use crate::agent::CliError;
use saccade_core::wave7::models::{self, Registry, VisionError};
use std::path::{Path, PathBuf};
pub(crate) fn error(e: VisionError) -> CliError {
let code = match &e {
VisionError::RuntimeIncompatible { .. } => "runtime_incompatible",
VisionError::Unavailable(_) => "vision_unavailable",
VisionError::Integrity(_) => "model_integrity",
VisionError::Invalid(_) => "invalid_vision_input",
VisionError::Io(_) => "io_error",
VisionError::Json(_) => "invalid_json",
};
CliError::new(code, e.to_string())
}
fn home() -> Result<PathBuf, CliError> {
std::env::var_os("HOME")
.map(PathBuf::from)
.ok_or_else(|| CliError::usage("HOME unavailable; supply registry/cache paths"))
}
pub(crate) fn registry(path: Option<&Path>) -> Result<Registry, CliError> {
let p = match path {
Some(p) => p.to_path_buf(),
None => home()?.join(".config/saccade/models.json"),
};
if !p.exists() && path.is_none() {
Registry::pinned_wave7().map_err(error)
} else {
Registry::load(&p).map_err(error)
}
}
pub(crate) fn cache(path: Option<&Path>) -> Result<PathBuf, CliError> {
match path {
Some(p) => Ok(p.to_path_buf()),
None => Ok(home()?.join(".cache/saccade/models")),
}
}
fn emit<T: serde::Serialize>(value: &T, json: bool) -> Result<u8, CliError> {
let text = if json {
serde_json::to_string(value)?
} else {
serde_json::to_string_pretty(value)?
};
crate::emit(&format!("{text}\n"))?;
Ok(0)
}
#[derive(clap::Args)]
pub(crate) struct ModelsArgs {
#[command(subcommand)]
operation: ModelsOperation,
}
#[derive(clap::Subcommand)]
enum ModelsOperation {
List {
#[arg(long)]
registry: Option<PathBuf>,
#[arg(long)]
cache: Option<PathBuf>,
#[arg(long)]
json: bool,
},
Pull {
id: String,
#[arg(long)]
registry: Option<PathBuf>,
#[arg(long)]
cache: Option<PathBuf>,
#[arg(long)]
json: bool,
},
}
pub(crate) fn models(args: ModelsArgs) -> Result<u8, CliError> {
match args.operation {
ModelsOperation::List {
registry: p,
cache: c,
json,
} => {
let cache = cache(c.as_deref())?;
let status = registry(p.as_deref())?.status(&cache);
#[cfg(feature = "local-models")]
let status = {
let mut status = status;
status["runtime"] = saccade_core::wave7::runtime_install::status(&cache);
status
};
emit(&status, json)
}
ModelsOperation::Pull {
id,
registry: p,
cache: c,
json,
} => {
if id == "runtime" {
#[cfg(feature = "local-models")]
{
let cache = cache(c.as_deref())?;
let library =
saccade_core::wave7::runtime_install::pull(&cache).map_err(error)?;
return emit(
&serde_json::json!({"schema":models::MODELS_SCHEMA,
"runtime":saccade_core::wave7::runtime_install::status(&cache),
"ORT_DYLIB_PATH":library}),
json,
);
}
#[cfg(not(feature = "local-models"))]
return Err(error(VisionError::Unavailable(
"compile local-models to provision the runtime".into(),
)));
}
let r = registry(p.as_deref())?;
let m = r.model(&id).map_err(error)?;
let paths = models::ensure(m, &cache(c.as_deref())?, true).map_err(error)?;
emit(
&serde_json::json!({"schema":models::MODELS_SCHEMA,"model":m,"status":"cached_verified","artifact_count":paths.len(),"source_parity":m.parity_sha256.is_some()}),
json,
)
}
}
}
#[derive(clap::Args)]
pub(crate) struct ModelOptions {
#[arg(long)]
registry: Option<PathBuf>,
#[arg(long)]
cache: Option<PathBuf>,
#[arg(long)]
runtime_library: Option<PathBuf>,
#[arg(long)]
allow_download: bool,
}
impl ModelOptions {
fn preflight(&self, requested: &[&str]) -> Result<(), CliError> {
#[cfg(not(feature = "local-models"))]
{
let _ = requested;
