Makerfabs MaTouch ESP32-S3 2.8インチカメラ ESP32-S3 AIビジョン：カメラを向けて、それが何を見ているか尋ねる

/*
 * ===========================================================================
 *  05_Vision_AI  —  MaTouch AI ESP32-S3 2.8" TFT ST7789V
 * ===========================================================================
 *
 ----------
 *  ROBOJAX.COM  -  MaTouch AI ESP32-S3 2.8" project series
 *
 *    WATCH THE VIDEO
 *        https://youtu.be/6AL3g3tC_Hk
 *
 *    WRITTEN TUTORIALS - every project, with photos and full explanation
 *        Camera and touchscreen.... https://robojax.com/RTJ849
 *        Offline face recognition.. https://robojax.com/RTJ850
 *        AI voice assistant........ https://robojax.com/RTJ851
 *        AI vision................. https://robojax.com/RTJ852
 *
 *    GET THE BOARD - SAVE $5 with coupon code:  Robojax_Makerfab
 *        https://www.makerfabs.com/matouch-ai-esp32s3-2-8-tft-st7789v.html
 *        (enter the code at checkout)
 *
 *  All of this code is free. If it helped you, a subscribe on YouTube is
 *  the best way to support more of it.
 *  
 *
 *  Watching the video first will save you time - it shows the Arduino IDE
 *  settings and the library versions being set up step by step.
 *  ---------------------------------------------------------------------------
 *
 *  The board SEES. Point the camera at something, touch a button, and an
 *  OpenAI vision model tells you what it is looking at - drawn on the screen
 *  and (optionally) spoken out of the board's own speaker via Azure TTS.
 *
 *  Two modes, two buttons:
 *
 *      OBJECTS   "What do you see?" - names the things in front of the lens.
 *      PERSON    Describes the person in frame: glasses, expression, what
 *                they are doing. DESCRIPTION ONLY - it will not and must not
 *                try to say WHO someone is. Identifying people by face is
 *                against OpenAI's usage policies, and it is the right call:
 *                say this in the video, it is worth 15 honest seconds.
 *
 *  WHY OPENAI FOR THIS DEMO AND NOT DEEPSEEK: DeepSeek's API is text-only.
 *  It cannot accept an image at all. Azure OpenAI could do it, but plain
 *  OpenAI is one endpoint with no deployment setup - simplest to follow.
 *
 *  HOW THE IMAGE TRAVELS: the OV3660 gives us a JPEG directly (800x600,
 *  ~40 KB). We base64-encode it in PSRAM (~55 KB of text) and embed it in
 *  the JSON request as a data: URI. The 8 MB PSRAM makes this trivial.
 *
 *  ---------------------------------------------------------------------------
 *  FILL IN secrets.h BEFORE FLASHING
 *  (needs WIFI_*, OPENAI_*; AZURE_* only if SPEAK_REPLIES is 1).
 *
 *  BOARD SETTINGS (Tools menu - EVERY line matters, wrong = black screen
 *  or compile errors. These reset when you switch cores - recheck them!)
 *
 *      Board            : ESP32S3 Dev Module
 *      ESP32 core       : 2.0.17
 *      PSRAM            : OPI PSRAM        <-- required, image buffers live there
 *      Flash Size       : 16MB (128Mb)
 *      Partition Scheme : 16M Flash (3MB APP/9.9MB FATFS)
 *      USB CDC On Boot  : Disabled         <-- speaker shares pins with native USB
 *      Upload Speed     : 921600
 *      Port             : the CH340K USB-C port (the one near RESET)
 *
 *  LIBRARIES
 *      GFX Library for Arduino   v1.5.6   (NOT 1.6.x - that pairs with core 3)
 *      bb_captouch               v1.3.1
 *      ArduinoJson               v7.x
 *      Adafruit NeoPixel         any recent
 *
 *  ---------------------------------------------------------------------------
 *  FUNCTIONS IN THIS SKETCH
 *      led(r,g,b)          set the RGB status LED colour
 *      getTouch(&x,&y)     read the touch panel, mapped to screen coordinates
 *      camStart(...)       start the camera in a given format/size
 *      camStartPreview()   start the 240x240 RGB565 live-view camera
 *      captureJpeg(&len)   take one 800x600 JPEG into PSRAM (camera stays OFF
 *                          afterwards - caller restarts the preview)
 *      readHttpResponse()  read an HTTPS reply, de-chunking it properly
 *      askVision(...)      send photo + prompt to OpenAI, return the answer
 *      spkInit()           configure the I2S speaker output
 *      readExact(...)      read exactly N bytes from a TLS connection
 *      speak(text)         Azure TTS -> download voice to PSRAM -> play it
 *      playLastAnswer()    replay the kept voice from PSRAM (REPLAY button)
 *      drawButton(...)     draw one side-column button
 *      drawWifi()          WiFi signal bars + dBm readout
 *      drawButtons()       normal side column (the two ask buttons)
 *      drawClearSide()     answer-mode side column (CLEAR + REPLAY)
 *      showAnswer(text)    word-wrapped answer overlay on the viewfinder
 *      lookAndTell(...)    one full cycle: capture -> ask -> show -> speak
 *      setup() / loop()    boot sequence / viewfinder + touch handling
 *
 *  Robojax.com
 * ===========================================================================
 */

