Add multi-version protocol types: two wire-surface fact blocks

One superset type set in mcp.types plus two declarative wire-fact
blocks interpreted by a single engine at the wire boundary: an empty
block serving every protocol version up to and including 2025-11-25,
and a strict block for 2026-07-28 carrying the required-field
injections (resultType, the ttlMs/cacheScope don't-cache defaults, the
reserved protocolVersion _meta entry) and the refusals for constructs
that revision removed.

Serialization is additive-only: nothing is ever removed from a model's
dump at any version. The empty block makes emission for 2025-11-25 and
earlier the plain dump by construction; the schemas through 2025-11-25
evolve strictly additively, and a structural coverage test proves the
newest one covers them all.

Author: maxisbey

docs/migration.md

@@ -1168,9 +1168,9 @@ In practice, replace direct `ServerSession` use with `Server.run(read_stream, wr
 
 ### Experimental Tasks support removed
 
-Tasks (SEP-1686) have been removed from the MCP specification and are no longer part of this SDK. The `mcp.client.experimental`, `mcp.server.experimental`, `mcp.shared.experimental`, and `mcp.server.lowlevel.experimental` modules have been removed, along with all `Task*` types, the `tasks` capability fields, `Tool.execution`, and the `experimental` properties on `ClientSession`, `ServerSession`, `Server`, and `ServerRequestContext`.
+The experimental task execution machinery (SEP-1686) is no longer part of this SDK. The `mcp.client.experimental`, `mcp.server.experimental`, `mcp.shared.experimental`, and `mcp.server.lowlevel.experimental` modules have been removed, along with the `experimental` properties on `ClientSession`, `ServerSession`, `Server`, and `ServerRequestContext`. Task execution is expected to return as a separate MCP extension in a future release.
 
-Tasks are expected to return as a separate MCP extension in a future release.
+The 2025-11-25 protocol's task **types** remain available from `mcp.types` (`Task`, `TaskMetadata`, `TaskStatus`, the `tasks/*` request and result types, the `tasks` capability fields, and `Tool.execution`), so task payloads from 2025-11-25 peers can be constructed and parsed. Compared with the v1 experimental module these are plain protocol types: attributes are snake_case (`task_id`, `created_at` — wire names are unchanged), timestamps are plain `str` fields, extra fields are ignored rather than preserved, and the `tasks/*` methods are not members of the request/notification unions — serving them requires registering a handler for the method explicitly.
 
 ## Deprecations
 
@@ -1238,6 +1238,12 @@ app = server.streamable_http_app(
 
 The lowlevel `Server` also now exposes a `session_manager` property to access the `StreamableHTTPSessionManager` after calling `streamable_http_app()`.
 
+### Newer protocol fields are modeled and round-trip
+
+Fields introduced by the 2026-07-28 protocol revision are now typed on the existing payload models: `result_type` on results, `ttl_ms`/`cache_scope` on cacheable list/read results, `input_responses`/`request_state` on interactive request params, and `extensions` on the capability objects. Inbound payloads carrying these fields now retain them, and user-level `model_dump()` output includes them, where they were previously dropped as unknown fields. This is a strictly-more-lenient change: nothing that parsed before fails now.
+
+Wire output to peers on protocol versions that predate a field is unaffected — the version-aware boundary (`mcp.types.wire.serialize_for`) drops these fields when emitting for 2025-11-25 and earlier, so existing peers see byte-identical frames.
+
 ## Need Help?
 
 If you encounter issues during migration:

pyproject.toml

@@ -143,6 +143,12 @@ venv = ".venv"
 # those private functions instead of testing the private functions directly. It makes it easier to maintain the code source
 # and refactor code that is not public.
 executionEnvironments = [
+    # Generated spec-oracle modules are committed verbatim from the generator;
+    # a handful of generated patterns (recursive JSONValue alias, error `data`
+    # overrides without defaults) are not strict-clean and must not be hand-edited.
+    { root = "tests/spec_oracles", extraPaths = [
+        ".",
+    ], reportUnusedFunction = false, reportPrivateUsage = false, reportGeneralTypeIssues = false, reportUnknownVariableType = false, reportInvalidTypeForm = false },
     { root = "tests", extraPaths = [
         ".",
     ], reportUnusedFunction = false, reportPrivateUsage = false },
@@ -177,6 +183,10 @@ max-complexity = 24 # Default is 10
 
 [tool.ruff.lint.per-file-ignores]
 "__init__.py" = ["F401"]
+# Generated spec-oracle modules: docstrings come verbatim from the spec schema
+# and ruff format does not wrap long docstring lines.
+"tests/spec_oracles/v*.py" = ["E501"]
+"tests/spec_oracles/ext_tasks.py" = ["E501"]
 "tests/server/mcpserver/test_func_metadata.py" = ["E501"]
 "tests/shared/test_progress_notifications.py" = ["PLW0603"]
 
@@ -234,6 +244,11 @@ omit = [
     "src/mcp/server/__main__.py",
     "src/mcp/os/posix/utilities.py",
     "src/mcp/os/win32/utilities.py",
+    # Generated spec-oracle modules and their comparison harness, vendored
+    # unchanged from the oracle generator's output. The harness carries
+    # defensive comparison branches and assertion-failure formatters that a
+    # green suite never executes, and the files are not hand-edited here.
+    "tests/spec_oracles/*",
 ]
 
 # https://coverage.readthedocs.io/en/latest/config.html#report