Err(error(VisionError::Unavailable("optional model runtime missing; fix: install a saccade build with local-models; saccade models list --json, then saccade models pull MODEL_ID and saccade models pull runtime".into())))
}
#[cfg(feature = "local-models")]
{
let reg = registry(self.registry.as_deref())?;
let cache = cache(self.cache.as_deref())?;
let library = saccade_core::wave7::runtime_install::resolve(
self.runtime_library.as_deref(),
&cache,
)
.map_err(error)?;
saccade_core::optional::require_library(&library)?;
for id in requested {
let fallback = match *id {
"grounding-dino-tiny" => "owlv2-base",
"sam-2.1-tiny" => "efficientsam-ti",
_ => id,
};
let model = reg
.model(id)
.or_else(|_| reg.model(fallback))
.map_err(error)?;
models::ensure(model, &cache, self.allow_download).map_err(|_|error(VisionError::Unavailable(format!("optional model {} missing or corrupt; fix: saccade models pull {} (use the same --registry and --cache); saccade models list --json",model.id,model.id))))?;
}
Ok(())
}
}
}
#[derive(clap::Args)]
pub(crate) struct LocateArgs {
image: PathBuf,
phrase: String,
#[arg(long)]
segment: bool,
#[arg(long, default_value = "grounding-dino-tiny")]
detector: String,
#[arg(long, default_value = "sam-2.1-tiny")]
segmenter: String,
#[arg(long)]
observations: Option<PathBuf>,
#[arg(long)]
overlay: Option<PathBuf>,
#[command(flatten)]
model: ModelOptions,
#[arg(long)]
json: bool,
}
pub(crate) fn write_png(path: &Path, pixels: &image::RgbImage) -> Result<(), CliError> {
use std::io::Write;
let file = std::fs::OpenOptions::new()
.write(true)
.create_new(true)
.open(path)
.map_err(|e| CliError::io(e.to_string()))?;
let mut writer = std::io::BufWriter::new(file);
image::DynamicImage::ImageRgb8(pixels.clone())
.write_to(&mut writer, image::ImageFormat::Png)
.map_err(|e| CliError::io(e.to_string()))?;
writer.flush().map_err(|e| CliError::io(e.to_string()))
}
pub(crate) fn locate(a: LocateArgs) -> Result<u8, CliError> {
let json = a.json;
let r = locate_measure(a)?;
emit(&r, json)
}
fn locate_measure(a: LocateArgs) -> Result<saccade_core::wave7::vision::LocateReport, CliError> {
use saccade_core::wave7::vision::{LocateReport, VisionImage};
if a.observations.is_none() {
let mut ids = vec![a.detector.as_str()];
if a.segment {
ids.push(a.segmenter.as_str());
}
a.model.preflight(&ids)?;
}
let image = VisionImage::load(&a.image).map_err(error)?;
let r = if let Some(p) = &a.observations {
if a.model.allow_download {
return Err(CliError::usage(
"observation replay does not download models",
));
}
LocateReport::replay(p, &image, &a.phrase, a.segment).map_err(error)?
} else {
let reg = registry(a.model.registry.as_deref())?;
#[cfg(feature = "local-models")]
{
use saccade_core::wave7::{
native::{EfficientSam, TextDetector},
vision,
};
let cache = cache(a.model.cache.as_deref())?;
let library = saccade_core::wave7::runtime_install::resolve(
a.model.runtime_library.as_deref(),
&cache,
)
.map_err(error)?;
let detector_id =
if a.detector == "grounding-dino-tiny" && reg.model(&a.detector).is_err() {
"owlv2-base"
} else {
&a.detector
};
let mut detector = TextDetector::load(
reg.model(detector_id).map_err(error)?,
&cache,
&library,
a.model.allow_download,
)
.map_err(error)?;
let mut segmenter = if a.segment {
let id = if a.segmenter == "sam-2.1-tiny" && reg.model(&a.segmenter).is_err() {
"efficientsam-ti"
} else {
&a.segmenter
};
Some(
EfficientSam::load(
reg.model(id).map_err(error)?,
&cache,
&library,
a.model.allow_download,
)
.map_err(error)?,
)
} else {
None
};
vision::locate(
&image,
&a.phrase,
&mut detector,
segmenter.as_mut().map(|s| s as &mut dyn vision::Segmenter),
)
.map_err(error)?