#define SPEAK_REPLIES  1     // 1 = read the answer aloud with Azure TTS, 0 = screen only

/* Status LED brightness, 0-255. The WS2812 runs from the power rail and is
 * uncomfortably bright at full power - 25 is plenty visible on camera. */
#define LED_BRIGHTNESS 15

/* CAMERA ORIENTATION
 * 1 = camera faces the SAME way as the screen (how the board ships) - the
 *     image is mirrored so it looks natural when you point it at yourself.
 * 0 = you folded the ribbon so the lens faces AWAY from the screen
 *     (phone-style, screen to you / camera to the subject).
 * If the picture looks left-right reversed, flip this number. */
#define CAMERA_FACES_USER  1

#include <Arduino_GFX_Library.h>
#include <bb_captouch.h>
#include <Adafruit_NeoPixel.h>
#include <ArduinoJson.h>
#include <Wire.h>
#include <WiFi.h>
#include <WiFiClientSecure.h>
#include <HTTPClient.h>
#include "mbedtls/base64.h"
#include "esp_camera.h"
#include "pins.h"
#include "secrets.h"

#if SPEAK_REPLIES
#include "driver/i2s.h"
#define WAV_HEADER_LEN 44
#endif

Arduino_ESP32SPI *bus = new Arduino_ESP32SPI(
    TFT_DC, TFT_CS, TFT_SCLK, TFT_MOSI, TFT_MISO, HSPI, true);
Arduino_GFX *gfx = new Arduino_ST7789(bus, TFT_RES, 1, true);

BBCapTouch bbct;
Adafruit_NeoPixel rgb(RGB_LED_NUM, RGB_LED_PIN, NEO_GRB + NEO_KHZ800);

/* --- the two vision prompts ------------------------------------------------
 * Keep answers short: they must fit a 320x240 screen and, if spoken, must not
 * leave the presenter waiting awkwardly on camera. */
const char *PROMPT_OBJECTS =
    "Look at this photo from a small camera. In ONE short sentence of at most "
    "20 words, plain text only, name the main object(s) you see.";

const char *PROMPT_PERSON =
    "Look at this photo. If there is a person, describe them in ONE short "
    "sentence of at most 20 words: mood, glasses or not, what they are doing. "
    "Never guess who they are. If no person is visible, say so. Plain text only.";

/* --- layout: viewfinder left, buttons right ------------------------------- */
#define BTN_X   242
#define BTN_W    78
#define BTN_H    52
#define BTN_OBJ_Y     4
#define BTN_PERSON_Y 62

void led(uint8_t r, uint8_t g, uint8_t b) { rgb.setPixelColor(0, rgb.Color(r, g, b)); rgb.show(); }


/* ===========================================================================
 *  Touch
 * =========================================================================== */
bool getTouch(uint16_t *x, uint16_t *y) {
  TOUCHINFO ti;
  if (!bbct.getSamples(&ti)) return false;
  if (ti.count < 1) return false;
  *x = ti.y[0];
  *y = (ti.x[0] > 240) ? 0 : (240 - ti.x[0]);
  return true;
}


/* ===========================================================================
 *  Camera  —  RGB565 for the live view; JPEG capture happens by restarting
 *  the driver, same technique as sketch 02.
 * =========================================================================== */
bool mirrored = CAMERA_FACES_USER;   // see the define at the top of the file

/* The viewfinder is OFF while a cloud call runs and while the answer is on
 * screen. Two reasons, both learned on the bench:
 *  1. RAM: the live camera driver eats the internal memory a TLS handshake
 *     needs - with the preview running, connecting to OpenAI fails (HTTP -1).
 *  2. Readability: the viewfinder repaints 25x/s and would wipe the answer.
 * The answer STAYS on screen until the user taps CLEAR - no timer. */
bool preview_on = false;

