v1.0.2

Changelog

All notable changes to this project will be documented in this file.

[1.0.2] - 2026-07-08

Security

• Path Traversal Prevention: rms_read and rms_write now reject paths containing .. components, preventing escape from the vault directory.
• LanceDB Filter Injection: All filter strings in store.rs (delete_document, read_document) now escape single quotes via escape_filter(), preventing potential data corruption.

Fixed

• OpenCode MCP Payload Format: Replaced hardcoded if ide.name == "OpenCode" branching with a PayloadBuilder dependency-injection architecture. The opencode_payload function now produces OpenCode's native McpLocalConfig schema ({"type": "local", "command": [array], "enabled": true}), which was previously missing the type field and using a string command instead of an array.
• Zed JSONC Silent Skip: The installer now falls back to strip_json_comments() when serde_json::from_str fails on config files containing // comments (e.g. Zed's settings.json). Previously, Zed configs were silently skipped. On parse failure, tracing::warn! now logs a clear diagnostic message.
• Missing enabled: true for Standard Payload: The standard_payload function now injects "enabled": true for all non-OpenCode IDEs. This is required by Zed's context_servers schema and is a harmless no-op for Claude, Cursor, and VSCode.
• Hybrid Search Activated: The LanceDB search() method now uses combined vector search + Tantivy full-text search (FullTextSearchQuery). Previously, only vector search was performed despite the FTS index being created. Falls back to vector-only gracefully if FTS index is unavailable.
• Write create Mode: Now correctly rejects overwriting existing files. Previously, create fell through to the catch-all _ branch which behaved as a full overwrite.
• links_resolved No Longer Stub: Both sync_vault and index_vault_full now store actual normalized link paths in links_resolved instead of the placeholder "[]".
• Safe Unsafe: Added // SAFETY: documentation comment for the TMPDIR environment variable override in Indexer::new().

Added

• rms-memory doctor (Full Diagnostics): Implements 5-point vault health check: directory structure, missing document IDs, broken cross-document links, LanceDB store accessibility, and registry coherence.
• rms-memory uninstall: New command to remove rms-memory entries from all discovered IDE configuration files. Uses patcher::remove_key() for safe JSONC-aware key removal with automatic .bak backups.
• Graceful Shutdown: Added SIGINT/Ctrl+C handler in main.rs via tokio::signal::ctrl_c(), ensuring clean log flush on exit instead of immediate process termination.
• Zombie Process Prevention: MCP server now signals background file-watcher tasks to stop when stdin closes (EOF on disconnect). std::process::exit(0) in main.rs guarantees the process terminates even if tokio runtime has lingering tasks.
• llms.txt Export Compliance: export-llms now generates a proper llms.txt spec format with clickable links, frontmatter-derived titles, and content summaries, in addition to the full vault contents section.

Changed

• Installer DI Architecture: Introduced PayloadBuilder type alias (fn(exe: &str) -> serde_json::Value) and attached it to IdeConfig as a build_payload field. Adding a new IDE format no longer requires modifying run_installer() — just one line in get_ide_registry().
• ExportLlms Optimization: Fixed double find_markdown_files() call — now computed once.

v1.0.1

Drop macOS Intel (x86_64) builds

v1.0.0

Changelog

All notable changes to this project will be documented in this file.

[1.0.0] - 2026-07-07

Fixed

• MCP Initialization Bug: Changed MCP workspace provisioning to fallback to the process's current working directory (cwd) when the IDE (like Zed) fails to pass a valid rootUri in the JSON-RPC initialize handshake. Furthermore, strict validation now rejects root paths (/) to entirely prevent the generation of orphaned UnknownProject vaults.
• Config Directory Isolation: Removed the directories::ProjectDirs abstraction which defaulted to messy system paths (e.g., ~/Library/Application Support/). All configs, databases, and logs are now strictly bound to a cross-platform ~/.rms-memory/ directory to ensure clean disk usage.
• Import Routing Categorization: Fixed an issue where unmatched markdown files were erroneously dumped into guides/. They now default to docs/. Added strict mapping for task.md, walkthrough.md, changelog, history and implementation_plan to route directly to artifacts/.
• Infinite Sync Loop Fix (CPU Hog): Resolved a critical issue where the notify file watcher would recursively trigger itself indefinitely by ignoring changes to .lancedb, store.json, and .log files, ensuring 0% CPU consumption during idle times.
• macOS Sandboxing Crashes: rms-memory install now automatically copies a safe binary, builds entitlements (disable-library-validation), and runs codesign atomically to bypass macOS terminating the running process with SIGKILL.
• Agent Rule Templates: Updated all MCP injected templates (general_mcp_guide.md, cursor_rules.md, claude_code_rules.md, zed_assistant_rules.md) to explicitly use IDE-prefixed tools (rms-memory_rms_search) and document the artifacts/, docs/, and api/ directories.
• Repository Pollution: Removed the generation of .bak files during agent rule injection (rules_injector.rs) to prevent flooding user workspaces with backup files.
• Batched Vector Indexing (OOM Fix): Replaced monolithic embedding calls with a batched chunking architecture (batch_size = 32) in sync_vault and index_vault_full. This entirely resolves extreme CPU/RAM spikes and process deadlocks (OOM) when indexing exceptionally large files, allowing the fastembed ONNX Runtime to efficiently ingest 100% of the file content without truncation.

