Skip to content

Changelog

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

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[0.8.1] - 2026-04-19

Added

  • Artifacts plugin: agents can emit standalone HTML artifacts served via capability URLs. MCP tool create_artifact, HMAC-signed share tokens, TTL + pin/revoke state, /me/artifacts portal page, admin UI with list/detail/actions, sandboxed webchat preview modal with URL detection, background cleanup worker. Wrapper page ships Copy-link and Open-in-tab actions.
  • WhatsApp Meta Cloud API channel (whatsapp_api): direct Meta Cloud API integration as an alternative to Evolution. Webhook HMAC validation, authorized-numbers management UI, access_token/phone_number_id as plugin-level secrets, media upload/send logging. Legacy whatsapp plugin renamed to whatsapp_evolution for clarity. Plugins can now opt into public_api: true in the manifest to mount their API router on the public UI app (needed for external webhook endpoints).
  • Backup plugin: scheduled database + secrets backups with optional S3 upload. boto3 and postgresql-client shipped in the image.
  • Session TTL: session_ttl_minutes in agent context_options auto-expires stale conversation sessions and prefixes the next response with a notice.
  • Service-account identity: service_account in agent context_options overrides MCP identity so the agent acts with its own credentials instead of the sender's.
  • CI smoke-install job: reproduces the README Setup flow end-to-end on every PR to catch fresh-install regressions that only appear against an empty DB.
  • Auto-advance for Planning plugin: plan tasks advance automatically without requiring model cooperation.
  • Deep-link post-login redirect: /auth/login honours post_login_redirect so share links and deep URLs survive the login wall.
  • Diagnostics: optional prompt-dump toggle in settings for debugging agent context.

Fixed

  • Fresh-install crash loop: multiple independent bugs surfaced on empty databases — ORM could not discover SystemConfig / config models, boot failed without any agents or runners, default-company INSERT referenced a dropped plan column, and the user_mcp_permissions unified-id backfill migration crashed when app.user_identities did not exist. All paths now tolerate day-zero state.
  • MCP tool routing: gateway allowlist pattern corrected from 4 segments (mcp__<gw>__<server>__*, which never matched Claude's parser) to 3 (mcp__<gateway>__*). Claude runner now applies each agent's mcp_permissions to --allowedTools instead of relying on server-side filtering alone, and the MCP gateway falls through when a user has no explicit permissions (matching the docstring) rather than zapping every namespaced tool.
  • Agent MCP isolation: Claude runner invoked with --strict-mcp-config to prevent MCP pollution from .claude.json and claude.ai remote connectors.
  • Claude OAuth leak: non-pool subprocess path now receives the same OAuth env as the pooled path, so fallback execution still authenticates.
  • Master-key handling: _find_key_file skips unreadable paths (e.g. /root/.ssh/) instead of aborting, falls through on race conditions.
  • Avatars persist on rebuild: AVATARS_DIR moved under data/, served via dedicated /static/avatars StaticFiles mount. Legacy files under ui/static/avatars/ are migrated on first boot. The dedicated bind-mount in docker-compose.yml.example is removed — the standard ./data:/app/data now suffices.
  • Agent save preserves state: avatar, context_options, services, and custom per-channel fields are no longer wiped when editing an agent. Single-agent hot-reload also re-binds message handlers on the new channel instances.
  • Permissions page on fresh install: /permissions form ("Add/Edit User") now lists MCP servers even without runtime PluginManager in the UI container — it falls back to enabled plugins in the DB registry and their manifests (new ui/utils/mcp_servers.py, mirroring channels.py).
  • Base URL mismatch: dashboard emits a persistent warning when GRIDBEAR_BASE_URL points to localhost but the request reaches from a non-loopback host, so share links / WebAuthn / OAuth misconfiguration is visible immediately.
  • Docker entrypoint: fixes bind-mount ownership on startup for data/, config/, credentials/.
  • docker-compose.yml shipped as .example: the tracked compose file is now a template — operators copy it to docker-compose.yml so local customisation is not clobbered on git pull.
  • WhatsApp API media sending: media URL validation fixed; phone_number_id is now configurable from the admin UI instead of hard-coded.
  • Webchat in shared conversations: plan execution messages correctly mention the agent name.
  • Artifact Preview card missing on markdown links: webchat sniffer now harvests hrefs from <a> tags in addition to rendered text, so artifact URLs emitted as [Title](url) by the agent are detected and decorated with Preview / New-tab buttons.