/* The last spoken answer is KEPT in PSRAM so the REPLAY button can play it
 * again without another cloud call. Overwritten by the next answer. */
uint8_t *last_audio = nullptr;
size_t   last_audio_len = 0;

bool camStart(pixformat_t fmt, framesize_t size, int fb_count, int quality) {
  camera_config_t c;
  c.ledc_channel = LEDC_CHANNEL_0;
  c.ledc_timer   = LEDC_TIMER_0;
  c.pin_d0 = CAM_PIN_D0;  c.pin_d1 = CAM_PIN_D1;
  c.pin_d2 = CAM_PIN_D2;  c.pin_d3 = CAM_PIN_D3;
  c.pin_d4 = CAM_PIN_D4;  c.pin_d5 = CAM_PIN_D5;
  c.pin_d6 = CAM_PIN_D6;  c.pin_d7 = CAM_PIN_D7;
  c.pin_xclk     = CAM_PIN_XCLK;
  c.pin_pclk     = CAM_PIN_PCLK;
  c.pin_vsync    = CAM_PIN_VSYNC;
  c.pin_href     = CAM_PIN_HREF;
  /* Share the I2C bus Wire already drives (the touch panel lives there too)
   * instead of letting the camera install a second driver on the same pins -
   * that kills touch. Wire.begin() must run before this function. */
  c.pin_sccb_sda = -1;
  c.pin_sccb_scl = -1;
  c.sccb_i2c_port = 0;               // Wire = I2C port 0
  c.pin_pwdn     = CAM_PIN_PWDN;
  c.pin_reset    = CAM_PIN_RESET;
  c.xclk_freq_hz = 20000000;
  c.frame_size   = size;
  c.pixel_format = fmt;
  c.grab_mode    = CAMERA_GRAB_WHEN_EMPTY;
  c.fb_location  = CAMERA_FB_IN_PSRAM;
  c.jpeg_quality = quality;
  c.fb_count     = fb_count;

  if (esp_camera_init(&c) != ESP_OK) return false;

  sensor_t *s = esp_camera_sensor_get();
  if (s) {
    s->set_hmirror(s, mirrored ? 1 : 0);
    s->set_vflip(s,   mirrored ? 1 : 0);
    s->set_brightness(s, 1);
  }
  return true;
}

bool camStartPreview() { return camStart(PIXFORMAT_RGB565, FRAMESIZE_240X240, 2, 12); }

/* Capture one SVGA JPEG into a PSRAM buffer the caller owns. */
uint8_t *captureJpeg(size_t *len_out) {
  esp_camera_deinit();
  delay(120);
  if (!camStart(PIXFORMAT_JPEG, FRAMESIZE_SVGA, 1, 12)) { *len_out = 0; return nullptr; }

  // a few warm-up frames so exposure settles
  for (int i = 0; i < 3; i++) {
    camera_fb_t *w = esp_camera_fb_get();
    if (w) esp_camera_fb_return(w);
    delay(100);
  }

  uint8_t *copy = nullptr;
  *len_out = 0;
  camera_fb_t *fb = esp_camera_fb_get();
  if (fb && fb->len > 0) {
    copy = (uint8_t *)ps_malloc(fb->len);
    if (copy) { memcpy(copy, fb->buf, fb->len); *len_out = fb->len; }
  }
  if (fb) esp_camera_fb_return(fb);

  /* Deliberately leave the camera OFF here. The caller restarts the preview
   * after the cloud call - a running camera driver starves the TLS handshake
   * of internal RAM (that was the "Vision HTTP -1" bug). */
  esp_camera_deinit();
  delay(120);
  return copy;
}


/* ===========================================================================
 *  HTTP response reader  —  shared by the cloud calls below.
 *  Returns the status code and fills body_out. Handles chunked transfer
 *  encoding PROPERLY: the chunk-size markers must be stripped, or they end
 *  up embedded inside the JSON and the parse fails. (OpenAI pretty-prints
 *  its replies so they span several chunks - this bit us on the bench.)
 * =========================================================================== */
/* Block until the connection has data (or the budget runs out). Every read
 * below goes through this, because the server can go quiet for many seconds
 * while it analyses the image - and a bare read() would simply time out. */
static bool waitData(WiFiClientSecure &c, uint32_t ms) {
  uint32_t t0 = millis();
  while (!c.available()) {
    if (!c.connected()) return false;
    if (millis() - t0 > ms) return false;
    delay(10);
  }
  return true;
}

static int readHttpResponse(WiFiClientSecure &client, String &body_out, uint32_t idle_ms) {
  body_out = "";

  if (!waitData(client, idle_ms)) { Serial.println("HTTP: no response at all"); return 0; }
  String status_line = client.readStringUntil('\n');
  int code = 0;
  sscanf(status_line.c_str(), "HTTP/%*s %d", &code);

