@obversa/teams

Three functions, each a team: give it your brief, your workspace, your test command and one engine per seat, and it returns a job for run.

npm install @obversa/runtime @obversa/teams

What you get

writerReviewerPair. One model writes the files your brief names, your test command runs, and a model from a different family reviews the result. A rejection sends the work back to the writer.
thresholdPanel. One model implements, your test command runs, and several reviewers read the change at the same time. The change passes when at least the threshold number of them accept.
featureDelivery. Analyse, implement, test, review, approve. The analyse step writes a delivery note, the reviewers can send the work back to the implementer, and the approve step writes an approval note.

Every team is a graph of named steps with a desc and a gate sentence on each. Every seat is checked before a run: the implementer and each reviewer must be different model families. A step that promises a file fails by name when the file is missing or empty. The test step passes on the command’s exit code.

Run one

import { ClaudeCliEngine } from '@obversa/engine-claude-cli';
import { CodexEngine } from '@obversa/engine-codex';
import { run } from '@obversa/runtime';
import { writerReviewerPair } from '@obversa/teams';

const workspace = process.cwd();
const writer = {
  engine: new ClaudeCliEngine({
    defaultModel: 'claude-sonnet-4-5',
    permissionMode: 'bypassPermissions',
  }),
  identity: {
    adapter: 'claude-cli',
    provider: 'anthropic',
    modelFamily: 'claude',
    model: 'claude-sonnet-4-5',
  },
};
const reviewer = {
  engine: new CodexEngine({
    defaultModel: 'gpt-5.6-luna',
    permissionMode: 'bypassPermissions',
  }),
  identity: {
    adapter: 'codex',
    provider: 'openai',
    modelFamily: 'gpt',
    model: 'gpt-5.6-luna',
  },
};
const team = writerReviewerPair({
  brief: 'Write a pure add(a, b) function in src/add.mjs with a Node test in test/add.test.mjs.',
  workspace,
  files: ['src/add.mjs', 'test/add.test.mjs'],
  test: { command: 'node', args: ['--test', 'test/add.test.mjs'] },
  writer,
  reviewer,
});
const result = await run(team, { cwd: workspace });

console.log(JSON.stringify(result.outcome, null, 2));

A real run of this file, and the files it left behind, are on the writer and reviewer page. The review panel and feature delivery pages do the same for the other two teams.

What every team takes

field	what it is
`brief`	The work, as text. Every model in the team reads it.
`workspace`	The directory the team works in. Files are written here.
`files`	The paths, relative to the workspace, that the brief expects written. At least one.
`test`	`{ command, args, timeoutMs? }`, run in the workspace. Exit 0 passes.
`maxKickbacks`	How many times a rejection may send work back. Default 1.

A seat is { engine, identity }: an engine plugin instance and the identity it runs under, { adapter, provider, modelFamily, model }. The identity is what the run record and the family check use, so it must name what runs.

team	seats
`writerReviewerPair`	`writer`, `reviewer`
`thresholdPanel`	`implement`, `reviewers` (a list of `{ name, seat }`), `threshold`
`featureDelivery`	`analyse`, `implement`, `reviewers`, `reviewThreshold`, `approve`

Gotchas

It ships no command. You import a team and run it from your own file.
Engines are yours to construct. The package chooses no model. Each seat is a plugin instance you build, with the permission mode the role needs: a writer must be allowed to write files.
The distinct-family check reads your declaration. Declaring one family for two different models defeats it, and declaring two families for one model lies to the record.

Source

packages/teams in the repository. The package depends on @obversa/runtime and nothing else.

Start

Concepts

Examples

Reviewing

Recording runs

Driving runs

Writing workflow shapes

Memory

Workspace

Core packages

Plugins

What you get

Run one

What every team takes

Gotchas

Source

​What you get

​Run one

​What every team takes

​Gotchas

​Source

What you get

Run one

What every team takes

Gotchas

Source