| 1 | //! Byte-level canonicalization of JSON Schema for prefix-cache stability. |
| 2 | //! |
| 3 | //! When MCP servers return tool schemas, the field order within each schema |
| 4 | //! object and the order of entries in `required` / `dependentRequired` arrays |
| 5 | //! can vary across reconnections. This module normalizes those orderings so |
| 6 | //! that two logically equivalent schemas always produce identical bytes after |
| 7 | //! serialization. |
| 8 | //! |
| 9 | //! The approach mirrors `reasonix/internal/provider/schema_canonicalize.go`: |
| 10 | //! |
| 11 | //! 1. Sort every `"required"` array alphabetically. |
| 12 | //! 2. Sort every `"dependentRequired"` sub-array alphabetically. |
| 13 | //! 3. Recurse into all nested objects and arrays. |
| 14 | //! |
| 15 | //! `serde_json::Value::Object` uses `IndexMap` when `preserve_order` is |
| 16 | //! enabled (which this crate does). We therefore rebuild the map with sorted |
| 17 | //! keys to guarantee deterministic key ordering. |
| 18 | |
| 19 | use serde_json::Value; |
| 20 | |
| 21 | /// Recursively canonicalize a JSON Schema value in-place. |
| 22 | /// |
| 23 | /// After canonicalization, two schemas that are semantically equivalent |
| 24 | /// (same keys, same `required` set, same `dependentRequired` sets) will |
| 25 | /// serialize to byte-identical JSON regardless of the original field or |
| 26 | /// array order. |
| 27 | pub fn canonicalize_schema(value: &mut Value) { |
| 28 | match value { |
| 29 | Value::Object(map) => { |
| 30 | // Sort `required` arrays (they are sets per JSON Schema spec). |
| 31 | if let Some(Value::Array(req)) = map.get_mut("required") { |
| 32 | sort_string_array(req); |
| 33 | } |
| 34 | // Sort `dependentRequired` sub-arrays. |
| 35 | if let Some(Value::Object(deps)) = map.get_mut("dependentRequired") { |
| 36 | for dep_value in deps.values_mut() { |
| 37 | if let Value::Array(arr) = dep_value { |
| 38 | sort_string_array(arr); |
| 39 | } |
| 40 | } |
| 41 | } |
| 42 | // Recurse into every child value. |
| 43 | for v in map.values_mut() { |
| 44 | canonicalize_schema(v); |
| 45 | } |
| 46 | // Rebuild the map with sorted keys so serialization is deterministic. |
| 47 | // serde_json::Map backed by IndexMap (preserve_order) doesn't have |
| 48 | // drain(), so we swap to a temporary and rebuild. |
| 49 | let old = std::mem::take(map); |
| 50 | let mut entries: Vec<(String, Value)> = old.into_iter().collect(); |
| 51 | entries.sort_by(|a, b| a.0.cmp(&b.0)); |
| 52 | for (k, v) in entries { |
| 53 | map.insert(k, v); |
| 54 | } |
| 55 | } |
| 56 | Value::Array(arr) => { |
| 57 | for v in arr.iter_mut() { |
| 58 | canonicalize_schema(v); |
| 59 | } |
| 60 | } |
| 61 | _ => {} |
| 62 | } |
| 63 | } |
| 64 | |
| 65 | /// Sort a JSON array of string values alphabetically in-place. |
| 66 | /// |
| 67 | /// Non-string entries are left at the end in their original relative order. |
| 68 | fn sort_string_array(arr: &mut [Value]) { |
| 69 | arr.sort_by(|a, b| match (a.as_str(), b.as_str()) { |
| 70 | (Some(x), Some(y)) => x.cmp(y), |
| 71 | (Some(_), None) => std::cmp::Ordering::Less, |
| 72 | (None, Some(_)) => std::cmp::Ordering::Greater, |
| 73 | (None, None) => std::cmp::Ordering::Equal, |
| 74 | }); |
| 75 | } |
| 76 | |
| 77 | #[cfg(test)] |
| 78 | mod tests { |
| 79 | use super::*; |
| 80 | use serde_json::json; |
| 81 | |
| 82 | #[test] |
| 83 | fn sorts_required_array() { |
| 84 | let mut schema = json!({ |
| 85 | "type": "object", |
| 86 | "required": ["z", "a", "m"], |
| 87 | "properties": {} |
