Implement pure Rust visual snapshots (#132)

Cargo.lock

@@ -2270,10 +2270,14 @@ dependencies = [
 name = "metacrate-grid-agent"
 version = "0.0.1"
 dependencies = [
+ "base64",
  "crossterm",
  "libremetaverse",
+ "libremetaverse-imaging",
+ "libremetaverse-rendering-simple",
  "libremetaverse-types",
  "metacrate-lsl-tools",
+ "png",
  "reqwest",
  "rustls",
  "serde",

crates/metacrate-grid-agent/Cargo.toml

@@ -9,10 +9,14 @@ description = "Bounded pure-Rust OpenSim grid-agent service foundation"
 publish = false
 
 [dependencies]
+base64 = "0.22"
 crossterm = "0.29"
 libremetaverse = { version = "0.0.1", path = "../libremetaverse", default-features = false, optional = true }
+libremetaverse-imaging = { version = "0.0.1", path = "../libremetaverse-imaging", default-features = false, features = ["rust-j2k"], optional = true }
+libremetaverse-rendering-simple = { version = "0.0.1", path = "../libremetaverse-rendering-simple", optional = true }
 libremetaverse-types = { version = "0.0.1", path = "../libremetaverse-types" }
 metacrate-lsl-tools = { version = "0.0.1", path = "../metacrate-lsl-tools" }
+png = "0.17"
 reqwest = { version = "0.13.4", default-features = false, features = ["rustls"] }
 rustls = { version = "0.23.43", default-features = false, features = ["aws_lc_rs", "std", "tls12"] }
 serde = { version = "1", features = ["derive"] }
@@ -31,7 +35,7 @@ tokio = { version = "1.53.1", features = ["io-util", "net", "rt-multi-thread", "
 
 [features]
 default = []
-live-grid = ["dep:libremetaverse"]
+live-grid = ["dep:libremetaverse", "dep:libremetaverse-imaging", "dep:libremetaverse-rendering-simple"]
 
 [lints]
 workspace = true

crates/metacrate-grid-agent/src/control_plane.rs

@@ -202,6 +202,9 @@ pub struct RuntimeView {
     pub active_build_transaction: Option<String>,
     pub build_progress: Option<String>,
     pub build_orphan_ids: Vec<String>,
+    pub active_visual_capture: Option<String>,
+    pub visual_progress: Option<String>,
+    pub visual_image_sha256: Option<String>,
 }
 
 #[derive(Clone, Debug, Eq, PartialEq, Serialize, Deserialize)]

crates/metacrate-grid-agent/src/control_plane_tests.rs

@@ -56,6 +56,9 @@ impl ControlTarget for FakeTarget {
                     active_build_transaction: None,
                     build_progress: None,
                     build_orphan_ids: Vec::new(),
+                    active_visual_capture: None,
+                    visual_progress: None,
+                    visual_image_sha256: None,
                 })),
                 ControlRequest::ListSessions { page } => Ok(ControlPayload::Sessions(page_values(
                     &page,

crates/metacrate-grid-agent/src/control_runtime.rs

@@ -85,6 +85,7 @@ pub struct AgentControlTarget {
     audit: Arc<MemoryPolicyAudit>,
     observability: Mutex<Option<Arc<Observability>>>,
     build: Mutex<Option<Arc<dyn crate::build::BuildControl>>>,
+    vision: Mutex<Option<Arc<dyn crate::vision::VisionControl>>>,
     behavior: BehaviorIngress,
     commands: mpsc::Sender<RuntimeControlCommand>,
     command_capacity: usize,
@@ -129,6 +130,7 @@ impl AgentControlTarget {
                 audit,
                 observability: Mutex::new(None),
                 build: Mutex::new(None),
+                vision: Mutex::new(None),
                 behavior,
                 commands,
                 command_capacity,
@@ -149,6 +151,11 @@ impl AgentControlTarget {
         *lock(&self.build) = Some(build);
     }
 
