address review: charge batch error entries against the byte budget

crates/mcp/src/server.rs

@@ -16,6 +16,12 @@ 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;
@@ -93,8 +99,28 @@ 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 replies: Vec<Value> = batch.into_iter().filter_map(|m| self.handle(m)).collect();
+                    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
+                            });
+                        }
+                    }
                     (!replies.is_empty()).then_some(Value::Array(replies))
                 }
             }

crates/mcp/src/tools.rs

@@ -357,18 +357,30 @@ 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![];
-            for c in a.get("commands").and_then(Value::as_array).cloned().unwrap_or_default() {
+            let mut used = 0usize;
+            for c in commands {
                 let id = c.get("command").and_then(Value::as_str).unwrap_or("");
-                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 };
-                        }
-                    }
+                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 };
                 }
             }
             ToolResult::json(&Value::Array(out))

crates/mcp/tests/it/batch_budget.rs

@@ -0,0 +1,70 @@
+//! 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

@@ -1,3 +1,5 @@
 // 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/media/src/audio.rs

@@ -81,9 +81,7 @@ pub fn decode(bytes: &Bytes) -> Result<Pcm> {
             Some(r) => r,
             None => contained(|| decode_symphonia(bytes, "mkv")),
         },
-        Container::Mp4 | Container::Unknown | Container::Mp3 if !crate::mp4guard::probe_fits(bytes) => {
-            Err(MediaError::Corrupt("the MP4's sample tables do not fit the file".into()))
-        }
+        Container::Mp4 if !crate::mp4guard::tables_fit(bytes) => Err(MediaError::Corrupt("the MP4's sample tables do not fit the file".into())),
         c => contained(|| decode_symphonia(bytes, c.name())),
     }
 }

crates/media/src/mp4guard.rs

@@ -12,29 +12,6 @@ pub(crate) fn tables_fit(b: &[u8]) -> bool {
     boxes_fit(b, b.len() as u64, 0)
 }
 
-/// Bytes symphonia's probe scans for a format marker before giving up.
-const PROBE_SEARCH_LIMIT: usize = 1024 * 1024;
-
-/// Whether symphonia can be given `b`, whatever the sniffer made of it: its probe takes the first
-/// `ftyp` within its search window as the start of an MP4 (the size field 4 bytes before it), so
-/// the table check must start where that parser will, not at offset 0.
-pub(crate) fn probe_fits(b: &[u8]) -> bool {
-    if crate::Container::sniff(b) == crate::Container::Mp4 && !tables_fit(b) {
-        return false;
-    }
-    // The probe reads leading ID3v2 tags as metadata and searches again after the last one.
-    let mut rest = b;
-    while rest.starts_with(b"ID3") {
-        let size = rest.get(6..10).map(|s| s.iter().fold(0usize, |a, &x| (a << 7) | (x & 0x7f) as usize)).unwrap_or(0);
-        rest = rest.get(10 + size..).unwrap_or(&[]);
-    }
-    let head = &rest[..rest.len().min(PROBE_SEARCH_LIMIT + 16)];
-    match head.windows(4).position(|w| w == b"ftyp") {
-        Some(p) => tables_fit(&rest[p.saturating_sub(4)..]),
-        None => true,
-    }
-}
-
 fn be32(b: &[u8], at: usize) -> Option<u64> {
     b.get(at..at.checked_add(4)?).map(|s| u32::from_be_bytes([s[0], s[1], s[2], s[3]]) as u64)
 }
@@ -97,29 +74,3 @@ fn table_fits(kind: &[u8], body: &[u8], file_len: u64) -> bool {
     };
     be32(body, 4).is_some_and(|count| count <= body.len().saturating_sub(8) as u64 / entry)
 }
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-
-    /// An MP4 whose `stsc` count asks for far more entries than the file holds.
-    fn mp4_with_huge_stsc() -> Vec<u8> {
-        let stsc = [&24u32.to_be_bytes()[..], b"stsc", &[0, 0, 0, 0], &u32::MAX.to_be_bytes(), &[0; 8]].concat();
-        let stbl = [&(stsc.len() as u32 + 8).to_be_bytes()[..], b"stbl", &stsc].concat();
-        let moov = [&(stbl.len() as u32 + 8).to_be_bytes()[..], b"moov", &stbl].concat();
-        [&16u32.to_be_bytes()[..], b"ftyp", b"isom", &[0; 4], &moov].concat()
-    }
-
-    #[test]
-    fn a_prefixed_mp4_is_checked_where_symphonia_starts_it() {
-        let mp4 = mp4_with_huge_stsc();
-        assert!(!tables_fit(&mp4));
-        assert!(!probe_fits(&mp4));
-        let prefixed = [&b"junk before the file "[..], &mp4].concat();
-        assert_eq!(crate::Container::sniff(&prefixed), crate::Container::Unknown);
-        assert!(tables_fit(&prefixed), "offset 0 sees no MP4 boxes");
-        assert!(!probe_fits(&prefixed));
-        let far = [&vec![0u8; PROBE_SEARCH_LIMIT + 16][..], &mp4].concat();
-        assert!(probe_fits(&far), "past the probe window symphonia never reads it as an MP4");
-    }
-}

crates/media/src/probe.rs

@@ -71,8 +71,7 @@ pub fn probe_headers(bytes: &Bytes) -> Result<MediaInfo> {
     };
     if info.audio.is_none() && info.video.is_none() || info.audio.as_ref().is_some_and(|a| a.sample_rate == 0) {
         // Plain audio files (and anything our demuxers didn't describe): symphonia's probe.
-        // Symphonia opens an MP4 wherever its probe finds an `ftyp` (after leading ID3v2 tags too); other known containers (a WAV's audio bytes can spell an `ftyp`) skip the check.
-        if matches!(c, Container::Mp4 | Container::Unknown | Container::Mp3) && !crate::mp4guard::probe_fits(bytes) {
+        if c == Container::Mp4 && !crate::mp4guard::tables_fit(bytes) {
             return Err(MediaError::Corrupt("the MP4's sample tables do not fit the file".into()));
         }
         let (codec, rate, channels, frames) = std::panic::catch_unwind(|| crate::audio::probe_symphonia(bytes, c.name()))