Eigen RadarAI
Archive
Analysis

MeshXFS creator says Claude wrote its new shared-storage code

MeshXFS creator Steve Shoecraft says Claude Opus generated the code that lets the project coordinate several computers using the same storage. He says he set the goals without writing or reviewing that new code. Existing XFS code remains the foundation; the account describes an AI coding experiment rather than an independent assessment of its reliability.

Artificial Intelligence··Midday
Cables from three differently sized computers converge on a shared storage enclosure on a workbench, with a network switch in the foreground.

Shared storage starts from existing XFS code

Steve Shoecraft, creator of MeshXFS, says Claude Opus wrote the new code that adds clustered operation to XFS. MeshXFS is a file-system project intended to let multiple computers read and write the same storage concurrently. The existing XFS implementation provides the foundation. Shoecraft's account concerns the added clustering code, while the underlying file system was developed by people.[1], [2]

He describes Claude Opus working autonomously inside Claude Code, a coding-agent tool, and says his own role was to set goals and criteria. He told Phoronix he did not write, edit or review the new code. Shoecraft supplied this account of the development process; an independent code audit has not been published. He says the agent consulted Claude Fable, GPT, Gemini and Astra about lock protocols, concurrent faults during recovery and correctness of fencing. According to him, AI also ran the tests, diagnosed failures and made corrections. The developer set the goals against which that work was carried out.[1], [2]

Locks coordinate computers using one disk

The design describes two distributed lock-manager paths, one using a network and another keeping lock state on shared storage through SCSI COMPARE AND WRITE, a storage command. These locks coordinate access to files and their metadata. Cache coherence is intended to keep the computers' temporary data copies consistent. Fencing through SCSI persistent reservations restricts storage access by a node that encounters trouble. After failure, a surviving node is assigned recovery of the transaction journal. The project also includes tools for creating, checking and resizing file systems.[1], [2]

Integrity and runtime remain developer-set criteria

Shoecraft says the priorities are data integrity, avoiding host lockups and keeping execution time within twice that of native XFS on the same storage. The runtime criterion is Shoecraft’s development target. His figures put the code tree's growth at about 226,000 to 314,000 lines; the count alone does not assess the new code's quality.[1], [2]

The creator describes about 30 tests per release on real multi-node clusters. Individual outcomes, pass criteria and detailed workload measurements have not been supplied. MeshXFS declares GPL-2.0 licensing and describes a functioning clustered file system compatible with POSIX, the portable operating-system interface standard. Shoecraft presents the work as an experiment in AI development of low-level software.[2]

References

  1. News sourcePhoronixMeshXFS developer says Claude wrote new clustering code↩1↩2↩3↩4
  2. News sourceDiarioBitcoinMeshXFS developer attributes clustering code to Claude Opus↩1↩2↩3↩4↩5
AI safety and regulationChip wars and the supply chainThe frontier model race

More analysis from this desk

View all Brief analyses from the Artificial Intelligence desk →