# Haochuan's Blog

> An independent engineering journal about personal intelligence systems,
> agentic software, local AI, wearable computing, and human agency.

## Site

- [Home](https://blog.haochuanz.net/): Latest writing and selected projects.
- [About](https://blog.haochuanz.net/about): Haochuan's background and interests.
- [Projects](https://blog.haochuanz.net/projects): Public projects, source code, and demos.
- [Applets](https://blog.haochuanz.net/applets): Opt-in CyberHUD interactive previews.
- [Archive](https://blog.haochuanz.net/archive): All published posts.
- [Tags](https://blog.haochuanz.net/tags): Browse writing by topic.
- [Browsing guide](https://blog.haochuanz.net/skill.md): Where to start, what to fetch, and the limits of this public snapshot.
- [Site map](https://blog.haochuanz.net/llms.txt): Public pages and the full post index, with short descriptions.
- [Full portfolio text](https://blog.haochuanz.net/llms-full.txt): The complete public About page and project index in one Markdown document. Journal articles are linked, not bundled.
- [Public profile](https://blog.haochuanz.net/api/profile.json): Versioned JSON with a source-derived introduction and links to the public biography and work.
- [Project catalog](https://blog.haochuanz.net/api/projects.json): Versioned JSON with project descriptions, categories, source links, and demos.
- [RSS](https://blog.haochuanz.net/feed.xml): Journal subscription feed.

