I'm migrating ~150 components from the legacy CMS. I started by planning the work in OpenCode and OpenSpec, but hit a wall. As context accumulated across a long session, things got progressively slower and eventually stalled out entirely.
So over the weekend I built a framework to solve this: Nightshift.
Nightshift is a batch processing system that breaks long-running work into discrete items and delegates each one to specialized subagents. Three agents with strict role separation handle the work:
- A manager that reads the queue, delegates tasks, and tracks state
- A dev agent that executes the steps for each item
- A QA agent that independently verifies the results
Each agent invocation starts with fresh context (no accumulated bloat), and the dev agent self-improves its own step definitions as it goes — meaning it actually gets faster and more reliable as it processes more items, not slower.
I tested it on the component migration and it worked great. It ran independently (while I watched the Olympics), processing items through the queue without intervention. The self-improvement loop meant later items completed faster than earlier ones.
The tool is general-purpose — it's not specific to CMS migrations. Any batch of similar tasks that can be defined as steps with validation criteria can be run as a "shift." If you've got an idea for how to use this on one of your projects, let me know!
I'm migrating ~150 components from the legacy CMS. I started by planning the work in OpenCode and OpenSpec, but hit a wall. As context accumulated across a long session, things got progressively slower and eventually stalled out entirely.
So over the weekend I built a framework to solve this: Nightshift.
Nightshift is a batch processing system that breaks long-running work into discrete items and delegates each one to specialized subagents. Three agents with strict role separation handle the work: - A manager that reads the queue, delegates tasks, and tracks state - A dev agent that executes the steps for each item - A QA agent that independently verifies the results
Each agent invocation starts with fresh context (no accumulated bloat), and the dev agent self-improves its own step definitions as it goes — meaning it actually gets faster and more reliable as it processes more items, not slower.
I tested it on the component migration and it worked great. It ran independently (while I watched the Olympics), processing items through the queue without intervention. The self-improvement loop meant later items completed faster than earlier ones.
The tool is general-purpose — it's not specific to CMS migrations. Any batch of similar tasks that can be defined as steps with validation criteria can be run as a "shift." If you've got an idea for how to use this on one of your projects, let me know!