Replace vMCP Modern kill-switch with a capability gate

deploy/charts/operator-crds/files/crds/toolhive.stacklok.dev_virtualmcpservers.yaml

@@ -1819,6 +1819,10 @@ spec:
                       When enabled, vMCP exposes only find_tool and call_tool operations to clients
                       instead of all backend tools directly. This reduces token usage by allowing
                       LLMs to discover relevant tools on demand rather than receiving all tool definitions.
+                      Enabling the optimizer currently pins this instance to MCP 2025-11-25: the
+                      find_tool/call_tool meta-tools are session-scoped, so Modern-capable (2026-07-28)
+                      clients are negotiated down to the Legacy revision (see
+                      pkg/vmcp/server/modern_gate.go; Modern parity is tracked in stacklok/toolhive#6089).
                     properties:
                       embeddingHeaders:
                         additionalProperties:
@@ -5266,6 +5270,10 @@ spec:
                       When enabled, vMCP exposes only find_tool and call_tool operations to clients
                       instead of all backend tools directly. This reduces token usage by allowing
                       LLMs to discover relevant tools on demand rather than receiving all tool definitions.
+                      Enabling the optimizer currently pins this instance to MCP 2025-11-25: the
+                      find_tool/call_tool meta-tools are session-scoped, so Modern-capable (2026-07-28)
+                      clients are negotiated down to the Legacy revision (see
+                      pkg/vmcp/server/modern_gate.go; Modern parity is tracked in stacklok/toolhive#6089).
                     properties:
                       embeddingHeaders:
                         additionalProperties:

deploy/charts/operator-crds/templates/toolhive.stacklok.dev_virtualmcpservers.yaml

@@ -1822,6 +1822,10 @@ spec:
                       When enabled, vMCP exposes only find_tool and call_tool operations to clients
                       instead of all backend tools directly. This reduces token usage by allowing
                       LLMs to discover relevant tools on demand rather than receiving all tool definitions.
+                      Enabling the optimizer currently pins this instance to MCP 2025-11-25: the
+                      find_tool/call_tool meta-tools are session-scoped, so Modern-capable (2026-07-28)
+                      clients are negotiated down to the Legacy revision (see
+                      pkg/vmcp/server/modern_gate.go; Modern parity is tracked in stacklok/toolhive#6089).
                     properties:
                       embeddingHeaders:
                         additionalProperties:
@@ -5269,6 +5273,10 @@ spec:
                       When enabled, vMCP exposes only find_tool and call_tool operations to clients
                       instead of all backend tools directly. This reduces token usage by allowing
                       LLMs to discover relevant tools on demand rather than receiving all tool definitions.
+                      Enabling the optimizer currently pins this instance to MCP 2025-11-25: the
+                      find_tool/call_tool meta-tools are session-scoped, so Modern-capable (2026-07-28)
+                      clients are negotiated down to the Legacy revision (see
+                      pkg/vmcp/server/modern_gate.go; Modern parity is tracked in stacklok/toolhive#6089).
                     properties:
                       embeddingHeaders:
                         additionalProperties:

docs/arch/10-virtual-mcp-architecture.md

@@ -412,6 +412,85 @@ The four Modern list verbs (`tools/list`, `resources/list`, `resources/templates
 
 The completion handler is a single global handler installed via `WithCompletionHandler`, so it recovers the session from the SDK request context rather than a per-session closure. Setting it makes the shim auto-advertise the `completions` capability at initialize.
 
