Apply review fixes

- crates/pptx/src/read.rs

crates/mcp/src/server.rs

@@ -16,12 +16,6 @@ const INVALID_REQUEST: i64 = -32600;
 const METHOD_NOT_FOUND: i64 = -32601;
 const INVALID_PARAMS: i64 = -32602;
 const INTERNAL_ERROR: i64 = -32603;
-
-/// Entries one batch (a JSON-RPC batch, or the `batch` tool's commands) may hold.
-pub(crate) const MAX_BATCH_ITEMS: usize = 1000;
-/// Bytes of results one batch may gather before the rest of it is refused: each result can copy
-/// much of the document (an inspection holds every note), so a batch holds at most this much.
-pub(crate) const MAX_BATCH_BYTES: usize = 64 << 20;
 const RESOURCE_NOT_FOUND: i64 = -32002;
 /// Keep request frames in step with the connected backend's reply limit.
 const MAX_FRAME_BYTES: usize = 16 * 1024 * 1024;
@@ -99,28 +93,8 @@ impl Server {
                 // Batches were removed in 2025-06-18; still answer older clients sensibly.
                 if batch.is_empty() {
                     Some(error(Value::Null, INVALID_REQUEST, "empty batch"))
-                } else if batch.len() > MAX_BATCH_ITEMS {
-                    Some(error(Value::Null, INVALID_REQUEST, format!("a batch holds at most {MAX_BATCH_ITEMS} messages")))
                 } else {
-                    let mut used = 0usize;
-                    let mut replies: Vec<Value> = Vec::new();
-                    for m in batch {
-                        if used > MAX_BATCH_BYTES {
-                            // Not run: answer each remaining request with a short error.
-                            if let Some(id) = m.get("id").filter(|id| id.is_string() || id.is_number()) {
-                                replies.push(error(id.clone(), INTERNAL_ERROR, "the batch's replies passed 64 MiB; not run"));
-                            }
-                            continue;
-                        }
-                        if let Some(r) = self.handle(m) {
-                            used = used.saturating_add(r.to_string().len());
-                            replies.push(if used > MAX_BATCH_BYTES {
-                                error(r.get("id").cloned().unwrap_or(Value::Null), INTERNAL_ERROR, "the batch's replies passed 64 MiB; reply dropped")
-                            } else {
-                                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

@@ -357,30 +357,18 @@ pub fn call_tool(b: &mut dyn Backend, name: &str, a: &Value) -> ToolResult {
             wrap(exec(b, c, a.get("params").cloned().unwrap_or(json!({}))))
         }
         "batch" => {
-            use crate::server::{MAX_BATCH_BYTES, MAX_BATCH_ITEMS};
             let keep = a.get("keepGoing").and_then(Value::as_bool).unwrap_or(false);
-            let commands = a.get("commands").and_then(Value::as_array).map(Vec::as_slice).unwrap_or_default();
-            if commands.len() > MAX_BATCH_ITEMS {
-                return ToolResult::error(format!("a batch holds at most {MAX_BATCH_ITEMS} commands"));
-            }
             let mut out = vec![];
-            let mut used = 0usize;
-            for c in commands {
+            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 (entry, failed) = match exec(b, id, c.get("params").cloned().unwrap_or(json!({}))) {
-                    Ok(v) => (json!({"command": id, "ok": true, "result": v}), false),
-                    Err(e) => (json!({"command": id, "ok": false, "error": e}), true),
-                };
-                // What the reply carries: the entry as text, escaped again inside the JSON-RPC reply.
-                // Failures count too: with keepGoing, a remote backend's large errors add up.
-                used = used.saturating_add(json!(entry.to_string()).to_string().len());
-                if used > MAX_BATCH_BYTES {
-                    out.push(json!({"command": id, "ok": false, "error": "the batch's results passed 64 MiB; stopped here"}));
-                    return ToolResult { content: vec![json!({"type": "text", "text": Value::Array(out).to_string()})], is_error: true };
-                }
-                out.push(entry);
-                if failed && !keep {
-                    return ToolResult { content: vec![json!({"type": "text", "text": Value::Array(out).to_string()})], is_error: true };
+                match exec(b, id, c.get("params").cloned().unwrap_or(json!({}))) {
+                    Ok(v) => out.push(json!({"command": id, "ok": true, "result": v})),
+                    Err(e) => {
+                        out.push(json!({"command": id, "ok": false, "error": e}));
+                        if !keep {
+                            return ToolResult { content: vec![json!({"type": "text", "text": Value::Array(out).to_string()})], is_error: true };
+                        }
+                    }
                 }
             }
             ToolResult::json(&Value::Array(out))

