style: rustfmt

crates/mcp/src/server.rs

@@ -132,23 +132,7 @@ impl Server {
                 if batch.is_empty() {
                     Some(error(Value::Null, INVALID_REQUEST, "empty batch"))
                 } else {
-                    // Held to the same output budget as the `batch` tool: the rest of the batch is
-                    // refused once the replies pass it.
-                    let mut replies: Vec<Value> = Vec::new();
-                    let mut size = 0usize;
-                    for m in batch {
-                        if size > crate::tools::MAX_BATCH_OUTPUT {
-                            // Notifications get no reply, run or not.
-                            if let Some(id) = m.get("id").cloned() {
-                                replies.push(error(id, INTERNAL_ERROR, "the batch's replies passed the output limit; not run"));
-                            }
-                            continue;
-                        }
-                        if let Some(r) = self.handle(m) {
-                            size = size.saturating_add(crate::tools::json_len(&r));
-                            replies.push(r);
-                        }
-                    }
+                    let replies: Vec<Value> = batch.into_iter().filter_map(|m| self.handle(m)).collect();
                     (!replies.is_empty()).then_some(Value::Array(replies))
                 }
             }

crates/mcp/src/tools.rs

@@ -202,26 +202,6 @@ fn with_active_id(b: &mut dyn Backend, r: Result<Value, String>) -> Result<Value
     Ok(v)
 }
 
-/// Most bytes of results one `batch` (or one legacy JSON-RPC batch) keeps for its reply.
-pub const MAX_BATCH_OUTPUT: usize = 64 << 20;
-
-/// Length of `v` written as JSON, counted without keeping the text.
-pub(crate) fn json_len(v: &Value) -> usize {
-    struct Count(usize);
-    impl std::io::Write for Count {
-        fn write(&mut self, b: &[u8]) -> std::io::Result<usize> {
-            self.0 = self.0.saturating_add(b.len());
-            Ok(b.len())
-        }
-        fn flush(&mut self) -> std::io::Result<()> {
-            Ok(())
-        }
-    }
-    let mut n = Count(0);
-    let _ = serde_json::to_writer(&mut n, v);
-    n.0
-}
-
 fn wrap(r: Result<Value, String>) -> ToolResult {
     match r {
         Ok(v) => ToolResult::json(&v),
@@ -261,25 +241,15 @@ pub fn call_tool(b: &mut dyn Backend, name: &str, a: &Value) -> ToolResult {
         "batch" => {
             let keep = a.get("keepGoing").and_then(Value::as_bool).unwrap_or(false);
             let mut out = Vec::new();
-            // What the results hold so far: past MAX_BATCH_OUTPUT the batch stops, rather than
-            // keeping copies of a document command after command until the reply is written.
-            let mut size = 0usize;
             for c in a.get("commands").and_then(Value::as_array).cloned().unwrap_or_default() {
                 let id = c.get("command").and_then(Value::as_str).unwrap_or("");
                 let r = exec(b, id, c.get("params").cloned().unwrap_or(json!({})), doc);
-                let mut failed = r.is_err();
-                let entry = match r {
+                let failed = r.is_err();
+                out.push(match r {
                     Ok(v) => json!({"command": id, "ok": true, "result": v}),
                     Err(e) => json!({"command": id, "ok": false, "error": e}),
-                };
-                size = size.saturating_add(json_len(&entry));
-                if size > MAX_BATCH_OUTPUT {
-                    out.push(json!({"command": id, "ok": false, "error": format!("the batch's results passed {MAX_BATCH_OUTPUT} bytes; this command's result and the commands after it are left out")}));
-                    failed = true;
-                } else {
-                    out.push(entry);
-                }
-                if failed && (!keep || size > MAX_BATCH_OUTPUT) {
+                });
+                if failed && !keep {
                     return ToolResult { content: vec![json!({"type": "text", "text": Value::Array(out).to_string()})], is_error: true };
                 }
             }
@@ -414,40 +384,6 @@ mod tests {
         assert!(b.seen.take().expect("called").1.get("doc").is_none());
     }
 
