agent: add NodeBody::Image for Qwen3-VL vision input

Images are rendered as `<|vision_start|>` + N × `<|image_pad|>` +
`<|vision_end|>` where N is computed from the image dimensions using
Qwen3-VL's smart_resize rules (patch_size=16, merge_size=2, min=64K,
max=16M pixels). The token count matches what vLLM will produce at
request time, so budget accounting stays accurate.

Bytes are stored inline on the leaf and base64-encoded in the JSON
form. Token IDs are hand-assembled instead of re-running the tokenizer
on a potentially-huge placeholder string.

Follow-ups: view_image tool rewrite, multi_modal_data on the vLLM
request, API-layer plumbing from leaf bytes to request body.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>

src/agent/context.rs

@@ -81,10 +81,33 @@ pub enum NodeBody {
     Memory { key: String, text: String, score: Option<f64> },
     Dmn(String),
 
+    // Vision input — rendered as <|vision_start|> <|image_pad|>×N <|vision_end|>.
+    // `token_count` is N, the count vLLM will compute for this image's grid.
+    Image {
+        #[serde(with = "b64_bytes")]
+        bytes: Vec<u8>,
+        mime: String,
+        orig_height: u32,
+        orig_width: u32,
+        token_count: u32,
+    },
+
     // Non-visible (0 tokens in prompt)
     Log(String),
 }
 
+mod b64_bytes {
+    use base64::{Engine, engine::general_purpose::STANDARD};
+    use serde::{Serializer, Deserializer, Deserialize};
+    pub fn serialize<S: Serializer>(bytes: &[u8], s: S) -> Result<S::Ok, S::Error> {
+        s.serialize_str(&STANDARD.encode(bytes))
+    }
+    pub fn deserialize<'de, D: Deserializer<'de>>(d: D) -> Result<Vec<u8>, D::Error> {
+        let s = String::deserialize(d)?;
+        STANDARD.decode(s).map_err(serde::de::Error::custom)
+    }
+}
+
 /// A leaf node: typed content with cached token IDs.
 /// Token IDs are not serialized — they're recomputed on deserialization.
 #[derive(Debug, Clone, Serialize)]
@@ -103,11 +126,7 @@ impl<'de> Deserialize<'de> for NodeLeaf {
             timestamp: DateTime<Utc>,
         }
         let raw = Raw::deserialize(deserializer)?;
-        let token_ids = if raw.body.is_prompt_visible() {
+        let token_ids = raw.body.compute_token_ids();
-            tokenizer::encode(&raw.body.render())
-        } else {
-            vec![]
-        };
         Ok(NodeLeaf { body: raw.body, token_ids, timestamp: raw.timestamp })
     }
 }
@@ -221,6 +240,13 @@ impl NodeBody {
                 out.push_str(text);
                 out.push_str("<|im_end|>\n");
             }
+            Self::Image { token_count, .. } => {
+                out.push_str("<|vision_start|>");
+                for _ in 0..*token_count {
+                    out.push_str("<|image_pad|>");
+                }
+                out.push_str("<|vision_end|>");
+            }
         }
     }
 
@@ -235,6 +261,26 @@ impl NodeBody {
         !matches!(self, Self::Thinking(_) | Self::Log(_))
     }
 
+    /// Hand-assemble token IDs for body types where running the tokenizer
+    /// on the rendered text would be needlessly expensive (Image). Falls
+    /// back to encoding the rendered text for everything else.
+    fn compute_token_ids(&self) -> Vec<u32> {
+        if !self.is_prompt_visible() {
+            return Vec::new();
+        }
+        match self {
+            Self::Image { token_count, .. } => {
+                let mut ids = Vec::with_capacity(*token_count as usize + 2);
+                ids.push(tokenizer::VISION_START);
+                ids.extend(std::iter::repeat(tokenizer::IMAGE_PAD)
+                    .take(*token_count as usize));
+                ids.push(tokenizer::VISION_END);
+                ids
+            }
+            _ => tokenizer::encode(&self.render()),
+        }
+    }
+
     /// The text content of this leaf (for display, not rendering).
     pub fn text(&self) -> &str {
         match self {
@@ -242,17 +288,14 @@ impl NodeBody {
                 | Self::ToolResult(t) | Self::Dmn(t) => t,
             Self::ToolCall { name, .. } => name,
             Self::Memory { text, .. } => text,
+            Self::Image { mime, .. } => mime,
         }
     }
 }
 
 impl NodeLeaf {
     fn new(body: NodeBody) -> Self {
-        let token_ids = if body.is_prompt_visible() {
+        let token_ids = body.compute_token_ids();
-            tokenizer::encode(&body.render())
-        } else {
-            vec![]
-        };
         Self { body, token_ids, timestamp: Utc::now() }
     }
 
@@ -305,6 +348,24 @@ impl AstNode {
         Self::Leaf(NodeLeaf::new(NodeBody::Log(text.into())))
     }
 
