feat(core,server,client): SEP-2106 JSON Schema 2020-12 — $ref guard, per-dialect validation, non-object structuredContent

.changeset/client-response-cache-substrate.md

@@ -4,4 +4,4 @@
 
 `Client.listTools()` / `listPrompts()` / `listResources()` / `listResourceTemplates()` now **auto-aggregate every page** when called without a `cursor` and return the complete result with `nextCursor: undefined` (matching the C#, Java, and mcp.d SDKs). Pass an explicit `{ cursor }` string to fetch a single page; the per-page path is unchanged. Existing manual pagination loops keep working — the first iteration returns everything and the loop exits — but can be deleted. The aggregated result is written to the new pluggable `ResponseCacheStore` (default: a fresh per-instance `InMemoryResponseCacheStore`); a `ClientResponseCache` collaborator owns the eviction-generation guard and the derived `tools/list` index that `callTool`'s output validation and SEP-2243 `Mcp-Param-*` mirroring read. New exports: `ResponseCacheStore`, `CacheKey`, `CacheEntry`, `CacheScope`, `MaybePromise`, `InMemoryResponseCacheStore`; new `ClientOptions.responseCacheStore` / `ClientOptions.listMaxPages` (caps the auto-aggregate walk at 64 pages by default; throws `SdkError` with `SdkErrorCode.ListPaginationExceeded` on overrun so a partial aggregate is never cached). The store interface is async-ready (`MaybePromise<…>`); the in-memory default stays synchronous. Entries are automatically scoped by the connected server's identity and (when set) the consumer-supplied `cachePartition`, so a shared store does not collide across servers or principals; evictions are likewise scoped to the connected server's partitions.
 
-**Behavior change (every era):** output-schema validator compilation is now lazy — validators are compiled on the first `callTool()` against the cached `tools/list` entry, not eagerly inside `listTools()` — and non-throwing: an uncompilable `outputSchema` is `console.warn`-ed and validation is skipped for that tool only (previously `listTools()` threw). A pluggable `jsonSchemaValidator` provider therefore observes compilation at `callTool` time, not `listTools` time. The legacy-era `listTools()` path is unchanged at the wire level but is observably different at the validator-lifecycle level.
+**Behavior change (every era):** output-schema validator compilation is now lazy — validators are compiled on the first `callTool()` against the cached `tools/list` entry, not eagerly inside `listTools()`. `listTools()` no longer throws on an uncompilable `outputSchema` (every tool stays listed; the compile failure is captured per-tool); calling `callTool()` on the affected tool throws `ProtocolError(InvalidParams, "Tool 'X' has an invalid outputSchema: …")` before the request is sent — output-schema validation is never silently skipped. A pluggable `jsonSchemaValidator` provider therefore observes compilation at `callTool` time, not `listTools` time. The legacy-era `listTools()` path is unchanged at the wire level but is observably different at the validator-lifecycle level.

.changeset/sep-2106-dialect-posture.md

@@ -0,0 +1,7 @@
+---
+'@modelcontextprotocol/core': minor
+'@modelcontextprotocol/client': major
+'@modelcontextprotocol/server': major
+---
+
+SEP-1613 / SEP-2106 (JSON Schema 2020-12 posture): the Node default JSON Schema validator is now `Ajv2020` (true draft 2020-12) instead of the draft-07 `Ajv` class — `$defs`/`prefixItems`/`unevaluatedProperties`/`dependentRequired` are now enforced where they were previously silently ignored; to opt back, construct the draft-07 instance with the v1 defaults — `const ajv = new Ajv({ strict: false, validateFormats: true, validateSchema: false, allErrors: true }); addFormats(ajv);` — and pass `new AjvJsonSchemaValidator(ajv)`. Schemas declaring a `$schema` other than 2020-12 are rejected with a clear error rather than mis-validating. `outputSchema` may now have a non-object root and `CallToolResult.structuredContent` is widened to `unknown` (a deliberate source-level break for typed consumers — see the migration guide for the narrowing pattern). Toward 2025-era clients McpServer wraps a non-object `outputSchema` (and the matching `structuredContent`) in a `{result: …}` envelope so the tool stays callable, with same-document `$ref`/`$dynamicRef` pointers rewritten to keep resolving — low-level `Server` users (those bypassing `McpServer` and registering a `tools/call` handler directly) get the same wrap by routing the result through the new `Server.projectCallToolResult(result, advertisedOutputSchema)`. Independently, on every era (the SEP's MUST applies regardless of client version), McpServer auto-appends a `TextContent` JSON serialisation when a handler returns non-object `structuredContent` without its own text block. The `structuredContent` presence check is `!== undefined` (not falsy) on both sides. Thanks @mattzcarey (#2249).

docs/client.md

@@ -269,17 +269,21 @@ const result = await client.callTool({
 console.log(result.content);
 ```
 
-Tool results may include a `structuredContent` field — a machine-readable JSON object for programmatic use by the client application, complementing `content` which is for the LLM:
+Tool results may include a `structuredContent` field — a machine-readable JSON value (any JSON type per SEP-2106) for programmatic use by the client application, complementing `content` which is for the LLM:
 
 ```ts source="../examples/guides/clientGuide.examples.ts#callTool_structuredOutput"
 const result = await client.callTool({
     name: 'calculate-bmi',
     arguments: { weightKg: 70, heightM: 1.75 }
 });
 