+    /// Attaches synthetic viewport cancellation and metadata to operator control.
+    pub fn attach_vision_control(&self, vision: Arc<dyn crate::vision::VisionControl>) {
+        *lock(&self.vision) = Some(vision);
+    }
+
     pub fn update_session(&self, session: SessionStatus) {
         let mut state = lock(&self.state);
         state.session = session;
@@ -245,6 +252,7 @@ impl AgentControlTarget {
                 let state = lock(&self.state).clone();
                 let usage = self.policy.global_budget_usage();
                 let build = lock(&self.build);
+                let vision = lock(&self.vision);
                 Ok(ControlPayload::Runtime(RuntimeView {
                     grid_state: state.session.state.as_str().to_owned(),
                     generation: state.session.generation,
@@ -265,6 +273,13 @@ impl AgentControlTarget {
                     build_orphan_ids: build
                         .as_ref()
                         .map_or_else(Vec::new, |build| build.build_orphans()),
+                    active_visual_capture: vision
+                        .as_ref()
+                        .and_then(|vision| vision.active_capture()),
+                    visual_progress: vision.as_ref().and_then(|vision| vision.capture_progress()),
+                    visual_image_sha256: vision
+                        .as_ref()
+                        .and_then(|vision| vision.last_image_sha256()),
                 }))
             }
             ControlRequest::ListSessions { page } => {
@@ -345,6 +360,12 @@ impl AgentControlTarget {
             ControlRequest::PauseAutonomy => {
                 let response = self.enqueue("pause", RuntimeControlCommand::Pause)?;
                 self.behavior.pause();
+                let vision = lock(&self.vision).clone();
+                if let Some((vision, active)) =
+                    vision.and_then(|vision| vision.active_capture().map(|active| (vision, active)))
+                {
+                    let _ = vision.cancel_capture(&active);
+                }
                 Ok(response)
             }
             ControlRequest::ResumeAutonomy => {
@@ -356,10 +377,13 @@ impl AgentControlTarget {
                 let build_cancelled = lock(&self.build)
                     .as_ref()
                     .is_some_and(|build| build.cancel_build(&action_id));
+                let vision_cancelled = lock(&self.vision)
+                    .as_ref()
+                    .is_some_and(|vision| vision.cancel_capture(&action_id));
                 let cancelled = self.behavior.cancel_action(&action_id).map_err(|_| {
                     control_error(ControlErrorCode::InvalidRequest, "invalid action ID", false)
                 })?;
-                if !cancelled && !build_cancelled {
+                if !cancelled && !build_cancelled && !vision_cancelled {
                     return Err(control_error(
                         ControlErrorCode::NotFound,
                         "active behavior action not found",

crates/metacrate-grid-agent/src/control_runtime_tests.rs

@@ -7,15 +7,39 @@ use crate::control_runtime::{AgentControlTarget, RuntimeControlCommand};
 use crate::conversation::{ConversationLimits, ConversationStore};
 use crate::perception::WorldPosition;
 use crate::policy::{MemoryPolicyAudit, PolicyAuditSink, PolicyGateway, PolicyLimits};
-use crate::{AgentConfig, EmbodiedPose, Observability, ObservabilityLimits, SecretString};
+use crate::{
+    AgentConfig, EmbodiedPose, Observability, ObservabilityLimits, SecretString, VisionControl,
+};
 use libremetaverse_types::UUID;
 use libremetaverse_types::compat::CancellationToken;
 use std::collections::BTreeSet;
-use std::sync::Arc;
+use std::sync::{Arc, Mutex};
 use std::time::Duration;
 
 struct Sink;
 
+struct FakeVision(Mutex<Option<String>>);
+impl crate::VisionControl for FakeVision {
+    fn cancel_capture(&self, correlation_id: &str) -> bool {
+        let mut active = self.0.lock().unwrap();
+        if active.as_deref() == Some(correlation_id) {
+            *active = None;
+            true
+        } else {
+            false
+        }
+    }
+    fn active_capture(&self) -> Option<String> {
+        self.0.lock().unwrap().clone()
+    }
+    fn capture_progress(&self) -> Option<String> {
+        self.active_capture().map(|_| "rendering".into())
+    }
+    fn last_image_sha256(&self) -> Option<String> {
+        Some("abc123".into())
+    }
+}
+
 impl EmbodimentSink for Sink {
     fn current_pose(
         &self,
@@ -99,6 +123,7 @@ fn envelope(id: &str, request: ControlRequest) -> ControlRequestEnvelope {
 }
 
 #[tokio::test]
+#[allow(clippy::too_many_lines)]
 async fn production_target_projects_state_and_routes_real_mutations() {
     let settings = AgentConfig::offline("https://llm.invalid/chat", "llm-key")
         .expect("offline config")
@@ -125,6 +150,8 @@ async fn production_target_projects_state_and_routes_real_mutations() {
             .expect("target");
     let observability = Observability::memory(ObservabilityLimits::default()).expect("events");
     target.attach_observability(observability);
+    let vision = Arc::new(FakeVision(Mutex::new(Some("visual-1".into()))));
+    target.attach_vision_control(vision.clone());
     target.update_region(
         Some("00000000-0000-4000-8000-000000000001".into()),
         Some("Test Region".into()),
@@ -146,6 +173,9 @@ async fn production_target_projects_state_and_routes_real_mutations() {
     };
     assert_eq!(runtime.region_name.as_deref(), Some("Test Region"));
     assert_eq!(runtime.control_queue_capacity, 8);
+    assert_eq!(runtime.active_visual_capture.as_deref(), Some("visual-1"));
+    assert_eq!(runtime.visual_progress.as_deref(), Some("rendering"));
+    assert_eq!(runtime.visual_image_sha256.as_deref(), Some("abc123"));
     let metrics = operator
         .request(envelope("metrics", ControlRequest::Metrics))
         .await;
@@ -165,6 +195,7 @@ async fn production_target_projects_state_and_routes_real_mutations() {
         commands.recv().await,
         Some(RuntimeControlCommand::Pause)
     ));
+    assert!(vision.active_capture().is_none(), "pause cancels capture");
 