  bool chunked = false;
  while (waitData(client, idle_ms)) {
    String h = client.readStringUntil('\n');
    if (h == "\r" || h.length() <= 1) break;         // blank line = end of headers
    h.toLowerCase();
    if (h.startsWith("transfer-encoding:") && h.indexOf("chunked") >= 0) chunked = true;
  }

  if (chunked) {
    int blanks = 0;
    while (true) {
      /* Waiting here - instead of letting read() time out - is essential: a
       * timed-out read looks exactly like "0" (final chunk), which silently
       * truncates the body to nothing while the server is still thinking. */
      if (!waitData(client, idle_ms)) {
        Serial.println("HTTP: timed out waiting for the next chunk");
        break;
      }
      String szline = client.readStringUntil('\n');
      szline.trim();
      if (szline.length() == 0) {
        if (++blanks > 4) break;
        continue;
      }
      blanks = 0;

      long sz = strtol(szline.c_str(), NULL, 16);
      if (sz <= 0) break;                            // genuine final chunk

      long got = 0;
      while (got < sz) {
        if (!waitData(client, idle_ms)) break;
        while (client.available() && got < sz) { body_out += (char)client.read(); got++; }
      }
      if (waitData(client, 3000)) client.readStringUntil('\n');
      if (got < sz) { Serial.println("HTTP: short chunk"); break; }
    }
  } else {
    while (waitData(client, idle_ms))
      while (client.available()) body_out += (char)client.read();
  }
  return code;
}


/* ===========================================================================
 *  OpenAI vision call
 *  The request body is built by hand in PSRAM rather than through ArduinoJson,
 *  because embedding a 55 KB base64 string in a JSON document would mean
 *  holding two copies. Base64 text never needs JSON escaping, so this is safe.
 * =========================================================================== */
bool askVision(const uint8_t *jpg, size_t jpg_len, const char *prompt, String &answer_out) {
  // --- base64 encode the image into PSRAM ---
  size_t b64_cap = ((jpg_len + 2) / 3) * 4 + 16;
  unsigned char *b64 = (unsigned char *)ps_malloc(b64_cap);
  if (!b64) return false;
  size_t b64_len = 0;
  if (mbedtls_base64_encode(b64, b64_cap, &b64_len, jpg, jpg_len) != 0) {
    free(b64);
    return false;
  }

  // --- assemble the JSON request around it ---
  /* Current OpenAI models reject the old "max_tokens" name - it must be
   * "max_completion_tokens" (verified against the live API, July 2026). */
  const char *head_fmt =
      "{\"model\":\"%s\",\"max_completion_tokens\":%d,\"messages\":[{\"role\":\"user\","
      "\"content\":[{\"type\":\"text\",\"text\":\"%s\"},"
      "{\"type\":\"image_url\",\"image_url\":{\"url\":\"data:image/jpeg;base64,";
  const char *tail = "\"}}]}]}";

  char head[640];
  int head_len = snprintf(head, sizeof(head), head_fmt, OPENAI_MODEL, LLM_MAX_TOKENS, prompt);

  size_t body_len = head_len + b64_len + strlen(tail);
  char *body = (char *)ps_malloc(body_len + 1);
  if (!body) { free(b64); return false; }
  memcpy(body, head, head_len);
  memcpy(body + head_len, b64, b64_len);
  strcpy(body + head_len + b64_len, tail);
  free(b64);

  // --- send it: manual HTTP, chunk-wise upload (proven pattern from 04) ---
  WiFiClientSecure client;
  client.setInsecure();
  client.setTimeout(25);

  if (!client.connect(OPENAI_HOST, 443)) {
    Serial.println("Vision: TLS connect failed (is the camera still running?)");
    free(body);
    return false;
  }

  client.print(String("POST /v1/chat/completions HTTP/1.1\r\n"
               "Host: " OPENAI_HOST "\r\n"
               "Authorization: Bearer " OPENAI_KEY "\r\n"
               "Content-Type: application/json\r\n"
               "Connection: close\r\n"
               "Content-Length: ") + String(body_len) + "\r\n\r\n");

  size_t sent = 0;
  while (sent < body_len) {
    size_t n = min((size_t)4096, body_len - sent);
    size_t w = client.write((uint8_t *)body + sent, n);
    if (w == 0) {
      delay(50);
      w = client.write((uint8_t *)body + sent, n);
      if (w == 0) break;
    }
    sent += w;
    yield();
  }
  free(body);
  if (sent < body_len) {
    Serial.printf("Vision: upload stalled at %u/%u bytes\n",
                  (unsigned)sent, (unsigned)body_len);
    client.stop();
    return false;
  }