-// Machine-readable output for the client application
-if (result.structuredContent) {
-    console.log(result.structuredContent); // e.g. { bmi: 22.86 }
+// Machine-readable output for the client application. SEP-2106: structuredContent is
+// `unknown` (any JSON value). Check for presence with `!== undefined` and narrow before use.
+if (result.structuredContent !== undefined) {
+    const sc: unknown = result.structuredContent; // e.g. { bmi: 22.86 }
+    if (typeof sc === 'object' && sc !== null && 'bmi' in sc) {
+        console.log(sc.bmi);
+    }
 }
 ```
 

docs/migration-SKILL.md

@@ -574,7 +574,7 @@ side: auto-fulfilment is on by default (`ClientOptions.inputRequired`, `maxRound
 
 `Client.listTools()` / `listPrompts()` / `listResources()` / `listResourceTemplates()` / `readResource()` now honour the server-stamped SEP-2549 `ttlMs`/`cacheScope`: a still-fresh cached entry is returned without a round trip. Opt-in by server hint — a server that sends `ttlMs: 0` (the SDK's default stamp) sees no behaviour change. Per-call override: pass `{ cacheMode: 'refresh' }` (always fetch and re-store) or `{ cacheMode: 'bypass' }` (fetch without touching the cache). Server `ttlMs` is clamped at 24 h (`MAX_CACHE_TTL_MS`). Entries are automatically scoped by connected-server identity; new `ClientOptions.cachePartition` (per-principal slot for `'private'`-scoped entries on a shared `responseCacheStore`; default `''`) and `ClientOptions.defaultCacheTtlMs` (TTL when the result lacks one, e.g. legacy-era responses; default `0`). `ResponseCacheStore` gained `delete(key)` (driven by `notifications/resources/updated`); `InMemoryResponseCacheStore` is now bounded (`{ maxEntries }`, default 512).
 
-Output-schema validator compilation is now lazy (first `callTool()` against the cached `tools/list` entry) and non-throwing (an uncompilable `outputSchema` is `console.warn`-ed and validation is skipped for that tool only); `listTools()` no longer throws on an uncompilable `outputSchema`. Applies on every era — the legacy-era `listTools()` path is unchanged at the wire level only.
+Output-schema validator compilation is now lazy (first `callTool()` against the cached `tools/list` entry); `listTools()` no longer throws on an uncompilable `outputSchema` — every tool stays listed and the compile failure is captured per-tool. Calling `callTool()` on the affected tool throws `ProtocolError(InvalidParams, "Tool 'X' has an invalid outputSchema: …")` before the request is sent (validation is never silently skipped). Applies on every era — the legacy-era `listTools()` path is unchanged at the wire level only.
 
 New (no v1 equivalent): `Client.connect(transport, { prior: DiscoverResult })` — zero-round-trip connect (2026-07-28+ only; throws `EraNegotiationFailed` otherwise). Probe once, persist `client.getDiscoverResult()` (`JSON.stringify`), feed to every worker. New exported type:
 `ConnectOptions` (extends `RequestOptions` with `prior?: DiscoverResult`).
@@ -664,6 +664,18 @@ Validator behavior:
   `@modelcontextprotocol/{client,server}/validators/cf-worker` for `CfWorkerJsonSchemaValidator`. Importing from a subpath means the corresponding peer dep must be in your `package.json`.
 - To replace validation entirely, pass `jsonSchemaValidator: myCustomValidator` with your own implementation of the `jsonSchemaValidator` interface.
 
+JSON Schema 2020-12 posture (SEP-1613 / SEP-2106): the default validator supports JSON Schema 2020-12 only (the spec's only MUST) — on Node it is now `Ajv2020` instead of draft-07 `Ajv`. Schemas declaring a different `$schema` are rejected with a clear `Error("…unsupported
+dialect…")`; to validate other dialects, pass a pre-configured Ajv instance: `new AjvJsonSchemaValidator(new Ajv({...}))`. `CallToolResult.structuredContent` is typed `unknown` (was `{ [k: string]: unknown }`). The presence check is `!== undefined`, not falsy. External `$ref`
+is not dereferenced (unchanged from v1; Ajv throws `MissingRefError` at compile, surfaced per-tool on `callTool`). Toward 2025-era clients a non-object `outputSchema`/`structuredContent` is wrapped in a `{result:…}` envelope.
+
+| v1 pattern                                                         | Mechanical fix                                                                                                                                                                                                 |
+| ------------------------------------------------------------------ | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| `result.structuredContent.<key>` / `result.structuredContent?.<k>` | narrow first: `const sc = result.structuredContent; if (typeof sc === 'object' && sc !== null && '<k>' in sc) { sc.<k> }`                                                                                      |
+| `if (!result.structuredContent)`                                   | `if (result.structuredContent === undefined)`                                                                                                                                                                  |
+| relying on default `Ajv` being draft-07                            | `const ajv = new Ajv({ strict: false, validateFormats: true, validateSchema: false, allErrors: true }); addFormats(ajv); new AjvJsonSchemaValidator(ajv)` (import `Ajv`, `addFormats` from `…/validators/ajv`) |
+| draft-07 idioms via `fromJsonSchema(schema)`                       | `fromJsonSchema(schema, new AjvJsonSchemaValidator(ajv))` — the `McpServer`/`Client` `jsonSchemaValidator` option does **not** reach `fromJsonSchema`-authored schemas                                         |
+| `outputSchema` or `inputSchema` with absolute-URI `$ref`           | inline under `$defs` and reference with `#/$defs/Name`                                                                                                                                                         |
+
 ## 15. Migration Steps (apply in this order)
 
 1. Update `package.json`: `npm uninstall @modelcontextprotocol/sdk`, install the appropriate v2 packages