     let observer = plane.connect("observer-token").expect("observer");
     let denied = observer

crates/metacrate-grid-agent/src/interaction.rs

@@ -296,6 +296,7 @@ pub enum InteractionModelError {
     Timeout,
     Failed,
     PolicyRejected,
+    MultimodalUnsupported,
 }
 
 impl fmt::Display for InteractionModelError {
@@ -305,6 +306,7 @@ impl fmt::Display for InteractionModelError {
             Self::Timeout => "interaction model request timed out",
             Self::Failed => "interaction model request failed",
             Self::PolicyRejected => "interaction request was rejected by policy",
+            Self::MultimodalUnsupported => "configured endpoint does not support visual context",
         })
     }
 }
@@ -1152,7 +1154,9 @@ async fn process_batch(
         Ok(_) => InferenceOutcome::Completed,
         Err(InteractionModelError::Cancelled) => InferenceOutcome::Cancelled,
         Err(InteractionModelError::Timeout) => InferenceOutcome::TimedOut,
-        Err(InteractionModelError::Failed) => InferenceOutcome::Failed,
+        Err(InteractionModelError::Failed | InteractionModelError::MultimodalUnsupported) => {
+            InferenceOutcome::Failed
+        }
         Err(InteractionModelError::PolicyRejected) => InferenceOutcome::PolicyRejected,
     };
     observe(
@@ -1249,6 +1253,7 @@ async fn process_batch(
                 InteractionModelError::Cancelled | InteractionModelError::Failed => {
                     DeliveryOutcome::Failed
                 }
+                InteractionModelError::MultimodalUnsupported => DeliveryOutcome::Failed,
             };
             let delivered = if error == InteractionModelError::Cancelled {
                 delivery_observation(
@@ -1822,6 +1827,9 @@ impl InteractionResponder for PolicyLlmResponder {
                     crate::tool_loop::ToolLoopError::WallClockTimeout => {
                         InteractionModelError::Timeout
                     }
+                    crate::tool_loop::ToolLoopError::Transport(
+                        crate::llm::LlmError::MultimodalUnsupported,
+                    ) => InteractionModelError::MultimodalUnsupported,
                     _ => InteractionModelError::Failed,
                 })?;
             visible_text(&outcome.final_message)

crates/metacrate-grid-agent/src/lib.rs

@@ -23,6 +23,7 @@ pub mod session;
 pub mod tool_loop;
 pub mod tui;
 pub mod types;
+pub mod vision;
 
 #[cfg(test)]
 mod behavior_tests;
@@ -50,6 +51,8 @@ mod script_delivery_tests;
 mod session_tests;
 #[cfg(test)]
 mod tui_tests;
+#[cfg(test)]
+mod vision_tests;
 
 pub use backend::{
     AuthorizedToolBackend, BackendError, BackendFuture, GridBackend, OfflineGridBackend,
@@ -172,3 +175,11 @@ pub use types::{
     GridEvent, GridEventKind, LlmRequest, LlmResult, MessageRole, ObservableEvent,
     ProposedToolCall, ToolCallOutcome,
 };
+#[cfg(feature = "live-grid")]
+pub use vision::LibremetaverseSceneSource;
+pub use vision::{
+    CameraPose, SceneEntity, SceneEntityKind, SceneSnapshot, SceneSource, SceneTriangle,
+    SnapshotCompleteness, VisionAugmentedResponder, VisionCapture, VisionControl, VisionError,
+    VisionFuture, VisionLimits, VisionObservation, VisionObservationKind, VisionService,
+    visual_question,
+};

crates/metacrate-grid-agent/src/llm.rs

@@ -301,6 +301,7 @@ pub enum LlmError {
     HttpStatus(u16),
     MalformedJson,
     UnsupportedResponse,
+    MultimodalUnsupported,
     DuplicateToolCallId,
 }
 