+    /// Build an Image leaf. `token_count` is computed from the image
+    /// dimensions using Qwen3-VL's resizing rules.
+    pub fn image(
+        bytes: Vec<u8>,
+        mime: impl Into<String>,
+        orig_height: u32,
+        orig_width: u32,
+    ) -> Self {
+        let token_count = qwen3_image_token_count(orig_height, orig_width);
+        Self::Leaf(NodeLeaf::new(NodeBody::Image {
+            bytes,
+            mime: mime.into(),
+            orig_height,
+            orig_width,
+            token_count,
+        }))
+    }
+
     // -- Branch constructors --------------------------------------------------
 
     pub fn branch(role: Role, children: Vec<AstNode>) -> Self {
@@ -334,11 +395,7 @@ impl AstNode {
     pub fn retokenize(self) -> Self {
         match self {
             Self::Leaf(leaf) => {
-                let token_ids = if leaf.body.is_prompt_visible() {
+                let token_ids = leaf.body.compute_token_ids();
-                    tokenizer::encode(&leaf.body.render())
-                } else {
-                    vec![]
-                };
                 Self::Leaf(NodeLeaf { token_ids, ..leaf })
             }
             Self::Branch { role, children, timestamp, memory_scores } => Self::Branch {
@@ -397,6 +454,8 @@ impl AstNode {
                     None => format!("mem: {}", key),
                 },
                 NodeBody::Dmn(_) => "dmn".into(),
+                NodeBody::Image { orig_height, orig_width, token_count, .. } =>
+                    format!("image: {}x{} ({} tokens)", orig_width, orig_height, token_count),
                 NodeBody::Log(t) => format!("log: {}", truncate_preview(t, 60)),
             },
         }
@@ -857,11 +916,7 @@ impl ContextState {
         let node = &mut nodes[index];
         match node {
             AstNode::Leaf(leaf) => {
-                let token_ids = if body.is_prompt_visible() {
+                let token_ids = body.compute_token_ids();
-                    tokenizer::encode(&body.render())
-                } else {
-                    vec![]
-                };
                 leaf.body = body;
                 leaf.token_ids = token_ids;
             }
@@ -991,6 +1046,58 @@ impl ContextState {
     }
 }
 
+// ---------------------------------------------------------------------------
+// Qwen3-VL image token count
+//
+// Port of Qwen2VLImageProcessor.smart_resize + image_token_count. We need the
+// exact same answer that vLLM's Qwen3VL processor will produce, because the
+// token stream in our context must match what vLLM expands `<|image_pad|>`
+// to at request time. Constants come from Qwen3.5-27B's preprocessor_config.
+// ---------------------------------------------------------------------------
+
+const QWEN3_PATCH_SIZE: u32 = 16;
+const QWEN3_MERGE_SIZE: u32 = 2;
+const QWEN3_MIN_PIXELS: u64 = 65_536;
+const QWEN3_MAX_PIXELS: u64 = 16_777_216;
+
+fn smart_resize(h: u32, w: u32, factor: u32, min_pixels: u64, max_pixels: u64) -> (u32, u32) {
+    let max_s = h.max(w) as f64;
+    let min_s = h.min(w) as f64;
+    assert!(max_s / min_s <= 200.0, "aspect ratio too extreme: {}x{}", h, w);
+
+    let fh = h as f64;
+    let fw = w as f64;
+    let ff = factor as f64;
+
+    let h_bar = ((fh / ff).round() as u32) * factor;
+    let w_bar = ((fw / ff).round() as u32) * factor;
+    let total = (h_bar as u64) * (w_bar as u64);
+
+    if total > max_pixels {
+        let beta = ((fh * fw) / max_pixels as f64).sqrt();
+        let hf = ((fh / beta / ff).floor() as u32) * factor;
+        let wf = ((fw / beta / ff).floor() as u32) * factor;
+        (hf.max(factor), wf.max(factor))
+    } else if total < min_pixels {
+        let beta = (min_pixels as f64 / (fh * fw)).sqrt();
+        let hc = ((fh * beta / ff).ceil() as u32) * factor;
+        let wc = ((fw * beta / ff).ceil() as u32) * factor;
+        (hc, wc)
+    } else {
+        (h_bar, w_bar)
+    }
+}
+
+/// Compute how many `<|image_pad|>` tokens vLLM will emit for an image of
+/// the given dimensions. Matches Qwen3VL's feature-size calculation exactly:
+///   (grid_h * grid_w) / merge_size^2
+/// where (grid_h, grid_w) = resized dims / patch_size.
+fn qwen3_image_token_count(orig_h: u32, orig_w: u32) -> u32 {
+    let factor = QWEN3_PATCH_SIZE * QWEN3_MERGE_SIZE;
+    let (rh, rw) = smart_resize(orig_h, orig_w, factor, QWEN3_MIN_PIXELS, QWEN3_MAX_PIXELS);
+    (rh / QWEN3_PATCH_SIZE) * (rw / QWEN3_PATCH_SIZE) / (QWEN3_MERGE_SIZE * QWEN3_MERGE_SIZE)
+}
+
 pub fn context_window() -> usize {
     let app = crate::config::app();
     app.backends.get(&app.default_backend)
@@ -1370,6 +1477,82 @@ mod tests {
         assert!(serde_json::from_str::<AstNode>(json).is_err());
     }
 
