_{contribution graph rendered with Platane/snk}

• CodeInsights — an LLM-powered multi-agent platform for open-source contribution workflows. It is built around repository understanding, issue/PR scouting, agent coordination, and turning scattered open-source work into a repeatable engineering loop.
• yuansheng-kit — an agent-oriented toolkit for RISC-V optimization knowledge extraction, pattern mining, real-hardware performance analysis, root-cause diagnosis, and code generation.
• ClawPerch — a lightweight desktop perch for watching and managing OpenClaw / Codex-style agents. It focuses on keeping long-running agent work visible, interruptible, and operationally sane.
• openclaw-apex-team — a customizable OpenClaw engineering team framework for building, coordinating, and shipping with specialized AI agents.
• hermes-hip — a Hermes + Whip event-to-channel notification router that bypasses gateway sessions to avoid context pollution.
• Ferrovisor — a memory-safe, high-performance type-1 hypervisor built in Rust.
• serenity-alpha-lab — a research-only lab for stock, sector, and theme analysis inspired by Serenity-style reasoning. The goal is not to chase signals blindly, but to build clearer research workflows around narratives, fundamentals, and market structure.

Linux Work Plan — RISC-V upstream candidates

Source map: linux-riscv-docs/patch-work. These are snapshot-based candidates; before starting a patch, I re-check mainline, linux-next, maintainer trees, and lore.

• KVM / G-stage memory — build toward better RISC-V virtualization observability and lifecycle correctness: VIRT-01 G-stage / IOMMU ptdump, VIRT-02 lockless and reschedulable teardown, and VIRT-04 KVM_PRE_FAULT_MEMORY. Sources: ranked roadmap, KVM 2026H1 candidates.
• MMU / TLB / DMA correctness — focus on high-impact memory-management gaps such as MM-02 batched non-coherent DMA sync, MM-06 precise pte_needs_flush(), and MM-11 hot-remove range TLB batching. Source: MMU, memory, and TLB gap analysis.
• ISA-optimized kernel primitives — land small but measurable RISC-V-specific primitives first: ISA-01 Zbb memcmp, ISA-02 Zbb memchr, and ISA-03 Zvkg POLYVAL hooks. Sources: ISA optgap overview, string asm notes, crypto asm notes.
• Hardening and observability — help unblock higher-level tooling through CORE-01 reliable unwinder / livepatch review work, BPF stack-walk exception coverage, and CORE-16 ARCH_HAS_EXECMEM_ROX. Source: core ABI, observability, and hardening roadmap.
• Platform, ACPI, and RAS enablement — advance system-level readiness with PLAT-01 ACPI CPU physical hotplug, PLAT-06 CPPC FIE / RV32 READ_HI, and BOOT-01 crashkernel CMA wiring. Sources: platform, ACPI, NUMA, power, and RAS roadmap, patch-work overview.

• Linux kernel / RISC-V KVM — I have contributed across RISC-V KVM, guest statistics, interrupt reporting, guest extension enablement, dirty memory tracking, gstage mapping, VMID / nested virtualization exploration, and related kernel paths.
• Patch-first engineering — I treat upstream kernel work as a long-running trail of small, reviewable patches: some land, some evolve through review, and some become the next iteration of the design.
• Community PRs — I also count public PR work across major communities, whether merged or not, because the engineering value is in the attempt, discussion, review, and iteration. My PR trail spans projects such as CopilotKit, OpenHands, ragflow, SWE-agent, pydantic-ai, LlamaIndex, Haystack, Dify, anything-llm, Flowise, and related AI / agent ecosystems.
• Contribution surface — low-level systems, open-source AI infrastructure, agent frameworks, developer tooling, and workflows that make complex repositories easier to understand and improve.

The important layer is not the generated files. Code is the artifact; the system that produces, reviews, retries, and ships the code is the real thing to study.

My current engineering lens:

• Human value moves upward: direction, constraints, taste, judgment, and knowing what is worth building.
• Agents should not be treated as faster autocomplete. The real jump is the orchestration layer: planning, execution, review, notification, retry, and recovery.
• Good AI systems are closed loops, not one-shot prompts. They keep working when nobody is staring at the terminal.
• The bottleneck is no longer typing speed. It is problem decomposition, architectural clarity, feedback design, and the ability to distinguish signal from noise.
• The best systems amplify clear thinking. They do not replace it.

_{systems at the bottom · agents at the edge · orchestration in the loop}

• zcxGGmu's Blog — long-form notes on systems, AI, open-source work, investing, and durable workflows.

Latest posts, refreshed daily at 06:00 CST:

• Anti Aging Life Blood Vessel Glycation Sleep Exercise — 2026-07-31
• Liaofan Four Lessons Fate Cultivation Change Destiny — 2026-07-31
• Reverse Thinking Break Inertia One Percent Life — 2026-07-31
• Wealth Of Nations Market Division Labor Money Logic — 2026-07-30
• Employment Crisis Wealth Concentration Common Prosperity Distribution — 2026-07-28

_{Auto-refreshed daily at 06:00 CST · public memory details are privacy-redacted}

• Latest snapshot: 2026-07-31 06:02 CST; archive date 2026-07-30.
• Skills: 153 tracked / 153 active; today +1 Δ14 -0; activity +40, patches +5.
• Memory: 14 durable entries; today +1 -0; Memory map +1 added / 0 removed · details privacy-redacted.
• Signal: + open-source-contribution-camp… · new skill; Δ windows-process-management · patch +4, use +3, view +3