}
#[cfg(not(feature = "local-models"))]
{
let _ = reg;
return Err(error(VisionError::Unavailable(
"compile local-models for native detection".into(),
)));
}
};
let out = a
.overlay
.clone()
.unwrap_or_else(|| a.image.with_extension("locate.png"));
write_png(
&out,
&saccade_core::wave7::vision::overlay(&image, &r.detections).map_err(error)?,
)?;
Ok(r)
}
#[cfg(feature = "local-vlm")]
#[derive(clap::Args)]
pub(crate) struct ObserveArgs {
request: PathBuf,
#[arg(long)]
endpoint: String,
#[arg(long)]
runtime_revision: String,
#[arg(long)]
response: Option<PathBuf>,
#[arg(long)]
json: bool,
}
#[cfg(feature = "local-vlm")]
pub(crate) fn observe(a: ObserveArgs) -> Result<u8, CliError> {
use saccade_core::wave7::{
local_vlm::LocalVlm,
observation::{ObservationProvider, ObservationRequest},
};
let r: ObservationRequest = serde_json::from_slice(
&models::read_bounded(&a.request, 24 * 1024 * 1024).map_err(error)?,
)?;
let mut p = LocalVlm {
endpoint: a.endpoint,
runtime_revision: a.runtime_revision,
};
let mut report = if let Some(path) = &a.response {
p.decode(&r, &models::read_bounded(path, 1024 * 1024).map_err(error)?)
.map_err(error)?
} else {
p.observe(&r).map_err(error)?
};
if report.provenance.runtime == "external-http" && a.response.is_some() {
report.provenance.runtime = "replay".into();
}
emit(&report, a.json)
}
#[derive(clap::Args)]
pub(crate) struct QualityArgs {
image: PathBuf,
#[arg(long)]
reference: Option<PathBuf>,
#[arg(long, default_value = "musiq")]
metric: String,
#[arg(long)]
observations: Option<PathBuf>,
#[command(flatten)]
model: ModelOptions,
#[arg(long)]
json: bool,
}
fn learned_metric(name: &str) -> Result<saccade_core::wave7::quality::LearnedMetric, CliError> {
use saccade_core::wave7::quality::LearnedMetric as M;
match name {
"lpips" => Ok(M::LpipsAlexV01),
"dists" => Ok(M::Dists),
"musiq" => Ok(M::MusiqTechnical),
_ => Err(CliError::usage("metric must be lpips, dists or musiq")),
}
}
pub(crate) fn quality(a: QualityArgs) -> Result<u8, CliError> {
use saccade_core::wave7::{
quality::{self, QualityReport},
vision::VisionImage,
};
let metric = learned_metric(&a.metric)?;
if a.observations.is_none() {
a.model.preflight(&[metric.model_id()])?;
}
let image = VisionImage::load(&a.image).map_err(error)?;
let reference = a
.reference
.as_deref()
.map(VisionImage::load)
.transpose()
.map_err(error)?;
let r = if let Some(p) = &a.observations {
let mut r: QualityReport =
serde_json::from_slice(&models::read_bounded(p, 1024 * 1024).map_err(error)?)?;
r.validate(metric, &image, reference.as_ref())
.map_err(error)?;
for m in &mut r.named_metrics {
m.provenance.runtime = "replay".into();
m.provenance.source_parity = false;
}
r
} else {
let reg = registry(a.model.registry.as_deref())?;
let m = reg.model(metric.model_id()).map_err(error)?;
#[cfg(feature = "local-models")]
{
let model_cache = cache(a.model.cache.as_deref())?;
let library = saccade_core::wave7::runtime_install::resolve(
a.model.runtime_library.as_deref(),
&model_cache,
)
.map_err(error)?;
let mut runtime = saccade_core::wave7::runtime::OnnxModel::load(
m,
&cache(a.model.cache.as_deref())?,
&library,
a.model.allow_download,
)
.map_err(error)?;
quality::measure(metric, &image, reference.as_ref(), &mut runtime).map_err(error)?
