jev-codex-factory
Multi-agent Codex factory: Jev decides task shape, decomposition and routing, and failures are classified, not blindly retried.
Overview
Jev Codex Factory turns one engineering request into a routed, parallel, verified and resumable swarm of Codex coding agents. Its premise is that a single big model call works for small tasks but becomes messy with independent subtasks, varying difficulty, missing dependencies, flaky verification, parallel file edits, partial failures and interruptions. Jev does not write code; it decides how the code should be written.
The flow begins with Jev deciding the task's shape, whether a single agent suffices or the task should be decomposed into a dependency graph of subtasks. A Jev router then picks model and effort for each agent, Codex agents work in parallel, verification gates every change, failures are classified rather than blindly retried, and blocked work can resume later. It targets OpenAI Codex and is Apache-2.0 licensed. The README excerpt is largely conceptual, so measured results are not stated.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Should this task run as a single agent or be decomposed into a DAG? | single · decompose |
| Choice | Which model and effort should this agent use? | — |
| Choice | What kind of failure is this (retry, fix, block)?(inferred) | — |
How it works
- Jev decides single-agent or decomposition
- Build a task DAG if decomposing
- Jev routes each subtask to a model and effort
- Codex agents work in parallel; verification gates changes
- Classify failures; resume blocked work later
Strengths
- Parallel execution of independent subtasks
- Failure classification instead of blind retries
- Resumable work
Caveats
- No measured results in the README excerpt
- Complex orchestration to operate
At a glance
- Type
- Project
- Status
- prototype
- Language
- Python
- Runtime
- orchestrator for Codex agents
- Has code
- yes
- License
- Apache-2.0
Use cases and integrations
How curated lists describe it
Route smarter. Code in parallel. Resume what breaks. A Jev-powered multi-agent factory for Codex.
logicrw/awesome-jev-projects
Frequently asked questions
Does Jev write code here?
No, it decides how work is split, routed and handled on failure.
What happens when an agent fails?
The failure is classified, not blindly retried, and blocked work can resume.
Which agent does it orchestrate?
OpenAI Codex.
Related projects
One skill and MCP server for Claude Code, Codex and Hermes where Jev gates large-model turns, routes models and guards Git steps.
MCP serverbetaPythonClaude Code skill that splits work into PBIs for Codex, Antigravity or Claude sub-agents, using Jev for model choice and triage.
ProjectprototypePythonRuns Claude Code on Opus 5.5 with Jev re-picking reasoning effort after every tool batch, via per-message effort so the cache survives.
ProxybetaTypeScriptTerminal coding agent where an LLM reads, edits and runs code, and Jev makes a few optional side decisions.
ProjectbetaTypeScriptSelf-improving multi-platform agent harness with an optional Jev companion for typed judgments and per-session model routing.
ServicebetaTypeScriptSupervisor-Lead-Peer roles for a team of Paseo coding agents, with opt-in Jev routing recorded as a receipt in shadow mode.
Pluginbeta