dsh-ls125-uv-probe

A DeepSeek Harness skill plugin for the 林上 LinShang LS125 紫外辐照计 and its UVALED-X3 UV probe: the private UART protocol between the meter and the probe, reverse-engineered end to end, plus the tooling to capture and decode it and a verified host firmware that replaces the original meter.

中文说明：README.zh.md

Install it and your agent gains a ls125-uv-probe skill — the protocol table, the commands, the toolchain usage, and the traps that make a working setup look dead.

What it answers

Question	Answer
Can I read this probe from my own MCU, without the original meter?	Yes — verified on an STM32F103: 21/21 frames, zero CRC errors, zero timeouts
What is the protocol?	9600 8N1, polled every 500 ms, 28-byte replies, CRC-16/MODBUS
Is there a hidden handshake, challenge, or ID check?	No — the poll command is a constant 8-byte string
Can I reset the probe's records?	Yes — a second command; it clears energy, minimum, and the elapsed-time counter
Do I have to send the meter's startup configuration?	No — polling alone is enough (verified)

Install

dsh plugin --profile web add dsh-ls125-uv-probe

Or straight from this repository:

dsh plugin --profile web add github:2333hutao/dsh-ls125-uv-probe

Protocol at a glance

主机 → 探头  轮询（每 500 ms）  AB 20 60 00 14 00 89 07
             清零             AB 21 20 00 02 00 01 00 3D 8A

探头 → 主机  28 字节帧，帧头 AB 20 60 00 14 00
  偏移  6–9   float LE   实时辐照度          µW/cm²（表显 mW/cm² = ÷1000）
  偏移 10–13  float LE   累积能量            µJ/cm²
  偏移 14–15  uint16 LE  距上次清零的时间     ×0.1 s
  偏移 16–17  2 字节     恒 0（含义未定）
  偏移 18–21  float LE   最小值              µW/cm²
  偏移 22–25  float LE   平均辐照度          µW/cm²
  偏移 26–27  uint16 LE  CRC-16/MODBUS（over 0–25，poly 0xA001 / init 0xFFFF / LSB-first）

Frame values are always µW/cm² and µJ/cm² regardless of the meter's display unit setting, so a host never has to track that setting. The maximum is not in the frame — the meter computes it, and so should you.

What ships in here

skill/SKILL.md      the skill body (protocol quick table, commands, toolchain usage, hard rules)
skill/PROTOCOL.md   full specification, command families, reference C implementation
skill/LESSONS.md    transferable methodology + 13 measured pitfalls with their verdicts
skill/AGENTS.md     a takeover briefing for an agent landing in the workspace
skill/tools/        fx2cap.cjs (capture, official libusb + fx2lafw) and four decode/analysis scripts
skill/firmware/     f103_probe (verified host) and f103_rgb (controllable light source)
skill/captures/     a sample capture so the toolchain can be exercised with no hardware

The toolchain bypasses a long-standing sigrok problem: every Windows sigrok build statically links a 2016 libusb whose backend dispatch fails on some systems, so fx2lafw reports LIBUSB_ERROR_NOT_SUPPORTED while --scan still lists the device. Reinstalling drivers does not help. fx2cap.cjs talks to the device with an official libusb instead, and adds what sigrok lacks: no capture duration cap and a soft trigger.

Evidence, not assertions

Every claim above was measured, and the repository keeps the measurements: three rounds of "frame value ↔ meter display" matched to all four decimals, a 65536-polynomial brute force that found the checksum, a halt test that proved the probe never speaks unless polled, and four independent experiments that ruled D4 out as a reset line.

How this was built — AI authorship disclosure

All of the code, firmware, tooling and documentation in this repository was written by an AI agent (DeepSeek Harness), working from measurements taken on real hardware.

The protocol knowledge here was not copied from a datasheet, a vendor document, or a third-party implementation — it was captured off the wire, cross-checked against the meter's own display, and every claim is backed by a recorded experiment. Nothing in skill/ is asserted without a measurement behind it, and the few open items are labelled as open.

A human operator did the parts an AI cannot: wiring, power, probes, pressing the meter's own keys, and reporting what its display showed.

Review it as you would any machine-written code before relying on it. The measurement evidence is all here, so the claims can be re-checked rather than taken on trust.

License

MIT