crates/mcp/tests/it/batch_budget.rs

@@ -1,70 +0,0 @@
-//! A batch (JSON-RPC or the `batch` tool) is bounded in entries and in the bytes of results it
-//! gathers, so a short request cannot copy a large document into memory hundreds of times.
-
-use deckcraft_mcp::{Headless, Server};
-use serde_json::{Value, json};
-
-fn server_with_notes(bytes: usize) -> Server {
-    let mut server = Server::new(Box::new(Headless::with_document()));
-    let line = json!({"jsonrpc": "2.0", "id": 1, "method": "tools/call", "params": {"name": "run_command", "arguments": {"command": "slide.notes", "params": {"text": "n".repeat(bytes)}}}});
-    let reply: Value = serde_json::from_str(&server.handle_line(&line.to_string()).unwrap()).unwrap();
-    assert_ne!(reply["result"]["isError"], true, "{reply}");
-    server
-}
-
-#[test]
-fn json_rpc_batches_are_bounded_in_entries_and_reply_bytes() {
-    let mut server = server_with_notes(1 << 20);
-    let ping = |i: usize| json!({"jsonrpc": "2.0", "id": i, "method": "ping"});
-    let too_many: Vec<Value> = (0..1001).map(ping).collect();
-    let reply: Value = serde_json::from_str(&server.handle_line(&Value::Array(too_many).to_string()).unwrap()).unwrap();
-    assert_eq!(reply["error"]["code"], -32600, "{reply}");
-    // 100 inspections of a slide with 1 MB of notes: about 64 are answered, the rest refused.
-    let inspect = |i: usize| json!({"jsonrpc": "2.0", "id": i, "method": "tools/call", "params": {"name": "inspect_slide", "arguments": {}}});
-    let line = Value::Array((0..100).map(inspect).collect()).to_string();
-    let reply = server.handle_line(&line).unwrap();
-    assert!(reply.len() < (64 << 20) + (4 << 20), "{} bytes", reply.len());
-    let replies: Vec<Value> = serde_json::from_str(&reply).unwrap();
-    assert_eq!(replies.len(), 100, "every request is answered");
-    assert!(replies.last().unwrap()["error"]["message"].as_str().unwrap().contains("64 MiB"));
-}
-
-#[test]
-fn the_batch_tool_is_bounded_in_commands_and_result_bytes() {
-    let mut server = server_with_notes(1 << 20);
-    let call = |server: &mut Server, commands: Vec<Value>| {
-        let line = json!({"jsonrpc": "2.0", "id": 1, "method": "tools/call", "params": {"name": "batch", "arguments": {"commands": commands, "keepGoing": true}}});
-        let reply: Value = serde_json::from_str(&server.handle_line(&line.to_string()).unwrap()).unwrap();
-        (reply["result"]["isError"] == true, reply["result"]["content"][0]["text"].as_str().unwrap_or_default().to_string())
-    };
-    let (err, text) = call(&mut server, vec![json!({"command": "document.inspect"}); 1001]);
-    assert!(err && text.contains("at most 1000"), "{text}");
-    let (err, text) = call(&mut server, vec![json!({"command": "slide.inspect"}); 100]);
-    assert!(err && text.contains("64 MiB"), "{}", &text[text.len().saturating_sub(200)..]);
-    assert!(text.len() < (64 << 20) + (4 << 20));
-}
-
-#[test]
-fn the_batch_tool_charges_the_escaped_reply() {
-    // Quotes escape twice on the way out (`\\\"`), so 30 results of ~2 MiB serialize to ~120 MiB.
-    let mut server = Server::new(Box::new(Headless::with_document()));
-    let notes = json!({"jsonrpc": "2.0", "id": 1, "method": "tools/call", "params": {"name": "run_command", "arguments": {"command": "slide.notes", "params": {"text": "\"".repeat(1 << 20)}}}});
-    server.handle_line(&notes.to_string()).unwrap();
-    let commands = vec![json!({"command": "slide.inspect"}); 30];
-    let line = json!({"jsonrpc": "2.0", "id": 2, "method": "tools/call", "params": {"name": "batch", "arguments": {"commands": commands, "keepGoing": true}}});
-    let reply = server.handle_line(&line.to_string()).unwrap();
-    assert!(reply.len() < (64 << 20) + (4 << 20), "{} bytes", reply.len());
-    assert!(reply.contains("64 MiB"));
-}
-
-#[test]
-fn the_batch_tool_charges_failures_too() {
-    // Each unknown command's error echoes its ~1 MiB id; with keepGoing none of them is a result.
-    let mut server = Server::new(Box::new(Headless::with_document()));
-    let id = "x".repeat(1 << 20);
-    let commands = vec![json!({"command": id}); 70];
-    let line = json!({"jsonrpc": "2.0", "id": 2, "method": "tools/call", "params": {"name": "batch", "arguments": {"commands": commands, "keepGoing": true}}});
-    let reply = server.handle_line(&line.to_string()).unwrap();
-    assert!(reply.len() < (64 << 20) + (4 << 20), "{} bytes", reply.len());
-    assert!(reply.contains("64 MiB"));
-}