  /* read the reply with proper de-chunking */
  String resp;
  int code = readHttpResponse(client, resp, 30000);
  client.stop();

  if (code != 200) {
    Serial.printf("Vision HTTP %d: %s\n", code, resp.c_str());
    return false;
  }

  bool ok = false;
  JsonDocument doc;
  if (!deserializeJson(doc, resp)) {
    const char *content = doc["choices"][0]["message"]["content"];
    if (content) {
      answer_out = String(content);
      answer_out.trim();
      ok = answer_out.length() > 0;
    }
  } else {
    Serial.printf("Vision: JSON parse failed, %u bytes received\n", resp.length());
  }
  return ok;
}


/* ===========================================================================
 *  Azure TTS  —  same streaming trick as sketch 04: raw PCM into I2S.
 * =========================================================================== */
#if SPEAK_REPLIES
void spkInit() {
  i2s_config_t cfg = {
    .mode = (i2s_mode_t)(I2S_MODE_MASTER | I2S_MODE_TX),
    .sample_rate = 16000,
    .bits_per_sample = I2S_BITS_PER_SAMPLE_16BIT,
    .channel_format = I2S_CHANNEL_FMT_ONLY_LEFT,
    .communication_format = I2S_COMM_FORMAT_STAND_I2S,
    .intr_alloc_flags = ESP_INTR_FLAG_LEVEL1,
    .dma_buf_count = 8,
    .dma_buf_len = 256,
    .use_apll = false,
    .tx_desc_auto_clear = true,
    .fixed_mclk = 0
  };
  i2s_pin_config_t pins = {
    .mck_io_num = I2S_PIN_NO_CHANGE,
    .bck_io_num = I2S_SPK_BCLK,
    .ws_io_num = I2S_SPK_LRC,
    .data_out_num = I2S_SPK_DOUT,
    .data_in_num = I2S_PIN_NO_CHANGE
  };
  i2s_driver_install(I2S_SPK_PORT, &cfg, 0, NULL);
  i2s_set_pin(I2S_SPK_PORT, &pins);
  i2s_zero_dma_buffer(I2S_SPK_PORT);
}

/* Read exactly n bytes from a client (or until timeout). */
static size_t readExact(WiFiClientSecure &c, uint8_t *dst, size_t n) {
  size_t got = 0;
  uint32_t t0 = millis();
  while (got < n && millis() - t0 < 10000) {
    int r = c.read(dst + got, n - got);
    if (r > 0) { got += r; t0 = millis(); }
    else if (!c.connected() && !c.available()) break;
    else delay(2);
  }
  return got;
}

void playLastAnswer();   // defined below; explicit prototype for the IDE

void speak(const String &text) {
  String safe = text;
  safe.replace("&", "&amp;");
  safe.replace("<", "&lt;");
  safe.replace(">", "&gt;");
  String ssml = "<speak version='1.0' xml:lang='" AZURE_TTS_LANG "'>"
                "<voice name='" AZURE_TTS_VOICE "'>" + safe + "</voice></speak>";

  /* Manual HTTP with proper de-chunking: Azure sends this audio chunked, and
   * HTTPClient's raw stream leaks the ASCII chunk-size lines into the PCM -
   * each one plays as an audible KNOCK. */
  WiFiClientSecure client;
  client.setInsecure();
  client.setTimeout(20);

  const char *host = AZURE_REGION ".tts.speech.microsoft.com";
  if (!client.connect(host, 443)) { Serial.println("TTS: TLS connect failed"); return; }

  client.print(String("POST /cognitiveservices/v1 HTTP/1.1\r\n"
               "Host: ") + host + "\r\n"
               "Ocp-Apim-Subscription-Key: " AZURE_SPEECH_KEY "\r\n"
               "Content-Type: application/ssml+xml\r\n"
               "X-Microsoft-OutputFormat: riff-16khz-16bit-mono-pcm\r\n"
               "User-Agent: MaTouchRobojax\r\n"
               "Connection: close\r\n"
               "Content-Length: " + String(ssml.length()) + "\r\n\r\n");
  client.print(ssml);

  String status_line = client.readStringUntil('\n');
  int code = 0;
  sscanf(status_line.c_str(), "HTTP/%*s %d", &code);
  bool chunked = false;
  long content_len = -1;
  while (client.connected() || client.available()) {
    String h = client.readStringUntil('\n');
    if (h == "\r" || h.length() <= 1) break;
    h.toLowerCase();
    if (h.startsWith("transfer-encoding:") && h.indexOf("chunked") >= 0) chunked = true;
    if (h.startsWith("content-length:")) content_len = h.substring(15).toInt();
  }
  if (code != 200) { Serial.printf("TTS HTTP %d\n", code); client.stop(); return; }

  const size_t AUDIO_CAP = 1200 * 1024;
  uint8_t *audio = (uint8_t *)ps_malloc(AUDIO_CAP);
  if (!audio) { client.stop(); return; }
  size_t alen = 0;