@@ -320,6 +321,9 @@ impl fmt::Display for LlmError {
             Self::HttpStatus(status) => write!(formatter, "LLM endpoint returned HTTP {status}"),
             Self::MalformedJson => formatter.write_str("LLM endpoint returned malformed JSON"),
             Self::UnsupportedResponse => formatter.write_str("unsupported LLM response shape"),
+            Self::MultimodalUnsupported => {
+                formatter.write_str("LLM endpoint does not support image input")
+            }
             Self::DuplicateToolCallId => {
                 formatter.write_str("LLM response repeated a tool-call ID")
             }
@@ -390,6 +394,13 @@ impl LlmClient {
         for tool in tools {
             tool.validate()?;
         }
+        let has_image = messages.iter().any(|message| {
+            message
+                .content
+                .as_slice()
+                .iter()
+                .any(|part| matches!(part, ContentPart::Image { .. }))
+        });
         let body = request_body(messages, tools)?;
         if body.len() > self.limits.max_prompt_bytes {
             return Err(LlmError::PromptTooLarge);
@@ -411,9 +422,14 @@ impl LlmClient {
             drop(permit);
             result
         };
-        tokio::time::timeout(self.limits.total_timeout, operation)
+        let result = tokio::time::timeout(self.limits.total_timeout, operation)
             .await
-            .map_err(|_| LlmError::Timeout)?
+            .map_err(|_| LlmError::Timeout)?;
+        if has_image && matches!(result, Err(LlmError::HttpStatus(400 | 415 | 422))) {
+            Err(LlmError::MultimodalUnsupported)
+        } else {
+            result
+        }
     }
 
     async fn complete_with_retries(

crates/metacrate-grid-agent/src/main.rs

@@ -224,6 +224,7 @@ async fn run_live(
         config.limits.observable_queue,
     )?
     .start();
+    live.vision.set_generation(handle.status().generation);
 
     if run_once {
         let readiness = tokio::time::timeout(config.timeouts.startup, async {
@@ -232,19 +233,21 @@ async fn run_live(
                     Some(event @ SessionObservation::Transition { status, .. })
                         if status.agent_ready =>
                     {
+                        live.vision.set_generation(status.generation);
                         record_session_observation(&live.observability, &event);
                         return Ok::<(), CliError>(());
                     }
                     Some(
                         event @ SessionObservation::Transition {
                             status:
-                                metacrate_grid_agent::SessionStatus {
+                                status @ metacrate_grid_agent::SessionStatus {
                                     state: SessionState::AuthenticationBlocked,
                                     ..
                                 },
                             ..
                         },
                     ) => {
+                        live.vision.set_generation(status.generation);
                         record_session_observation(&live.observability, &event);
                         return Err(CliError(
                             "grid authentication/configuration requires operator action".into(),
@@ -292,6 +295,7 @@ async fn run_live(
     )?;
     control_target.attach_observability(live.observability.clone());
     control_target.attach_build_control(live.build_control.clone());
+    control_target.attach_vision_control(live.vision.clone());
     control_target.update_session(handle.status());
     let erased_target: Arc<dyn ControlTarget> = control_target.clone();
     let (control_plane, integrated_client, control_server) = match config.mode {
@@ -360,6 +364,7 @@ async fn run_live(
                 let Some(event) = event else { break; };
                 record_session_observation(&live.observability, &event);
                 if let SessionObservation::Transition { status, reason, retry_in } = event {
+                    live.vision.set_generation(status.generation);
                     control_target.update_session(status);
                     live.landmark_roaming
                         .update_pause(|pause| pause.degraded = !status.agent_ready);
@@ -492,6 +497,8 @@ struct LiveInteractions {
     _landmark_intake: metacrate_grid_agent::LibremetaverseLandmarkIntake,
     landmark_roaming: metacrate_grid_agent::LandmarkRoamingHandle,
     build_control: Arc<dyn metacrate_grid_agent::BuildControl>,
+    vision:
+        Arc<metacrate_grid_agent::VisionService<metacrate_grid_agent::LibremetaverseSceneSource>>,
 }
 
 #[cfg(feature = "live-grid")]
@@ -504,12 +511,13 @@ fn start_live_interactions(
         AuthorizedBackendRouter, AuthorizedToolBackend, BehaviorBackend, BehaviorController,
         BuildLimits, BuildService, BuildToolBackend, ConversationStore, InteractionCoordinator,
         LandmarkLimits, LandmarkService, LandmarkToolBackend, LibremetaverseBuildGrid,
-        LibremetaverseLandmarkGrid, LibremetaverseLandmarkIntake, LibremetaverseScriptInventory,
-        LlmClient, LlmTransportLimits, MemoryPolicyAudit, Observability, ObservabilityLimits,
-        PerceptionBackend, PolicyAuditSink, PolicyGateway, PolicyLimits, PolicyLlmResponder,
-        ScriptDeliveryBackend, ScriptDeliverySettings, SystemRoamingRandom, ToolLoopLimits,
-        UnifiedPolicyAudit, behavior_policy_tools, build_policy_tools, landmark_policy_tools,
-        perception_policy_tools, script_delivery_policy_tool,
+        LibremetaverseLandmarkGrid, LibremetaverseLandmarkIntake, LibremetaverseSceneSource,
+        LibremetaverseScriptInventory, LlmClient, LlmTransportLimits, MemoryPolicyAudit,
+        Observability, ObservabilityLimits, PerceptionBackend, PolicyAuditSink, PolicyGateway,
+        PolicyLimits, PolicyLlmResponder, ScriptDeliveryBackend, ScriptDeliverySettings,
+        SystemRoamingRandom, ToolLoopLimits, UnifiedPolicyAudit, VisionAugmentedResponder,
+        VisionLimits, VisionService, behavior_policy_tools, build_policy_tools,
+        landmark_policy_tools, perception_policy_tools, script_delivery_policy_tool,
     };
 
