Skip to content

CHANGELOG.md

Latest commit

 

History

History
241 lines (200 loc) · 18.9 KB

CHANGELOG.md

File metadata and controls

241 lines (200 loc) · 18.9 KB

Changelog

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

1.5.0 - 2026-03-22

Security

  • Registration disabled — Public registration endpoint (POST /api/auth/register) now returns 403. Only admin-seeded users can access the system. Prevents unauthorized account creation on public-facing deployments.
  • RBAC enforcement on 20 endpoints — Added requireRole middleware to plugins, skills, plugin-registry, sandbox, proxy, tasks, and delegation routes. Viewers can no longer upload plugins, deploy skills, execute sandbox code, restart agents, or dispatch tasks.
  • Content Security Policy enabled — Helmet now enforces a full CSP directive set (default-src 'self', script-src 'self', style-src 'self' 'unsafe-inline', img-src 'self' data: blob:, connect-src 'self', font-src 'self', frame-ancestors 'none').
  • Security headers hardened — Added X-Frame-Options: DENY and Referrer-Policy: strict-origin-when-cross-origin via Helmet.
  • Version info removed from public health endpoints — GET /health on both dashboard and agent no longer returns the version string. Version remains available on the authenticated /health/detailed endpoint.
  • Upload filename sanitization — Skill upload filenames are now sanitized with path.basename() and null byte stripping to prevent path traversal.
  • Upload file size limits — Both skill and plugin upload endpoints now enforce a 50MB per-file size limit via multer.
  • ZIP path traversal hardening — Added path.resolve() prefix validation to skill ZIP extraction, ensuring extracted files cannot escape the destination directory.
  • Shell injection eliminated in agent spawn — Replaced execSync("ln -s ...") with fs.symlinkSync() in agent spawn to prevent command injection via crafted paths.
  • Spawn directory validation — Agent deletion now validates that spawnDir is within ~/.multiclaw/agents/ before executing rmSync, preventing arbitrary directory deletion via database manipulation.
  • Spawned agents bound to localhost — Local agents now bind to 127.0.0.1 instead of 0.0.0.0, ensuring they are only accessible through the dashboard proxy.
  • Sensitive fields stripped from API responses — GET /api/agents no longer returns spawnDir, spawnPid, containerId, or containerImage to prevent infrastructure detail leakage.
  • SSE ticket exchange — SSE connections now use short-lived, single-use tickets (30s TTL) instead of passing the long-lived JWT in URL query parameters. Prevents token leakage in server logs, browser history, and Referer headers.
  • Log sanitization — Tool arguments, cron commands, and HTTP response bodies are no longer logged in plaintext. Admin seed password output is masked.

Changed

  • Documentation version updated to 1.5
  • Plugins section moved from README.md to dedicated PLUGINS.md

1.4.0 - 2026-03-22

Added

  • Task Polling in Dashboard UI — The agent tasks tab now automatically polls for status updates every 2 seconds while any task is in "running" or "queued" state. Tasks transition to their final status (completed/failed) in real time without requiring a manual page refresh. Polling stops automatically once all tasks reach a terminal state.
  • Agent Task ID Tracking — The dashboard now stores the agent-side task_id in the agent_tasks database table (agent_task_id column), enabling future re-checks of task status and better debugging of stuck tasks.
  • Solidity Auditor Plugin — Added the Pashov Audit Group's Solidity security auditor to the plugin registry. Runs 8 specialized agents in parallel (vector scan, math precision, access control, economic security, execution trace, invariant, periphery, first principles) to audit Solidity smart contracts. Installs as a skill-based git plugin from https://github.com/pashov/skills.
  • Orchestrator Error Recovery — When orchestrator task polling fails or times out, the agent_tasks database record is now properly updated to "failed" instead of being left permanently in "running" state.

