop's steps. Skills live in skills/, agents in agents/, the keyword hook, setup and forge view in extensions/node-loop/.

License

MIT

About

Multi-agent coding loop for omp: scouts, planner, isolated implementers, cross-model review and a model council.

Topics

Resources

Stars

16 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages