COSMO / PERSONAL PROJECT · OPEN SOURCE · 2026
Verify before integrating.
An orchestrator for development agents that requires executable evidence before integrating their changes.
Python · OpenSpec · Playwright · Docker · SQLite
Context
An agent’s success report does not prove the software works. Cosmo organizes development so implementations undergo real checks before reaching the integration branch.
Solution
Each task uses an isolated Git worktree. The workflow combines specification, implementation, builds, tests and review; failures may trigger retries or block a task for human review. Events and results are stored in SQLite and an event log.
Cosmo in action
The original README demo shows the console during a run. Open the panel to play it and close it to hide it.
Scope and evidence
The experiment log records that in batch B in September 2026, 21 of 24 tasks completed without manual state intervention and three required it. These findings describe that experiment; they do not establish general autonomy or the impact of later fixes.