Fixed

  • Agent tasks sent from the dashboard stayed in "running" status in the UI because the tasks tab only fetched once on mount and once after sending — now polls automatically
  • Dashboard task proxy timed out after 120 seconds, prematurely marking long-running tasks (browser automation, SEO analysis) as failed — increased to 10 minutes
  • Orchestrator pollTaskCompletion had the same 120-second timeout issue — increased to 10 minutes
  • Tool schema generation crash when tools lacked callable handlers — now silently filters invalid tools instead of crashing the entire agentic loop

Changed

  • Task proxy poll interval changed from 1 second to 2 seconds to reduce unnecessary requests
  • AgentTasksTab component uses silent polling (no loading spinner flash) during background refreshes
  • Delete button on tasks now visible for all task statuses (was restricted to queued/running only)

Database

  • New column agent_task_id (text, nullable) added to agent_tasks table

1.3.0 - 2026-03-21

Added

  • Plugin Uninstall Steps — Manifests now support an uninstallSteps field alongside postInstallSteps. When a plugin is removed, the agent runs declared cleanup commands before deleting the plugin folder — unregistering daemons, purging packages, removing data directories, and cleaning up installed CLIs.
  • Uninstall Coverage for All Plugins — Every plugin in the registry now has a complete uninstall procedure:
    • Docker — stops and disables the Docker daemon, purges all docker-ce packages via apt, removes /var/lib/docker, /var/lib/containerd, and /etc/docker
    • Portainer — removes the portainer container, deletes the portainer_data volume, and removes the image
    • Tailscale — disconnects from the Tailscale network (tailscale down), purges the tailscale package, removes /var/lib/tailscale and /etc/tailscale
    • Browser Control — removes Chromium browser files from ~/.cache/ms-playwright/, uninstalls the Playwright Python package
    • Nmap Unleashed — runs pipx uninstall nmapUnleashed to remove the nu CLI
    • Bankr Agent — runs npm uninstall -g @bankr/cli
    • AgentPay SDK — runs the repo's uninstall:cli-launcher script or falls back to npm uninstall -g agentpay
    • AgentPay Skill Pack — removes all skills copied to ~/.claude/skills/ during installation
    • Superpowers — removes all skills copied to ~/.claude/skills/ during installation
  • Per-Agent Environment Isolation — Each spawned agent instance now uses its own .env file under ~/.multiclaw/agents/<name>/.env. The systemd service EnvironmentFile directive points to the per-agent file, so MULTICLAW_PLUGINS_DIR and MULTICLAW_SKILLS_DIR are scoped per agent without any global overrides.
  • Agent Connection Retry — On startup, the agent retries its registration handshake with the dashboard using exponential backoff (0 / 2 / 5 / 10 / 20 seconds). If the dashboard is temporarily unavailable (e.g., restarting), the agent reconnects automatically instead of failing silently.
  • Built-in Plugin Source Directories — hello-plugin, docker, and portainer now have source directories in the agent source tree (multi-claw-agent/plugins/), enabling the activate endpoint to copy them into per-agent plugin directories correctly.
  • Agent-Side Delete Error Surfacing — When a plugin deletion request to an agent returns an error, the dashboard now captures and returns the agent's error message in the response. The UI shows a warning when the plugin was removed from the dashboard registry but the agent reported a problem.

Fixed

  • Git plugins (Superpowers, Nmap Unleashed, etc.) now show as active based on the enabled flag in plugin.json rather than requiring a Python class to be loaded in memory — they were incorrectly showing as inactive after each agent restart
  • Plugin folders in the per-agent plugins/ directory are now fully deleted when a plugin is removed — previously the folder persisted because the agent was reading from the wrong plugins_dir path
  • Manifest uninstall steps for built-in plugins (non-git) are now executed on deletion — the built-in delete path previously skipped manifest cleanup and jumped directly to shutil.rmtree

Changed

  • PluginManifest dataclass gains uninstall_steps: list[PluginPostStep] field
  • parse_manifest() reads uninstallSteps array from manifest JSON
  • ManifestRunner gains run_uninstall_steps() method, mirroring run_post_install_steps()
  • GitPluginManager.uninstall() runs manifest uninstall steps (if declared) before deleting the plugin directory; falls back to uninstall.sh only when no manifest steps are defined
  • uninstall_plugin API endpoint reads the stored manifest from plugin.json and runs uninstall_steps before calling shutil.rmtree
  • uninstall.sh fallback now inherits the real HOME and PATH from the process environment (was using a hardcoded restricted env with HOME set to the repo directory)

