jackie-cqzFIELD NOTES ON SOFTWARE & AI
中文RSS

BEHIND THE NOTEBOOK

Hello, I'm jackie-cqz.

Building things. Taking notes. Following the interesting questions.

I work on AI systems, autonomous agents, and privacy-preserving machine learning. My interests span agent memory and collaboration, vision-language models, robotics, and the engineering that connects models to working systems.

What this notebook is for

A place to explain design decisions and share practical guides from building agent tools. Notes connect the implementation to setup, everyday use, and the evidence behind a result.

Internship experience

Agent engineeringINTERNSHIP

Z.ai

My internship work at Z.ai focuses on the AutoClaw agent harness.

AutoClaw

Agent harness

AutoClaw is a general-purpose AI agent from Z.ai. It connects models, context, and tools to carry out multi-step tasks across files and the browser.

BigModel platformZ.ai Chat

Open-source contributions

oceanbase/powercontext

PowerContext

Persistent context and memory for AI agents.

My contributions connect PowerContext to ZCode and Kimi Code, make memory behavior observable, and improve recall across sessions.

  • ZCode integration: native MCP access, prompt capture, workspace Scope workflows, setup and diagnostics, plus bounded lifecycle handling.
  • Kimi Code integration: a managed plugin, native memory tools, Hook observations, and real-host acceptance on Windows and Linux.
  • Memory and validation: preserve source language and exact identifiers, improve focused lexical recall, and automate CLI acceptance and Skill behavior evaluation.

loopx-project/loopx

LoopX

A local-first control plane for long-horizon agents and personal agent teams.

My contributions focus on native host control, traceable execution evidence, and reliable collaboration and recovery on Windows.

  • ZCode Goal control: explicitly managed CLI sessions with model selection, start/pause/resume/stop, status readback, and Core quota admission and revocation.
  • Execution evidence: reconcile native Codex sub-agent events with Turns; execute opt-in, SHA-pinned public GitHub research with independent admission and ledger readback.
  • Reliability fixes: Windows peer handoffs and long paths, configuration backup/restore, chat append recovery, lock sidecars, blocked-Todo channel notices, public-safety validation, and shared CI contracts.

Personal projects

AgentMeld

A Pi-powered, local-first group-chat workspace for AI coding teams. Supports peer consultation, confirmed task handoffs, and task coordination as needed, with files, tool calls, approvals, and artifacts kept alongside the conversation.

Elsewhere

Find the code and public work on GitHub, or follow new notes through the RSS feed.