}
#[cfg(not(feature = "local-models"))]
{
let _ = m;
let _ = quality::QUALITY_SCHEMA;
return Err(error(VisionError::Unavailable(
"compile local-models for learned quality inference".into(),
)));
}
};
emit(&r, a.json)
}
#[derive(clap::Args)]
pub(crate) struct WatermarkArgs {
image: PathBuf,
#[arg(long)]
expected_payload: Option<String>,
#[arg(long, default_value_t = 36.)]
quantization_step: f32,
#[arg(long, default_value_t = 0.9)]
minimum_agreement: f32,
#[arg(long)]
observations: Option<PathBuf>,
#[arg(long)]
trustmark: bool,
#[command(flatten)]
model: ModelOptions,
#[arg(long)]
json: bool,
}
pub(crate) fn unhex(s: &str) -> Result<Vec<u8>, CliError> {
if s.is_empty()
|| s.len() > 128
|| !s.len().is_multiple_of(2)
|| !s.bytes().all(|b| b.is_ascii_hexdigit())
{
return Err(CliError::usage("payload needs 1..64 hex-encoded bytes"));
}
(0..s.len())
.step_by(2)
.map(|i| {
u8::from_str_radix(&s[i..i + 2], 16).map_err(|_| CliError::usage("invalid hex payload"))
})
.collect()
}
pub(crate) fn watermark(a: WatermarkArgs) -> Result<u8, CliError> {
use saccade_core::wave7::{
vision::VisionImage,
watermark::{self, DwtConfig, WatermarkReport},
};
if a.trustmark && a.observations.is_none() {
a.model.preflight(&["trustmark"])?;
}
let image = VisionImage::load(&a.image).map_err(error)?;
let r = if let Some(p) = a.observations {
let mut r: WatermarkReport =
serde_json::from_slice(&models::read_bounded(&p, 1024 * 1024).map_err(error)?)?;
r.validate(&image).map_err(error)?;
for f in &mut r.findings {
if let Some(p) = &mut f.provenance {
p.runtime = "replay".into();
p.source_parity = false;
}
f.interpretation = format!("Explicit observation replay. {}", f.interpretation);
}
r
} else {
let legacy = a
.expected_payload
.as_ref()
.map(|s| {
unhex(s).map(|expected_payload| DwtConfig {
expected_payload,
quantization_step: a.quantization_step,
minimum_agreement: a.minimum_agreement,
})
})
.transpose()?;
if a.trustmark {
#[cfg(feature = "local-models")]
{
let reg = registry(a.model.registry.as_deref())?;
let cache = cache(a.model.cache.as_deref())?;
let library = saccade_core::wave7::runtime_install::resolve(
a.model.runtime_library.as_deref(),
&cache,
)
.map_err(error)?;
let mut decoder = saccade_core::wave7::trustmark::TrustMarkQ::load(
reg.model("trustmark").map_err(error)?,
&cache,
&library,
a.model.allow_download,
)
.map_err(error)?;
watermark::inspect(&image, Some(&mut decoder), legacy.as_ref()).map_err(error)?
}
#[cfg(not(feature = "local-models"))]
return Err(error(VisionError::Unavailable(
"compile local-models for TrustMark Q".into(),
)));
} else {
watermark::inspect(&image, None, legacy.as_ref()).map_err(error)?
