Changelog

v0.6.0

MCP server fixes, per-server mode configuration, @ mention system for referencing resources, parallel file inspection tool, conversation title generation improvements, and security hardening.

MCP Server Fixes

• Celery worker MCP loading -- MCP servers are now loaded in the Celery background worker path, matching the inline session path. Previously MCP tools were completely unavailable when using background jobs
• Multi-server dispatch -- Tools from multiple MCP servers now dispatch to their correct originating client. Previously self._mcp_client was overwritten by each server, causing tools from earlier servers to fail at execution time
• Plan mode MCP access -- MCP tools are no longer blocked by the plan-mode whitelist. Each tool is tracked with its own mode configuration and bypasses the built-in tool restrictions
• Exception logging -- register_mcp_tools now logs warnings on failure instead of silently swallowing exceptions with bare except: pass
• Tool name collision logging -- MCP tools that attempt to shadow built-in tool names are logged with a warning for security observability
• Connection timeout -- MCP server connections in the Celery worker are wrapped with a 30-second timeout to prevent hanging workers
• Cleanup -- MCP connections are properly disconnected in the Celery worker's finally block

Per-Server Mode Configuration

• Mode checkboxes -- Each MCP server can be configured to be available in Plan mode, Execute mode, or both (default: both) via checkboxes in Settings > MCP Servers
• Backend enforcement -- The modes field is stored in the server config, passed through MCPManager.connect_servers to register_mcp_tools, and enforced by ToolRouter.is_tool_allowed per-tool
• Status endpoint -- GET /api/mcp/status now includes modes in each server's response

@ Mention System

• MentionPopover component -- Type @ in the chat input to open a dropdown showing MCP servers and workspace files. Supports directory browsing (typing @code/ lists files in code/), keyboard navigation (arrows, Enter, Tab, Escape), and filtering by name
• Mention chips -- Active mentions are displayed as colored chips above the input area (blue for MCP servers, amber for files)
• Resource references -- Mentions are sent as lightweight structured references ({type, value}) alongside the message. The backend prepends reference hints that instruct the agent to use appropriate tools (read, inspect_files, MCP tools) to interact with the referenced resources
• Mention model -- New Mention Pydantic model with type: Literal["server", "file"] and value: str (max 1024 chars). Added mentions field to MessageSend
• Input sanitization -- Mention values are sanitized (backticks, newlines, control characters stripped, length capped at 256) before interpolation into prompt text to prevent LLM prompt injection

inspect_files Tool

• Parallel file reading -- New inspect_files tool reads multiple files or directories concurrently via asyncio.gather and scores each file for relevance against a user query
• Keyword relevance scoring -- Files are scored by keyword overlap between their content and the query, sorted by relevance, and returned within a configurable token budget (100K chars default)
• Directory expansion -- Directory paths are expanded to their file listings; hidden files (dotfiles) are excluded
• Safety limits -- Max 50 files per call, 200 lines per file for scoring, 2MB file-size gate (large files skipped before reading), negative max_files clamped to 1
• Security -- Each child file in expanded directories is re-validated via _validate_path to catch symlinks escaping the workspace
• Plan mode access -- Added to the plan-mode allowlist for read-only context gathering

Conversation Title Generation

• Deferred generation -- Title generation no longer triggers after the 1st user message. It now triggers after the 3rd user message or on page refresh, whichever comes first
• No re-updates -- Once a title is set, it is not overwritten by subsequent triggers. A race-condition guard in _auto_title re-checks the current title from DB before persisting
• Trigger guard -- The send_message endpoint checks conv.title == "New conversation" before triggering, preventing redundant generation

Security Hardening

• Symlink traversal protection in inspect_files -- each child entry in expanded directories is validated via _validate_path before reading
• File-size gate in inspect_files -- files over 2MB are skipped before read_text() to prevent OOM
• asyncio.get_running_loop() used instead of deprecated get_event_loop() in async contexts
• MCP tool name shadowing logged as a warning for security observability
• Mention values sanitized to strip backticks, newlines, and control characters before prompt interpolation
• Mention.value field constrained to max 1024 characters via Pydantic Field
• MCP connection timeout (30s) in Celery worker prevents indefinite worker stalls

UI Fixes

• Layout gap fix -- Fixed 1px gap between the main content area and the right sidebar caused by paddingRight being 1px larger than the RightPanel's rendered width (289px → 288px, 49px → 48px)
• MCP live connection status -- MCP server dots in the right panel now turn green when connected. Previously the status was hardcoded to connected: false in the REST endpoint. An mcp_status SSE event is now broadcast from both the session manager and Celery worker after MCPManager.connect_servers() succeeds, and the frontend handles it to update the dots in real time
• Pre-existing lint fixes -- Removed extraneous f-string prefix in papers.py, removed unused AsyncMock import in test_tools_papers.py

Testing

• 34 new backend tests -- MCP multi-client dispatch (9), inspect_files tool (12), mention enrichment (7), Mention model validation (3), title generation (3 from prior session)
• Total: 915 backend + 223 frontend = 1,138 tests
• All ruff checks pass, frontend eslint 0 errors

v0.5.0

Project-scoped conversations, unified file workspace, Monaco code viewer, TODO approval flow, comprehensive agent guidance, and test infrastructure improvements.

Project-Scoped Conversations

• Mandatory projects -- Every conversation must belong to a project; the "All Conversations" concept has been removed entirely
• Onboarding project creation -- 3-step onboarding flow: Providers > Model > Create First Project. Users name their project and understand the concept from the start
• Orphan cleanup -- Existing conversations without a project are automatically deleted when the conversation list is loaded
• Project selector -- No more null state; always shows a real project. Dropdown removed the "All" option

Unified File Workspace

• Resources materialized as files -- PLAN.md, completion reports, and paper drafts are now written as real files to the project workspace (.project-meta/plans/, .project-meta/reports/, papers/) and appear in the Files panel
• Resources panel removed -- The separate Resources section in the right panel has been replaced by the unified FileTree. Tasks tab renamed to Todos and displayed alongside Files as stacked collapsible panels (no more tab buttons)
• workspace_files_changed SSE event -- FileTree auto-refreshes when resources are written to the workspace
• File badges -- Pin icon for PLAN.md, clipboard icon for reports, book icon for paper drafts in the FileTree

Monaco Code Viewer

• Agent/Editor tabs -- Main content area now has two tabs: Agent (chat) and Editor (read-only code viewer)
• Monaco Editor integration -- VS Code engine (@monaco-editor/react) with full syntax highlighting, minimap, line numbers, bracket pair colorization, and word wrap. Supports 15+ languages (Python, TypeScript, JSON, Markdown, YAML, LaTeX, shell, etc.)
• Multi-tab file viewer -- Open multiple files as tabs in the Editor, close individually. Closing the last tab switches back to the Agent tab
• FileTree integration -- Clicking a file in the Files panel opens it in the Editor tab; inline preview removed

Agent Improvements

• Message persistence fix -- Assistant text that precedes tool calls is now persisted to the database via an assistant_message event, surviving page refreshes
• Plan-mode research budget -- Warns after 5+ research tool calls in Plan mode; guides the agent to save comprehensive research for Execute mode tasks
• TODO approval in Execute mode -- Creating or adding tasks in Execute mode requires user approval via a dedicated review UI with current vs. proposed diff view and inline task editing
• Blank section enforcement -- Writing tool warns about incomplete sections after each write; get_draft appends a WARNING listing all [Not yet written] placeholders

Workspace Targeting

• Project workspace auto-targeting -- read, write, edit, and bash tools automatically target the active project workspace directory. Files created by the agent appear in the Files panel
• Workspace context wiring -- Both inline and Celery worker sessions resolve the project workspace and set context for workspace tools (knowledge graph, notes) and local tools (file operations)

Comprehensive Agent Guidance

• System prompt v7 -- Full rewrite with a Tool Selection Decision Tree, Workspace Structure reference, expanded Code Execution constraints, and strengthened Plan/Execute mode rules
• Expanded tool descriptions -- bash (Docker constraints, common patterns), read (offset/limit usage), write (workspace targeting, anti-pattern for papers), edit (find-and-replace semantics), research (available tools, iteration limits), writing (full workflow, placeholder enforcement), plan_tool (enforcement rules, workspace file locations)
• Research sub-agent prompt -- Now lists available tools, constraints (60 iterations, ~190k token budget), stop conditions, error handling guidance, and clarifies that the sub-agent cannot write to the workspace

