Files
AvianVisitors/avian/api/cutout.php
T
Twarner491 2ec45b9138 AvianVisitors hardening pass — install via fork's newinstaller
Restructures the install so users run ONE command. The fork's
newinstaller.sh clones this repo (instead of upstream Nachtzuster) and
runs BirdNET-Pi's installer. install_services.sh symlinks the avian/
overlay into the Caddy site root, so http://birdnet.local/avian/ comes
up after reboot with no Caddy config needed.

Reviewer-flagged fixes:

Frontend (avian/frontend/apt.js)
- Switch all API paths from /api/*.json (which didn't route) to relative
  action-based URLs (./api/birdnet-api.php?action=recent etc.).
- Fetch ./masks.json + ./dims.json relatively (were /avian/*.json).
- Replace every innerHTML user-content sink with createElement +
  textContent + setAttribute — species labels are user-editable in
  BirdNET-Pi, so untrusted.
- Add esc() helper for the few remaining innerHTML attribute contexts.
- fetchJson throws on non-200 (was silently parsing 404 HTML as JSON).
- Defensive null-check in loadMask if MASKS hasn't loaded.

Backend (avian/api/*.php)
- birdnet-api.php now derives DB_PATH from dirname(__DIR__, 3) so any
  BirdNET-Pi install user works (no hardcoded /home/birdnet).
- recording/spectrogram/cutout use getenv('HOME') for the same reason.
- cutout.php reworked as the canonical image resolver: bundled
  illustration → bundled cutout → cached rembg → fresh
  Wikipedia+rembg, with the binomial regex guard and atomic rename
  on cache write. Wikipedia URL host pinned to wikimedia.org /
  wikipedia.org to block SSRF.
- Stale 'X-BirdNET-Proxy-Token' auth claims removed everywhere.

Pregen (avian/scripts/pregen.py)
- Gemini key moved from URL query to x-goog-api-key header.
- Model bumped to gemini-2.5-flash-image-preview.
- Bounded retry on 429 + 5xx with Retry-After honoured.
- Default --sleep raised from 1s to 4s (under free-tier RPM).
- Unified parser handles |/_/comma in all three input modes.
- responseModalities now [TEXT, IMAGE] (was IMAGE-only).
- --poses validated against POSES keys.
- ASCII [ok]/[fail] markers (was ✓/✗ — Windows console crash).
- Surfaces finishReason + blockReason on safety blocks.

Prompt template
- 'Two placeholders' typo → three.
- 'warm paper' / 'transparent background' contradiction resolved.

Forwarding
- mqtt-bridge.py: paho-mqtt 2.x CallbackAPIVersion compat shim.
- avian-mqtt.service: dropped broken %i, set User=birdnet + %h path.
- forwarding/README.md uses the actual action-based URLs.

README
- One-line install via curl … newinstaller.sh.
- No more 'install BirdNET-Pi separately' step (our fork does it).
- License section uses absolute github.com links.
- bird.onethreenine.net replaces the not-yet-existent project writeup.
- docs/thumb.png replaced with the 24-bird Twitter capture.

Removed
- avian/caddy/ and avian/scripts/install.sh — no longer needed; the
  overlay symlink in install_services.sh handles everything.
2026-05-28 10:49:41 -07:00

