feat(schema): retire schema_agent_metadata JSON in favor of catalog-only delivery

Co-authored-by: Cursor <cursoragent@cursor.com>

scripts/policy/check-generated-drift.sh

@@ -12,12 +12,10 @@ policy_prepare_runtime "$ROOT"
 
 tmp="$(mktemp -d)"
 exec_tmp="$(policy_runtime_mktemp_dir dws-generated-drift)"
-metadata_generator="$exec_tmp/schema-agent-metadata"
 catalog_generator="$exec_tmp/schema-catalog"
 param_aliases_generator="$exec_tmp/param-aliases"
 trap 'rm -rf "$tmp" "$exec_tmp"' EXIT HUP INT TERM
 
-go build -o "$metadata_generator" ./internal/generator/cmd_schema_agent_metadata
 go build -a -o "$catalog_generator" ./internal/generator/cmd_schema_catalog
 go build -o "$param_aliases_generator" ./internal/generator/cmd_param_aliases
 
@@ -45,30 +43,16 @@ fi
 selection_guard="$tmp/selection-hints"
 cp -R internal/cli/schema_hints/selection "$selection_guard"
 
-metadata_tmp="$tmp/metadata"
-audit_tmp="$tmp/audit.json"
 catalog_tmp="$tmp/schema_catalog"
 catalog_tmp_second="$tmp/schema_catalog-second"
 param_aliases_tmp="$tmp/param_aliases_generated.go"
 param_aliases_tmp_second="$tmp/param_aliases_generated-second.go"
 
-"$metadata_generator" \
-  -root . \
-  -registry internal/cli/schema_command_registry \
-  -output-dir "$metadata_tmp" \
-  -audit-output "$audit_tmp"
-
-if ! diff -qr internal/cli/schema_agent_metadata "$metadata_tmp" >/dev/null; then
-	printf '%s\n' 'generated drift: internal/cli/schema_agent_metadata is stale' >&2
-	printf '%s\n' 'run: make generate-schema' >&2
-	diff -ru internal/cli/schema_agent_metadata "$metadata_tmp" || true
-	exit 1
-fi
-
-if ! cmp -s internal/cli/schema_agent_metadata_audit.json "$audit_tmp"; then
-	printf '%s\n' 'generated drift: internal/cli/schema_agent_metadata_audit.json is stale' >&2
-	printf '%s\n' 'run: make generate-schema' >&2
-	diff -u internal/cli/schema_agent_metadata_audit.json "$audit_tmp" || true
+# Retired intermediate: schema_agent_metadata/ must not reappear as a delivery
+# artifact. Catalog generation injects Agent metadata in-memory only.
+if [ -e internal/cli/schema_agent_metadata ] || [ -e internal/cli/schema_agent_metadata_audit.json ]; then
+	printf '%s\n' 'generated drift: retired schema_agent_metadata delivery artifact is present' >&2
+	printf '%s\n' 'remove internal/cli/schema_agent_metadata/ and schema_agent_metadata_audit.json' >&2
 	exit 1
 fi
 

scripts/policy/check-schema-catalog.sh

@@ -28,15 +28,17 @@ registry_count="$(jq -r '.tools | length' "$catalog")"
 catalog_count="$registry_count"
 catalog_product_count="$(jq -r '.catalog.count' "$catalog")"
 mcp_snapshot_registry_count="$(jq -r '.coverage.surface_tools' internal/cli/schema_mcp_metadata.json)"
-agent_registry_count="$(jq -r '.coverage.surface_tools' internal/cli/schema_agent_metadata/index.json)"
-agent_product_count="$(jq -r '.coverage.products_with_metadata' internal/cli/schema_agent_metadata/index.json)"
-agent_selection_count="$(jq -r '[.coverage.tools_with_use_when, .coverage.tools_with_avoid_when, .coverage.tools_with_examples, .coverage.tools_with_interface_mode] | min' internal/cli/schema_agent_metadata/index.json)"
+# Agent metadata is no longer a shipped intermediate. Selection completeness is
+# proven on the Catalog itself (every tool/product carries Agent prose).
+agent_registry_count="$registry_count"
+agent_product_count="$catalog_product_count"
+agent_selection_count="$(jq -r '[.tools[] | select((.use_when|type)=="array" and (.avoid_when|type)=="array" and (.examples|type)=="array" and ((.interface_mode//"")|length)>0)] | length' "$catalog")"
 if [ "$agent_registry_count" != "$registry_count" ] ||
 	[ "$agent_product_count" != "$catalog_product_count" ] ||
 	[ "$agent_selection_count" != "$registry_count" ]; then
-	printf 'generated schema counts disagree: registry=%s catalog=%s products=%s agent=%s/%s\n' \
+	printf 'generated schema counts disagree: registry=%s catalog=%s products=%s selection=%s\n' \
 		"$registry_count" "$catalog_count" "$catalog_product_count" \
-		"$agent_product_count" "$agent_registry_count" >&2
+		"$agent_selection_count" >&2
 	exit 1
 fi
 
