Post

How Swades Works? 1️⃣ Dynamic Decomposing When you issue a task, Swades classifies its complexity. If it's high, it decomposes the prompt into independent tasks, spawning parallel subagents running inside isolated git worktree sandboxes. No overlap, no code pollution. 2️⃣ Automated Git Merging Once parallel subagents complete their code changes, Swades gathers the patches and runs a 3-way merge engine to merge changes back to main, automatically spawning resolution agents if conflicts arise. 3️⃣ Desktop Control & Supervision Beyond terminal commands, Swades features a native GNOME Wayland CUA (Computer Use Agent) for visual browser/GUI automation, all supervised by a 24/7 "Director AI" loop that reviews agent states and self-corrects until done. Standard Mode runs a ReAct (Reasoning/Action) loop with real-time stream parsing. At boot, it ingests .agent_index.json (a pre-indexed AST map of classes and imports) to minimize workspace token overhead. When writing files, it executes static syntax guardrails (matching bracket pairs, running node --check or python linting) before writing to disk. If validation fails, it pipes errors back to the context, forcing the model to self-correct before executing terminal actions. When task complexity is classified as HIGH, Orchestrated Mode triggers: Sandboxing: Spawns concurrent tasks in separate git worktree workspaces to prevent state pollution. 3-Way Merging: Merges patches using git apply --3way. If a merge conflict is detected, it programmatically spawns a dedicated merge-resolution agent. Simulation: Spawns a final sandbox simulation thread (runSimulated) to run test suites and verify workspace integration. Designed for raw output speed and lower token overhead. It follows the orchestrated workflow (complexity check -> subtask decomposition -> parallel git worktrees -> 3-way merge) but bypasses the post-merge simulation. Diffs are written directly to the real workspace upon clean merge verification, serving as a fast, multi-file code generator. A state-machine supervisor loop. A worker agent executes the first phase of the task. Upon completion, a secondary "Director" model inspects the full git diff, files, and terminal execution logs. The Director updates task.md, writes the prompt for the next phase, and spawns a new worker session. This worker-director cycle loops continuously until the Director verifies the target goal matches 100%. Built for Wayland GUI environments where legacy X11 automation (like xdotool) fails. A Python helper (cua_helper.py) communicates over OS D-Bus interfaces using GLib variants to inject native keystrokes and cursor clicks into the GNOME Shell. It captures real-time screenshots for visual models, backed by strict Click Recurrence Guardrails in JS that block duplicate coordinate requests to break visual feedback loops. By decoupling the CLI workspace, sandboxing parallel tasks in Git worktrees, communicating via native D-Bus layers, and using a Director-Worker supervisor hierarchy, Swades Agent handles everything from localized edits to entire system architectures. 🔗 Explore the code: github.com/Electroiscoding/Swades-Agent