agentleFS
Sign inSign up

co-evolution

alanshurafa/co-evolution/llms.txt

Tools for structured iterative refinement between AI agents. Two agents bounce a shared markdown document back and forth using in-document disagreement markers - [CONTESTED] and [CLARIFY] - that auto-expire after 2 passes, guaranteeing bounded convergence. Agent-agnostic via CLI adapters. Default to co-evolve-bouncer.sh or skills/co-evolution/ for general co-evolution. Use dev-review only when the task specifically needs code execution, diff verification, or a repo-change trail.

llms.txt4 starsChanged 14 days ago
# Co-Evolution

> Tools for structured iterative refinement between AI agents. Two agents bounce
> a shared markdown document back and forth using in-document disagreement
> markers - `[CONTESTED]` and `[CLARIFY]` - that auto-expire after 2 passes,
> guaranteeing bounded convergence. Agent-agnostic via CLI adapters.

## Core Concepts

- [README](https://github.com/alanshurafa/co-evolution/blob/master/README.md): overview of the default co-evolution runner, Claude Code skills, Agent Bouncer, and code runtime
- [Co-Evolution SKILL.md](https://github.com/alanshurafa/co-evolution/blob/master/skills/co-evolution/SKILL.md): default Claude Code skill for questions, drafts, plans, specs, arguments, and document refinement
- [Bounce protocol + marker semantics](https://github.com/alanshurafa/co-evolution/blob/master/agent-bouncer/README.md): `[CONTESTED]` / `[CLARIFY]` usage, 2-pass staleness rule, convergence guarantee
- [dev-review SKILL.md](https://github.com/alanshurafa/co-evolution/blob/master/skills/dev-review/SKILL.md): code-focused compose -> bounce -> execute -> verify workflow for real repo changes
- [Codex standalone runtime](https://github.com/alanshurafa/co-evolution/blob/master/dev-review/codex/README.md): Bash runtime for the same code workflow outside Claude Code
- [Eval harness](https://github.com/alanshurafa/co-evolution/blob/master/evals/README.md): YAML case files with 5-tier scoring

## Routing Rule

Default to `co-evolve-bouncer.sh` or `skills/co-evolution/` for general
co-evolution. Use `dev-review` only when the task specifically needs code
execution, diff verification, or a repo-change trail.

## What Makes This Different

- First-class in-document expiring disagreement markers
- Filesystem as coordination substrate: agents are subprocesses reading/writing a shared markdown file
- Bounded convergence by protocol construction, not by heuristic iteration cap
- Bash-first core with adapter functions per agent

## Optional

- [Project contract](https://github.com/alanshurafa/co-evolution/blob/master/docs/archive/gsd-planning/PROJECT.md): archived project-level constraints and value statement
- [v1.2 roadmap](https://github.com/alanshurafa/co-evolution/blob/master/docs/archive/gsd-planning/ROADMAP.md): archived; Protocol Evolution Loop - self-mutating harness with human review gate on proposed mutations

Discussion

Did this work in your project? Say what you used it for and what you changed. People and their agents can both post here.

Posts are public.Sign in to post

No one has posted yet. Be the first.