     let transport_limits = LlmTransportLimits {
@@ -635,6 +643,12 @@ fn start_live_interactions(
         loop_limits,
         now,
     )?);
+    let vision_limits = VisionLimits::default();
+    let vision = Arc::new(VisionService::new(
+        Arc::new(LibremetaverseSceneSource::new(owner, vision_limits)),
+        vision_limits,
+    )?);
+    let responder = Arc::new(VisionAugmentedResponder::new(vision.clone(), responder));
     let sink = Arc::new(owner.interaction_sink());
     let pacer: Arc<dyn metacrate_grid_agent::ResponsePacer> = Arc::new(behavior_ingress);
     let interaction = InteractionCoordinator::new(
@@ -662,6 +676,7 @@ fn start_live_interactions(
         _landmark_intake: landmark_intake,
         landmark_roaming,
         build_control,
+        vision,
     })
 }
 

crates/metacrate-grid-agent/src/tui.rs

@@ -717,6 +717,14 @@ fn render_overview(s: &OperatorSnapshot, out: &mut Vec<String>) {
                 v.build_orphan_ids.len()
             ));
         }
+        if v.active_visual_capture.is_some() || v.visual_image_sha256.is_some() {
+            out.push(format!(
+                "vision={} progress={} last-image-sha256={}",
+                v.active_visual_capture.as_deref().unwrap_or("idle"),
+                v.visual_progress.as_deref().unwrap_or("none"),
+                v.visual_image_sha256.as_deref().unwrap_or("none")
+            ));
+        }
     }
 }
 fn render_queues(s: &OperatorSnapshot, out: &mut Vec<String>) {

crates/metacrate-grid-agent/src/tui_tests.rs

@@ -46,6 +46,9 @@ impl TuiTransport for FakeTransport {
                     active_build_transaction: Some("build-7".into()),
                     build_progress: Some("configured".into()),
                     build_orphan_ids: Vec::new(),
+                    active_visual_capture: None,
+                    visual_progress: None,
+                    visual_image_sha256: None,
                 }),
                 ControlRequest::ListSessions { .. } => ControlPayload::Sessions(crate::Page {
                     items: Vec::new(),