crates/mcp/tests/it/main.rs

@@ -1,5 +1,3 @@
 // Integration tests: unwrapping and panicking on failure is fine here, unlike in shipped code (AGENTS.md).
 #[allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
 mod automation_roots;
-#[allow(clippy::unwrap_used, clippy::expect_used, clippy::panic)]
-mod batch_budget;

crates/pptx/src/read.rs

@@ -100,8 +100,7 @@ impl Part {
     }
     /// Like [`Part::keep`], also declaring prefixes that only an enclosing wrapper binds.
     pub fn keep_inner(&self, el: &El) -> String {
-        let ns: Vec<(String, String)> = self.ns.iter().chain(&self.inner_ns).cloned().collect();
-        keep_xml(el, &ns)
+        keep_with(el, &[&self.ns, &self.inner_ns])
     }
 }
 
@@ -136,36 +135,69 @@ fn inner_decls(root: &El, ns: &[(String, String)]) -> Vec<(String, String)> {
     out
 }
 
-/// Serialize `el`, declaring every prefix it uses that the source root declared.
-pub fn keep_xml(el: &El, ns: &[(String, String)]) -> String {
-    let mut used = HashSet::new();
-    fn rec(e: &El, used: &mut HashSet<String>, depth: usize) {
+/// Longest namespace name copied onto a kept fragment or a resolved `mc` branch. Real ones are
+/// URIs of well under 100 bytes; a longer one is not copied, so one declaration cannot be
+/// multiplied by every fragment that uses its prefix.
+const MAX_NS_VALUE: usize = 2048;
+
+/// Prefixes `el` and its descendants use in element and attribute names, and those listed in
+/// `mc:Ignorable` or `mc:MustUnderstand` (which must stay declared for the markup-compatibility rules to hold),
+/// and those starting a QName-like attribute value.
+fn used_prefixes(el: &El) -> HashSet<&str> {
+    fn rec<'a>(e: &'a El, used: &mut HashSet<&'a str>, depth: usize) {
         if depth > crate::xml::MAX_DEPTH {
             return;
         }
         if let Some((p, _)) = e.name.split_once(':') {
-            used.insert(p.to_string());
+            used.insert(p);
         }
-        for (k, _) in &e.attrs {
-            if let Some((p, _)) = k.split_once(':')
+        for (k, v) in &e.attrs {
+            if let Some((p, local)) = k.split_once(':')
                 && p != "xmlns"
                 && p != "xml"
             {
-                used.insert(p.to_string());
+                used.insert(p);
+                if local == "Ignorable" || local == "MustUnderstand" {
+                    used.extend(v.split_whitespace());
+                }
+            }
+            // QName-valued attributes (`xsi:type="q:Item"`) need their prefix declared too.
+            if k != "xmlns" && !k.starts_with("xmlns:") {
+                for t in v.split_whitespace() {
+                    if let Some((p, _)) = t.split_once(':')
+                        && !p.is_empty()
+                    {
+                        used.insert(p);
+                    }
+                }
             }
         }
         for c in e.elements() {
             rec(c, used, depth + 1);
         }
     }
+    let mut used = HashSet::new();
     rec(el, &mut used, 0);
-    let mut el = el.clone();
-    for (k, v) in ns {
+    used
+}
+
+/// Serialize `el`, declaring every prefix it uses that the source root declared.
+pub fn keep_xml(el: &El, ns: &[(String, String)]) -> String {
+    keep_with(el, &[ns])
+}
+
+fn keep_with(el: &El, decls: &[&[(String, String)]]) -> String {
+    let used = used_prefixes(el);
+    let mut have: HashSet<&str> = el.attrs.iter().map(|(a, _)| a.as_str()).collect();
+    let mut extra = Vec::new();
+    for (k, v) in decls.iter().flat_map(|d| d.iter()) {
         let p = k.strip_prefix("xmlns:").unwrap_or("");
-        if !p.is_empty() && used.contains(p) && !el.attrs.iter().any(|(a, _)| a == k) {
-            el.attrs.push((k.clone(), v.clone()));
+        if !p.is_empty() && used.contains(p) && v.len() <= MAX_NS_VALUE && have.insert(k) {
+            extra.push((k.clone(), v.clone()));
         }
     }
+    let mut el = el.clone();
+    el.attrs.extend(extra);
     el.to_xml()
 }
 