| 88 | }); |
| 89 | canonicalize_schema(&mut schema); |
| 90 | assert_eq!(schema["required"], json!(["a", "m", "z"])); |
| 91 | } |
| 92 | |
| 93 | #[test] |
| 94 | fn equivalent_ordering_matches() { |
| 95 | // Two schemas that differ only in field order and required order |
| 96 | // must serialize to identical bytes. |
| 97 | let mut a = json!({ |
| 98 | "required": ["b", "a"], |
| 99 | "properties": {"x": {}, "y": {}}, |
| 100 | "type": "object" |
| 101 | }); |
| 102 | let mut b = json!({ |
| 103 | "type": "object", |
| 104 | "properties": {"y": {}, "x": {}}, |
| 105 | "required": ["a", "b"] |
| 106 | }); |
| 107 | canonicalize_schema(&mut a); |
| 108 | canonicalize_schema(&mut b); |
| 109 | assert_eq!( |
| 110 | serde_json::to_string(&a).unwrap(), |
| 111 | serde_json::to_string(&b).unwrap(), |
| 112 | "logically equivalent schemas must produce identical bytes" |
| 113 | ); |
| 114 | } |
| 115 | |
| 116 | #[test] |
| 117 | fn sorts_dependent_required() { |
| 118 | let mut schema = json!({ |
| 119 | "type": "object", |
| 120 | "dependentRequired": { |
| 121 | "x": ["z", "a"], |
| 122 | "y": ["m", "b"] |
| 123 | } |
| 124 | }); |
| 125 | canonicalize_schema(&mut schema); |
| 126 | assert_eq!(schema["dependentRequired"]["x"], json!(["a", "z"])); |
| 127 | assert_eq!(schema["dependentRequired"]["y"], json!(["b", "m"])); |
| 128 | } |
| 129 | |
| 130 | #[test] |
| 131 | fn recursive_into_properties() { |
| 132 | let mut schema = json!({ |
| 133 | "type": "object", |
| 134 | "properties": { |
| 135 | "nested": { |
| 136 | "type": "object", |
| 137 | "required": ["z", "a"], |
| 138 | "properties": {} |
| 139 | } |
| 140 | } |
| 141 | }); |
| 142 | canonicalize_schema(&mut schema); |
| 143 | assert_eq!( |
| 144 | schema["properties"]["nested"]["required"], |
| 145 | json!(["a", "z"]) |
| 146 | ); |
| 147 | } |
| 148 | |
| 149 | #[test] |
| 150 | fn preserves_non_required_array_order() { |
| 151 | // Arrays that are not `required` or `dependentRequired` should |
| 152 | // keep their semantic order (e.g. enum values, oneOf items). |
| 153 | let mut schema = json!({ |
| 154 | "type": "string", |
| 155 | "enum": ["z", "a", "m"] |
| 156 | }); |
| 157 | canonicalize_schema(&mut schema); |
| 158 | assert_eq!(schema["enum"], json!(["z", "a", "m"])); |
| 159 | } |
| 160 | |
| 161 | #[test] |
| 162 | fn handles_empty_schema() { |
| 163 | let mut schema = json!({}); |
| 164 | canonicalize_schema(&mut schema); |
| 165 | assert_eq!(schema, json!({})); |
| 166 | } |
| 167 | |
| 168 | #[test] |
| 169 | fn handles_deeply_nested() { |
| 170 | let mut schema = json!({ |
| 171 | "type": "object", |
| 172 | "properties": { |
| 173 | "level1": { |
| 174 | "type": "object", |
| 175 | "properties": { |
| 176 | "level2": { |
| 177 | "type": "object", |
| 178 | "required": ["z", "a"] |
| 179 | } |
| 180 | } |
| 181 | } |
| 182 | } |
| 183 | }); |
| 184 | canonicalize_schema(&mut schema); |
| 185 | assert_eq!( |
| 186 | schema["properties"]["level1"]["properties"]["level2"]["required"], |
| 187 | json!(["a", "z"]) |
| 188 | ); |
| 189 | } |
| 190 | |
| 191 | #[test] |
| 192 | fn key_order_is_alphabetical_after_canonicalize() { |
| 193 | let mut schema = json!({ |
| 194 | "z_field": 1, |
| 195 | "a_field": 2, |
| 196 | "m_field": 3 |
| 197 | }); |
| 198 | canonicalize_schema(&mut schema); |
| 199 | let keys: Vec<&str> = schema |
| 200 | .as_object() |
| 201 | .unwrap() |
| 202 | .keys() |
| 203 | .map(|s| s.as_str()) |
| 204 | .collect(); |
| 205 | assert_eq!(keys, vec!["a_field", "m_field", "z_field"]); |
| 206 | } |
| 207 | } |
| 208 |