crates/metacrate-grid-agent/src/vision_tests.rs

@@ -0,0 +1,376 @@
+use crate::vision::*;
+use crate::{
+    BoundedText, CompletionMessage, InteractionChannel, InteractionIntent, InteractionModelError,
+    InteractionOrigin, InteractionResponder, MessageRole, ResponderFuture, ResponseRequest,
+    VisibleResponse,
+};
+use libremetaverse_types::{
+    UUID,
+    compat::{CancellationToken, CancellationTokenSource},
+};
+use std::sync::{Arc, Mutex};
+
+struct FakeScene {
+    scene: SceneSnapshot,
+    pending: Mutex<bool>,
+}
+impl SceneSource for FakeScene {
+    fn capture_scene(
+        &self,
+        _: u64,
+        cancellation: CancellationToken,
+    ) -> VisionFuture<'_, SceneSnapshot> {
+        Box::pin(async move {
+            if *self.pending.lock().unwrap() {
+                tokio::select! {()=cancellation.cancelled()=>Err(VisionError::Cancelled),()=std::future::pending()=>unreachable!()}
+            } else {
+                Ok(self.scene.clone())
+            }
+        })
+    }
+}
+fn uuid(value: u32) -> UUID {
+    UUID::new_with_string(format!("00000000-0000-4000-8000-{value:012}")).unwrap()
+}
+fn triangle(z: f32, color: [u8; 4]) -> SceneTriangle {
+    SceneTriangle {
+        vertices: [[-1.0, -1.0, z], [1.0, -1.0, z], [0.0, 1.0, z]],
+        color_srgb: color,
+    }
+}
+fn scene() -> SceneSnapshot {
+    SceneSnapshot {
+        generation: 7,
+        observed_unix_millis: 1_700_000_000_000,
+        region_id: uuid(1),
+        region_name: "Fixture Region".into(),
+        camera: CameraPose {
+            position: [0.0, 0.0, 0.0],
+            forward: [0.0, 0.0, 1.0],
+            up: [0.0, 1.0, 0.0],
+            vertical_fov_degrees: 60.0,
+        },
+        entities: vec![
+            SceneEntity {
+                id: uuid(3),
+                kind: SceneEntityKind::Resident,
+                display_name: "Private Resident".into(),
+                triangles: vec![triangle(4.0, [255, 0, 0, 255])],
+                texture_available: false,
+            },
+            SceneEntity {
+                id: uuid(2),
+                kind: SceneEntityKind::Object,
+                display_name: "Near".into(),
+                triangles: vec![triangle(2.0, [0, 255, 0, 255])],
+                texture_available: true,
+            },
+        ],
+        completeness: SnapshotCompleteness {
+            textures_missing: 1,
+            terrain_available: false,
+            ..SnapshotCompleteness::default()
+        },
+        texture_fetches: 1,
+        texture_bytes: 128,
+        decoded_texture_pixels: 16,
+    }
+}
+
+#[tokio::test]
+async fn golden_scene_is_deterministic_depth_ordered_and_privacy_marked() {
+    let source = Arc::new(FakeScene {
+        scene: scene(),
+        pending: Mutex::new(false),
+    });
+    let service = VisionService::new(
+        source,
+        VisionLimits {
+            width: 64,
+            height: 64,
+            minimum_interval: std::time::Duration::ZERO,
+            ..VisionLimits::default()
+        },
+    )
+    .unwrap();
+    service.set_generation(7);
+    let first = service
+        .capture("one", 7, CancellationToken::default())
+        .await
+        .unwrap();
+    let second = service
+        .capture("two", 7, CancellationToken::default())
+        .await
+        .unwrap();
+    assert_eq!(first.png, second.png);
+    assert_eq!(first.image_sha256, second.image_sha256);
+    assert_eq!(
+        first.image_sha256,
+        "e943b14142da05fcc1471ee0b639ffc1c11c74bfde83432ad3a26f994e4e665a"
+    );
+    assert!(first.summary.contains("privacy-marked residents"));
+    assert!(!first.summary.contains("Private Resident"));
+    assert!(first.summary.contains("textures_missing=1"));
+    assert!(first.data_url.starts_with("data:image/png;base64,"));
+    let decoder = png::Decoder::new(first.png.as_slice());
+    let mut reader = decoder.read_info().unwrap();
+    assert_eq!(
+        reader.info().srgb,
+        Some(png::SrgbRenderingIntent::Perceptual)
+    );
+    let mut pixels = vec![0; reader.output_buffer_size()];
+    let info = reader.next_frame(&mut pixels).unwrap();
+    let center = ((info.height as usize / 2) * info.width as usize + info.width as usize / 2) * 4;
+    assert_eq!(&pixels[center..center + 4], &[0, 255, 0, 255]);
+}
+
+#[tokio::test]
+async fn invalid_camera_huge_scene_stale_generation_and_size_limit_fail_closed() {
+    let mut bad = scene();
+    bad.camera.forward = [0.0, 0.0, 0.0];
+    let source = Arc::new(FakeScene {
+        scene: bad,
+        pending: Mutex::new(false),
+    });
+    let service = VisionService::new(
+        source,
+        VisionLimits {
+            width: 64,
+            height: 64,
+            minimum_interval: std::time::Duration::ZERO,
+            ..VisionLimits::default()
+        },
+    )
+    .unwrap();
+    service.set_generation(7);
+    assert_eq!(
+        service
+            .capture("bad", 7, CancellationToken::default())
+            .await
+            .unwrap_err(),
+        VisionError::InvalidScene
+    );
+    assert_eq!(
+        service
+            .capture("stale", 6, CancellationToken::default())
+            .await
+            .unwrap_err(),
+        VisionError::StaleGeneration
+    );
+    let source = Arc::new(FakeScene {
+        scene: scene(),
+        pending: Mutex::new(false),
+    });
+    let service = VisionService::new(
+        source,
+        VisionLimits {
+            width: 64,
+            height: 64,
+            max_png_bytes: 4096,
+            minimum_interval: std::time::Duration::ZERO,
+            ..VisionLimits::default()
+        },
+    )
+    .unwrap();
+    service.set_generation(7);
+    let result = service
+        .capture("size", 7, CancellationToken::default())
+        .await;
+    assert!(matches!(result, Ok(_) | Err(VisionError::ResourceLimit)));
+
+    let mut excessive_texture_work = scene();
+    excessive_texture_work.texture_fetches = VisionLimits::default().max_texture_fetches + 1;
+    let source = Arc::new(FakeScene {
+        scene: excessive_texture_work,