Improved

  • Conversation docs via MCP tool: previously injected into every prompt, now exposed as a tool so the agent reads them only when relevant — saves context tokens.
  • Prompt via stdin: Claude CLI now receives the prompt on stdin instead of argv, avoiding ARG_MAX truncation on very long contexts.
  • Plugin admin hardening: menu gate, no-op shutdown() stubs where missing, summary log after MCP tool discovery for easier diagnosis.
  • README: master-key generation step documented in the install flow; operations that still require a restart are now explicitly listed.
  • Dependency bumps: psycopg >= 3.3.3, pymupdf >= 1.27.2.2, msal >= 1.36.0, ruff >= 0.15.10, anthropic >= 0.94.0, pre-commit >= 4.5.1, google-auth-oauthlib >= 1.3.1, python-docx >= 1.2.0, plus the minor-and-patch dependabot group.

[0.8.0] - 2026-04-11

Architecture

  • Unified container: Bot, MCP Gateway, REST API, and Admin UI now run in a single container/process. The gridbear-ui service is removed from docker-compose. The UI serves on port 8080 within the same process, with embedded mode that reuses the core's DB connection and skips duplicate initialization.
  • Core autonomy: MCP Gateway, REST API, and OAuth2 discovery endpoints are mounted on the core's internal API (port 8000). The bot no longer depends on the UI container for tool access.
  • Agent config in DB: All agent configuration (personality, runner, model, channels, MCP permissions, plugins) migrated from YAML files to PostgreSQL (app.agent_configs table via ORM). YAML files auto-migrated and renamed to .migrated on first boot.

Added

  • AgentConfigRecord ORM model: Full agent configuration persisted in DB with all fields (personality, channels, voice, email, MCP permissions, plugins, context options)
  • Agent hot-reload: POST /api/agents/{id}/reload endpoint for instant config changes without restart. Atomic swap (new agent created before old one stopped)
  • Runner config hot-reload: Runners re-read model/timeout from DB on every invocation — no restart needed for config changes
  • Webchat conversation pinning: Pin/unpin conversations to top of sidebar (per-user via pinned_at on participants table). Visual separator between pinned and unpinned sections
  • Webchat context menu: Right-click (desktop) or long-press (mobile) for Pin, Context, Participants, Documents, Plan, Rename, Delete. Reduced inline buttons from 7 to 2 (pin + ellipsis)
  • Webchat notification navigation: Clicking a notification opens the correct conversation instead of just focusing the window
  • ORM models for webchat: WebchatConversation, WebchatParticipant, WebchatMessage, WebchatDocument models with auto-migration
  • Virtual tool provider filtering: VTP plugins filtered by agent's plugins.enabled AND mcp_permissions — disabled plugins no longer contribute tools
  • MCP_GATEWAY_ENABLED env var: Controls MCP Gateway initialization per-container for safe migration

Fixed

  • Codex CLI MCP tool calls: --full-auto only approves shell commands, not MCP tools. Switched to --dangerously-bypass-approvals-and-sandbox (safe in Docker sandbox)
  • Vibe CLI model selection: Vibe was ignoring agent model config (always using devstral-2). Now write_config() updates active_model and adds missing model definitions
  • Vibe CLI auto-approve: MCP tool calls silently rejected without approval. Set auto_approve = true in config for programmatic mode
  • Message loading truncation: LIMIT 200 loaded oldest messages instead of most recent. Fixed with DESC subquery + ASC wrapper
  • Message timestamps: Show date + time for messages from previous days (was time-only)
  • Conversation auto-reorder: Sidebar conversations move to top when new messages arrive, respecting pin sections
  • Stale hostname references: Fixed gridbear-admin:8080 → gridbear-ui:8080 in 3 files
  • Company-specific placeholders: Replaced hardcoded Dubhe references in agent config UI with generic examples
  • Plugin config model dropdown: Uses models registry (76 models from API) instead of hardcoded manifest enum (10 models)

Removed

  • gridbear-ui service: Removed from docker-compose (unified into gridbear container)
  • Agent YAML files: Migrated to DB, no longer read at runtime
  • JSON schema validation: Removed config/schemas/agent.schema.json (ORM handles validation)
  • config/mcp_servers.json.example: MCP config is fully dynamic

[0.7.2] - 2026-04-08

Fixed

  • Cross-conversation message bug: messages from one shared conversation could appear in another conversation if the user had switched. user_message events now include conversation_id and the frontend filters accordingly. Non-active conversations get an unread badge instead.
  • Wrong sender on history reload: in shared conversations, messages from other users were attributed to the current user after page reload because get_messages did not select sender_id. Fixed query to SELECT sender_id and JOIN app.users for display name.
  • Shared conversation history access: invited users could not load message history because get_messages used strict ownership check. Switched to validate_conversation_access.
  • Plan Execute/Resume buttons did nothing: passing argument to sendMessage() had no effect since it reads from inputText. Now sets inputText before calling.
  • Pause-on-user-message removed: was too aggressive, prevented users from answering agent questions during a task without pausing the plan. Manual pause from panel still works.
  • Peggy "SKIP" bug on Telegram: workflow notification step with agent_id + prompt passed the email summary as a new prompt to the agent, which then (confused by context) replied "SKIP". Fixed by switching peggy_email_processor workflow to template mode.