  if (chunked) {
    while (true) {
      String szline = client.readStringUntil('\n');
      long sz = strtol(szline.c_str(), NULL, 16);
      if (sz <= 0) break;
      if (alen + sz > AUDIO_CAP) break;
      size_t got = readExact(client, audio + alen, sz);
      alen += got;
      client.readStringUntil('\n');
      if (got < (size_t)sz) break;
    }
  } else if (content_len > 0) {
    alen = readExact(client, audio, min((size_t)content_len, AUDIO_CAP));
  } else {
    uint32_t idle = millis();
    while ((client.connected() || client.available()) && millis() - idle < 5000) {
      int r = client.read(audio + alen, min((size_t)2048, AUDIO_CAP - alen));
      if (r > 0) { alen += r; idle = millis(); }
      else delay(5);
    }
  }
  client.stop();

  if (alen > WAV_HEADER_LEN) {
    /* keep this answer for the REPLAY button (replacing the previous one),
     * then play it */
    if (last_audio) free(last_audio);
    last_audio = audio;
    last_audio_len = alen;
    playLastAnswer();
  } else {
    free(audio);
  }
}

/* Play the kept answer from PSRAM - used right after download AND by REPLAY. */
void playLastAnswer() {
  if (!last_audio || last_audio_len <= WAV_HEADER_LEN) return;
  static const uint8_t lead_in[640] = {0};     // 20 ms silence pre-roll
  size_t w = 0;
  i2s_write(I2S_SPK_PORT, lead_in, sizeof(lead_in), &w, portMAX_DELAY);
  i2s_write(I2S_SPK_PORT, last_audio + WAV_HEADER_LEN,
            last_audio_len - WAV_HEADER_LEN, &w, portMAX_DELAY);
  i2s_write(I2S_SPK_PORT, lead_in, sizeof(lead_in), &w, portMAX_DELAY);
  delay(150);
  i2s_zero_dma_buffer(I2S_SPK_PORT);
}
#endif  // SPEAK_REPLIES


/* ===========================================================================
 *  UI
 * =========================================================================== */
void drawButton(int y, const char *l1, const char *l2, uint16_t colour) {
  gfx->fillRoundRect(BTN_X, y, BTN_W, BTN_H, 6, colour);
  gfx->drawRoundRect(BTN_X, y, BTN_W, BTN_H, 6, WHITE);
  gfx->setTextSize(1);
  gfx->setTextColor(WHITE);
  gfx->setCursor(BTN_X + 8, y + 14);
  gfx->print(l1);
  gfx->setCursor(BTN_X + 8, y + 28);
  gfx->print(l2);
}

/* WiFi bars + dBm at the bottom of the side column, refreshed from the loop */
void drawWifi() {
  gfx->fillRect(BTN_X, 188, BTN_W, 52, BLACK);
  bool up = (WiFi.status() == WL_CONNECTED);
  long rssi = up ? WiFi.RSSI() : -100;
  int bars = rssi > -55 ? 4 : rssi > -65 ? 3 : rssi > -75 ? 2 : rssi > -85 ? 1 : 0;

  for (int b = 0; b < 4; b++) {
    int bh = 6 + b * 6;
    uint16_t col = (b < bars) ? GREEN : gfx->color565(60, 60, 60);
    gfx->fillRect(BTN_X + 4 + b * 9, 216 - bh, 7, bh, col);
  }
  gfx->setTextSize(1);
  gfx->setCursor(BTN_X + 44, 196);
  if (up) {
    gfx->setTextColor(CYAN);
    gfx->printf("%ld", rssi);
  } else {
    gfx->setTextColor(RED);
    gfx->print("DOWN");
  }
  gfx->setCursor(BTN_X + 44, 208);
  gfx->setTextColor(gfx->color565(120, 120, 120));
  gfx->print("dBm");
  gfx->setCursor(BTN_X + 4, 228);
  gfx->setTextColor(gfx->color565(120, 120, 120));
  gfx->print("WiFi signal");
}

