fix(repos)!: stop mirroring file bodies in get_file_contents structured output

docs/typed-tool-schemas.md

@@ -70,3 +70,12 @@ request-dependent output such as CSV. This is for intentional non-DTO responses
 such as resources, errors, or CSV; ordinary declared structured success must
 retain shared DTO serialization, with modern JSON text equal to
 `structuredContent`, even when legacy handler text was a plain mutation message.
+That text is encoded without HTML escaping (`<`, `>` and `&` stay literal), so it
+matches a standard re-serialization such as `JSON.stringify(structuredContent)`
+and clients can recognize it as a duplicate.
+
+Content blocks are kept as-is whenever any block is not text. Structured output
+for such results should describe the blocks rather than repeat their payloads:
+`get_file_contents` keeps file bodies in the embedded resource and reports only
+its `uri` and `mimeType` in `structuredContent`, so clients that forward both
+fields do not send each file to the model twice.

pkg/github/__toolsnaps__/get_file_contents.snap

@@ -181,61 +181,11 @@
                   "properties": {
                     "resource": {
                       "additionalProperties": false,
-                      "oneOf": [
-                        {
-                          "not": {
-                            "required": [
-                              "blob"
-                            ]
-                          },
-                          "required": [
-                            "text"
-                          ]
-                        },
-                        {
-                          "not": {
-                            "required": [
-                              "text"
-                            ]
-                          },
-                          "required": [
-                            "blob"
-                          ]
-                        },
-                        {
-                          "not": {
-                            "anyOf": [
-                              {
-                                "required": [
-                                  "text"
-                                ]
-                              },
-                              {
-                                "required": [
-                                  "blob"
-                                ]
-                              }
-                            ]
-                          }
-                        }
-                      ],
+                      "description": "Metadata for the embedded resource. The file body is returned only in the result content, not repeated here.",
                       "properties": {
-                        "blob": {
-                          "description": "Base64-encoded file bytes.",
-                          "type": [
-                            "null",
-                            "string"
-                          ]
-                        },
                         "mimeType": {
                           "type": "string"
                         },
-                        "text": {
-                          "type": [
-                            "null",
-                            "string"
-                          ]
-                        },
                         "uri": {
                           "type": "string"
                         }

pkg/github/repository_tool_types.go

@@ -1,7 +1,6 @@
 package github
 
 import (
-	"encoding/base64"
 	"encoding/json"
 	"fmt"
 	"maps"
@@ -212,8 +211,9 @@ func (out ForkRepositoryOutput) MarshalJSON() ([]byte, error) {
 }
 
 // Repository contents can be a projected directory listing or a heterogeneous
-// MCP content response. The latter is represented losslessly as typed content
-// blocks, including the human-readable status and embedded text/blob or link.
+// MCP content response. The latter is represented as typed content blocks,
+// including the human-readable status and resource or link metadata. File
+// bodies stay in the result content and are not repeated in structured output.
 type RepositoryContentsOutput struct {
 	Directory []*RepositoryDirectoryEntryOutput
 	Content   *RepositoryContentOutput
@@ -324,14 +324,15 @@ type RepositoryTextBlock struct {
 
 type RepositoryResourceBlock struct {
 	Type     string                   `json:"type"`
-	Resource RepositoryResourceOutput `json:"resource"`
+	Resource RepositoryResourceOutput `json:"resource" jsonschema:"Metadata for the embedded resource. The file body is returned only in the result content, not repeated here."`
 }
 
+// RepositoryResourceOutput describes an embedded file resource without its
+// body, so clients that forward both content and structured output to a model
+// do not receive every file twice.
 type RepositoryResourceOutput struct {
-	URI      string  `json:"uri"`
-	MIMEType string  `json:"mimeType,omitempty"`
-	Text     *string `json:"text,omitempty"`
-	Blob     *string `json:"blob,omitempty" jsonschema:"Base64-encoded file bytes."`
+	URI      string `json:"uri"`
+	MIMEType string `json:"mimeType,omitempty"`
 }
 