@@ -110,10 +112,8 @@ if ! jq -e '
 fi
 
 catalog_registry_hash="$(jq -r '.surface_hash' "$catalog")"
-agent_registry_hash="$(jq -r '.surface_hash' internal/cli/schema_agent_metadata/index.json)"
-if [ "$catalog_registry_hash" != "$agent_registry_hash" ]; then
-	printf 'schema CommandRegistry hashes disagree: catalog=%s agent=%s\n' \
-		"$catalog_registry_hash" "$agent_registry_hash" >&2
+if [ -z "$catalog_registry_hash" ] || [ "$catalog_registry_hash" = "null" ]; then
+	printf 'schema Catalog is missing surface_hash\n' >&2
 	exit 1
 fi
 
@@ -256,13 +256,17 @@ if policy_search_paths 'mcp-gw\.dingtalk\.com|mcp\.dingtalk\.com/server|Authoriz
 	"$catalog" \
 	internal/cli/schema_mcp_metadata.json \
 	internal/cli/schema_mcp_service_review.json \
-	internal/cli/schema_agent_metadata \
 	internal/cli/schema_parameter_bindings.json \
 	internal/cli/schema_hints; then
 	printf '%s\n' 'schema assets contain endpoint or credential material' >&2
 	exit 1
 fi
 
+if [ -e internal/cli/schema_agent_metadata ] || [ -e internal/cli/schema_agent_metadata_audit.json ]; then
+	printf '%s\n' 'retired schema_agent_metadata delivery artifact must not be present' >&2
+	exit 1
+fi
+
 if policy_search_go '\.ListTools\(' internal/app internal/cli; then
 	printf '%s\n' 'startup/schema packages must not call MCP tools/list' >&2
 	exit 1

scripts/policy/check-schema-command-registry.sh

@@ -39,8 +39,12 @@ fi
 
 # The registry argument is validation-only and all go:generate outputs must be
 # downstream assets. Reject any directive that targets the reviewed input.
-if ! grep -Eq '^//go:generate .*cmd_schema_agent_metadata .* -registry internal/cli/schema_command_registry ' internal/cli/gen.go; then
-	printf '%s\n' 'go generate must validate the reviewed CommandRegistry explicitly' >&2
+if ! grep -Eq '^//go:generate .*cmd_schema_catalog ' internal/cli/gen.go; then
+	printf '%s\n' 'go generate must register the Catalog generator' >&2
+	exit 1
+fi
+if grep -E '^//go:generate' internal/cli/gen.go | grep -Eq 'cmd_schema_agent_metadata|schema_agent_metadata'; then
+	printf '%s\n' 'go generate must not regenerate retired schema_agent_metadata/' >&2
 	exit 1
 fi
 if policy_search_go '^//go:generate .*-(output|output-dir|audit-output)(=|[[:space:]]+)([^[:space:]]*/)?schema_command_registry\.json([[:space:]]|$)' \
@@ -51,11 +55,20 @@ if policy_search_go '^//go:generate .*-(output|output-dir|audit-output)(=|[[:spa
 	exit 1
 fi
 
-# Embedded Agent/MCP/parameter metadata is intentionally expensive and must be
+# Embedded MCP/parameter metadata is intentionally expensive and must be
 # parsed only through its sync.Once accessor. Each raw loader is allowed at
 # exactly two production locations: its declaration and the assignment inside
 # that accessor. Any third reference is an eager initializer or an accessor
 # bypass and fails this static check.
+# Agent metadata embed/loader is retired; production must not reopen it.
+if policy_search_production_go 'go:embed schema_agent_metadata' internal/cli; then
+	printf '%s\n' 'schema_agent_metadata must not be re-embedded' >&2
+	exit 1
+fi
+if policy_search_production_go 'loadEmbeddedAgentMetadata\(' internal/cli; then
+	printf '%s\n' 'retired loadEmbeddedAgentMetadata must not remain in production code' >&2
+	exit 1
+fi
 check_schema_loader_references() {
 	loader="$1"
 	allowed="$2"
@@ -73,9 +86,6 @@ check_schema_loader_references() {
 	fi
 }
 
-check_schema_loader_references \
-	'loadEmbeddedAgentMetadata' \
-	'^internal/cli/schema_agent_metadata\.go:[0-9]+:(func loadEmbeddedAgentMetadata\(\) embeddedAgentMetadata \{|[[:space:]]*runtimeEmbeddedAgentMetadataLazy\.metadata = loadEmbeddedAgentMetadata\(\))$'
 check_schema_loader_references \
 	'loadEmbeddedMCPMetadata' \
 	'^internal/cli/runtime_schema\.go:[0-9]+:(func loadEmbeddedMCPMetadata\(\) embeddedMCPMetadata \{|[[:space:]]*runtimeEmbeddedMCPMetadataLazy\.metadata = loadEmbeddedMCPMetadata\(\))$'