}
};
emit(&r, a.json)
}
#[derive(clap::Args)]
pub(crate) struct FacesArgs {
image: PathBuf,
#[arg(long, default_value = "yunet-2026may")]
detector: String,
#[arg(long)]
observations: Option<PathBuf>,
#[arg(long)]
blur_faces: Option<PathBuf>,
#[command(flatten)]
model: ModelOptions,
#[arg(long)]
json: bool,
}
fn face_report(
a: &FacesArgs,
image: &saccade_core::wave7::vision::VisionImage,
) -> Result<saccade_core::wave7::faces::FaceReport, CliError> {
use saccade_core::wave7::faces::{self, FaceReport};
if let Some(p) = &a.observations {
let mut r: FaceReport =
serde_json::from_slice(&models::read_bounded(p, 1024 * 1024).map_err(error)?)?;
r.validate(image).map_err(error)?;
r.provenance.runtime = "replay".into();
r.provenance.source_parity = false;
return Ok(r);
}
let reg = registry(a.model.registry.as_deref())?;
let m = reg.model(&a.detector).map_err(error)?;
#[cfg(feature = "local-models")]
{
let model_cache = cache(a.model.cache.as_deref())?;
let library = saccade_core::wave7::runtime_install::resolve(
a.model.runtime_library.as_deref(),
&model_cache,
)
.map_err(error)?;
let mut runtime = saccade_core::wave7::runtime::OnnxModel::load(
m,
&cache(a.model.cache.as_deref())?,
&library,
a.model.allow_download,
)
.map_err(error)?;
faces::detect(image, &mut runtime).map_err(error)
}
#[cfg(not(feature = "local-models"))]
{
let _ = m;
let _ = faces::FACES_SCHEMA;
Err(error(VisionError::Unavailable(
"compile local-models for face inference".into(),
)))
}
}
pub(crate) fn faces(a: FacesArgs) -> Result<u8, CliError> {
if a.observations.is_none() {
a.model.preflight(&[a.detector.as_str()])?;
}
let image = saccade_core::wave7::vision::VisionImage::load(&a.image).map_err(error)?;
let r = face_report(&a, &image)?;
if let Some(p) = &a.blur_faces {
write_png(
p,
&saccade_core::wave7::faces::blur_faces(&image, &r).map_err(error)?,
)?;
}
emit(&r, a.json)
}
#[derive(clap::Args)]
pub(crate) struct CropArgs {
#[command(flatten)]
faces: FacesArgs,
#[arg(long, required = true)]
crop: Vec<String>,
#[arg(long)]
focal_point: Option<String>,
}
fn numbers(s: &str, separator: char, count: usize) -> Result<Vec<f32>, CliError> {
let values = s
.split(separator)
.map(|s| {
s.parse::<f32>()
.map_err(|_| CliError::usage("crop coordinates must be numbers"))
})
.collect::<Result<Vec<_>, _>>()?;
if values.len() != count || values.iter().any(|v| !v.is_finite()) {
return Err(CliError::usage("invalid crop coordinate count/value"));
}
Ok(values)
}
pub(crate) fn crop(a: CropArgs) -> Result<u8, CliError> {
use saccade_core::wave7::{
faces::{self, CropSpec},
vision::{Rect, VisionImage},
};
if a.faces.observations.is_none() {
a.faces.model.preflight(&[a.faces.detector.as_str()])?;
}
let image = VisionImage::load(&a.faces.image).map_err(error)?;
let r = face_report(&a.faces, &image)?;
let specs = a
.crop
.iter()
.map(|s| {
if s.contains(':') {
numbers(s, ':', 2).map(|v| CropSpec::Ratio {
width: v[0],
height: v[1],
})
} else {
numbers(s, ',', 4).map(|v| CropSpec::Rectangle {
rect: Rect {
x: v[0],
y: v[1],
width: v[2],
height: v[3],
},
})
}
})
.collect::<Result<Vec<_>, _>>()?;
let focal = a
.focal_point
.as_deref()
.map(|s| numbers(s, ',', 2).map(|v| [v[0], v[1]]))
.transpose()?;
let report = faces::crop_check(&image, &r, &specs, focal).map_err(error)?;