Testing & Infrastructure

• Pytest hang fix -- Removed deprecated event_loop session fixture, added _dispose_engine_at_exit for async engine cleanup, set asyncio_mode = "auto" in pytest config. Full backend suite now exits cleanly in ~50s
• 47 new backend tests -- Message persistence, research budget, project workspace targeting, incomplete section warnings, orphan cleanup, workspace path resolution, TODO approval Redis helpers, report persistence
• 22 new frontend tests -- EditorPanel (12 tests), FileTree (10 tests)
• ProjectSelector tests -- 12 tests verifying no-null project selection
• TodoReviewDrawer tests -- 10 tests for the approval UI
• Total: 858 backend + 226 frontend = 1,084 tests

Bug Fixes

• Fixed silent exception swallowing in _wire_persistence (now logs errors via logger.exception)
• Fixed inline-mode SSE events missing conversation_uuid (prevented cross-conversation event leakage)
• Fixed pnpm-lock.yaml not including @monaco-editor/react for Docker builds
• Fixed test_list_conversations and test_conversations_isolated_by_user for project-scoped queries
• Fixed test_conversations.py integration tests to create conversations with a project

---

v0.4.0

Projects, workspaces, interactive terminal, multi-provider model picker, security hardening, and centralized version management.

Projects & Workspaces

• Project management — Create, rename, archive, and manage projects via the sidebar or Settings. Each project gets a persistent workspace directory with standard subdirectories (code/, data/, papers/, research/, etc.)
• Default project — Every user gets an auto-created "All Conversations" project; conversations can be moved between projects
• File browser — Interactive file tree panel for browsing project workspace files, with lazy-loading directories, file-type icons, inline text preview, and file sizes on hover
• File operations API — Browse, read, write, delete, and upload files within project workspaces via REST endpoints
• Knowledge graph — Per-project persistent knowledge graph (backed by networkx) for cross-conversation memory. Supports 10 entity types (paper, concept, method, dataset, etc.) and 13 relationship types (cites, implements, extends, etc.). Context is injected into new conversations automatically
• Workspace persistence — Saves search results, research notes, parsed papers, tool failure logs, compute probes, experiment logs, and cross-conversation state to the project workspace
• Workspace agent tools — New workspace tool with 8 operations: status, search, note, knowledge_add, knowledge_relate, knowledge_query, knowledge_summary, and recent_failures
• Project manage modal — Bulk project management UI with inline rename, conversation counts, and delete with confirmation

Interactive Terminal

• WebSocket PTY terminal — Full interactive bash shell in the browser via xterm.js, connected to a real server-side PTY process scoped to the project workspace
• Environment scrubbing — Terminal sessions receive a minimal allowlisted environment; server secrets (API keys, DATABASE_URL, JWT_SECRET_KEY, etc.) are never exposed
• Terminal UI — Maximize/minimize toggle, connection status indicator, manual reconnect, 5,000-line scrollback, JetBrains Mono font

Model Picker & Providers

• Multi-provider model catalog — Browse models from OpenAI, Anthropic, OpenRouter, OpenCode Go, Ollama, and LM Studio in a unified picker. Live model list fetched from models.dev with hardcoded fallbacks
• Custom providers — Register arbitrary OpenAI-compatible, Anthropic SDK, OpenRouter, or LiteLLM endpoints with custom base URLs and API keys. Fetch and cache model lists from custom endpoints
• Model picker UI — Two-tab modal with search/filter, provider logos, recently used models (up to 5), and a custom model ID input for arbitrary model strings
• Provider settings — Tabbed settings page (Models, Search, Papers, Compute, Others) with per-provider status indicators, inline API key management, and an "Add Custom Provider" modal
• Automatic SDK routing — Models are routed to the correct SDK (OpenAI or Anthropic) based on provider and model name. Custom providers specify their SDK type explicitly
• Recently used models — Last 5 used models tracked per-user and shown at the top of the picker