1.2.0 - 2026-03-21

Added

  • Plugin Manifest System — Every plugin in the registry now declares its environment variables, dependencies, post-install steps, and health checks in a structured manifest. This is the standard for all current and future plugins.
  • Environment Variable Prompting — Deploying a plugin that requires configuration (API keys, network settings, auth tokens) now shows a configuration modal before deployment. Required fields are validated, optional fields are collapsible, and secret values are masked.
  • Auto-Generate Support — Plugins like Bankr Agent and AgentPay SDK can auto-generate secrets, wallets, and API keys during installation. The UI offers an "Auto-generate" toggle for supported variables.
  • Plugin Dependency Validation — The deploy flow checks that all plugin dependencies are installed on the target agent before proceeding. Portainer requires Docker, AgentPay Skill Pack requires AgentPay SDK, etc. Missing dependencies are shown with clear error messages.
  • Plugin Health Check Endpoints — New GET /api/plugins/{slug}/health endpoint on each agent runs manifest-defined health checks (command, HTTP, Python import, file existence). Dashboard proxies health checks via GET /api/plugin-registry/:id/health/:agentId.
  • Health Indicators in UI — The Plugins tab now shows a Health column with green/red status dots. Click to run health checks and see detailed results for failed checks.
  • Manifest-Driven Post-Install — When a manifest is provided, the agent runs specific post-install steps instead of generic auto-detection. Steps run sequentially with per-step timeouts, progress tracking, and error reporting.
  • Per-Plugin Setup Scripts — Docker, Portainer, Browser Control (Playwright), and Tailscale have dedicated setup scripts for system-level installation with multi-distro support.
  • 10 Complete Plugin Manifests — Superpowers, Shannon, AgentPay SDK, AgentPay Skill Pack, Bankr Agent, Docker, Portainer, Browser Control, Tailscale, and Hello Plugin all have full manifests with env vars, dependencies, post-deploy steps, and health checks.

Changed

  • Plugin deploy timeout increased from 60s to 600s for manifest-driven installs (system packages like Docker can take minutes)
  • Plugin registry API responses now include parsed manifest objects
  • Deploy route accepts envVars alongside agentIds and forwards both manifest and env vars to agents
  • agentRegistryPlugins status enum expanded with "configuring" and "installing" states
  • Undeploy and update routes now use resolveAgentUrl() for Tailscale compatibility (was using agent.url directly)
  • Seed registry rewritten with upsertPlugin() helper that backfills manifests on existing entries

Security

  • SSL verification only disabled for localhost health checks (self-signed certs), not globally
  • Python import paths validated against ^[a-zA-Z_][a-zA-Z0-9_.]*$ to prevent code injection
  • File existence checks use expanduser() for tilde expansion
  • Environment variable values injected via subprocess env (not shell interpolation) to prevent injection
  • Env var values never logged — only key names are logged

Database

  • New migration 0016_plugin_manifests.sql: adds manifest text column to plugin_registry table

1.1.0 - 2026-03-21

Added

  • Dashboard Brain — The dashboard now has its own LLM-powered administrator profile that answers queries directly when no agents are @mentioned. Customizable personality in Settings with auto-injected live system state (agent list, statuses, task counts).
  • @Mention Agent Tagging — Replaced agent selection buttons with Slack-style @mention autocomplete in the task prompt. Type @ to see a dropdown of agents with status indicators. Agents appear as inline blue pills and in a tag bar above the editor.
  • Parallel Task Execution — When multiple agents are @mentioned, tasks dispatch to all agents simultaneously (not sequentially). Each agent works independently and results stream in real-time as they complete.
  • Dashboard Synthesis — After all parallel agents complete, the dashboard automatically synthesizes a unified summary of all results, noting agreements, contradictions, and gaps.
  • Centralized Memory Integration — Relevant knowledge base entries are injected into agent prompts before dispatch. Agent results and dashboard syntheses are written back to the knowledge base for future reference.
  • Orchestrations Database Table — Task orchestrations are now persisted to SQLite (previously in-memory only), enabling task history queries and surviving server restarts.
  • Dashboard Profile Settings — New "Dashboard Profile" section in Settings page with custom instructions textarea and live context preview showing auto-injected system state.
  • Keyword Fallback Search — Memory context search falls back to keyword matching when OpenAI embeddings are unavailable, ensuring memory injection works with only an Anthropic API key.
  • New SSE/WebSocket events: dashboard_answer_start, dashboard_answer, synthesis_start, synthesis_complete, synthesis_error