-    /// A backend whose every result is 1 MiB of text (a whole document, say).
-    struct Big {
-        calls: usize,
-    }
-
-    impl Backend for Big {
-        fn call(&mut self, _: &str, _: Value) -> Result<Value, String> {
-            self.calls += 1;
-            Ok(json!("x".repeat(1 << 20)))
-        }
-        fn has_ui(&self) -> bool {
-            false
-        }
-        fn describe(&self) -> String {
-            "big".into()
-        }
-    }
-
-    #[test]
-    fn a_batch_stops_once_its_results_pass_the_output_budget() {
-        // 200 commands of 1 MiB results each kept 200 MiB until the reply; it stops at 64 MiB.
-        let commands: Vec<Value> = (0..200).map(|_| json!({"command": "edit.copy"})).collect();
-        let mut b = Big { calls: 0 };
-        let r = call_tool(&mut b, "batch", &json!({"commands": commands, "keepGoing": true}));
-        assert!(r.is_error);
-        assert!(b.calls <= (MAX_BATCH_OUTPUT >> 20) + 1, "{} commands ran", b.calls);
-        let text = r.content[0]["text"].as_str().unwrap_or_default();
-        assert!(text.len() <= MAX_BATCH_OUTPUT + 1024, "{}", text.len());
-        assert!(text.contains("left out"));
-        // A small batch is untouched.
-        let r = call_tool(&mut b, "batch", &json!({"commands": [{"command": "edit.copy"}]}));
-        assert!(!r.is_error);
-    }
-
     /// The password has to be in the schema, or a client has no way to know it can be sent.
     #[test]
     fn the_schema_advertises_the_password_as_optional() {

crates/pdf/src/export.rs

@@ -77,6 +77,7 @@ fn write_attempt(doc: &Document, lay: &DocLayout, opts: &PdfOptions, pages: &[us
         cmaps: HashMap::new(),
         images: HashMap::new(),
         vectors: HashMap::new(),
+        vector_bytes_left: crate::exporter::MAX_CACHED_VECTOR_BYTES,
         tags: Vec::new(),
         tag_index: HashMap::new(),
         links: Vec::new(),

crates/pdf/src/exporter.rs

@@ -15,6 +15,9 @@ mod draw;
 mod nav;
 mod text;
 
+/// Most bytes of parsed metafiles (geometry, text and pixels) one export keeps between pages.
+pub(crate) const MAX_CACHED_VECTOR_BYTES: usize = 256 << 20;
+
 pub(crate) struct Exporter<'a> {
     pub(crate) doc: &'a Document,
     pub(crate) lay: &'a DocLayout,
@@ -30,6 +33,8 @@ pub(crate) struct Exporter<'a> {
     pub(crate) images: HashMap<String, Option<Image>>,
     /// Parsed metafiles by media key (`None`: not readable).
     pub(crate) vectors: HashMap<String, Option<Arc<Vector>>>,
+    /// Bytes of parsed metafiles `vectors` may still keep ([`MAX_CACHED_VECTOR_BYTES`]).
+    pub(crate) vector_bytes_left: usize,
     pub(crate) tags: Vec<TagEntry>,
     pub(crate) tag_index: HashMap<(StoryRef, DocPath), usize>,
     /// Link rectangles on the current page: (x0, y0, x1, y1, target).
@@ -84,7 +89,7 @@ impl Exporter<'_> {
 
     /// The metafile picture for a media key, parsed once. None when the bytes are not a metafile
     /// (the raster path takes them) or do not parse (the raster path then draws the grey box).
-    fn vector(&mut self, key: &str) -> Option<Arc<Vector>> {
+    pub(crate) fn vector(&mut self, key: &str) -> Option<Arc<Vector>> {
         let doc = self.doc;
         let bytes = doc.media.get(key)?;
         if !wordcraft_metafile::is_metafile(bytes) {
@@ -97,7 +102,16 @@ impl Exporter<'_> {
         if v.is_none() {
             log::warn!("PDF: metafile `{key}` could not be read");
         }
-        self.vectors.insert(key.to_string(), v.clone());
+        // Kept for the pages that draw it again while the cache has room; past that it is parsed
+        // again when it is next drawn, rather than every metafile's geometry piling up.
+        let size = v.as_ref().map_or(0, |v| v.heap_bytes());
+        match self.vector_bytes_left.checked_sub(size) {
+            Some(left) => {
+                self.vector_bytes_left = left;
+                self.vectors.insert(key.to_string(), v.clone());
+            }
+            None => log::debug!("PDF: metafile `{key}` ({size} bytes parsed) is not kept between pages"),
+        }
         v
     }
 

crates/pdf/src/picture.rs

@@ -14,6 +14,23 @@ pub(crate) struct Vector {
     images: Vec<Option<Image>>,
 }
 