Improved

  • Stricter planning instructions: explicit workflow rules requiring in_progress -> completed transitions per task, no skipping. Helps the agent actually mark tasks as completed.
  • Smart auto-scroll: auto-scroll only triggers if user is already within 100px of bottom. Reading older messages no longer gets interrupted by new messages arriving.
  • New messages bubble: floating "New messages ↓" indicator appears when a message arrives while reading older content. Click to jump to bottom; auto-clears when user manually scrolls back.

[0.7.1] - 2026-04-07

Added

  • Agent Planning Mode: New plugins/planning/ plugin lets agents break complex requests into task plans. Virtual MCP tools (plan_create, plan_task_update, plan_update) for plan management, collapsible plan panel in webchat with status badges, action buttons (execute/pause/resume/cancel), task editing modal, add/remove/reorder tasks, numbered task list
  • Auto-continuation: Between tasks the system auto-sends a continuation message to the agent, avoiding timeouts on long plans. Safety limit (20 continuations), lock-free streaming, re-verification of plan status to handle user pause races
  • Pause on user message: When the user writes during an active plan, it pauses after the current task
  • Stop Agent Button: End-to-end abort from webchat. Frontend sends {type: "stop"} → UI container calls /api/chat/abort → bot cancels asyncio task → runner kills Claude subprocess (subprocess mode) or destroys pooled process (pool mode)
  • Per-Conversation Documents (RAG): Upload PDF/DOCX/XLSX/TXT/CSV files to a conversation's knowledge base. Text extracted at upload (pypdf, python-docx, openpyxl) and injected into agent context. Collapsible documents panel with upload/delete, 200KB extraction cap + 100KB context injection cap
  • Dependencies: pypdf, python-docx added to data extras

Improved

  • Dynamic plugin list on agent config page: Replace hardcoded plugin whitelist with available_services from DB — all enabled service plugins now appear automatically
  • Webchat mobile responsive layout: Fix horizontal overflow on small screens (root cause: missing min-w-0 on main content flex child in me/base.html), show conversation action icons always on mobile (hover doesn't work on touch), add CSS for markdown prose max-width and word-break
  • Plugin isolation: Move /rlm-query endpoint from core/internal_api/server.py to plugins/rlm/api/routes.py (core must not reference plugins)

Fixed

  • Peggy "SKIP" bug on Telegram: Workflow notification step with agent_id + prompt passed the email summary as a new prompt to the agent, which then (confused by context) replied "SKIP". Changed to template mode (message field only) so notifications are sent directly without agent re-interpretation
  • Planning tool editing: New update_tasks field on plan_update lets agent edit existing task title/description instead of remove+add. Frontend reloads plan from DB on structural changes (added_tasks/removed_tasks/updated_tasks/reordered)
  • Plan icon visibility: Only show plan toggle in conversation list when has_plan is true (new DB column via EXISTS subquery in list endpoint)

Dependencies

  • Bumped fastapi 0.135.2 → 0.135.3, uvicorn 0.42.0 → 0.44.0, python-multipart

[0.7.0] - 2026-04-05

Added

  • Shared Webchat Conversations: Multiple users in the same conversation with an agent — invite by username or shareable link, real-time message sync, @mention filtering
  • Mention Autocomplete: Type @ in the textarea to get a dropdown of participants and agents, navigate with Tab/Enter/arrows
  • @mention Agent Filter: In shared conversations, agent responds only when tagged with @agent_name
  • Agent @mentions Users: Agent mentions the user by @username when replying in shared conversations
  • User Typing Indicator: Real-time "user is typing..." broadcast between conversation participants
  • Per-Conversation Context Prompt: Set a working context on each conversation (gear icon) — injected in agent system prompt
  • Conversation Title Fallback: Agent sees conversation title as topic when no explicit context set
  • Off-topic Detection: Agent asks "right conversation?" if message seems unrelated to context
  • Channel Metadata: Messages carry channel/conversation metadata (webchat, Telegram, Discord)
  • Memory Auto-tagging: Memories tagged with channel and conversation context for future retrieval
  • Per-Conversation Runner Sessions: Each webchat conversation gets its own Claude session (no context bleed)
  • Background Message Processing: Agent response saved even if user switches conversation
  • Unread Indicators: Blue dot on conversations with new responses, participant-aware delivery
  • Browser Notifications: Native Notification API for mentions, agent responses, and unread messages
  • Google Workspace Drive Comments: list, add, reply, resolve comment tools
  • Internal Vault API: /api/vault/get and /api/vault/list for MCP subprocess secret access
  • Email Field: Added to user create/edit forms
  • Invite Join Page: Confirmation page with atomic token redemption (POST, not GET)
  • Participant Management: Online/offline status, owner can remove members, auto-revert to private