+### Served MCP revisions: the Modern capability gate
+
+vMCP serves two client-facing MCP revisions: **Legacy** (2025-11-25, the SDK
+session path) always, and **Modern** (2026-07-28, `classifyingHandler →
+dispatchModern`, stateless) **conditionally** — only when every enabled feature
+of the instance is servable by the stateless dispatch path. The condition is
+`modernDispatchBlockers` (`pkg/vmcp/server/modern_gate.go`), an explicit
+enumeration that replaced the temporary `TOOLHIVE_VMCP_MODERN_STATELESS`
+env-var kill-switch (#5959): instead of a global "don't serve Modern", the
+instance serves Modern exactly when it can serve it correctly.
+
+Features that currently gate Modern off, and why:
+
+| Feature | Why the stateless path cannot serve it |
+|---------|----------------------------------------|
+| Optimizer (`find_tool`/`call_tool`) | The meta-tools are Serve-layer and **session-scoped** (`serve_optimizer.go`): each session builds an FTS5 index over its advertised set and swaps the two meta-tools in. The index is deliberately not in the stateless core, and `dispatchModern` serves `tools/*` straight from `core.ListTools`/`core.CallTool` — a Modern client would silently receive the raw aggregated tool set and `tools/call find_tool` would fail. Parity needs an identity- or instance-scoped index |
+
+"Cannot serve" means a Modern client would silently get different behavior than
+the feature promises — not merely that the feature is session-flavored.
+Redis-backed session sharing, for example, does **not** gate Modern: Legacy
+clients keep their shared, reconstructible sessions while Modern clients are
+sessionless by design and store nothing, a coexistence asserted end-to-end by
+`test/e2e/thv-operator/virtualmcp/virtualmcp_dual_era_redis_test.go`. Rate
+limiting does not gate Modern either: the limiter is a core decorator
+(`pkg/vmcp/ratelimit`), so both eras meter the same `CallTool` seam, and the
+Modern dispatcher preserves the limiter's coded error on the wire — a real
+JSON-RPC error object, `-32029` with `data.retryAfterSeconds`
+(`writeModernCodedError`, at HTTP 200 because go-sdk rejects a non-200
+response before decoding the body, so on a 429 the error object — and its
+`retryAfterSeconds` — would be discarded unread) — where the Legacy SDK seam
+can only smuggle the same code and data into an `IsError` tool result's
+`structuredContent` (`conversion.ErrorToToolResult`). Note that `-32029` sits
+inside the draft spec's reserved band (`-32020`..`-32099`, reserved for
+spec-defined codes); reallocating it is tracked in #6101.
+
+"Does not gate Modern" is not "costs the same", though. The limiter wraps only
+the `CallTool` seam, so the list verbs and `server/discover` are unmetered on
+both eras — but Legacy aggregates once per session registration, while Modern
+re-runs the full backend fan-out on every request with no cache, by design
+(`core_vmcp.go`'s `aggregatedView`). A Modern client can therefore loop
+unmetered, uncached fan-outs — reachable unauthenticated when incoming auth is
+anonymous. Rate-limiting the list/discover verbs, or a short-TTL per-identity
+capability cache, is deferred until profiling shows the per-request fan-out
+cost matters (#5761 — the same deferral recorded in `dispatchModernDiscover`'s
+doc comment).
+
+Wire behavior when the gate is closed:
+
+- **`server/discover` falls through to the SDK** instead of dispatching. go-sdk's
+  `filterSupportedVersions` keeps every version the transport's
+  `SupportsProtocolVersion` accepts, and the stateful transport excludes only
+  >= 2026-07-28 (those require `Stateless`), so the probe answer is the
+  transport-filtered list — everything except Modern: `[2025-11-25,
+  2025-06-18, 2025-03-26, 2024-11-05]`. This matters because go-sdk v1.7+
+  clients are Modern-first: `Connect` probes `server/discover` **before**
+  `initialize` and upgrades to whatever the server advertises. The
+  fall-through answer is what lands them on the Legacy handshake — where
+  sessions, and every gated feature, work. That answer over-advertises,
+  though: the `-32022` refusal below lists only 2025-11-25, and the refusal is
+  the accurate one — `mcpcompat`'s `handleInitialize` always responds with
+  `LATEST_PROTOCOL_VERSION` regardless of what the client requests, so the
+  three older revisions in the discover answer are not actually servable. The
+  mismatch is the SDK's discover answer to narrow, not the `-32022` list to
+  widen.
+- **Every other well-formed Modern request** is refused with a conformant
+  400 + `-32022 UnsupportedProtocolVersionError` whose data lists the Legacy
+  version, the shape a client negotiates down from. It is answered in
+  `classifyingHandler` rather than falling through, because go-sdk's stateful
+  rejection for Modern traffic is a plain-text 400 carrying no version list.
+- **Legacy traffic is untouched** either way; the gate only ever affects
+  requests that classified Modern.
+
+The gate is derived from construction-time configuration, logged once at
+startup ("MCP 2026-07-28 (Modern) dispatch disabled…"), and pinned by
+`TestModernDispatchBlockers`, `TestClassifyingHandler_ModernCapabilityGate`,
+and the full-handler pair in `modern_gate_integration_test.go`. Achieving
+Modern parity for a feature means deleting its entry and updating those tests —
+nothing else needs to change.
+
 ### Subscription limitation (ack-level)
 
 vMCP advertises `resources.subscribe: true` and answers `resources/subscribe` / `resources/unsubscribe` at **ack level**: the request is accepted (enforcing session binding and validating the URI is an advertised, admitted resource), and go-sdk records the subscription. vMCP does **not** currently propagate backend `notifications/resources/updated` to the subscribed client — doing so requires persistent per-session backend connections, which is out of scope. Clients that subscribe will receive a success ack but no update stream yet.