170 lines
5.8 KiB
PHP

<?php
// AvianVisitors — bird image resolver.
//
// Lookup chain for /avian/api/cutout.php?sci=Calypte+anna:
// 1. ../assets/illustrations/<slug>.png (450+ bundled kachō-e renders)
// 2. ../assets/cutouts/<slug>.png (background-removed photo)
// 3. cached rembg of a Wikipedia photo at $HOME/BirdSongs/Extracted/cutouts/
// 4. fresh Wikipedia → rembg → cache (skipped gracefully if rembg unset)
//
// The frontend's <img src> points here for every species — bundled
// hits return instantly; cold misses fall through to the dynamic path.
//
// Default LAN deploy ships without auth. To expose publicly, gate
// /avian/api/* with basic_auth in your Caddyfile — see avian/forwarding/.
declare(strict_types=1);
$sci = trim((string)($_GET['sci'] ?? ''));
if ($sci === '') {
http_response_code(400);
echo 'sci required';
exit;
}
// Binomial / trinomial pattern. Rejects path-traversal payloads and
// junk before any filesystem or upstream lookup.
if (!preg_match('/^[A-Za-z]{2,40}(?:[ ][a-z]{2,40}){1,3}$/', $sci)) {
http_response_code(400);
echo 'invalid sci';
exit;
}
// Slugify scientific name for filename + cache key.
$slug = preg_replace('/[^a-z0-9]+/', '-', strtolower($sci));
$slug = trim((string)$slug, '-');
function serve_png(string $path): void {
header('Content-Type: image/png');
header('Cache-Control: public, max-age=86400');
header('Content-Length: ' . (string)filesize($path));
readfile($path);
exit;
}
// 1. Bundled illustration (the kachō-e PNG the repo ships with).
$bundled = dirname(__DIR__) . "/assets/illustrations/$slug.png";
if (is_file($bundled) && filesize($bundled) > 1024) {
serve_png($bundled);
}
// 2. Bundled cutout (background-removed photo, fallback for species
// without an illustration).
$cutout = dirname(__DIR__) . "/assets/cutouts/$slug.png";
if (is_file($cutout) && filesize($cutout) > 1024) {
serve_png($cutout);
}
// 3. Dynamic cache from a previous Wikipedia + rembg run.
$cacheDir = getenv('HOME') . '/BirdSongs/Extracted/cutouts';
$cachePath = "$cacheDir/$slug.png";
if (is_file($cachePath) && filesize($cachePath) > 1024) {
serve_png($cachePath);
}
// 4. Fresh Wikipedia fetch + rembg. Skipped if rembg-cli isn't on
// PATH — the resolver simply returns a 404 in that case rather
// than burning a Wikipedia request we can't use.
$rembg = '/usr/local/bin/rembg-cli';
if (!is_executable($rembg)) {
http_response_code(404);
echo 'no illustration bundled for ' . htmlspecialchars($sci) . ' (install rembg-cli to enable Wikipedia fallback)';
exit;
}
if (!is_dir($cacheDir)) @mkdir($cacheDir, 0755, true);
// Wikipedia's REST API asks for a contact-able identifier. Override
// via the AV_USER_AGENT env var (set in /etc/php/*/fpm/pool.d/www.conf
// or your shell) if your install hammers their endpoint at scale.
$ua = getenv('AV_USER_AGENT') ?: 'AvianVisitors/1.0 (+https://github.com/Twarner491/AvianVisitors)';
$ctx = stream_context_create([
'http' => ['header' => "User-Agent: $ua\r\n", 'timeout' => 12],
]);
$wpUrl = 'https://en.wikipedia.org/api/rest_v1/page/summary/' . rawurlencode($sci);
$wpJson = @file_get_contents($wpUrl, false, $ctx);
$srcUrl = null;
if ($wpJson !== false) {
$j = json_decode($wpJson, true);
$srcUrl = $j['originalimage']['source'] ?? $j['thumbnail']['source'] ?? null;
}
// Defensive: only follow URLs on Wikimedia / Wikipedia hosts so a
// poisoned summary endpoint can't redirect us to arbitrary servers.
if ($srcUrl !== null) {
$host = parse_url((string)$srcUrl, PHP_URL_HOST) ?: '';
if (!preg_match('/(?:^|\.)(?:wikimedia\.org|wikipedia\.org)$/i', $host)) {
$srcUrl = null;
}
}
if (!$srcUrl) {
http_response_code(404);
echo 'no Wikipedia photo for ' . htmlspecialchars($sci);
exit;
}
$imgBytes = @file_get_contents($srcUrl, false, $ctx);
if (!$imgBytes || strlen($imgBytes) < 1024) {
http_response_code(503);
echo 'failed to fetch source image';
exit;
}
// rembg via the wrapper. u2netp = lightweight model (~50MB peak RAM —
// matters on the Pi 3B+). Temp files because rembg's CLI prefers
// real paths.
$tmpInBase = tempnam(sys_get_temp_dir(), 'rembg-in-');
$tmpOutBase = tempnam(sys_get_temp_dir(), 'rembg-out-');
@unlink($tmpInBase); @unlink($tmpOutBase);
$tmpIn = $tmpInBase . '.jpg';
$tmpOut = $tmpOutBase . '.png';
file_put_contents($tmpIn, $imgBytes);
$cmd = sprintf(
'%s i -m u2netp -ppm %s %s 2>&1',
escapeshellarg($rembg),
escapeshellarg($tmpIn),
escapeshellarg($tmpOut)
);
$out = shell_exec($cmd);
@unlink($tmpIn);
if (!is_file($tmpOut) || filesize($tmpOut) < 1024) {
@unlink($tmpOut);
http_response_code(500);
header('Content-Type: text/plain');
echo "rembg failed (see your Pi's logs for details)";
error_log("rembg failed for $sci: " . ($out ?? '(no output)'));
exit;
}
// Tight-crop to the bird's bounding box + downscale to 800px max edge
// so cache stays small.
$im = @imagecreatefrompng($tmpOut);
if ($im !== false) {
$cropped = @imagecropauto($im, IMG_CROP_TRANSPARENT);
if ($cropped !== false) {
imagedestroy($im);
$im = $cropped;
}
$w = imagesx($im); $h = imagesy($im);
$max = 800;
if ($w > $max || $h > $max) {
$scale = $max / max($w, $h);
$nw = (int)($w * $scale); $nh = (int)($h * $scale);
$resized = imagecreatetruecolor($nw, $nh);
imagealphablending($resized, false);
imagesavealpha($resized, true);
imagecopyresampled($resized, $im, 0, 0, 0, 0, $nw, $nh, $w, $h);
imagedestroy($im);
$im = $resized;
}
imagealphablending($im, false);
imagesavealpha($im, true);
imagepng($im, $tmpOut, 6);
imagedestroy($im);
}
// Atomic install: rename is atomic on the same filesystem, so any
// concurrent reader either sees the old cached file or the new one,
// never a half-written PNG.
@rename($tmpOut, $cachePath);
serve_png($cachePath);