/* normal side column: the two ask buttons */
void drawButtons() {
  gfx->fillRect(BTN_X, 0, BTN_W, 188, BLACK);
  drawButton(BTN_OBJ_Y,    "WHAT DO",  "YOU SEE?", gfx->color565(0, 90, 160));
  drawButton(BTN_PERSON_Y, "DESCRIBE", "PERSON",   gfx->color565(120, 60, 140));

  gfx->setTextColor(CYAN);
  gfx->setCursor(BTN_X + 2, 128);
  gfx->print("OpenAI eyes");
  gfx->setCursor(BTN_X + 2, 140);
  gfx->print("Azure voice");
  gfx->setCursor(BTN_X + 2, 152);
  gfx->print("Robojax.com");

  drawWifi();
}

/* answer-mode side column: CLEAR (back to camera) + REPLAY (say it again).
 * The answer stays on screen until CLEAR is tapped. */
void drawClearSide() {
  gfx->fillRect(BTN_X, 0, BTN_W, 188, BLACK);
  gfx->fillRoundRect(BTN_X, BTN_OBJ_Y, BTN_W, BTN_H, 6, gfx->color565(110, 35, 35));
  gfx->drawRoundRect(BTN_X, BTN_OBJ_Y, BTN_W, BTN_H, 6, WHITE);
  gfx->setTextSize(2);
  gfx->setTextColor(WHITE);
  gfx->setCursor(BTN_X + 10, BTN_OBJ_Y + 14);
  gfx->print("CLEAR");

#if SPEAK_REPLIES
  if (last_audio_len > 0) {
    gfx->fillRoundRect(BTN_X, BTN_PERSON_Y, BTN_W, BTN_H, 6, gfx->color565(0, 110, 60));
    gfx->drawRoundRect(BTN_X, BTN_PERSON_Y, BTN_W, BTN_H, 6, WHITE);
    gfx->setTextSize(1);
    gfx->setTextColor(WHITE);
    gfx->setCursor(BTN_X + 16, BTN_PERSON_Y + 20);
    gfx->print("REPLAY");
  }
#endif

  gfx->setTextSize(1);
  gfx->setTextColor(YELLOW);
  gfx->setCursor(BTN_X + 2, 124);
  gfx->print("CLEAR = camera");
  gfx->setCursor(BTN_X + 2, 136);
  gfx->print("REPLAY = again");

  drawWifi();
}

/* Word-wrapped answer overlay across the bottom of the viewfinder. */
void showAnswer(const String &text) {
  const int max_chars = 38;
  int lines = (text.length() + max_chars - 1) / max_chars;
  if (lines > 6) lines = 6;
  int h = lines * 10 + 8;
  int y = 240 - h;

  gfx->fillRect(0, y, 240, h, gfx->color565(0, 0, 0));
  gfx->drawRect(0, y, 240, h, CYAN);
  gfx->setTextSize(1);
  gfx->setTextColor(WHITE);
  for (int i = 0; i < lines; i++) {
    gfx->setCursor(4, y + 5 + i * 10);
    gfx->print(text.substring(i * max_chars, min((int)text.length(), (i + 1) * max_chars)));
  }
}


/* ===========================================================================
 *  One full "look and tell" cycle
 * =========================================================================== */
void lookAndTell(const char *prompt, const char *label) {
  led(255, 120, 0);                                        // amber: working
  preview_on = false;                                      // camera goes OFF here

  gfx->fillRect(0, 0, 240, 240, BLACK);
  gfx->setTextColor(YELLOW);
  gfx->setTextSize(1);
  gfx->setCursor(30, 110);
  gfx->printf("capturing photo (%s)...", label);

  size_t jpg_len = 0;
  uint32_t t_cap = millis();
  uint8_t *jpg = captureJpeg(&jpg_len);
  t_cap = millis() - t_cap;

  if (!jpg || jpg_len == 0) {
    if (jpg) free(jpg);
    showAnswer("Capture failed - tap CLEAR to retry.");
    led(255, 0, 0);
    drawClearSide();
    return;
  }
  Serial.printf("Captured %u KB in %lu ms\n", (unsigned)(jpg_len / 1024), (unsigned long)t_cap);

  gfx->setCursor(30, 124);
  gfx->printf("asking OpenAI (%uKB)...", (unsigned)(jpg_len / 1024));

  String answer;
  uint32_t t_ai = millis();
  bool ok = askVision(jpg, jpg_len, prompt, answer);
  t_ai = millis() - t_ai;
  free(jpg);

  if (!ok) {
    showAnswer("No answer - see serial monitor for the reason.");
    led(255, 0, 0);
    drawClearSide();
    return;
  }

  Serial.printf("Vision (%lu ms): %s\n", (unsigned long)t_ai, answer.c_str());
  showAnswer(answer);
  drawClearSide();                                         // CLEAR replaces the ask buttons

#if SPEAK_REPLIES
  led(0, 255, 40);                                         // green: speaking
  speak(answer);                                           // answer stays on screen
#endif
  led(0, 0, 0);
  /* the answer now stays until the user taps CLEAR */
}