docs/arch/vmcp-local.md

@@ -103,6 +103,8 @@ thv vmcp serve --config vmcp.yaml
 | 2 | `--optimizer-embedding` | FTS5 + TEI semantic | Managed TEI container | `find_tool`, `call_tool` only |
 | 3 | `optimizer.embeddingService` in config YAML | FTS5 + external embedding service | User-managed | `find_tool`, `call_tool` only |
 
+Any tier >= 1 keeps clients on MCP 2025-11-25 (Legacy): the `find_tool`/`call_tool` meta-tools are session-scoped, so Modern-capable (2026-07-28) clients are negotiated down to Legacy by the capability gate (`pkg/vmcp/server/modern_gate.go`; Modern parity is tracked in #6089).
+
 Tier 2 (`--optimizer-embedding`) implies `--optimizer`. The TEI container is started automatically and stopped on server shutdown.
 
 **Implementation**: `pkg/vmcp/optimizer/optimizer.go`, `pkg/vmcp/cli/embedding_manager.go`

pkg/ratelimit/errors.go

@@ -12,7 +12,11 @@ import (
 
 const (
 	// CodeRateLimited is the JSON-RPC error code for rate-limited requests.
-	// Per RFC THV-0057: implementation-defined code in the -32000 to -32099 range.
+	// It was chosen (RFC THV-0057) when -32000..-32099 was uniformly
+	// implementation-defined, but the draft MCP spec now reserves
+	// -32020..-32099 exclusively for spec-defined codes, which this is not.
+	// The value is retained for wire compatibility; reallocation outside
+	// -32768..-32000 is tracked in #6101.
 	CodeRateLimited int64 = -32029
 
 	// MessageRateLimited is the error message for rate-limited requests.

pkg/vmcp/cli/serve.go

@@ -15,7 +15,6 @@ import (
 	"net"
 	"os"
 	"path/filepath"
-	"strconv"
 	"time"
 
 	"go.opentelemetry.io/otel/trace"
@@ -52,12 +51,6 @@ import (
 	vmcpstatus "github.com/stacklok/toolhive/pkg/vmcp/status"
 )
 
-// modernDispatchEnvVar is the kill-switch env var for direct-to-core dispatch
-// of well-formed MCP 2026-07-28 ("Modern") stateless requests (default off;
-// see server.Config.ModernDispatchEnabled). Env-only ahead of any CLI-flag or
-// CRD wiring.
-const modernDispatchEnvVar = "TOOLHIVE_VMCP_MODERN_STATELESS"
-
 // ServeConfig holds all parameters needed to start the vMCP server.
 // Populated by the caller from Cobra flag values or equivalent.
 // At least one of ConfigPath or GroupRef must be non-empty; ConfigPath takes
@@ -419,21 +412,6 @@ func Serve(ctx context.Context, cfg ServeConfig) error {
 		}()
 	}
 
-	// Read the Modern-stateless-dispatch kill-switch once, here at the
-	// composition root. Unset is the deliberate "off" default. A non-empty
-	// value that fails to parse as a bool is an operator typo (a misspelled
-	// "true"), not a request to enable the feature — warn and stay disabled
-	// rather than silently treating it as false.
-	modernDispatchEnabled := false
-	if raw := os.Getenv(modernDispatchEnvVar); raw != "" {
-		var err error
-		if modernDispatchEnabled, err = strconv.ParseBool(raw); err != nil {
-			slog.Warn(fmt.Sprintf("%s has an unrecognized value %q; Modern stateless dispatch stays disabled",
-				modernDispatchEnvVar, raw))
-			modernDispatchEnabled = false
-		}
-	}
-
 	// Resolve transport defaults once here at the composition root: the
 	// vMCP config edge is the single place flags/CRD/YAML become a fully-resolved
 	// Config, so server.New, Serve, and the derive* helpers downstream are pure
@@ -446,7 +424,6 @@ func Serve(ctx context.Context, cfg ServeConfig) error {
 		Host:                    cfg.Host,
 		Port:                    cfg.Port,
 		SessionTTL:              cfg.SessionTTL,
-		ModernDispatchEnabled:   modernDispatchEnabled,
 		AuthMiddleware:          authMiddleware,
 		AuthzMiddleware:         authzMiddleware,
 		AuthInfoHandler:         authInfoHandler,

