Geek · incurable tinkerer · dad of a lovely little girl
I take things apart until they tell me how they work — laptops, phone kernels, AI agents.
Lately: teaching software that already knows its craft to talk to AI, in ways people can trust.

极客，爱搞机，家有可爱小女一枚
喜欢把东西拆开，直到它说出自己是怎么运转的：笔记本、手机内核、AI Agent。
最近在做的事：让懂行的老软件学会和 AI 对话，并且让人放心。

• Making existing software agent-ready — exposing what a product already does well as task-level MCP tools that any agent can call, rather than building yet another agent. The domain assets stay the moat; the agent is interchangeable.
• The trust layer — identity carried through to the person behind the agent, permissions that never exceed theirs, audit trails, answers traceable to the source page, and human sign-off where it matters.
• Agent execution security — what an agent can read, run and send out, and how to enforce that at the OS boundary instead of trusting the vendor's promise.
• Agent runtimes — streaming lifecycles, tool-call recovery, loop guards, context compaction and MCP isolation, mostly in Rust.
• Financial document systems — a production platform for regulatory filings and annual reports that I owned for a long stretch: parsing, rule checks, and a Python stack modernized end to end. On top of it, I led the build of the AI-powered annual report platform an exchange launched for its listed issuers.
• Linux & low-level fixes — touchpad gestures, laptop ACPI quirks, phone kernels: when the hardware misbehaves, I go down to the driver.
• Pragmatic Python tooling — small tools that remove a daily annoyance and then get out of the way.

RunSeal · Rust
An OS-native sandbox layer and stable execution protocol for AI agents. Policy-governed filesystem and network boundaries, structured audit events, and fail-closed behavior whenever a request can't be fully enforced. Windows is the reference backend; macOS and Linux are experimental. → Why I built it · 做 RunSeal 的初衷

ZCode silent upload disclosure · Sep 2026
Found that an AI coding tool with 1M+ users was packing workspaces — including full .git history — encrypting them and uploading them to cloud storage, on by default with no way to turn it off. The disclosure led to the feature being removed, the client being open-sourced, and third-party security audits. → 扒一扒 ZCode 静默上传全量 Git 历史的骚操作

• 为什么说“套壳 Agent 就能做 SaaS”是个荒谬的幻想 — why wrapping an agent isn't a product
• Agent 死循环的分级治理阶梯（Warn → Block → Halt） — graded loop guards instead of a hard kill
• 跨 Chunk 流式状态机的 DSML 抢救实录 — recovering tool calls the model emitted as text
• Agent 长会话的两阶段预热与可回捞设计 — moving context compaction off the hot path
• MCP 服务的启动隔离与平滑降级 — optional plugins must never block the core session
• 消除桌面 Agent 输入框假死与流式卡顿 — decoupling heavy IO from UI end states

• 4090 24G 本地部署 Qwen3.8-Flash-Next：从 Ollama 换到 Strata
• 给女儿做了个作业打印助手，顺便榨了榨小鸡
• 祖传系统无损迁机：btrfs send/receive 实操记录
• 两台 DGX Spark 跑 Agent：跑分一时爽，干活卡成狗
• 做 RunSeal 的初衷：给本地 Agent 设一道命令执行边界

• check-requirements-txt ⭐12 — pre-commit hook that catches packages missing from requirements.txt / pyproject.toml
• pot-app-recognize-plugin-rapid-py ⭐15 — offline RapidOCR plugin for Pot App
• fast-context ⭐24 — codebase context locator for coding agents: local chunk prefetch plus remote symbol reasoning
• dsh-tool-ocr — local OCR so text-only LLMs can read images

• esengine/DeepSeek-Reasonix (⭐35k) — Windows 11 session-switch freeze, prompt history, per-arch macOS builds
• BYVoid/OpenCC (⭐10k) — manylinux2014 wheels, macOS ARM64 CI
• tiann/hapi (⭐5k) — terminal quick keys, mobile input and teardown fixes
• zurawiki/gptcommit (⭐2.4k) — custom OpenAI-compatible API base URLs
• h2non/filetype.py (⭐775) — zstd and CDFV2 Word detection, file-like reader fixes

_{Code is cheap, let's talk.}