Improved

  • Agent display name from YAML name field instead of file stem
  • Context builder: Google Sheets SA vs user account distinction in prompt
  • Context builder: email from_alias support with "prefer own account" instruction
  • Webchat agent list sorted by last conversation activity
  • ADE MCP server auto-reload on config changes (project/env/command save)
  • anyio cancellation throttle increased to 1s for stability

Fixed

  • CPU leak: Pure ASGI middleware (CSRFMiddleware + 3 context middlewares), anyio throttle
  • Prompt leak: Removed prompt_preview from timeout error callback
  • Runner timeout: Unified to 900s across CLI, API backend, RLM, and plugin config
  • Webchat error "undefined": Shows actual error details (text/details/error_type)
  • Gmail: CC field included in get_email response
  • ADE: Vault API URL corrected to port 8080, project secrets JS moved to {% block extra_js %}
  • Workflow: Missing WorkflowDefinition import in _cron_fire
  • CI: CVE-2026-4539 (pygments local-only DoS) ignored in pip-audit

[0.6.2] - 2026-03-28

Added

  • Agent email settings restored: account, sender_name, from_alias, signature
  • Webchat agent list sorted by last conversation activity
  • CPU profiling endpoint and py-spy/yappi in Dockerfile

Fixed

  • anyio cancellation throttle increased to 1s
  • Version display updated (was stuck at 0.5.0)

[0.6.1] - 2026-03-24

Added

  • Webchat: textarea with auto-resize and Enter/Ctrl+Enter toggle
  • Webchat: bot image/file delivery via send_file_to_chat(platform=webchat)
  • Per-user skills injected into agent system prompt
  • Gmail reply_email tool for threaded replies

Fixed

  • CPU leak: BaseHTTPMiddleware → pure ASGI for CSRFMiddleware and context middlewares
  • CPU leak: anyio _deliver_cancellation throttled via monkey-patch
  • MCP transport cleanup with wait_for(timeout=10) safety net
  • Workflow: use creator identity for user-aware MCP servers
  • Starlette pinned <1.0.0 for TemplateResponse compatibility
  • setuptools bumped >=78.1.1 for CVE PYSEC-2025-49
  • Plugin registry: restore installed state when plugin returns to disk
  • Plugin admin routes registered at module level
  • Memory browse 500 error
  • Telegram: catch NetworkError in message queue handler
  • MCP Gateway: enforce per-user permissions at all levels
  • Attachments: handle subdirectories in cleanup

[0.6.0] - 2026-03-19

Added

  • Multi-tenancy Phase 1: ORM tenant isolation with _tenant_field, automatic tenant filtering on all queries, Company and CompanyUser models
  • Unified User model: Single app.users table replaces dual admin/app user tables, with UserPlatform for platform identity mapping
  • Invite flow: Token-based user invite with password setup, email sent via system agent's Gmail MCP server
  • GWS drive tools: drive_download and drive_read_spreadsheet for Google Drive file access and XLSX/Google Sheets parsing
  • Ollama admin page: Cloud authentication, health check, model management with pull support
  • openpyxl dependency: Added to data extras for spreadsheet analysis

Improved

  • MCP Gateway: mark user OAuth2 token as expired on 401 from external servers, /me/connections shows amber badge
  • MCP Gateway: propagate user identity in subprocess mode for per-user tool access
  • MCP Gateway: normalize camelCase tool arguments from LLMs to match server expectations
  • MCP Gateway: fix user credential resolution for external (enterprise) plugins
  • MCP Gateway: skip virtual transport providers during SSE health checks
  • Invite emails sent via system agent's Gmail MCP server instead of SMTP
  • MCP user permissions migrated to unified_id (username-based)