## Published posts

- [Testing a Recursive Self-Improvement Pipeline on a CrunchDAO Competition](https://blog.haochuanz.net/posts/rsi-research-pipeline): I am using a real-time structural-break competition as a bounded proving ground for a two-loop research system: human-supervised OpenCode outside, a DSH and DAL improvement harness inside.
- [An Execution Contract for Skill-Bundled Scripts](https://blog.haochuanz.net/posts/agent-skill-runtime): My experiment with a shared runtime and host adapters for skill-bundled scripts, and how this small piece fits into the broader capability-package idea.
- [Distributing Agentic Capabilities](https://blog.haochuanz.net/posts/distributing-agentic-capabilities): As models and agent harnesses become operating systems for agentic work, the durable enterprise layer is the governed capability package: the app that connects them to real business systems.
- [Can DeepSeek Harness Reproduce Six Other Agent Harnesses?](https://blog.haochuanz.net/posts/agent-harness-08-dsh-comparison): A final comparison of seven agent harnesses, with a strict distinction between DSH capabilities that are shipped, supported by partial plugin evidence, and suggested only by an architectural seam.
- [How DeepSeek Harness Uses Cordis for Data-Driven Assembly](https://blog.haochuanz.net/posts/deepseek-harness-cordis-assembly): DeepSeek Harness turns a small composition model into profiles, bundles, presets, activation rules, and user overlays. The important part is the semantics that keep those layers inspectable.
- [DeepSeek Harness: An Auditable, Composable Agent Runtime](https://blog.haochuanz.net/posts/agent-harness-07-deepseek-harness): How DeepSeek Harness turns a small agent loop into a configurable product: turn and step scheduling, waterfall tools, event-sourced sessions, and a carefully bounded fail-closed subprocess sandbox.
- [Hermes Agent's Durability Rule: Persist Before a Tool Can Act](https://blog.haochuanz.net/posts/agent-harness-06-hermes-agent): Hermes Agent centers its safety model on a simple ordering rule: write an assistant tool call to SQLite before executing it. Combined with iteration budgets, protocol repair, non-destructive compaction, and layered terminal controls, that rule supports long-running autonomous work.
- [OpenClaw's Agent Gateway: Sessions, Persistent Work, and Optional Isolation](https://blog.haochuanz.net/posts/agent-harness-05-openclaw): OpenClaw turns an agent loop into a persistent gateway for chat channels. Session keys, serialized lanes, heartbeats, cron jobs, and optional sandbox backends address the risks of an assistant that keeps running when you leave the terminal.
- [Pi's Small Agent Harness: Queues, Session Trees, and an Honest Security Boundary](https://blog.haochuanz.net/posts/agent-harness-04-pi): Pi keeps its agent runtime deliberately small: a two-queue loop, append-only session trees, and extensions in place of built-in product features. Its most important decision is equally explicit: it provides no sandbox.
- [opencode Under the Hood: Durable State, Live Streams, and Reversible Sessions](https://blog.haochuanz.net/posts/agent-harness-03-opencode): A source-level look at opencode's TypeScript harness: database-backed session projections, streamed tool events, layered compaction, reversible workspace snapshots, and blocking approvals.
- [Codex Under the Hood: Making the Sandbox a First-Class Agent Concern](https://blog.haochuanz.net/posts/agent-harness-02-codex): A source-level tour of Codex's Rust harness: multi-signal turn completion, PTY-backed execution, layered sandbox policy, replayable JSONL history, and an app-server interface.
- [Claude Code Under the Hood: A Recovery-Oriented Agent Harness](https://blog.haochuanz.net/posts/agent-harness-01-claude-code): An unverified source snapshot shows a harness designed around recovery: explicit loop transitions, streaming tool execution, layered context defenses, and replayable session history.
- [The Age of Intention](https://blog.haochuanz.net/posts/age-of-intention): Working with AI agents every day has changed the scarce part of my work. Intent, judgment, and the engineering systems that turn ideas into evidence now matter more than keystrokes.
- [Building CyberHUD, Month One: From a Clock in My Glasses to an iPhone Desktop](https://blog.haochuanz.net/posts/building-cyberhud-month-one): How a missing iOS app for my display glasses grew into an iPhone-hosted Desktop, a bounded applet platform, and a document-driven development system run with concurrent AI agents.
- [Cordis Runtime From First Principles](https://blog.haochuanz.net/posts/cordis-runtime-from-first-principles): A first-principles tour of Context, Services, Plugins, events, lifecycle ownership, and inject: the pieces Cordis needs once runtime capabilities can appear, disappear, and depend on one another.
- [I Ran DeepSeek V4 Flash 0731 Locally with 48 GB of RAM](https://blog.haochuanz.net/posts/running-deepseek-v4-flash-0731-with-48gb-ram): How I worked through Blackwell compatibility, NCCL deadlocks, host OOMs, and NVMe-backed MoE storage before pushing the official DeepSeek V4 Flash 0731 model to two exact 400K-context requests.
- [Gemma 4 vs Qwen Models: A Practical Local LLM Test](https://blog.haochuanz.net/posts/gemma-4-vs-qwen-models-local-llm-test): I tested several open-source local LLMs on a few short reasoning and trick questions to see which ones felt best in real local use, then added three more models to see how the rankings shifted.
- [Designing an OpenClaw Agency (Part 1): Why an Agent Company Is Now Achievable](https://blog.haochuanz.net/posts/openclaw-agency-part-1): Part 1 of my implementation series: why I believe a digital agent company is now feasible, and how I’m using research-driven architecture to design OpenClaw Agency with rigor.
- [I Spent $60+ "Reinvent" OpenClaw Memory System Locally — Here’s What Actually Worked](https://blog.haochuanz.net/posts/reinvent-memory-system): I tried to reinvent a MemOS-style three-layer memory stack locally for OpenClaw, compared hosted vs local behavior, and learned why engineering methodology matters more than brute-force agent rewrites.
- [Your Next Startup Is You: The Rise of the Personal Agentic Company](https://blog.haochuanz.net/posts/rise-of-the-personal-agentic-company): The next leap in personal leverage is not using one AI assistant better, but building a personal agentic company: a digital workforce you design, govern, and continuously improve.
- [Building an Agent That Can Grow: From Prompting to Self-Transformation](https://blog.haochuanz.net/posts/agent-self-transformation): From prompt tricks to durable architecture: four things I built to help my agent learn, debug itself, and expand tools without context bloat.
- [Ghost in My Machine: How I Built Miko with OpenClaw in a Few Days](https://blog.haochuanz.net/posts/openclaw-miko-first-days): I turned OpenClaw into a practical daily copilot: Codex for the main thread, qwen3.5 for cheap background tasks, fully local semantic memory, and Claude Code CLI when the work gets serious.
- [Stop Typing, Start Orchestrating: How IC Engineers Can Survive (and Own) the AI Revolution](https://blog.haochuanz.net/posts/stop-typing-start-orchestrating): The core shift for senior ICs: from writing syntax to designing and orchestrating AI systems with strong taste, safety, and context.
- [Why I Keep a Personal Blog in the Age of AI](https://blog.haochuanz.net/posts/why-i-keep-a-personal-blog-in-the-age-of-ai): In an era of algorithmic feeds and AI-generated noise, this blog is my way to preserve clear thinking, data sovereignty, and human judgment.
- [Dynamic Routing and Static Generation](https://blog.haochuanz.net/posts/dynamic-routing): Lorem ipsum dolor sit amet, consectetur adipiscing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua. Praesent elementum facilisis leo vel fringilla est ullamcorper eget. At imperdiet dui accumsan sit amet nulla facilities morbi tempus.
- [Learn How to Pre-render Pages Using Static Generation with Next.js](https://blog.haochuanz.net/posts/hello-world): Lorem ipsum dolor sit amet, consectetur adipiscing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua. Praesent elementum facilisis leo vel fringilla est ullamcorper eget. At imperdiet dui accumsan sit amet nulla facilities morbi tempus.
- [Preview Mode for Static Generation](https://blog.haochuanz.net/posts/preview): Lorem ipsum dolor sit amet, consectetur adipiscing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua. Praesent elementum facilisis leo vel fringilla est ullamcorper eget. At imperdiet dui accumsan sit amet nulla facilities morbi tempus.