user-agency-substrate

Discord-native research operations toolkit for ML experimentation with Codex/Claude agents: relay orchestration, human-in-the-loop plan/task/worktree workflows, callback-driven long-running jobs, and hypothesis-driven research skills.

Architecture, safety charter, quickstart, and philosophy for Starshard — a substrate for multi-agent personal AI where user-owned memory is the primary layer and agents are peripherals.

Self-hosting deployment kit for OpenClaw: scripts, config templates, and a Chinese-language guide for running your own Discord-fronted personal AI agent on your own server.

An open, user-owned communication substrate for personal AI agents: inbox, addressing, trust tiers, policy and receipts — plus SPEC-0003 on agent-to-agent identity, authentication, delegated authority and secure messaging with no central platform.

Public front door to a one-person, AI-assisted interdisciplinary research program on information and its future-reachability — how reachability is created, stored, degraded, and repaired in physical, living, and cognitive systems.

Reproducible re-analysis on two human EPIC datasets asking whether partial epigenetic reprogramming reverts the causal-damage (DamAge) or the adaptive (AdaptAge) methylation layer — the two layers dissociate, and the dissociation replicates.

A small standard plus helper for adding clean, honest provenance to an agent skill — who made it, where to learn more, how to report problems, what license — as installer-first hygiene, not advertising.

Smallest runnable proof of four multi-agent coordination primitives: shared cross-agent memory, a content-gated permission boundary, a receipt per action, and cross-session work-claims. Stdlib Python, runs in ~1s.

An open, agent-traversable knowledge graph of non-consensus claims about intelligence, learning, and information dynamics — every node carries an explicit epistemic state (survived / speculative / raw) and a JSON-LD manifest others can fork and link into.

Five installable Claude Code verification skills for multi-agent coding: a drift mirror, an adversarial review panel, a reuse compiler, a frontend smoke gate, and a delegation bound — for when both agents say done and nobody verified.

Public receipt log of a personal multi-agent stack (Starshard): each session's decisions, changes, and artifacts published as an auditable trail. No commentary, no thesis — just receipts.

One command turns Claude Code into a personal 智子 (Zhizi) assistant: a warm Lite mode for non-technical users, and a Pro mode adding agent-OS skills, hooks, subagents, and OS daemons. Bring-your-own-key, privacy-clean.

A public-safe, git-backed blackboard: durable inbox/outbox JSON state that a mobile capture app and an agent relay exchange as auditable commits, without handing a broad token to the client.

Fleet Coordination Protocol (FCP) v0 — a thin coordination wire above any agent runtime: shared-memory facts, work-claims, receipts before irreversible actions, and belief-conflict quarantine.

Public essays and notes by Macheng Shen, auto-published from an agent pipeline — AI infrastructure, human-AI partnership principles, and forecasting inside a computational universe.

Phase 0 reference server for the Starshard memory hub: FastAPI + SQLite + MCP bridge, giving any Claude Code / Codex / MCP agent persistent memory across sessions and machines.

An open, continuously-published log of small pre-registered continual-learning and credit-assignment experiments — how a learner decides how to learn — with negative and diagnostic results published by design.

Personal research site of Macheng Shen — essays toward a theory of intelligence and contemporary AI, information dynamics, and the user-agency substrate. Published agent-readable (llms.txt + JSON-LD), English and Chinese.