/* ===========================================================================
 *  SETUP
 * =========================================================================== */
void setup() {
  Serial.begin(115200);
  delay(400);
  Serial.println("\n=== 05 Vision AI  |  Robojax.com ===");

  pinMode(TFT_BLK, OUTPUT);
  digitalWrite(TFT_BLK, LOW);
  pinMode(SD_CS, OUTPUT);
  digitalWrite(SD_CS, HIGH);

  gfx->begin();
  gfx->fillScreen(BLACK);
  digitalWrite(TFT_BLK, HIGH);

  bbct.init(TOUCH_SDA, TOUCH_SCL, TOUCH_RST, TOUCH_INT);
  delay(50);

  // The camera's SCCB shares this bus, so Wire must be up before the camera
  Wire.begin(I2C_SDA, I2C_SCL, 100000);
  delay(20);

  rgb.begin();
  rgb.setBrightness(LED_BRIGHTNESS);
  led(0, 0, 0);

#if SPEAK_REPLIES
  spkInit();
#endif

  gfx->setTextSize(1);
  gfx->setTextColor(YELLOW);
  gfx->setCursor(4, 4);
  gfx->printf("Connecting to %s ...", WIFI_SSID);

  WiFi.mode(WIFI_STA);
  WiFi.begin(WIFI_SSID, WIFI_PASS);
  uint32_t t0 = millis();
  while (WiFi.status() != WL_CONNECTED && millis() - t0 < 20000) delay(300);

  gfx->setCursor(4, 16);
  if (WiFi.status() == WL_CONNECTED) {
    gfx->setTextColor(GREEN);
    gfx->print("WiFi ok");
  } else {
    gfx->setTextColor(RED);
    gfx->print("WiFi FAILED (2.4GHz only! check secrets.h)");
  }

  gfx->setCursor(4, 28);
  gfx->setTextColor(YELLOW);
  gfx->print("starting camera...");
  if (!camStartPreview()) {
    gfx->fillScreen(RED);
    gfx->setTextColor(WHITE);
    gfx->setTextSize(2);
    gfx->setCursor(20, 100);
    gfx->print("CAMERA FAILED");
    gfx->setTextSize(1);
    gfx->setCursor(20, 130);
    gfx->print("Press RESET (camera reset = board reset)");
    while (1) delay(1000);
  }

  delay(400);
  gfx->fillScreen(BLACK);
  drawButtons();
  preview_on = true;
  Serial.println("Running. Touch a button to have the AI look through the camera.");
}


/* ===========================================================================
 *  LOOP
 * =========================================================================== */
void loop() {
  static uint32_t last_touch = 0;

  if (preview_on) {
    camera_fb_t *fb = esp_camera_fb_get();
    if (fb) {
      gfx->draw16bitBeRGBBitmap(0, 0, (uint16_t *)fb->buf, fb->width, fb->height);
      esp_camera_fb_return(fb);
    }
  } else {
    delay(20);                               // answer on screen, waiting for CLEAR
  }

  /* live WiFi signal, refreshed every 2 s in both modes */
  static uint32_t last_wifi = 0;
  if (millis() - last_wifi > 2000) {
    last_wifi = millis();
    drawWifi();
  }

  /* Edge-detected touch: an action fires only on a NEW finger-down, and the
   * finger must fully lift (4 consecutive empty reads) before anything can
   * fire again. This is what stops one tap on CLEAR from also triggering the
   * capture button that appears in the same spot a moment later. */
  static bool    touch_down = false;
  static uint8_t release_count = 0;

  uint16_t x, y;
  if (getTouch(&x, &y)) {
    release_count = 0;
    if (!touch_down) {
      touch_down = true;                     // new tap - act exactly once

      if (!preview_on) {
        /* answer mode: CLEAR returns to the camera, REPLAY says it again */
        if (x >= BTN_X && y >= BTN_OBJ_Y && y < BTN_OBJ_Y + BTN_H) {
          if (camStartPreview()) {
            preview_on = true;
            drawButtons();
          } else {
            showAnswer("Camera restart failed - press RESET.");
          }
        }
#if SPEAK_REPLIES
        else if (x >= BTN_X && y >= BTN_PERSON_Y && y < BTN_PERSON_Y + BTN_H) {
          led(0, 255, 40);                   // green while speaking
          playLastAnswer();
          led(0, 0, 0);
        }
#endif
      } else if (x >= BTN_X && y >= BTN_OBJ_Y && y < BTN_OBJ_Y + BTN_H) {
        lookAndTell(PROMPT_OBJECTS, "objects");
      } else if (x >= BTN_X && y >= BTN_PERSON_Y && y < BTN_PERSON_Y + BTN_H) {
        lookAndTell(PROMPT_PERSON, "person");
      }
    }
  } else if (touch_down) {
    if (++release_count >= 4) { touch_down = false; release_count = 0; }
  }
  (void)last_touch;
}