fix(orchestration): resolve skill hallucination, brace crashes, and silent turns in swarm control ...

- Implemented strictly-worded 'Tool-belt' in system prompt to prevent hallucinating external skills.
- Replaced fragile .format() with robust .replace() for system prompts to prevent crashes on curly braces.
- Optimized TTS streaming: increased concurrency and reduced chunk sizes to eliminate mid-sentence pauses.
- Fixed 'silent turn' bug in Swarm Control by ensuring fallback content is yielded even if AI only produces reasoning/tools.
- Improved history formatting to handle assistant turns mapping to tool calls without main content.
- Corrected Gemini model fallback to gemini-1.5-flash.

refactor: remove dspy dependency and migrate to orchestration modules

feat: Implement real-time streaming for Swarm Chat and premium Antigravity UI

feat: Remove DSPy dependency and implement Swarm Control ...

- Removed DSPy-based pipelines (code assistant, file selector, question decider).
- Implemented Swarm Control for multi-node command execution and terminal handling.
- Updated gRPC services for batch command processing and improved perception.
- Replaced Coding Assistant page with Swarm Control interface in the UI.
- Updated Protobuf definitions to support interactive terminal sessions.
- Improved node registry and RAG services for better agent performance.

chore: backup local changes before investigation

feat(mesh): robust node connectivity and sync reliability fixes ...

- Fixed Node Registry persistence by adding missing db.commit() for online/offline status updates.
- Resolved 'Seed from Local' sync failures by correctly tracking mesh memberships in the TaskAssistant service.
- Enhanced frontend safety: disabled offline node selection and filtered source node candidates.
- Added 'Receiver Only' toggle for more granular control over multi-node synchronization.
- Corrected false-positive 'Source Node Disconnected' alert by verifying global connection status.
- Improved logging in Hub and Agent Nodes for better debugging of file transfer events.

feat(terminal): overhaul with xterm.js, raw TTY streaming, and dynamic resizing ...

- Replaced simulated terminal with xterm.js for native emulation
- Implemented raw TTY byte streaming over WebSockets
- Added support for persistent shell sessions and window resizing (ioctl TIOCSWINSZ)
- Improved SkillManager with a transparent TTY bypass to prevent thread pool saturation
- Added periodic heartbeat-based latency monitoring (RTT)
- Fixed skill_config parsing crash in AI Hub registry
- Enabled auto-scrolling and terminal-to-PTY grid synchronization

feat(M3+M4): persistent sessions with node attachment + invite token auth ...

M3 — Persistent Session Model:
  DB:
  - Session model: +sync_workspace_id, +attached_node_ids (JSON), +node_sync_status (JSON)
  - Migration: auto-adds 3 new columns to existing sessions table
  API (sessions router):
  - POST /sessions/        now auto-attaches user's default_node_ids from preferences
    on creation, generates sync_workspace_id, seeds 'pending' sync status
  - POST /sessions/{id}/nodes       attach nodes to an existing session
  - DELETE /sessions/{id}/nodes/{node_id}   detach a node
  - GET  /sessions/{id}/nodes       live sync status (merges DB + live registry)
  Schemas: NodeAttachRequest, NodeSyncStatusEntry, SessionNodeStatusResponse

M4 — Invite Token Validation:
  API (nodes router):
  - GET  /nodes/admin/{id}/config.yaml  generate ready-to-deploy agent_config.yaml
    (contains node_id, hub_url, grpc_endpoint, invite_token, skill toggles)
  - POST /nodes/validate-token          internal endpoint called by gRPC server;
    validates token against DB, returns skill_config for SandboxPolicy
  gRPC (grpc_server.py):
  - SyncConfiguration now calls POST /nodes/validate-token before registering node
  - HUB_API_URL env var controls this; empty = dev mode (skip validation)
  - Shell skill disabled by admin → empty allowed_commands in SandboxPolicy
  - Gracefully degrades if Hub is unreachable (logs warning, proceeds in dev)

feat(M1+M2): Agent Node API — admin-owned nodes, group access, DB persistence ...

DB Models:
- AgentNode: admin-owned, display_name, skill_config toggles (shell/browser/sync),
  invite_token, is_active, last_status/last_seen_at persisted on every connect
- NodeGroupAccess: admin grants groups access with access_level (view/use/admin)
- Migration updated for all new columns + node_group_access table

NodeRegistryService (M2 persistence):
- register() → upserts AgentNode in DB (status=online)
- deregister() → marks AgentNode offline in DB, keeps record for UI
- update_stats() → bumps last_seen_at on every heartbeat
- EventBus with 20 event types: per-node + per-user WS fan-out
- Uses get_db_session() context manager (no extra deps)

API (13 routes):
Admin:
  POST   /nodes/admin              — Create node + generate invite_token
  GET    /nodes/admin              — List all nodes (full detail)
  GET    /nodes/admin/{id}         — Node detail incl. invite_token
  PATCH  /nodes/admin/{id}         — Update skill_config toggles, description
  POST   /nodes/admin/{id}/access  — Grant group access
  DELETE /nodes/admin/{id}/access/{group_id} — Revoke group access
User:
  GET    /nodes/                   — List user-accessible nodes (group-filtered)
  GET    /nodes/{id}/status        — Quick probe
  POST   /nodes/{id}/dispatch      — Dispatch task
  PATCH  /nodes/preferences        — Set default_node_ids + data_source config
  GET    /nodes/preferences        — Read node preferences
WebSocket:
  WS     /nodes/{id}/stream        — Single node live execution stream
  WS     /nodes/stream/all         — Multi-node global bus (split-pane UI)

feat(M1): add Agent Node REST + WebSocket API to AI Hub ...

