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-Paper-Radar: reads new arXiv papers each morning and surfaces the few worth reading.systemone-compiler: declare a decision, compile typed Jev questions, measure them, ship a JSON program.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 a zero-dependency installer for eight harnesses. |
jev-prune-kit |
Capability-aware context-pruning installer and bounded Jev assessment engine (experimental, not live-host tested). | |
jev-sentinel |
Multi-harness reference implementation of JEV Sentinel. | |
jev-codex-approval |
Experimental typed JEV approval preflight for Codex with deterministic policy gates and Guardian fallback. | |
systemone-compiler |
Declare a decision, compile typed Jev questions, measure them, ship a JSON program. | |
| 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 with TypeSafe JEV: reproducible pipeline and interactive geographic visualizations. |
england-crime-atlas |
Interactive England crime map with regional conviction demographics and offence codes from official open data. | |
gemma4good |
Clean Drop: edge-AI water reporting and triage with Gemma 4. | |
| Research and papers | beyond-the-token-bottleneck |
Obsidian research wiki on latent-space reasoning and inter-agent communication. |
JEV-Paper-Radar |
Reads new arXiv papers each morning and surfaces the few worth reading. | |
computervision |
Computer vision library for Lean. | |
| Integrations | openclaw-mautic-plugin |
OpenClaw plugin for controlled Mautic CRM operations: typed API tools, webhooks and guarded maintenance. |
The organization also holds forks of upstream projects (such as llama.cpp and a Codex fork); those belong to their original authors.
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.