# CLIO Coder

https://akougkas.io/work/clio-coder/

Author: Anthony Kougkas

Updated: 2026-11-09

## Why I built it

Scientific software has an execution environment, a history, and standards that matter beyond the next plausible patch. A simulation kernel may depend on a compiler at a laboratory. A numerical change may need comparison with a reference result. A repository may not be allowed to send its contents to an arbitrary provider.

CLIO Coder is where I work on the runtime around that problem: context that has a useful lifetime, explicit authority, bounded delegation, and results a person can inspect. It sits inside [CLIO](https://akougkas.io/work/clio/), the context architecture I created for [IOWarp](https://akougkas.io/work/iowarp/).

## The road here

I began exploring this line in December 2023. Early wrappers around Claude Code helped me study the workflow. Warpio, a Gemini CLI fork, explored scientific personas and context handover. AWOC explored agent orchestration and shared memory. PanCode refined the control loop and worker model. In April 2026 I consolidated the lessons into the present CLIO Coder repository.

1. 2023
   
   
   
   ### [WarpGPT](https://akougkas.io/work/iowarp/)
   
   
   
   The natural-language interface I wrote into the IOWarp proposal.
2. Aug 2025
   
   
   
   ### [Warpio](https://github.com/akougkas/warpio-cli)
   
   
   
   Scientific personas and context handover inside an agent, first as extensions for Claude Code and Gemini CLI, then as Warpio CLI.
3. Sep 2025
   
   
   
   ### AWOC
   
   
   
   The fleet model. Coding agents as bounded workers to direct with typed intent, rather than a library to call.
4. Mar 2026
   
   
   
   ### [PanCode](https://akougkas.io/work/atlas/#project-pancode)
   
   
   
   The runtime. A control loop, a scored worker pool, and a reproducibility receipt for every run.
5. Apr 2026
   
   
   
   ### [CLIO Coder](https://akougkas.io/work/clio-coder/)
   
   
   
   The research line brought together under IOWarp: typed tools, bounded workers, sealed receipts, and local models.

The repository’s age is not the age of the idea. Those earlier systems exposed what the harness needs to own: how context reaches a model, which tools a task can use, how work is handed off, and what remains after a run. The present implementation still has to earn its use; a long design history does not remove its current limitations.

## The model is a choice

CLIO Coder treats the model as the user’s choice: a local or self-hosted server, a cloud API, a subscription, or an institutional inference gateway, with the conversation and its workers free to use different ones. Separating the harness from the model lets someone keep inference on approved infrastructure, or compare models on the same work under the same rules. A local interface does not make a cloud connection local, though: data sent to a provider follows that connection.

The [first-session guide](https://coder.iowarp.ai/tutorials/first-session.html) at [coder.iowarp.ai](https://coder.iowarp.ai/) covers installation and current commands.

## Context has a lifecycle

A working set, a handoff, and a durable lesson need different lifetimes. CLIO Coder gives each its own mechanism: a project handbook and code map for orientation, visible context use and compaction for the working set, branch-bound handoffs, and memory proposals a person reviews apart from the task. These organize what a model sees; they do not train its weights during the session.

Tools follow the same idea. The coordinator starts with a small toolkit and discovers further capabilities when a task needs them, and every tool call still passes admission. Disabling a resource cannot erase instructions the conversation has already read.

## Authority and results travel with the work

Authority is explicit. Workers and external peers run at the default permission level whatever the main session allows, and dispatched workers run inside an operating-system sandbox under immutable permits. Commands in the main session still execute with the user’s access. A managed run, by a native worker or by a peer agent such as Claude Code or Codex, ends in a result and a receipt.

Verification comes from the repository’s declared checks, assessed against fingerprints of the source, the check declarations, and the project’s quality policy, so a later edit can make an earlier pass stale. A denied check is not a validation result. Neither a passing test nor a sealed receipt proves scientific correctness.

## Work across machines

A local session can coordinate workers over SSH on a university cluster or a remote server, and the machine running inference and the machine running a worker’s tools are separate choices. The fleet needs matching runtimes, approved access, and the project at the same path on every node; the [fleet guide](https://coder.iowarp.ai/docs/guide/fleet-dispatch.html) lists the requirements.

## Where it stands

CLIO Coder is beta software, and interface and format changes arrive through its [release notes](https://github.com/iowarp/clio-coder/releases). Release gates do not establish scientific benchmark results, and model quality varies by task and configuration. Earlier versioned notes describe earlier designs; the guides at coder.iowarp.ai describe the supported workflow.

I create and maintain Coder; [Eneko Gonzalez](https://grc.iit.edu/center/team/eneko-gonzalez/) has contributed. The project builds on the Pi agent framework from Earendil Works, the Claude Agent SDK, the Agent Client Protocol, and other upstream work. Across the wider CLIO program, [Luke Logan](https://grc.iit.edu/center/team/luke-logan/) leads day-to-day Core engineering, [Jaime Cernuda](https://grc.iit.edu/center/team/jaime-cernuda/) leads Agent and Relay development, and students lead the scientific tool integrations in Kit.

The next question is how well these contracts hold across longer scientific work: context that remains useful, results that remain inspectable, and automation that leaves a person able to understand and intervene.