+    // -- Image leaf tests ---------------------------------------------------------
+
+    #[test]
+    fn test_smart_resize_within_bounds() {
+        // Typical case: 1024x768 → rounded to multiples of 32, under max.
+        let (h, w) = smart_resize(768, 1024, 32, 65_536, 16_777_216);
+        assert_eq!(h, 768);
+        assert_eq!(w, 1024);
+    }
+
+    #[test]
+    fn test_smart_resize_upscales_tiny() {
+        // 32x32 = 1024 pixels, below min_pixels=65536. Should scale up.
+        let (h, w) = smart_resize(32, 32, 32, 65_536, 16_777_216);
+        assert!((h as u64) * (w as u64) >= 65_536,
+            "resized {}x{} is under min_pixels", h, w);
+        assert_eq!(h % 32, 0);
+        assert_eq!(w % 32, 0);
+    }
+
+    #[test]
+    fn test_smart_resize_downscales_huge() {
+        // 8000x6000 = 48M pixels, above max_pixels=16M. Should scale down.
+        let (h, w) = smart_resize(8000, 6000, 32, 65_536, 16_777_216);
+        assert!((h as u64) * (w as u64) <= 16_777_216,
+            "resized {}x{} exceeds max_pixels", h, w);
+        assert_eq!(h % 32, 0);
+        assert_eq!(w % 32, 0);
+    }
+
+    #[test]
+    fn test_qwen3_token_count_matches_formula() {
+        // 512x512 → resized to 512x512 (already multiple of 32, within bounds).
+        // grid = 32x32, tokens = 32*32/4 = 256.
+        assert_eq!(qwen3_image_token_count(512, 512), 256);
+    }
+
+    #[test]
+    fn test_image_render_and_token_ids() {
+        let node = AstNode::image(vec![0u8, 1, 2, 3], "image/png", 512, 512);
+        let leaf = node.leaf().unwrap();
+        // 3 tokens of bookend + 256 image_pad tokens
+        assert_eq!(leaf.token_ids().len(), 258);
+        assert_eq!(leaf.token_ids()[0], tokenizer::VISION_START);
+        assert_eq!(leaf.token_ids()[257], tokenizer::VISION_END);
+        for pad in &leaf.token_ids()[1..257] {
+            assert_eq!(*pad, tokenizer::IMAGE_PAD);
+        }
+        // Rendered text has the expected bookends.
+        let rendered = leaf.body().render();
+        assert!(rendered.starts_with("<|vision_start|>"));
+        assert!(rendered.ends_with("<|vision_end|>"));
+    }
+
+    #[test]
+    fn test_image_serde_roundtrip() {
+        let node = AstNode::image(vec![0xDE, 0xAD, 0xBE, 0xEF], "image/png", 64, 64);
+        let json = serde_json::to_string(&node).unwrap();
+        // bytes must be base64-encoded in the JSON form
+        assert!(json.contains("3q2+7w=="));
+        let back: AstNode = serde_json::from_str(&json).unwrap();
+        let leaf = back.leaf().unwrap();
+        match leaf.body() {
+            NodeBody::Image { bytes, mime, orig_height, orig_width, token_count } => {
+                assert_eq!(bytes, &[0xDE, 0xAD, 0xBE, 0xEF]);
+                assert_eq!(mime, "image/png");
+                assert_eq!(*orig_height, 64);
+                assert_eq!(*orig_width, 64);
+                assert_eq!(*token_count, qwen3_image_token_count(64, 64));
+            }
+            other => panic!("expected Image, got {:?}", other),
+        }
+        // token_ids are recomputed on deserialization
+        assert_eq!(leaf.token_ids().len(), leaf.tokens());
+    }
+
     #[test]
     fn test_timestamp_present_accepted() {
         let json = r#"{"Leaf":{"body":{"Content":"hi"},"timestamp":"2026-04-16T12:00:00Z"}}"#;

src/agent/tokenizer.rs

@@ -16,6 +16,9 @@ static TOKENIZER: OnceLock<Tokenizer> = OnceLock::new();
 /// Special token IDs for Qwen 3.5
 pub const IM_START: u32 = 248045;
 pub const IM_END: u32 = 248046;
+pub const VISION_START: u32 = 248053;
+pub const VISION_END: u32 = 248054;
+pub const IMAGE_PAD: u32 = 248056;
 
 /// Initialize the global tokenizer from a file path.
 /// Call once at startup. Panics if the file can't be loaded.

src/user/chat.rs

@@ -486,6 +486,11 @@ impl InteractScreen {
                         if t.is_empty() { vec![] }
                         else { vec![(PaneTarget::ToolResult, text, Marker::None)] }
                     }
+                    NodeBody::Image { orig_height, orig_width, .. } => {
+                        vec![(PaneTarget::Conversation,
+                              format!("[image {}x{}]", orig_width, orig_height),
+                              Marker::None)]
+                    }
                 }
             }
             AstNode::Branch { role, children, .. } => {