Changed

  • Task dispatch UI completely redesigned — contenteditable Lexical editor replaces plain textarea
  • Send button dynamically shows "Ask Dashboard" (no tags) or "Dispatch to N agents" (with tags)
  • Sequential chain execution replaced by parallel dispatch as default mode
  • executeChain() preserved as deprecated for programmatic API use

Dependencies

  • Added @anthropic-ai/sdk (server — dashboard LLM calls)
  • Added lexical, @lexical/react, @lexical/plain-text (client — @mention editor)

1.0.1 - 2026-03-21

Added

  • Agent authentication for memory API endpoints (X-API-Key alongside JWT)
  • requireAuthOrAgent middleware for dual auth support
  • agent role in RBAC with task execution permissions
  • Python memory client (multi-claw-agent/src/memory.py) for shared state and knowledge
  • Auto-creation of management API keys when agents are registered, spawned, or docker-spawned
  • Auto-generated admin password on first startup when ADMIN_PASSWORD is not set
  • Improved agent stop/start with port-based process detection fallback

Fixed

  • Memory API endpoints blocked by delegation router middleware ordering
  • hasPermission return type coerced to boolean
  • Dashboard default port changed from 3000 to 3100 to avoid conflicts
  • Registration no longer grants first user admin role (admin seeded via seed-admin)
  • Agent stop button visibility for externally-started agents
  • Settings link moved to top-right nav for consistency

Changed

  • Default admin email set to admin@multiclaw.dev when not configured
  • Documentation updated for admin account seeding and CORS defaults
  • Tests directory excluded from git tracking

1.0.0 - 2026-03-20

Added

Core Platform

  • Distributed AI agent management platform with React dashboard and Python agents
  • Authentication system with register, login, JWT, and role-based access
  • Agent CRUD with health monitoring and real-time status updates
  • Task runner with queue management and REST API
  • Agent brain with Anthropic SDK integration
  • Agent-to-agent message relay through dashboard hub
  • SSE streaming for agent status and task updates
  • WebSocket support with JWT auth, reconnection, and dual-broadcast (SSE+WS)
  • Database schema with users, agents, tasks, skills, and plugins

Dashboard

  • React frontend with Tailwind CSS, routing, and API client
  • Login page with auth hook
  • Agent cards with real-time status, spawn UI, and overview controls
  • Agent detail page with per-agent Identity tab and system prompt management
  • Templates page with CRUD, import/export, and spawn integration
  • Workflows page with visual editor (React Flow) and run view
  • Delegations page with permissions matrix and history
  • Memory page with state browser and knowledge search
  • Audit Log admin page with filtering, pagination, and CSV export
  • Settings page with Allowed Origins management and Docker sandbox status
  • Error boundary, toast hooks, and SSE reconnect

Agent System

  • Plugin manager with install, load, activate lifecycle
  • Plugin REST API with install, activate, deactivate
  • Skill manager with install, execute, uninstall
  • Skill library with upload and deploy-to-agent
  • Plugin library with upload and deploy-to-agent
  • Browser Control plugin with Playwright-based automation engine
  • 19 agent-facing browser tools (navigation, forms, screenshots, etc.)
  • Cron job scheduling with APScheduler integration

Backend Services

  • Workflow DAG execution engine with conditional expression evaluator
  • Docker container agent spawn with lifecycle management (start/stop/logs)
  • Docker sandbox runner with resource limits and timeout
  • Delegation API with permissions and peer tokens
  • Knowledge base API with embedding search
  • Shared state key-value API with optimistic concurrency
  • Audit logger with admin API routes and CSV export
  • Dynamic CORS module with state management

Infrastructure

  • Interactive installer script for dashboard and agent setup
  • Automated dependency checking and installation
  • Complete uninstall option
  • HTTPS/TLS support for dashboard and agent
  • Docker agent base image (Dockerfile.agent)
  • Dashboard Dockerfile
  • manage.py CLI for service management

Documentation

  • README with technology, provider, and platform badges
  • SECURITY.md with comprehensive security documentation
  • DOCUMENTATION.md
  • In-app Help page
  • Roadmap design spec for 10 feature phases