docs: add ROADMAP.md with project direction for the next year

Cover planned work (JetBrains plugin, standard MCP integration,
Ultra mode, domain-specific long-term memory) and explicit non-goals.
Satisfies the OpenSSF Best Practices silver badge documentation_roadmap
criterion.

ROADMAP.md

@@ -0,0 +1,68 @@
+# Roadmap
+
+This document describes the planned direction for OpenCodeReview over the
+next year. It is a living document and will be updated as priorities evolve.
+
+Feedback is welcome via
+[GitHub Discussions](https://github.com/alibaba/open-code-review/discussions)
+or [Issues](https://github.com/alibaba/open-code-review/issues).
+
+## Current State (Mid-2026)
+
+OpenCodeReview currently provides:
+
+- A CLI tool (`ocr`) for AI-powered code review with deterministic
+  engineering and agent hybrid architecture.
+- Integration with coding agents: Claude Code (plugin/skill), Codex
+  (plugin), and Cursor (plugin).
+- A VSCode extension for in-editor code review.
+- CI/CD integration (GitHub Actions, GitLab CI, etc.).
+- Multi-provider LLM support (OpenAI-compatible, Anthropic, Google Gemini,
+  Amazon Bedrock, Azure OpenAI, etc.).
+- Review rules engine with per-file pattern matching.
+- Multi-language documentation (English, Chinese, Japanese, Korean, Russian).
+
+## Planned — H2 2026
+
+### IDE Plugins
+
+- **JetBrains plugin** — Bring AI code review to IntelliJ IDEA, GoLand,
+  PyCharm, and other JetBrains IDEs with the same capabilities as the
+  existing VSCode extension.
+
+### MCP Integration
+
+- **Standard MCP server** — Expose OpenCodeReview as a
+  [Model Context Protocol](https://modelcontextprotocol.io/) server,
+  allowing users to integrate external context tools (documentation
+  retrieval, issue trackers, internal knowledge bases) into the review
+  process through the standard MCP interface.
+
+### Ultra Mode
+
+- **Higher-recall review mode** — An opt-in mode that trades increased
+  token consumption and review time for significantly higher issue recall
+  rate. Designed for security-sensitive or high-risk changesets where
+  thoroughness is more important than speed.
+
+## Planned — H1 2027
+
+### Domain-Specific Long-Term Memory
+
+- **Persistent review knowledge** — Enable the review engine to accumulate
+  domain-specific knowledge over time (recurring patterns, past review
+  decisions, project-specific conventions) and apply it to future reviews,
+  improving relevance and reducing repeated feedback.
+
+## Not Planned
+
+The following are explicitly out of scope for the foreseeable future:
+
+- **Automated code fixing without human review** — OCR is a review tool,
+  not an auto-fix tool. While it can suggest fixes, applying changes always
+  requires human approval.
+- **General-purpose AI coding assistant** — OCR focuses exclusively on
+  code review. Features like code generation, refactoring, or chat-based
+  coding assistance are not planned.
+- **Self-hosted LLM bundling** — OCR connects to external LLM providers
+  but does not bundle or host models itself.