Edt MCP Architecture

EDT-MCP — target architecture

A 1C:EDT plugin with an MCP server (~62 tools). This skill maps how it should be (the project is mid-refactor toward shared helpers). If a helper below does not yet exist in code, it is a refactor task — see the card in .devtool/features/.

Package layout

Root: mcp/bundles/com.ditrix.edt.mcp.server/src/com/ditrix/edt/mcp/server

Canonical shared APIs (use, don't duplicate)

All helpers below exist in code. Verify the exact method name before calling (grep/Read).

• Project resolution → ProjectContext.of(projectName) (.exists(), .project(), .notFoundMessage(name)). Do not repeat ResourcesPlugin.getWorkspace()... by hand. (Metadata write tools get project+config via AbstractMetadataWriteTool.resolveProjectAndConfig.)
• BM model access → BmTransactions.read(model, name, op) / write(model, name, op). Reads in a read boundary, writes in a write boundary. Persist a metadata write to disk via BmTransactions.forceExportToDisk(project, [topFqn, configFqn]) (a bare write only enqueues the async export).
• Tool parameters → JsonUtils.extractStringArgument/extractIntArgument/extractBooleanArgument/extractObjectArray + JsonUtils.requireArguments(params, ...). Don't parse the Map by hand. (There is no ToolParams class.)
• Canonical key names → McpKeys (protocol/McpKeys.java) for the param/result keys shared across many tools: PROJECT_NAME, APPLICATION_ID, MODULE_PATH, LIMIT (input params) and MESSAGE, ACTION, PROJECT, ERROR (result keys). Use these in getInputSchema()/execute()/result-building instead of bare "projectName" literals; file-specific keys stay as private static final String constants in the tool. (error/success are owned by ToolResult — only hand-built error fields outside it use McpKeys.ERROR.)
• Metadata object/type resolution (bilingual) → MetadataTypeUtils (findObject, normalizeFqn); an existing node/member by FQN — MetadataNodeResolver.resolveExisting/resolveForCreate; a form — FormStructureReader.resolveMdForm. These are the shared bilingual resolvers — don't write your own.
• Synonym language → MetadataLanguageUtils.resolveLanguageCode(config, explicit) + getSynonymForLanguage(map, code). The synonym key is the language code (getLanguageCode()), NOT getName().
• Result/error → ToolResult: ToolResult.success().put(...) / ToolResult.error(...). Errors only via ToolResult.error(...).toJson(), never a bare string, never an exception escaping the tool.
• Pagination → the shared Pagination (utils/Pagination.java; one limit/offset, one "truncated" format).
• Markdown tables → the shared builder (escapes |/newline), not a per-tool StringBuilder.
• Tool registration → BuiltInToolRegistrar, not a list inside McpServer. EDT services — via Activator.getDefault().getXxx() or ServiceAccess.get(IFoo.class) (+ the package in MANIFEST Import-Package).

What NOT to do (anti-patterns)

• Don't copy project/module/BM resolution into each tool — call the shared helper.
• Don't put utilities or abstract bases in tools/impl/ — only IMcpTool classes (move-nontool-helpers-out-of-impl).
• Don't grow god-classes: RenameMetadataObjectTool/FindReferencesTool share MetadataReferenceLocator (extract-metadata-reference-locator); McpServer is being split (decompose-mcpserver).
• tags/* and groups/* — don't copy a third such stack; both features should inherit a common association-storage/service/refactoring base (extract-tags-groups-shared-base).

Related

• Two-language (ru/en) correctness — the edt-mcp-bilingual skill.
• Cross-tool contract (parameter naming, errors, output) — the edt-mcp-tool-conventions skill.
• Adding a new tool — the edt-mcp-new-tool skill.
• Build and tests — the edt-mcp-build-test skill.
• The full refactor task list — the .devtool/features/*.md board.