if let Some(p) = &a.faces.blur_faces {
write_png(p, &faces::blur_faces(&image, &r).map_err(error)?)?;
}
emit(&report, a.faces.json)
}
#[cfg(feature = "vision-providers")]
#[derive(clap::Args)]
pub(crate) struct ProviderArgs {
request: PathBuf,
#[arg(long)]
provider: String,
#[arg(long,value_parser=["openai-compatible","azure-openai"])]
endpoint_profile: Option<String>,
#[arg(long)]
response: Option<PathBuf>,
#[arg(long, default_value = "pixels")]
coordinates: String,
#[arg(long)]
json: bool,
}
#[cfg(feature = "vision-providers")]
pub(crate) fn provider(a: ProviderArgs) -> Result<u8, CliError> {
use saccade_core::wave7::{
observation::ObservationRequest,
providers::{Coordinates, Provider, ProviderAdapter},
};
let provider = match a.provider.as_str() {
"claude" => Provider::Claude,
"gpt" => Provider::Gpt,
_ => return Err(CliError::usage("provider must be claude or gpt")),
};
let coordinates = match a.coordinates.as_str() {
"pixels" => Coordinates::Pixels,
"unit" => Coordinates::Unit,
"thousand" => Coordinates::Thousand,
_ => {
return Err(CliError::usage(
"coordinates must be pixels, unit or thousand",
));
}
};
let r: ObservationRequest = serde_json::from_slice(
&models::read_bounded(&a.request, 24 * 1024 * 1024).map_err(error)?,
)?;
let adapter = ProviderAdapter {
provider,
coordinates,
};
if let Some(profile) = a.endpoint_profile {
use saccade_core::media::endpoints::{Endpoint, Kind};
let endpoint = Endpoint::load(if profile == "azure-openai" {
Kind::AzureOpenai
} else {
Kind::OpenaiCompatible
})
.map_err(error)?;
return if let Some(path) = a.response {
emit(
&endpoint
.decode(
&adapter,
&r,
&models::read_bounded(&path, 1024 * 1024).map_err(error)?,
)
.map_err(error)?,
a.json,
)
} else {
emit(&endpoint.mapping(&adapter, &r).map_err(error)?, a.json)
};
}
if let Some(path) = a.response {
emit(
&adapter
.decode(
&r,
&models::read_bounded(&path, 1024 * 1024).map_err(error)?,
)
.map_err(error)?,
a.json,
)
} else {
emit(
&serde_json::json!({"schema":"saccade-provider-mapping.v1","interface_only":true,"endpoint":provider.endpoint(),"request_sha256":r.hash().map_err(error)?,"body":adapter.request(&r).map_err(error)?}),
a.json,
)
}
}
#[cfg(test)]
#[allow(clippy::unwrap_used)]
mod tests {
use super::*;
use clap::Parser;
use saccade_core::wave7::{faces::*, quality::*, vision::*};
fn json(path: &Path, v: &impl serde::Serialize) {
std::fs::write(path, serde_json::to_vec(v).unwrap()).unwrap();
}
fn run(args: Vec<String>) -> u8 {
let c = crate::Cli::try_parse_from(args).unwrap();
crate::dispatch(c.command, false).unwrap()
}
#[test]
fn generated_commands_render_and_assess_receipts_without_models() {
let d = tempfile::tempdir().unwrap();
let p = d.path().join("image.png");
image::RgbImage::from_pixel(32, 24, image::Rgb([50, 100, 150]))
.save(&p)
.unwrap();
let i = VisionImage::load(&p).unwrap();
let face = FaceReport {
schema: FACES_SCHEMA.into(),
image_sha256: i.sha256.clone(),
image_size: i.size(),
faces: vec![Face {
bbox: Rect {
x: 2.,
y: 4.,
width: 8.,
height: 8.,
},
score: 0.9,
landmarks: vec![],
}],
provenance: Provenance::fixture("yunet-2026may"),
limitations: FACE_LIMIT.into(),
};
let fp = d.path().join("faces.json");
json(&fp, &face);
let loc = LocateReport {
schema: LOCATE_SCHEMA.into(),