 type RepositoryLinkBlock struct {
@@ -353,12 +354,6 @@ func repositoryContentOutput(result *mcp.CallToolResult) (*RepositoryContentsOut
 			out.Content = append(out.Content, RepositoryContentBlock{Text: &RepositoryTextBlock{Type: "text", Text: block.Text}})
 		case *mcp.EmbeddedResource:
 			resource := RepositoryResourceOutput{URI: block.Resource.URI, MIMEType: block.Resource.MIMEType}
-			if block.Resource.Blob != nil {
-				blob := base64.StdEncoding.EncodeToString(block.Resource.Blob)
-				resource.Blob = &blob
-			} else if block.Resource.Text != "" {
-				resource.Text = &block.Resource.Text
-			}
 			out.Content = append(out.Content, RepositoryContentBlock{Resource: &RepositoryResourceBlock{Type: "resource", Resource: resource}})
 		case *mcp.ResourceLink:
 			out.Content = append(out.Content, RepositoryContentBlock{Link: &RepositoryLinkBlock{
@@ -521,11 +516,6 @@ func repositoryContentsOutputSchema() *jsonschema.Schema {
 	text.Properties["type"].Const = new(any("text"))
 	resource := repositoryOutputSchema[RepositoryResourceBlock]()
 	resource.Properties["type"].Const = new(any("resource"))
-	resource.Properties["resource"].OneOf = []*jsonschema.Schema{
-		{Required: []string{"text"}, Not: &jsonschema.Schema{Required: []string{"blob"}}},
-		{Required: []string{"blob"}, Not: &jsonschema.Schema{Required: []string{"text"}}},
-		{Not: &jsonschema.Schema{AnyOf: []*jsonschema.Schema{{Required: []string{"text"}}, {Required: []string{"blob"}}}}},
-	}
 	link := repositoryOutputSchema[RepositoryLinkBlock]()
 	link.Properties["type"].Const = new(any("resource_link"))
 	content := &jsonschema.Schema{

pkg/github/typed_repository_outputs_test.go

@@ -373,7 +373,11 @@ func TestTypedRepositoryOutputs(t *testing.T) {
 					first := result.Content[0].(*mcp.TextContent).Text
 					switch {
 					case call.name == "get_file_contents" && !strings.HasPrefix(first, "["):
-						assert.JSONEq(t, `{"content":`+mustMarshalJSON(t, legacy.Content)+`}`, mustMarshalJSON(t, result.StructuredContent))
+						structured := mustMarshalJSON(t, result.StructuredContent)
+						assert.JSONEq(t, `{"content":`+legacyContentWithoutResourceBodies(t, legacy.Content)+`}`, structured,
+							"structured output repeats content metadata but not file bodies")
+						assert.NotContains(t, structured, `"text":"hello\n"`)
+						assert.NotContains(t, structured, `"blob"`)
 					case call.name == "get_file_contents":
 						var entries []*RepositoryDirectoryEntryOutput
 						require.NoError(t, json.Unmarshal([]byte(first), &entries))
@@ -439,6 +443,19 @@ func hasOnlyTextContent(content []mcp.Content) bool {
 	return true
 }
 
+func legacyContentWithoutResourceBodies(t *testing.T, content []mcp.Content) string {
+	t.Helper()
+	var blocks []map[string]any
+	require.NoError(t, json.Unmarshal([]byte(mustMarshalJSON(t, content)), &blocks))
+	for _, block := range blocks {
+		if resource, ok := block["resource"].(map[string]any); ok {
+			delete(resource, "text")
+			delete(resource, "blob")
+		}
+	}
+	return mustMarshalJSON(t, blocks)
+}
+
 func assertLegacyRepositoryText(t *testing.T, name string, args map[string]any, result *mcp.CallToolResult) {
 	t.Helper()
 	switch name {
@@ -546,9 +563,10 @@ func TestRepositoryOutputUnions(t *testing.T) {
 			"contents", repositoryContentsOutputSchema(),
 			[]string{`null`, `[]`, `[{"name":"file"}]`, `{"content":[]}`, `{"content":[{"type":"text","text":"status"}]}`,
 				`{"content":[{"type":"resource","resource":{"uri":"repo://file","mimeType":"text/plain"}}]}`,
-				`{"content":[{"type":"resource","resource":{"uri":"repo://file","blob":"AAEC"}}]}`,
 				`{"content":[{"type":"resource_link","uri":"repo://file","name":"file","size":1048576}]}`},
-			[]string{`{}`, `{"content":[{"type":"text"}]}`, `{"content":[{"type":"resource","resource":{"uri":"repo://file","text":"x","blob":"AA=="}}]}`,
+			[]string{`{}`, `{"content":[{"type":"text"}]}`,
+				`{"content":[{"type":"resource","resource":{"uri":"repo://file","text":"x"}}]}`,
+				`{"content":[{"type":"resource","resource":{"uri":"repo://file","blob":"AAEC"}}]}`,
 				`{"content":[{"type":"opaque","data":{}}]}`},
 		},
 	} {

pkg/inventory/typed_output.go

@@ -361,7 +361,7 @@ func applyModernStructuredText(result *mcp.CallToolResult) error {
 	if result == nil || result.IsError || result.InputRequests != nil || result.StructuredContent == nil {
 		return nil
 	}
-	encoded, err := json.Marshal(result.StructuredContent)
+	encoded, err := marshalStructuredText(result.StructuredContent)
 	if err != nil {
 		return fmt.Errorf("marshal structured tool output for text content: %w", err)
 	}
@@ -376,6 +376,58 @@ func applyModernStructuredText(result *mcp.CallToolResult) error {
 	return nil
 }
 