+        pending: Mutex::new(false),
+    });
+    let service = VisionService::new(
+        source,
+        VisionLimits {
+            width: 64,
+            height: 64,
+            minimum_interval: std::time::Duration::ZERO,
+            ..VisionLimits::default()
+        },
+    )
+    .unwrap();
+    service.set_generation(7);
+    assert_eq!(
+        service
+            .capture("texture-budget", 7, CancellationToken::default())
+            .await
+            .unwrap_err(),
+        VisionError::ResourceLimit
+    );
+}
+
+#[tokio::test]
+async fn cancellation_and_region_change_interrupt_nonblocking_capture() {
+    let source = Arc::new(FakeScene {
+        scene: scene(),
+        pending: Mutex::new(true),
+    });
+    let service = Arc::new(
+        VisionService::new(
+            source,
+            VisionLimits {
+                width: 64,
+                height: 64,
+                minimum_interval: std::time::Duration::ZERO,
+                ..VisionLimits::default()
+            },
+        )
+        .unwrap(),
+    );
+    service.set_generation(7);
+    let cancellation = CancellationTokenSource::new();
+    cancellation.cancel();
+    assert_eq!(
+        service
+            .capture("cancel", 7, cancellation.token())
+            .await
+            .unwrap_err(),
+        VisionError::Cancelled
+    );
+    let service2 = service.clone();
+    let task = tokio::spawn(async move {
+        service2
+            .capture("region", 7, CancellationToken::default())
+            .await
+    });
+    tokio::task::yield_now().await;
+    service.set_generation(8);
+    assert!(matches!(
+        task.await.unwrap(),
+        Err(VisionError::Superseded | VisionError::Cancelled)
+    ));
+
+    service.set_generation(9);
+    let first_service = service.clone();
+    let first = tokio::spawn(async move {
+        first_service
+            .capture("first", 9, CancellationToken::default())
+            .await
+    });
+    tokio::task::yield_now().await;
+    let second_cancel = CancellationTokenSource::new();
+    let second_service = service.clone();
+    let second_token = second_cancel.token();
+    let second =
+        tokio::spawn(async move { second_service.capture("second", 9, second_token).await });
+    assert_eq!(
+        tokio::time::timeout(std::time::Duration::from_secs(1), first)
+            .await
+            .unwrap()
+            .unwrap()
+            .unwrap_err(),
+        VisionError::Superseded
+    );
+    second_cancel.cancel();
+    assert_eq!(second.await.unwrap().unwrap_err(), VisionError::Cancelled);
+}
+
+#[test]
+fn explicit_visual_intent_is_narrow() {
+    assert!(visual_question("What do you see in front of you?"));
+    assert!(visual_question("Please look at this viewport"));
+    assert!(!visual_question("hello there"));
+    assert!(!visual_question("build a cube"));
+}
+
+struct FakeResponder {
+    parts: Mutex<usize>,
+    unsupported: bool,
+}
+impl InteractionResponder for FakeResponder {
+    fn respond(&self, request: ResponseRequest, _: CancellationToken) -> ResponderFuture<'_> {
+        Box::pin(async move {
+            *self.parts.lock().unwrap() = request.messages.last().unwrap().content.len();
+            if self.unsupported {
+                Err(InteractionModelError::MultimodalUnsupported)
+            } else {
+                VisibleResponse::new("seen").map_err(|_| InteractionModelError::Failed)
+            }
+        })
+    }
+}
+fn request(text: &str) -> ResponseRequest {
+    ResponseRequest {
+        delivery_id: BoundedText::new("id", "delivery").unwrap(),
+        sender_id: uuid(9),
+        session_id: BoundedText::new("session", "session").unwrap(),
+        channel: InteractionChannel::DirectIm,
+        origin: InteractionOrigin::AuthorizedIm,
+        intent: InteractionIntent::Informational,
+        messages: vec![CompletionMessage::text(MessageRole::Avatar, text).unwrap()],
+    }
+}
+
+#[tokio::test]
+async fn responder_adds_generic_image_payload_and_returns_summary_on_capability_rejection() {
+    let source = Arc::new(FakeScene {
+        scene: scene(),
+        pending: Mutex::new(false),
+    });
+    let vision = Arc::new(
+        VisionService::new(
+            source,
+            VisionLimits {
+                width: 64,
+                height: 64,
+                minimum_interval: std::time::Duration::ZERO,
+                ..VisionLimits::default()
+            },
+        )
+        .unwrap(),
+    );
+    vision.set_generation(7);
+    let endpoint = Arc::new(FakeResponder {
+        parts: Mutex::new(0),
+        unsupported: true,
+    });
+    let responder = VisionAugmentedResponder::new(vision, endpoint.clone());
+    let response = responder
+        .respond(request("What do you see?"), CancellationToken::default())
+        .await
+        .unwrap();
+    assert!(response.as_str().contains("does not support visual input"));
+    assert!(response.as_str().contains("Synthetic viewport"));
+    assert_eq!(*endpoint.parts.lock().unwrap(), 3);
+}
+
+#[tokio::test]
+async fn responder_never_captures_for_ordinary_text() {
+    let source = Arc::new(FakeScene {
+        scene: scene(),
+        pending: Mutex::new(false),
+    });
+    let vision = Arc::new(
+        VisionService::new(
+            source,
+            VisionLimits {
+                width: 64,
+                height: 64,
+                minimum_interval: std::time::Duration::ZERO,
+                ..VisionLimits::default()
+            },
+        )
+        .unwrap(),
+    );
+    vision.set_generation(7);
+    let endpoint = Arc::new(FakeResponder {
+        parts: Mutex::new(0),
+        unsupported: false,
+    });
+    let responder = VisionAugmentedResponder::new(vision.clone(), endpoint.clone());
+    assert_eq!(
+        responder
+            .respond(request("hello"), CancellationToken::default())
+            .await
+            .unwrap()
+            .as_str(),
+        "seen"
+    );
+    assert_eq!(*endpoint.parts.lock().unwrap(), 1);
+    assert!(vision.observations().is_empty());
+}

