SkillAgentSearch skills...

Converge

Converge is an open-source framework for coordinated, multi-repository software development using AI agents. It enables agents working across independent codebases to collaborate, agree, and ship changes through bounded negotiation, shared evidence, and human-in-the-loop governance without endless loops or unsafe automation.

Install / Use

/learn @sumanshusamarora/Converge
About this skill

Quality Score

0/100

Supported Platforms

Universal

README

Converge

Converge is a multi-repository coordination system that accepts tasks, runs bounded orchestration, escalates ambiguous decisions to HITL, and writes auditable run artifacts you can review or action.

Why Converge exists:

  • keep cross-repo ownership decisions explicit
  • provide a queue/worker model for repeatable orchestration runs
  • produce artifacts (summary, matrix, run payload, prompt packs) for humans to execute safely

Local quickstart (SQLite)

pip install -e ".[dev]"
export SQLALCHEMY_DATABASE_URI="sqlite:///./converge.db"
export CONVERGE_QUEUE_BACKEND="db"
export OPIK_TRACK_DISABLE="true"
converge server --port 8080
# in another terminal
converge worker

Docker quickstart (Postgres + API + Worker + UI)

docker compose up --build
  • API: http://localhost:8080
  • UI: http://localhost:3000
  • Postgres: internal-only by default (no host port published)

How Converge Works

Converge runs tasks through a queue-and-worker lifecycle: submit a task, generate plans and artifacts, optionally execute when policy allows, and escalate to HITL when human decisions are needed. This keeps automation bounded while preserving clear audit trails for every run.

See How Converge Runs for a complete explanation.

Codex planning defaults

Converge uses Codex CLI for repository planning when available and falls back automatically if a configured model is not accessible.

  • Default model candidates: gpt-5.3-codex,gpt-5,gpt-5-mini
  • Override model for one run: converge coordinate --coding-agent-model gpt-5 ...
  • Persisted env override: CONVERGE_CODING_AGENT_MODEL=gpt-5
  • Planning control mode: CONVERGE_CODING_AGENT_PLAN_MODE=auto|force|disable

Check what your environment will do:

converge doctor

Documentation

HITL and artifacts are first-class: when Converge cannot safely decide, tasks move to HITL_REQUIRED, and every run writes files under .converge/runs/<timestamp>/.

View on GitHub
GitHub Stars6
CategoryDevelopment
Updated1mo ago
Forks0

Languages

Python

Security Score

85/100

Audited on Mar 3, 2026

No findings