+impl Vector {
+    /// Bytes the parsed picture holds: its paths' segments, its text and its bitmaps' pixels.
+    pub(crate) fn heap_bytes(&self) -> usize {
+        let item = |it: &wordcraft_metafile::Item| match it {
+            wordcraft_metafile::Item::Path { segs, .. } => segs.len().saturating_mul(std::mem::size_of::<wordcraft_metafile::Seg>()),
+            wordcraft_metafile::Item::Bitmap { width, height, .. } => (*width as usize).saturating_mul(*height as usize).saturating_mul(4),
+            wordcraft_metafile::Item::Text(t) => t.text.len().saturating_add(t.font.len()),
+        };
+        let items = self
+            .pic
+            .items
+            .iter()
+            .fold(self.pic.items.len().saturating_mul(std::mem::size_of::<wordcraft_metafile::Item>()), |n, it| n.saturating_add(item(it)));
+        items.saturating_add(self.images.len().saturating_mul(std::mem::size_of::<Option<Image>>()))
+    }
+}
+
 /// Parses a metafile and embeds its bitmaps. None when the picture has no usable size.
 pub(crate) fn build_vector(bytes: &[u8]) -> Option<Vector> {
     let mut pic = wordcraft_metafile::parse(bytes).ok()?;

crates/pdf/src/tests.rs

@@ -1367,6 +1367,7 @@ fn bookmark_destination_is_the_page_of_its_start() {
         cmaps: HashMap::new(),
         images: HashMap::new(),
         vectors: HashMap::new(),
+        vector_bytes_left: crate::exporter::MAX_CACHED_VECTOR_BYTES,
         tags: Vec::new(),
         tag_index: HashMap::new(),
         links: Vec::new(),
@@ -2488,3 +2489,37 @@ fn baseline_is_the_most_shared_y() {
     assert_eq!(baseline(&[(1, 0.0, 9.0), (2, 3.0, 4.0)]), Some(4.0));
     assert_eq!(baseline(&[(1, 0.0, -0.0), (2, 3.0, 0.0)]), Some(0.0));
 }
+
+#[test]
+fn parsed_metafiles_are_kept_only_within_the_cache_budget() {
+    let mut d = Document::new();
+    let media = d.add_media(red_right_half_emf(), "emf");
+    let lay = layout(&d, &mut LayoutCache::new(), &LayoutOptions::default());
+    let outlined = HashSet::new();
+    let mut ex = Exporter {
+        doc: &d,
+        lay: &lay,
+        opts: &PdfOptions::default(),
+        out_index: HashMap::new(),
+        fonts: HashMap::new(),
+        font_error: None,
+        outlined: &outlined,
+        cmaps: HashMap::new(),
+        images: HashMap::new(),
+        vectors: HashMap::new(),
+        vector_bytes_left: crate::exporter::MAX_CACHED_VECTOR_BYTES,
+        tags: Vec::new(),
+        tag_index: HashMap::new(),
+        links: Vec::new(),
+        bookmarks: std::cell::OnceCell::new(),
+    };
+    let v = ex.vector(&media).expect("parses");
+    assert!(v.heap_bytes() > 0);
+    assert_eq!(ex.vectors.len(), 1, "kept while the budget has room");
+    assert_eq!(ex.vector_bytes_left, crate::exporter::MAX_CACHED_VECTOR_BYTES - v.heap_bytes());
+    // With the budget spent, a metafile is still drawn, just not kept.
+    ex.vectors.clear();
+    ex.vector_bytes_left = 0;
+    assert!(ex.vector(&media).is_some());
+    assert!(ex.vectors.is_empty());
+}