-
Notifications
You must be signed in to change notification settings - Fork 1
Features
Nebulus Atom includes 30+ features organized into core capabilities, developer tools, advanced features, and Overlord meta-orchestration.
Pin files to the active conversation so the agent always has them in context. Supports adding, removing, and listing pinned files.
Automatically creates checkpoints before risky file operations (writes, edits, command execution). Restore any checkpoint to undo changes.
Persistent, reusable capabilities that the agent can learn and execute. Skills are saved to disk and available across sessions.
Uses ChromaDB and sentence-transformers to build embeddings of your codebase. Search for code by meaning, not just text patterns.
Run the agent in autonomous mode where it executes multiple turns without waiting for confirmation. Includes safety limits.
Test-driven development automation. Give the agent a goal and it cycles through: write test, run test (fail), write code, run test (pass), refactor.
AST-based code analysis that builds a map of your codebase structure: classes, functions, imports, and their relationships.
Generate and save shell macros from natural language descriptions. Macros are stored for reuse.
Run shell commands in a sandboxed environment that restricts file system access to safe directories.
Model Context Protocol integration for connecting to external tool servers and expanding the agent's capabilities.
Automatic logging of session activities, decisions, and outcomes for review and learning.
"System 2" thinking that triggers deeper reasoning for complex tasks. The agent recognizes when a problem needs more careful analysis.
The agent learns your preferences over time: coding style, naming conventions, tool preferences. Stored per-session.
When requirements are ambiguous, the agent asks clarifying questions before proceeding rather than making assumptions.
Support for image and other non-text inputs when the underlying model supports it.
Structured task tracking with plans, subtasks, and progress monitoring.
Visualize task dependencies and progress in the TUI dashboard.
Structured logging with configurable levels and output formats (console, JSON, file).
Streamlit-based dashboard showing tool usage, response times, token consumption, and session history.
Tools for interacting with GitHub: reading issues, creating PRs, managing labels.
Multi-repo discovery via overlord.yml with dependency tracking, topological ordering (Kahn's algorithm), and circular dependency detection. Health scanner monitors git state, test health, and branch status across all projects.
DAG traversal for upstream/downstream analysis, blast radius calculation, affected project sets, and ASCII tree rendering. Used by dispatch and autonomy engines to understand change impact.
SQLite-backed observation store for remembering patterns, gotchas, and decisions across projects. Supports search by keyword/project/category, confidence tagging, and automatic pruning.
Confidence-based auto-dispatch with cautious/proactive/scheduled levels. Scores changes on scope, test health, and observation history. Per-project overrides and pre-approved action lists.
Executes tasks across projects in dependency order with rollback support. TaskParser converts natural language commands into DispatchPlans with scope analysis.
Multi-project Slack commands via @atom mentions: status, scan, merge, release, autonomy, memory, approve, deny, help. Bridges async Slack handlers to sync Phase 2 modules via asyncio.to_thread().
Approval workflow with Slack thread binding. Proposals flow through: pending → approved → executing → completed (or denied/expired). Thread replies control execution. SQLite-backed persistent store with auto-expiry.
Persistent process with croniter-based scheduling. Runs hourly health scans, nightly test sweeps, and weekly stale branch detection. Graceful shutdown via SIGINT/SIGTERM. Automatic Slack bot connection.
Three built-in detectors run during scheduled scans:
- StaleBranchDetector — branches with no activity for N days
- AheadOfMainDetector — develop branches with unmerged commits
- FailingTestDetector — projects with test failures or missing tests
Results are filtered by autonomy level and routed to Slack or the proposal system.
Dual-mode alert routing: urgent notifications post immediately to Slack, while routine events accumulate into a daily digest. Categories tracked: detections, proposals created/approved/denied, executions, health checks, test sweeps.
Detailed specifications for each feature are available in the repository at docs/features/. Each feature document includes:
- Overview and motivation
- Requirements
- Technical implementation details
- Verification plan
- Architecture - How features are organized
- CLI Reference - How to use features
- Nebulus Swarm - Multi-agent features