pkg/vmcp/client/modern_integration_test.go

@@ -95,10 +95,6 @@ func newModernVMCPServer(t *testing.T, backendURL string) *httptest.Server {
 			SessionTTL:     5 * time.Minute,
 			SessionFactory: vmcpsession.NewSessionFactory(authReg),
 			Aggregator:     agg,
-			// main re-added this kill-switch (default off, #5959); the harness
-			// must opt in so a well-formed Modern request reaches dispatchModern
-			// rather than falling through to the Legacy SDK path.
-			ModernDispatchEnabled: true,
 		},
 		router.NewSessionRouter(&vmcp.RoutingTable{}),
 		backendClient,

pkg/vmcp/config/config.go

@@ -166,6 +166,10 @@ type Config struct {
 	// When enabled, vMCP exposes only find_tool and call_tool operations to clients
 	// instead of all backend tools directly. This reduces token usage by allowing
 	// LLMs to discover relevant tools on demand rather than receiving all tool definitions.
+	// Enabling the optimizer currently pins this instance to MCP 2025-11-25: the
+	// find_tool/call_tool meta-tools are session-scoped, so Modern-capable (2026-07-28)
+	// clients are negotiated down to the Legacy revision (see
+	// pkg/vmcp/server/modern_gate.go; Modern parity is tracked in stacklok/toolhive#6089).
 	// +optional
 	Optimizer *OptimizerConfig `json:"optimizer,omitempty" yaml:"optimizer,omitempty"`
 

pkg/vmcp/conversion/errors.go

@@ -17,13 +17,17 @@ import (
 // Go errors: the SDK maps them to generic internal errors. For domain errors
 // that opt into CodedError, preserve the code/data in StructuredContent instead.
 func ErrorToToolResult(err error) *sdkmcp.CallToolResult {
+	// Denial first (matches writeModernDispatchError): an error that is both a
+	// CodedError and wraps ErrAuthorizationFailed must render as the denial,
+	// never as retry-shaped coded data (the sets are disjoint today; this
+	// ordering is the invariant).
+	if errors.Is(err, vmcp.ErrAuthorizationFailed) {
+		return sdkmcp.NewToolResultError(vmcp.DenyMessageToolCall)
+	}
 	var coded thvmcp.CodedError
 	if errors.As(err, &coded) {
 		return CodedErrorResult(err, coded)
 	}
-	if errors.Is(err, vmcp.ErrAuthorizationFailed) {
-		return sdkmcp.NewToolResultError(vmcp.DenyMessageToolCall)
-	}
 	return sdkmcp.NewToolResultError(err.Error())
 }
 
@@ -36,7 +40,7 @@ func CodedErrorResult(err error, coded thvmcp.CodedError) *sdkmcp.CallToolResult
 		"code":    coded.Code(),
 		"message": err.Error(),
 	}
-	if data := coded.Data(); data != nil {
+	if data := coded.Data(); len(data) > 0 {
 		structured["data"] = data
 	}
 	result.StructuredContent = structured

pkg/vmcp/server/authz_integration_test.go

@@ -178,15 +178,6 @@ func buildCedarAuthzServer(
 			Authz:          authzCfg,
 			AuditConfig:    auditCfg,
 			CodeModeConfig: codeModeCfg,
-			// Required for TestIntegration_CedarAuthzDenial_ModernPath_IsAudited to
-			// exercise what it claims: dispatchModern's re-homed call gate. Without
-			// it the kill switch refuses the Modern request in classifyingHandler and
-			// dispatchModern never runs. (Before the classifier learned to refuse an
-			// unserved revision, the request instead fell through to the SDK path, so
-			// that test passed on a denial from a different gate entirely.) Safe for
-			// the Legacy-shaped tests sharing this helper: classifyingHandler passes
-			// Legacy traffic through regardless of this flag.
-			ModernDispatchEnabled: true,
 		},
 		router.NewSessionRouter(&vmcp.RoutingTable{}),
 		backendClient,