Open research tools and experiments on AI coding agents, context management, and evaluation.
Stewarded by Complete Tech LLC.
complete.tech · @CompleteTechLLC on X · All repositories
- Context for coding agents. How to give an agent memory and a bounded context window without losing its sources: retrieval, pruning, and typed questions.
- Evaluation and measurement. Compiling decisions into measured programs, and running agents against games and environments to see what they actually do.
- Reading the frontier. Tracking new papers and mapping ideas such as latent-space reasoning and inter-agent communication.
- Data and tools. Open-data visualizations and integrations that put these ideas to work.
 Jev agent tooling |
 Games and agents |
 Data and visualization |
 Research and papers |
jev-context-fabric: local-first, source-backed memory layer for coding agents.jev-sentinel: multi-harness installer and security evaluator, a reference implementation.beyond-the-token-bottleneck: an Obsidian research wiki on latent-space reasoning and inter-agent communication.
| Area | Repository | What it is |
|---|---|---|
| Jev agent tooling | jev-context-fabric |
Local-first, source-backed memory layer for coding agents with an installer for eight harnesses and reversible /prune |
jev-prune-kit |
Capability-aware context-pruning installer and Jev assessment engine for coding agents; experimental, not live-host tested | |
jev-sentinel |
Multi-harness installer and security evaluator that keeps agent authorization with the host; reference implementation | |
jev-codex-approval |
Experimental typed JEV approval preflight for Codex with deterministic policy gates and Guardian fallback | |
| Codex tooling | codex-plaintext-collab |
Patch for OpenAI Codex CLI that disables encryption of inter-agent (subagent) messages, with build and update-safe install tooling |
| Games and agents | pokemon-harness |
Memory-first automation harness for Pokemon Red, Blue and Yellow with a typed Session API and MCP server |
ai-self-replication-study |
Modular AI agent system combining language models with shell command execution and file manipulation | |
| Data and visualization | jev-311-heatmap |
NYC 311 complaint heatmaps built with Jev: reproducible pipeline and interactive geographic visualizations |
england-crime-atlas |
Interactive England crime map with regional conviction demographics and detailed offence codes from official open data | |
gemma4good |
Clean Drop: edge-AI water reporting and triage built on Gemma 4 for mapping water stress | |
| Research and papers | beyond-the-token-bottleneck |
Obsidian research wiki on latent-space reasoning and inter-agent communication beyond the discrete token bottleneck |
| Integrations and notes | openclaw-mautic-plugin |
OpenClaw plugin for controlled Mautic CRM operations: typed API tools, webhooks and guarded maintenance |
clawexplorer.ai-changelog |
Public summary-only release notes and feed links for the ClawExplorer.ai events directory | |
| Organization | .github |
Organization profile and community health files for CompleteTech-LLC-AI-Research. |
These are forks of other authors' projects. The work belongs to the upstream authors; the forks are kept for our own reference and experiments.
JEV-Paper-Radar: fork of Eliot5566/JEV-Paper-Radar, which reads new arXiv papers each morning with Jev.codex-jev: fork of openai/codex.computervision: fork of vukimlien4706-cpu/computervision, a computer vision library for Lean.factorio-learning-environment: fork of JackHopkins/factorio-learning-environment.G0DM0D3: fork of elder-plinius/G0DM0D3.llama.cpp: fork of ggml-org/llama.cpp.
ml_breakout: archived and read-only.
Several of these projects are experimental. For example, jev-prune-kit is
marked experimental and not live-host tested, and pokemon-harness reports its
current release candidate as PARTIAL, not production-ready. Each repository's
README is the authority on its own status; this page makes no claims beyond them.