Spine lets one coding assistant organize nested work without creating another assistant for every subtask. A completed check returns a short result to its unfinished parent. The runtime keeps the full record separately and prepares a shorter next prompt.
Jiahong Xiang and colleagues implement this in Codex and Pi, the software that runs models’ tools, files and commands. They compare each variant with its native system while keeping the model fixed.
- With GPT-5.6 Sol on SWE-Milestone, fully resolved milestones rise from fourteen to nineteen of 98. The separate progress score rises from 38.1% to 42.6%.
- Model cost falls from $681.22 to $460.77, a 32.4% reduction. Recursive delegation reaches a similar 42.4% progress score but costs $1,749.73.
- On fifty ProgramBench reconstruction tasks, average test passing rises from 62.55% to 73.35%. Cost also rises, from $210.21 to $221.56.
- Three-project ablations weaken results when nesting is shallow or the model cannot see explicit work boundaries. Runtime state compilation remains enabled in the latter condition.
The 32.4% saving belongs to one task suite. These are author-run comparisons, and the reconstruction exception shows why shorter context and lower total cost need separate checks. An open-model Pi comparison also uses more total tokens while completing more milestones.
The Codex implementation and Pi extension make the design inspectable. Marking a subtask boundary is separate from deciding that another assistant should execute it.