crates/metacrate-grid-agent/tests/dependency_policy.rs

@@ -2,10 +2,14 @@ use std::collections::BTreeSet;
 use std::fs;
 use std::path::{Path, PathBuf};
 
-const ALLOWED_DEPENDENCIES: [&str; 13] = [
+const ALLOWED_DEPENDENCIES: [&str; 17] = [
+    "base64",
     "crossterm",
     "libremetaverse",
+    "libremetaverse-imaging",
+    "libremetaverse-rendering-simple",
     "metacrate-lsl-tools",
+    "png",
     "libremetaverse-types",
     "reqwest",
     "rustls",
@@ -48,10 +52,10 @@ fn package_has_only_reviewed_rust_dependencies_and_no_build_script() {
 #[test]
 fn runtime_source_has_no_subprocess_or_native_abi_escape_hatch() {
     let source = PathBuf::from(env!("CARGO_MANIFEST_DIR")).join("src");
-    let mut files = Vec::with_capacity(34);
+    let mut files = Vec::with_capacity(36);
     collect_rust_files(&source, &mut files);
     assert!(
-        files.len() <= 34,
+        files.len() <= 36,
         "source-file count needs a reviewed bound update"
     );
     for path in files {
@@ -82,6 +86,27 @@ fn runtime_source_has_no_subprocess_or_native_abi_escape_hatch() {
     }
 }
 
+#[test]
+fn vision_path_selects_only_the_pure_rust_renderer_and_codec() {
+    let root = PathBuf::from(env!("CARGO_MANIFEST_DIR"));
+    let manifest = fs::read_to_string(root.join("Cargo.toml")).expect("agent manifest");
+    assert!(manifest.contains("default-features = false, features = [\"rust-j2k\"]"));
+    for forbidden in ["imaging-skia", "openjpeg", "skia-safe"] {
+        assert!(!manifest.contains(forbidden), "vision enables {forbidden}");
+    }
+    let renderer = fs::read_to_string(root.join("../libremetaverse-rendering-simple/Cargo.toml"))
+        .expect("simple renderer manifest");
+    assert!(!renderer.contains("imaging-skia"));
+    assert!(!renderer.contains("openjpeg"));
+    let vision = fs::read_to_string(root.join("src/vision.rs")).expect("vision source");
+    for forbidden in ["screenshot crate", "gpu adapter", "std::process"] {
+        assert!(
+            !vision.to_ascii_lowercase().contains(forbidden),
+            "vision must remain synthetic and pure Rust: {forbidden}"
+        );
+    }
+}
+
 #[test]
 fn compatibility_crates_never_depend_on_or_reexport_metacrate_crates() {
     let workspace = PathBuf::from(env!("CARGO_MANIFEST_DIR"))

crates/metacrate-grid-agent/tests/llm_transport.rs

@@ -503,6 +503,45 @@ async fn retry_classification_retry_after_timeout_and_cancellation_are_bounded()
     assert_eq!(request.await.expect_err("cancelled"), LlmError::Cancelled);
 }
 
+#[tokio::test]
+async fn multimodal_capability_rejection_is_typed_and_never_retried() {
+    let mut server = fake_server(vec![ResponsePlan::status(415)]).await;
+    let mut retry_limits = limits();
+    retry_limits.max_retries = 2;
+    let image = CompletionMessage {
+        role: MessageRole::Avatar,
+        content: metacrate_grid_agent::BoundedVec::try_from_vec(
+            "message.content",
+            vec![ContentPart::Image {
+                url: metacrate_grid_agent::BoundedText::new(
+                    "image.url",
+                    "data:image/png;base64,iVBORw0KGgo=",
+                )
+                .unwrap(),
+                detail: ImageDetail::Low,
+            }],
+        )
+        .unwrap(),
+        tool_call_id: None,
+        proposed_calls: metacrate_grid_agent::BoundedVec::new(),
+    };
+    assert_eq!(
+        client(&server.url, retry_limits)
+            .complete(&[image], &[], &CancellationToken::default())
+            .await
+            .expect_err("multimodal rejection"),
+        LlmError::MultimodalUnsupported
+    );
+    server.requests.recv().await.expect("one image request");
+    assert!(
+        !matches!(
+            tokio::time::timeout(Duration::from_millis(100), server.requests.recv()).await,
+            Ok(Some(_))
+        ),
+        "capability rejection must not resend the large image"
+    );
+}
+
 #[tokio::test]
 async fn concurrent_slow_sessions_respect_the_shared_semaphore() {
     let plans = (0..6)