Changed

• Default Vault Structure: Vault initialization (cli.rs, workspace.rs) now explicitly creates docs/ and api/ directories alongside rules/, decisions/, architecture/, and artifacts/.

Added

• Monolith Refactoring & Dependency Injection: Completely dismantled cli.rs and mcp_server.rs monoliths into modular components (src/commands/ and src/tools/). Introduced AppContext for dependency injection of databases and models across the system.
• MCP Stdio Server: Full implementation of JSON-RPC protocol over standard I/O for rms-memory_rms_read, rms-memory_rms_write, and rms-memory_rms_search tooling.
• Global Vault Registry: Added registry.toml routing logic allowing the server to automatically detect the current code directory and isolate contextual documentation into a unified, secure system-level vault (~/.rms-memory/vaults/ProjectName).
• Hybrid LanceDB Retrieval: Deployed local embedded LanceDB (v0.31.0) configured with multi-threaded Vector Search and Tantivy FTS indices to guarantee zero-fail context hits.
• Multilingual-E5-Small FastEmbed Pipeline: Native ONNX AST embedding for dual-language (Russian & English) documentation support directly integrated into the ingest flow.
• Semantic AST Markdown Chunking: Precision boundary preservation (1500 chars limit) via pulldown-cmark. It attaches hierarchical headings to paragraphs and enforces smart sliding-window truncation for large code blocks.
• Dynamic IDE Auto-Installer (rms-memory install): Interactive recursive scanner spanning ~/.config/ and ~/Library/Application Support/ to auto-inject the mcpServers JSON object directly into Cursor, Zed, OpenCode, VS Code, and Claude Code configurations.
• Rules-as-Code IDE Patching: Non-destructive AST block-patching (<!-- RMS-MEMORY-START -->) to automatically inject contextual guide instructions into .cursorrules, .claude/CLAUDE.md, .zed/assistant.md, and RMS_MEMORY_GUIDE.md.
• Dry-Run & System Auditing (--dry-run): Execution previews for install and init commands to visualize JSON modifications and Markdown AST patches without corrupting host files. Generates automated .bak backups before any writes.
• Background Incremental Sync (rms-memory sync): Zero-latency startup sync task performing Delete-then-Insert vector replacement based on mtime bound metadata.
• Garbage Collection (rms-memory gc): Prunes orphaned LanceDB data caches matching deleted Vault boundaries.
• LLMs.txt Export Hook: Enables one-shot compilation of a project Vault into standard flat structures.
• Linked Documents Architecture: Introduced native support for "Link Files"—standard Markdown Vault entries containing a link: <source_path> frontmatter property. The LanceDB indexer, read and write endpoints seamlessly trace and operate on the original repository file while enforcing strict path constraints within the Vault directory structure.
• Interactive Documentation Importer (rms-memory import): Added a powerful CLI importer to scan project repositories for existing legacy documentation (README.md, CLAUDE.md, .cursorrules, docs/, ADR/). Features 5 user-selectable strategies including Link Only (creates zero-copy Link Files) and Import & Organize (deterministically categorizes files into the Vault).
• Auto-Import Integration: Added --auto-import flag to rms-memory config (auto_import_strategy in registry.toml). When auto_add_projects triggers in the background, the server can now silently resolve legacy documentation into the Vault using the pre-configured strategy without blocking the MCP connection.
• Write-Guard Snapshotting: Captures local fs::copy .bak state preservation automatically before an AI is permitted to execute JSON-RPC .md replace/append mutations.
• Dedicated Telemetry Logging: Integrated tracing streams routed directly to ~/.rms-memory/rms.log shielding MCP stdio channels from standard output noise.

Fixed

• Safe Auto-Inject Default: auto-inject configuration now defaults to false, preventing implicit modification of .cursorrules, CLAUDE.md, etc., upon first repository discovery without explicit user consent.
• Write-Guard Rolling Backups: Added max_backups parameter to prevent disk pollution. Ensures backup files are gracefully rotated and limited (default 5).
• Embedding Dimension Safety: Hardened Store::init to properly validate existing table schema dimension arrays against the current embedding model dimension. Returns a loud INDEX_REBUILD_REQUIRED error instead of a runtime vector search panic.
• Dependency Inversion (DIP) & Testing: Extracted VectorStore and Embedder traits from McpServer, establishing a robust Mock architecture for the JSON-RPC layer and cleanly eliminating LanceDB RecordBatch leaky abstractions.
• Bug Fix: Fixed server crash/early exit during initialize in strict sandboxes by routing hf-hub atomic download temp files to the .rms-memory cache directory instead of the system temp directory, preventing Read-only file system (os error 30) panics.
• Bug Fix: Fixed JSON-RPC stream corruption where background auto-initialization was writing [INFO] logs to stdout instead of using tracing or stderr, causing the MCP client to instantly disconnect the transport.