- NodeRegistryService: in-memory live node registry with rich EventBus
  - emit() delivers all execution events to per-node and per-user subscribers
  - EVENT_TYPES: 20 structured event types (task_start, sync_drift, retry, etc.)
- AgentNode DB model: persists node registrations + invite_token + last_status
- DB migration: auto-creates agent_nodes table + adds new columns
- REST API (/nodes):
  GET  /nodes                   - list all user's nodes (live+DB merged)
  GET  /nodes/{node_id}         - full live status
  GET  /nodes/{node_id}/status  - quick online/offline probe
  POST /nodes/{node_id}/dispatch - dispatch shell/browser task to a node
- WebSocket API:
  WS /nodes/{node_id}/stream    - single-node live execution stream
  WS /nodes/stream/all          - multi-node global bus (split-pane UI ready)

Refactor UI consistency, fix layout scrolling, and remove deprecated WSS file access features

feat: track STT/TTS models in session; fix audio progress bar; add token usage bar

chore: cleanup test scripts, secure deployment script, and implement Voice/TTS frontend refinements ...

- Removed all temporary root Python test scripts
- Secured remote_deploy.sh credentials using environment variables
- Added data/audio payloads directory to .gitignore
- Added automatic DB migration script (migrate.py) via uvicorn lifespan
- Applied major refactors to UI Voice Chat, Session handling, and unified React API configs
- Updated models and dependencies natively syncing with Nginx proxy features

feat: expose model info to users, multi-model provider cloning, cleaner profile UI ...

- Show model name under each provider button for all users (ProfilePage)
- Admin Settings: upgraded 'Register New Instance' panel with 4 fields:
  provider type, label suffix (with model-name hint), pre-set model
  dropdown, and 'Inherit API Key From' for same-type instance cloning
- Fix API key inheritance: use _clone_from marker instead of copying
  masked key; backend resolves real key from DB before saving
- Lock cloned-key instances: API key field is greyed out with lock icon
  and 'Inherited from <source>' display; Unlink button breaks the link
- Provider card header: show configured model name instead of
  'User Override / System Provided' labels
- ProfilePage: remove confusing 'Do Extra Work' override panel entirely;
  ServiceSelect is now a clean provider selector with no hidden admin UI
- SettingsPage: remove 'User Override' / 'System Provided' badge

Refactor STT model selection to dropdown, add YAML config import/export, and refine Voice/Coding Assistant UI

feat: session history sidebar with feature isolation and auto-titling ...

Backend changes:
- db/models.py: add feature_name column to Session model
- db/models.py: confirmed is_archived col for soft-delete support
- api/schemas.py: expose feature_name on SessionCreate and Session schemas
- core/services/session.py: accept feature_name in create_session()
- api/routes/sessions.py: pass feature_name on creation; add GET /sessions/
  (filtered by user + feature), DELETE /sessions/{id}, DELETE /sessions/
- core/services/rag.py: auto-title session from first user message;
  keep session.provider_name in sync with the model actually used
- core/services/workspace.py: same auto-title + provider_name sync
  on the WebSocket chat path

Frontend changes:
- services/apiService.js: createSession() now sends feature_name and
  provider_name; add getUserSessions(), deleteSession(), deleteAllSessions()
- services/websocket.js: getSessionId() accepts providerName and forwards
  it to createSession so the DB record reflects the real model
- hooks/useCodeAssistant.js: pass 'coding_assistant' namespace + 'deepseek'
  provider; expose handleSwitchSession + sessionId to page
- hooks/useVoiceChat.js: pass 'voice_chat' namespace + 'gemini' provider;
  expose handleSwitchSession + sessionId to page
- components/SessionSidebar.js: new collapsible sidebar (collapsed by default)
  showing history sessions per feature; click to switch, hover to see token
  usage tooltip, per-card delete and bulk 'Clear All'; shows smart relative
  date and provider badge per session
- components/SessionSidebar.css: theme-aware CSS (inherits Tailwind
  light/dark palette; no forced dark mode)
- pages/CodingAssistantPage.js: mount SessionSidebar with coding_assistant
- pages/VoiceChatPage.js: mount SessionSidebar with voice_chat

Tests:
- tests/core/services/test_session_service.py: new unit tests for
  SessionService covering feature_name namespace variants and DB rollback
- tests/api/routes/test_sessions.py: unit tests for new session mgmt routes
  (list by feature, single delete, bulk delete, token usage)
- integration_tests/test_sessions_api.py: end-to-end tests for feature
  isolation, single delete, 404 on missing session, and bulk delete
  confirming cross-feature sessions are unaffected

enable OIDC

move retriever logic out of rag for better flexiable and introduce file retriever tables

refactor faiss_store

use litellm underlying

further refactor

refactor db creation & centralize the config.py

refactor code

Refactor core structure and add new RAG modules

Add database and model to the project