Skip to main content
ArchWiz provides the autonomous execution layer of the Termux Monorepo. This guide shows you how to configure task variables, launch an autonomous runner, dispatch sandboxed tasks, monitor live execution, and recover from failures automatically.
1

Set task environment variables

Before starting any autonomous work, export the required variables so the runner knows what to target.
The runner reads tasks from ~/workspace/llm_map/master_tasks.json. Ensure your task entry exists there before you begin.
2

Start the autonomous runner

Launch the main loop that picks tasks from the registry and drives them to completion.
Expected output:
3

Dispatch a sandboxed task

Use dispatch_task.py to execute a single unit of work in isolation. This is the preferred way to run changes that need rollback protection.
Expected output:
4

Monitor with mirror

Open a live mirror of the runner state to watch progress, logs, and agent heartbeats in real time.
Expected output:
5

Auto-repair failures

If a task fails, invoke auto_repair.py to retry with adjusted parameters or fallback strategies.
Expected output:
Tasks must exist in docs/proposals/active/<id>/ITEMS.md before work begins. The runner validates this file at startup and refuses to proceed if the task is not registered.

​
Preferred execution loop

Follow this sequence to keep the monorepo stable and auditable:
  1. Registry: Add or update the task in docs/proposals/registry.yaml
  2. Pick item: Select the next open item from docs/proposals/active/<id>/ITEMS.md
  3. Branch: Create a feature branch for the change
  4. Implement: Run dispatch_task.py or autonomous_runner.py to apply the change
  5. PR: Open a pull request with the branch
  6. Gates green: Run python3 scripts/ci/repo_gate.py and python3 scripts/ci/termux_smoke.py
  7. Merge: Merge only after both gates pass
  8. Update ITEMS.md: Mark the item complete in the active proposal file
Run python3 archwiz/archwiz.py to open the cockpit dashboard and see all pending tasks, gate status, and agent health in one view.
For background on how proposals feed into this loop, see Proposals governance. To inspect what changed after a task runs, read the Forensics guide.