@@ -314,10 +346,18 @@ fn kept_parts(pkg: &Package, pres_path: &str) -> Vec<deckcraft_model::KeptPart>
             // The part's own relationships (customXml item → itemProps) and their targets.
             let rp = opc::rels_path(&name);
             if let Some(rels_xml) = text(&rp) {
+                let mut capped = false;
                 for sub in pkg.rels(&name).items.iter().filter(|s| !s.external && keepable(&s.target)) {
-                    if let Some(x) = text(&sub.target)
-                        && !out.iter().any(|k| k.name.eq_ignore_ascii_case(&sub.target))
-                    {
+                    // Checked before the part is read: repeated relationships to one large part
+                    // must not copy it again each time.
+                    if out.iter().any(|k| k.name.eq_ignore_ascii_case(&sub.target)) {
+                        continue;
+                    }
+                    if out.len() >= 400 {
+                        capped = true;
+                        continue;
+                    }
+                    if let Some(x) = text(&sub.target) {
                         out.push(KeptPart {
                             content_type: pkg.content_type(&sub.target).unwrap_or_default(),
                             name: sub.target.clone(),
@@ -326,7 +366,9 @@ fn kept_parts(pkg: &Package, pres_path: &str) -> Vec<deckcraft_model::KeptPart>
                         });
                     }
                 }
-                out.push(KeptPart { name: rp, xml: rels_xml, ..Default::default() });
+                if !capped {
+                    out.push(KeptPart { name: rp, xml: rels_xml, ..Default::default() });
+                }
             }
         }
     }
@@ -363,15 +405,39 @@ fn resolve_mc_rels(el: &mut El, depth: usize, rels: Option<&Rels>) {
                     });
                     let branch = choice.or_else(|| ac.child("Fallback"));
                     if let Some(b) = branch {
+                        // The branch's namespace declarations (the first usable one of each name,
+                        // in order), indexed by prefix once: each child then looks up the prefixes
+                        // it uses instead of scanning every declaration.
+                        let mut decls: Vec<(&String, &String)> = Vec::new();
+                        // An oversized declaration still shadows an outer one of its name: its
+                        // content is left unbound rather than bound to the outer namespace.
+                        let mut by_prefix: HashMap<&str, Option<usize>> = HashMap::new();
+                        for (k, v) in b.attrs.iter().chain(ac.attrs.iter()) {
+                            let prefix = if k == "xmlns" { Some("") } else { k.strip_prefix("xmlns:") };
+                            if let Some(p) = prefix
+                                && !by_prefix.contains_key(p)
+                            {
+                                by_prefix.insert(p, (v.len() <= MAX_NS_VALUE).then_some(decls.len()));
+                                if v.len() <= MAX_NS_VALUE {
+                                    decls.push((k, v));
+                                }
+                            }
+                        }
                         for n in &b.children {
                             if let Node::El(e) = n {
+                                // Carry the namespace declarations from the branch that its content
+                                // uses (all of them per child made children x declarations copies).
+                                let used = used_prefixes(e);
+                                let have: HashSet<&str> = e.attrs.iter().map(|(a, _)| a.as_str()).collect();
+                                let mut wanted: Vec<usize> = std::iter::once("")
+                                    .chain(used.iter().copied())
+                                    .filter_map(|p| by_prefix.get(p).copied().flatten())
+                                    .filter(|&i| !have.contains(decls[i].0.as_str()))
+                                    .collect();
+                                wanted.sort_unstable();
+                                wanted.dedup();
                                 let mut e = e.clone();
-                                // Carry namespace declarations from the branch onto its content.
-                                for (k, v) in b.attrs.iter().chain(ac.attrs.iter()) {