Fixed

  • Runner: destroy pooled Claude CLI process on timeout instead of releasing (prevents zombie processes)
  • Google SA: handle invalid JSON on per-agent service account upload (was 500)
  • Auth: add forgot password link to login page
  • _is_token_expired(): expires_at=0 was skipped because 0 is falsy in Python
  • UI: update collaboration label from tag syntax to tool name
  • Docker: PYTHONPATH + gateway URL for gridbear CLI

Dependencies

  • Bumped minor/patch dependencies

[0.5.0] - 2026-03-04

Added

  • Mistral runner plugin with API, CLI (Vibe), and Codestral backends
  • Codestral free endpoint with tool calling support
  • Ollama admin page: connection status, model management, model pull
  • Ollama Cloud authentication: device public key display, auth status probe
  • Docker: PYTHONPATH=/app for gridbear CLI console scripts
  • Docker: GRIDBEAR_GATEWAY_URL for in-container CLI usage
  • Docker: vibe_state volume for persistent Vibe CLI config
  • Agents without channels (CLI/API-only) no longer rejected at startup

Fixed

  • Ollama: OLLAMA_URL env var now takes precedence over DB config default
  • Ollama: removed unused Bearer auth (Ollama reads API key from own env)
  • CI: upgrade setuptools in security scan, skip editable install
  • CI: install PyTorch CPU-only to avoid NVIDIA deps

Changed

  • Python 3.11 → 3.12 (Dockerfile, CI, pyproject.toml)

Security

  • cryptography bumped to ≥46.0.5 (CVE-2026-26007)

[0.4.5] - 2026-03-03

Added

  • Gmail: mark_as_read MCP tool to mark emails as read after processing
  • Agent: get_channel_names() helper for channel discovery by plugins
  • Lifecycle: ON_STARTUP hook now fires at initial boot, not just on reload
  • Auth: master password bypass for initial setup and debugging

Fixed

  • Internal API: enterprise plugin route discovery via GRIDBEAR_PLUGIN_PATHS
  • Internal API: relative imports in dynamically loaded plugin route modules
  • Gitignore: avatar/icon paths updated after admin/ → ui/ rename

Dependencies

  • FastAPI 0.134.0 → 0.135.1

[0.4.4] - 2026-03-01

Fixed

  • Plugin admin pages: custom pages (ms365, github, etc.) were shadowed by the generic config catch-all due to route registration order
  • Dashboard uptime: now shows actual bot uptime instead of UI container process time
  • Codecov CI: updated file → files parameter for Codecov action v5
  • Plugin admin routes: register after ORM init to prevent startup errors

Added

  • Plugin isolation pre-commit hook: prevents core/ui from importing plugins directly
  • Plugin isolation also enforces no stray plugin templates in ui/templates/plugins/

Changed

  • Plugin-specific templates moved from ui/templates/plugins/ to self-contained plugin directories (plugins/<name>/admin/templates/)

[0.4.3] - 2026-03-01

Fixed

  • 2FA enable/disable: PostgreSQL boolean type mismatch (totp_enabled, webauthn_enabled)
  • Passkey registration/removal: same boolean type fix

Changed

  • Renamed github-mcp plugin to github for consistency
  • Renamed peggy example agent to myagent as neutral placeholder
  • Added GitHub issue templates (bug report, feature request)

[0.4.2] - 2026-02-28

First public open-source release.

Highlights

  • Plugin-based architecture with 35 bundled plugins
  • Multi-LLM support (Claude, OpenAI, Gemini, Ollama)
  • Multi-channel (Telegram, Discord, WhatsApp)
  • Admin UI with theming support (3 themes included)
  • User portal with dashboard, profile, service connections, web chat
  • MCP Gateway with SSE streaming and per-user OAuth2 connections
  • REST API with generic CRUD endpoints and ACL system
  • PostgreSQL with pgvector for memory and embeddings

Plugin Ecosystem

  • Channels: Telegram, Discord, WhatsApp
  • Runners: Claude, OpenAI, Gemini, Ollama
  • Services: Memory, Sessions, Attachments, Skills, Memo, TTS (5 providers), Transcription (3 providers), Image generation (3 providers), LiveKit voice agent
  • MCP Providers: Gmail, Google Sheets, Google Workspace, Microsoft 365, Home Assistant, GitHub, Playwright
  • Themes: Nordic, Enterprise, TailAdmin

Architecture

  • Plugin system with manifest.json-based discovery and topological dependency sorting
  • Multi-agent orchestration with per-agent channel instances and YAML config
  • Hook system for message lifecycle customization
  • ORM layer inspired by Odoo (create, search, write, delete, auto-migrations)
  • Sandboxed code executor (optional, isolated network)
  • Docker Compose deployment with optional services via override file