Security

• Path traversal prevention — All file operations use Path.relative_to() containment checking; symlinks pointing outside the workspace are blocked for both reads and deletes
• Resource limits — File uploads capped at 100 MB, writes at 10 MB, reads truncated at 500 KB; knowledge graph limited to 10,000 nodes / 50,000 edges; terminal input capped at 4 KB per message
• Process isolation — Terminal shells spawned with start_new_session=True, close_fds=True, --norc --noprofile; zombie prevention via SIGTERM/SIGKILL escalation with proper waitpid reaping
• Protected directories — Top-level workspace directories (code, data, papers, etc.) cannot be deleted via the API
• Config allowlist — POST /api/config only accepts a hardcoded list of 10 environment variable names

Version Management

• Single source of truth — New VERSION file at repo root; all version references derive from it
• Automatic version bumping — make version-patch, make version-minor, make version-major commands bump and propagate across the entire monorepo
• Explicit version setting — make version-set V=X.Y.Z for arbitrary version changes
• Version display — Current version shown in the webapp sidebar footer and docs site footer

Bug Fixes

• Fixed test_list_projects_api failing due to auto-created default project not being accounted for
• Fixed KeyManager path to use /app/.keys instead of /.keys
• Fixed dev entrypoint to always sync dependencies on startup
• Fixed eslint pre-commit hook path resolution for new files
• Added missing python-multipart dependency for file upload endpoint

Internal

• 65 new tests for projects, workspaces, knowledge graph, and workspace tools (740 backend tests, 186 frontend tests total)
• Unified version strings: removed hardcoded versions from app.py, health.py, test_app.py
• Changelog moved to repo root; synced into docs site at build time with VitePress frontmatter
• Pydantic request/response models for all API endpoints (backend/openmlr/models.py)

---

v0.3.0

Compute environments, UI improvements, and bug fixes.

Compute Environments

• Multi-backend compute — Execute code on local Docker, SSH remotes, or Modal cloud
• SSH key management — Generate Ed25519/RSA keys, upload existing keys via Settings > Compute
• Compute probing — Detect OS, GPUs (with CUDA version), Python versions, and disk space
• Compute selection — Switch between configured compute nodes mid-conversation
• Connection pooling — SSH connections are reused across tool calls for performance
• Docker-in-Docker detection — Worker container executes commands directly when already in Docker

UI Improvements

• Collapsible Tasks & Resources — Click section headers in the right panel to collapse/expand
• Fixed right panel layout — Right panel no longer causes page scroll issues when toggled
• Improved scroll behavior — Message list scrolls correctly without affecting page layout

Bug Fixes

• Fixed scrollIntoView causing entire page to scroll when RightPanel is open
• Fixed _get_draft database call using wrong function name
• Fixed test suite failures related to async database mocking

Internal

• Added _running_in_container() detection for Docker Compose worker environments
• Improved test coverage for compute and writing tools
• Updated Settings nav to reflect current menu structure

---

v0.2.0

Major rewrite of the mode system, paper writing, processing architecture, and UI routing.

Mode System

• Simplified from Plan/Research/Write to Plan + Execute (two modes)
• Plan mode: ask questions, gather context, create plans. No execution.
• Execute mode: all tools available. Follow the plan.
• Toggle with P/E button, Cmd+B (Plan), or Cmd+E (Execute)
• Amber border for Plan messages, blue border for Execute messages
• Three-layer mode enforcement: system prompt, tool filtering, runtime blocking

Paper Writing

• Writing tool with auto-save to database — survives across workers and restarts
• Paper preview in the Paper tab in the UI
• Client-side export to Markdown and LaTeX
• Outline, sections, and bibliography managed as structured data

Processing Architecture

• Per-conversation processing state — multiple conversations run in parallel
• Background jobs via Celery + Redis — close the browser, come back later
• Redis-based interrupt relay — actually kills running worker tasks, not just a flag check
• Sub-agent streaming — research tool shows nested tool calls in real-time

Settings & UI

• Settings as routed pages (/settings/providers, /settings/agent, /settings/sandbox, /settings/writing) — no longer a modal
• Sticky model selection — persisted per-user in the database
• Onboarding flow — guided setup when no LLM provider is configured
• Route restructure — app served from / instead of /app
• SSE reconnection catch-up — missed events replayed on reconnect
• PLAN.md auto-generated and pinned in resources panel

Testing & CI

• 149 backend tests + 29 frontend tests — comprehensive coverage
• GitHub CI — tests run on push and pull request
• make test runs all tests (backend + frontend + docs build)
• make test-backend, make test-frontend, make test-docs for targeted runs