Agent Code Generation Guide

[中文]

This guide is normative for AI agents that generate ESP-VISION application code. Treat the target as an ESP32 microcontroller running the ESP-VISION build of MicroPython v1.28.0, not as a PC running CPython. Use only APIs present in this knowledge pack, enabled for the selected board, or demonstrated by a repository example; never assume that a desktop standard-library module, pip package, native wheel, operating-system service, subprocess, or PC filesystem is available on the device.

Dependency Rules

The default ESP-VISION firmware cannot dynamically install dependencies from requirements.txt and does not include device-side mip. Generated target code must not invoke pip, parse requirements.txt, call mip.install(), or silently depend on an unverified third-party package. If a required module is not in the knowledge pack, either rewrite the solution with supported modules or explicitly state that it must be copied to /lib with a host tool, precompiled as compatible .mpy, frozen into custom firmware, or implemented as a native module before the script can run.

Host-side tooling is separate from target code. A PC utility may use CPython packages such as pyserial, but label that code as host-only and never import those packages from the ESP-VISION script.

All-in-One Output

Unless the user asks for a package or a multi-file architecture, return one self-contained main.py for the device. Put supported imports, constants, initialization, application logic, the main loop or bounded test, error handling, and cleanup in that file. Avoid helper modules that the user would have to create separately. Models and other binary assets may remain separate files, but list their exact device paths and deployment prerequisites.

Prefer bounded, observable verification code before an infinite product loop. Print a unique success marker only after the camera, model, storage, network, or other requested operation has really succeeded. The following target-side smoke test is an all-in-one MicroPython script and prints EVTEST:OK only after capturing a frame:

import gc
import sensor


def main():
    print("EVTEST:START")
    sensor.reset()
    try:
        sensor.set_pixformat(sensor.RGB565)
        sensor.set_framesize(sensor.QQVGA)
        sensor.skip_frames(time=500)
        image = sensor.snapshot()
        width = image.width()
        height = image.height()
    finally:
        sensor.shutdown()
    print("EVTEST:OK {}x{} heap={}".format(
        width, height, gc.mem_free()))


main()

Device Communication and Acceptance

Agent is a host application alongside IDE and CLI; use Device Core for device interaction rather than another EV-MUX client. The shared host/Core/protocol/firmware relationship is in Solution Architecture. Use Communication System and EV-MUX for acquisition, upload, execution and application evidence. The complete host-only camera closed-loop example is in Communication acceptance, closed-loop tests and troubleshooting. Handshake, upload acknowledgement, echo and REPL prompts do not prove application success.