+// marshalStructuredText encodes structured output for its JSON text block
+// without HTML escaping, so the text matches a standard re-serialization of
+// structuredContent (such as JSON.stringify) and clients can detect it as a
+// duplicate.
+func marshalStructuredText(value any) ([]byte, error) {
+	var buffer bytes.Buffer
+	encoder := json.NewEncoder(&buffer)
+	encoder.SetEscapeHTML(false)
+	if err := encoder.Encode(value); err != nil {
+		return nil, err
+	}
+	// Pre-encoded json.RawMessage values, including SDK typed output, keep
+	// their HTML escapes even when the encoder does not add new ones.
+	return unescapeJSONHTML(bytes.TrimSuffix(buffer.Bytes(), []byte("\n"))), nil
+}
+
+// unescapeJSONHTML replaces the \u003c, \u003e and \u0026 escapes that
+// encoding/json emits inside strings with the literal characters. In valid JSON
+// a backslash only starts a string escape, so consuming escapes pairwise
+// cannot misread an escaped backslash followed by "u003c".
+func unescapeJSONHTML(data []byte) []byte {
+	if !bytes.Contains(data, []byte(`\u00`)) {
+		return data
+	}
+	out := make([]byte, 0, len(data))
+	for i := 0; i < len(data); i++ {
+		if data[i] != '\\' || i+1 == len(data) {
+			out = append(out, data[i])
+			continue
+		}
+		if data[i+1] == 'u' && i+6 <= len(data) {
+			switch string(data[i+2 : i+6]) {
+			case "003c", "003C":
+				out = append(out, '<')
+				i += 5
+				continue
+			case "003e", "003E":
+				out = append(out, '>')
+				i += 5
+				continue
+			case "0026":
+				out = append(out, '&')
+				i += 5
+				continue
+			}
+		}
+		out = append(out, data[i], data[i+1])
+		i++
+	}
+	return out
+}
+
 func removeSDKOutputFallback(result *mcp.CallToolResult, state *typedCallState) error {
 	if result.StructuredContent == nil || len(result.Content) <= state.contentLengthBeforeSDK {
 		return nil

pkg/inventory/typed_output_test.go

@@ -1686,6 +1686,52 @@ func TestTypedScalarOutputProtocolGatePreservesHandlerText(t *testing.T) {
 	assert.Equal(t, "custom legacy text", result.Content[0].(*mcp.TextContent).Text)
 }
 
+func TestTypedModernTextDoesNotEscapeHTML(t *testing.T) {
+	type output struct {
+		HTML    string `json:"html"`
+		Literal string `json:"literal"`
+	}
+	tool := NewServerToolWithContextHandler(
+		mcp.Tool{Name: "typed_html_tool"},
+		testToolsetMetadata("test"),
+		func(context.Context, *mcp.CallToolRequest, struct{}) (*mcp.CallToolResult, output, error) {
+			return &mcp.CallToolResult{
+				Content: []mcp.Content{&mcp.TextContent{Text: "legacy text"}},
+			}, output{HTML: "<b>a & b</b>", Literal: `\u003c`}, nil
+		},
+	)
+	server := mcp.NewServer(&mcp.Implementation{Name: "test-server", Version: "v0.0.1"}, nil)
+	tool.RegisterFunc(server, nil)
+	modern := connectTypedTestClient(t, server, "")
+
+	result, err := modern.CallTool(context.Background(), &mcp.CallToolParams{Name: "typed_html_tool"})
+	require.NoError(t, err)
+	require.Len(t, result.Content, 1)
+	text := result.Content[0].(*mcp.TextContent).Text
+	assert.Equal(t, `{"html":"<b>a & b</b>","literal":"\\u003c"}`, text,
+		"modern text matches a standard re-serialization of structuredContent")
+
+	var reserialized bytes.Buffer
+	encoder := json.NewEncoder(&reserialized)
+	encoder.SetEscapeHTML(false)
+	require.NoError(t, encoder.Encode(result.StructuredContent))
+	assert.Equal(t, strings.TrimSuffix(reserialized.String(), "\n"), text)
+}
+
+func TestUnescapeJSONHTML(t *testing.T) {
+	for _, test := range []struct{ input, want string }{
+		{`{"a":"plain"}`, `{"a":"plain"}`},
+		{`"\u003cb\u003e \u0026 \u003C\u003E"`, `"<b> & <>"`},
+		{`"\\u003c"`, `"\\u003c"`},
+		{`"\\\u003c"`, `"\\<"`},
+		{`"\u00e9\u2028\"\u0026"`, `"\u00e9\u2028\"&"`},
+		{`"\u003`, `"\u003`},
+		{`"\`, `"\`},
+	} {
+		assert.Equal(t, test.want, string(unescapeJSONHTML([]byte(test.input))), test.input)
+	}
+}
+
 func TestTypedLegacyOutputGateRemovesFallbackWithEmptyHandlerContent(t *testing.T) {
 	tool := NewServerToolWithContextHandler(
 		mcp.Tool{Name: "typed_empty_content_tool"},