Switch-SWE tests whether a coding agent, a model running multi-step tasks with tools, can resume unfinished work after handling other requests. Each task keeps its own working copy. The challenge is recovering requirements and progress from a shared conversation.
- The authors pair 200 focus tasks from 19 Python repositories with uninterrupted controls, producing 400 session specifications. Interrupted sessions add work on four other unfinished tasks.
- Across seven model-and-agent configurations, interruptions lower focus-task completion by 11.5 percentage points on average. The gaps range from 6.0 to 15.0 points.
- Their OpenCode plugin selects task-relevant history and records requirements, completed work, uncertainties and next steps. Routing alone raises interrupted completion from 31.5% to 42.5% for DeepSeek-V4.1-Flash, and from 43.0% to 46.5% for GLM-5.3.
- Adding task-state memory raises those rates further, to 47.0% and 48.5%. Auxiliary model requests count toward the same execution budget and usage totals.
The released benchmark and plugin make this a concrete context-management experiment. The results support retaining a task’s specification and observed progress alongside its files.
Caveat: Only the final focus-task patch is scored, so the results do not measure successful completion of all the interleaved work. These are author-reported experiments, not an independent replication.