-                                    if k.starts_with("xmlns") && !e.attrs.iter().any(|(a, _)| a == k) {
-                                        e.attrs.push((k.clone(), v.clone()));
-                                    }
-                                }
+                                e.attrs.extend(wanted.into_iter().map(|i| (decls[i].0.clone(), decls[i].1.clone())));
                                 out.push(Node::El(e));
                             }
                         }
@@ -1414,3 +1480,72 @@ fn header_footer(p: &Presentation) -> HeaderFooter {
     }
     hf
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn mc_branches_carry_only_the_declarations_their_content_uses() {
+        // 2,000 declarations on the branch and 2,000 children: each child gets the one it uses
+        // (and the default), not all 2,000.
+        let decls: String = (0..2000).map(|i| format!(r#" xmlns:n{i}="urn:n{i}""#)).collect();
+        let kids: String = (0..2000).map(|i| format!("<n{i}:x/>")).collect();
+        let x = format!(
+            r#"<r xmlns:mc="http://schemas.openxmlformats.org/markup-compatibility/2006"><mc:AlternateContent><mc:Fallback xmlns="urn:d"{decls}>{kids}</mc:Fallback></mc:AlternateContent></r>"#
+        );
+        let mut d = crate::xml::parse(x.as_bytes()).unwrap();
+        resolve_mc(&mut d.root, 0);
+        let kids: Vec<&El> = d.root.elements().collect();
+        assert_eq!(kids.len(), 2000);
+        assert_eq!(kids[7].attrs, vec![("xmlns".to_string(), "urn:d".to_string()), ("xmlns:n7".to_string(), "urn:n7".to_string())]);
+    }
+
+    /// The branch's own declaration of a prefix wins over the `AlternateContent`'s; an oversized
+    /// one still shadows it (its content stays unbound); a child's own declaration is left alone.
+    #[test]
+    fn mc_branch_declarations_keep_their_precedence() {
+        let long = "u".repeat(MAX_NS_VALUE + 1);
+        let x = format!(
+            r#"<r xmlns:mc="http://schemas.openxmlformats.org/markup-compatibility/2006"><mc:AlternateContent xmlns:p="urn:ac" xmlns:q="urn:q-short"><mc:Fallback xmlns:p="urn:b" xmlns:q="{long}"><p:x/><q:y/><p:z xmlns:p="urn:own"/></mc:Fallback></mc:AlternateContent></r>"#
+        );
+        let mut d = crate::xml::parse(x.as_bytes()).unwrap();
+        resolve_mc(&mut d.root, 0);
+        let kids: Vec<&El> = d.root.elements().collect();
+        let attrs = |i: usize| kids[i].attrs.iter().map(|(k, v)| format!("{k}={v}")).collect::<Vec<_>>();
+        assert_eq!(attrs(0), ["xmlns:p=urn:b"]);
+        assert!(attrs(1).is_empty(), "{:?}", attrs(1));
+        assert_eq!(attrs(2), ["xmlns:p=urn:own"]);
+    }
+
+    #[test]
+    fn mc_branches_keep_prefixes_named_by_must_understand() {
+        let x = r#"<r xmlns:mc="http://schemas.openxmlformats.org/markup-compatibility/2006"><mc:AlternateContent><mc:Fallback xmlns:q="urn:q"><a mc:MustUnderstand="q"/></mc:Fallback></mc:AlternateContent></r>"#;
+        let mut d = crate::xml::parse(x.as_bytes()).unwrap();
+        resolve_mc(&mut d.root, 0);
+        let kid = d.root.elements().next().unwrap();
+        assert!(kid.attrs.iter().any(|(k, v)| k == "xmlns:q" && v == "urn:q"), "{:?}", kid.attrs);
+    }
+
+    #[test]
+    fn mc_branches_keep_prefixes_used_in_qname_values() {
+        let x = r#"<r xmlns:mc="http://schemas.openxmlformats.org/markup-compatibility/2006"><mc:AlternateContent><mc:Fallback xmlns:q="urn:q" xmlns:z="urn:z"><n xmlns:xsi="urn:xsi" xsi:type="q:Item"/></mc:Fallback></mc:AlternateContent></r>"#;
+        let mut d = crate::xml::parse(x.as_bytes()).unwrap();
+        resolve_mc(&mut d.root, 0);
+        let kid = d.root.elements().next().unwrap();
+        assert!(kid.attrs.iter().any(|(k, v)| k == "xmlns:q" && v == "urn:q"), "{:?}", kid.attrs);
+        assert!(!kid.attrs.iter().any(|(k, _)| k == "xmlns:z"), "unused declarations stay out: {:?}", kid.attrs);
+    }
+
+    #[test]
+    fn kept_fragments_skip_oversized_namespace_names() {
+        let big = "u".repeat(MAX_NS_VALUE + 1);
+        let ns = vec![("xmlns:a".to_string(), "urn:a".to_string()), ("xmlns:b".to_string(), big)];
+        let d = crate::xml::parse(br#"<a:x xmlns:a="urn:a" xmlns:b="urn:b"><b:y/></a:x>"#).unwrap();
+        let mut el = d.root.clone();
+        el.attrs.clear();
+        let out = keep_xml(&el, &ns);
+        assert!(out.contains(r#"xmlns:a="urn:a""#), "{out}");
+        assert!(!out.contains("uuuu"), "the oversized declaration is not copied");
+    }
+}