image_sha256: i.sha256.clone(),
image_size: i.size(),
phrase_sha256: models::digest(b"object"),
detections: vec![Detection {
bbox: face.faces[0].bbox,
score: 0.9,
mask: Some(Mask {
size: i.size(),
runs: vec![[130, 8]],
}),
}],
detector: Provenance::fixture("generated-detector"),
segmenter: Some(Provenance::fixture("generated-segmenter")),
};
let lp = d.path().join("locate.json");
json(&lp, &loc);
let overlay = d.path().join("overlay.png");
let a = vec![
"saccade".into(),
"locate".into(),
p.display().to_string(),
"object".into(),
"--segment".into(),
"--observations".into(),
lp.display().to_string(),
"--overlay".into(),
overlay.display().to_string(),
"--json".into(),
];
assert_eq!(run(a), 0);
assert!(overlay.is_file());
let redacted = d.path().join("redacted.png");
assert_eq!(
run(vec![
"saccade".into(),
"crop-check".into(),
p.display().to_string(),
"--crop".into(),
"16:9".into(),
"--crop".into(),
"0,0,16,24".into(),
"--observations".into(),
fp.display().to_string(),
"--blur-faces".into(),
redacted.display().to_string(),
"--json".into()
]),
0
);
assert!(redacted.is_file());
assert_eq!(
run(vec![
"saccade".into(),
"faces".into(),
p.display().to_string(),
"--observations".into(),
fp.display().to_string(),
"--json".into()
]),
0
);
let q = QualityReport {
schema: QUALITY_SCHEMA.into(),
image_sha256: i.sha256.clone(),
reference_sha256: None,
named_metrics: vec![QualityMeasurement {
metric: LearnedMetric::MusiqTechnical,
value: 50.,
direction: "higher_is_better".into(),
provenance: Provenance::fixture("musiq-technical"),
}],
affects_compare_verdict: false,
};
let qp = d.path().join("quality.json");
json(&qp, &q);
assert_eq!(
run(vec![
"saccade".into(),
"quality-score".into(),
p.display().to_string(),
"--observations".into(),
qp.display().to_string(),
"--json".into()
]),
0
);
assert_eq!(
run(vec![
"saccade".into(),
"watermark".into(),
p.display().to_string(),
"--json".into()
]),
0
);
}
#[test]
fn output_never_overwrites_an_original() {
let d = tempfile::tempdir().unwrap();
let p = d.path().join("original.png");
std::fs::write(&p, b"original").unwrap();
assert!(write_png(&p, &image::RgbImage::new(1, 1)).is_err());
assert_eq!(std::fs::read(p).unwrap(), b"original");
}
}
#[cfg(feature = "assist")]
#[derive(clap::Args, Default)]
pub(crate) struct Grounding {
#[arg(long)]
pub locate: bool,
#[arg(long)]
pub locate_observations: Option<PathBuf>,
#[arg(long)]
pub locate_registry: Option<PathBuf>,
#[arg(long)]
pub locate_cache: Option<PathBuf>,
#[arg(long)]
pub locate_runtime_library: Option<PathBuf>,
}
#[cfg(feature = "assist")]
pub(crate) fn locate_for_check(
image: &Path,
phrase: &str,
options: &Grounding,
) -> Result<serde_json::Value, CliError> {
let report = locate_measure(LocateArgs {
image: image.into(),
phrase: phrase.into(),
segment: false,
detector: "grounding-dino-tiny".into(),
segmenter: "sam-2.1-tiny".into(),
observations: options.locate_observations.clone(),
overlay: None,
model: ModelOptions {
registry: options.locate_registry.clone(),
cache: options.locate_cache.clone(),
runtime_library: options.locate_runtime_library.clone(),
allow_download: false,
},
json: true,
})?;
Ok(
serde_json::json!({"authority":"advisory model observation; never sufficient for a visible-condition verdict","report":report}),
)
}