research(tools): DeltaBox-style millisecond COW sandbox checkpoint/rollback (arXiv:2605.22781)

[bug-ops]

Description

A May 2026 arXiv paper, "DeltaBox: Scaling Stateful AI Agents with Millisecond-Level Sandbox Checkpoint/Rollback" (arXiv:2605.22781), presents change-based (copy-on-write) checkpoint/restore for agent sandboxes:

• DeltaFS: filesystem checkpoint/restore via layer freeze+insert, reducing writes to COW; rollback is a layer switch.
• DeltaCR: process-state checkpoint/restore via incremental dumps, rollback by fork() from a frozen template process.
• Reported ~14ms checkpoint / ~5ms rollback — orders of magnitude faster than full-state duplication.

How this applies to Zeph

zeph-tools' sandbox (crates/zeph-tools/src/sandbox/{macos,linux,noop}.rs) has no checkpoint/rollback primitive — every tool execution is a fresh process. No part of the agent loop (single turn, zeph-orchestration DAG, zeph-subagent fan-out) can cheaply "try, roll back, try again." If Zeph ever wants speculative/branching tool-execution exploration, a DeltaBox-style COW primitive is the enabling mechanism.

Distinct from closed #2409 (PASTE speculative predictive execution, arXiv:2603.18897): PASTE predicts-then-verifies a single path; DeltaBox cheaply explores many branches. No existing duplicate confirmed via gh issue list --search "DeltaBox" / "checkpoint rollback sandbox".

Expected Behavior

A documented team decision (adopt now / adopt later / explicitly decline) on whether a COW-based sandbox checkpoint/rollback primitive is worth building for any current or near-term Zeph subsystem (candidates: zeph-orchestration DAG retry, zeph-subagent fan-out).

Actual Behavior

No such primitive exists; not yet evaluated against a concrete Zeph consumer.

Environment

• Version: v0.22.2 (HEAD c0fc51e)

Logs / Evidence

• arXiv:2605.22781 — https://arxiv.org/abs/2605.22781
• crates/zeph-tools/src/sandbox/linux.rs, sandbox/mod.rs

Spec

.local/specs/073-deltabox-sandbox-checkpoint-rollback/spec.md