@hy-sde-org/dsh-fs-archive

A pure-TS multi-format archive engine — zip, tar, tar.gz, rar, 7z, iso, deb, rpm, cpio, cab, arj, asar — plus the codec layer (gzip, bzip2, ncompress LZW, xz, deflate, zstd) behind them. Ported from @oh-my-pi/pi-utils.

This is the durable core of the harness read tool's multi-format support: foo.zip lists an archive's root, foo.zip:dir lists a directory, and foo.zip:dir/file.txt reads one member as text. Published standalone as @hy-sde-org/dsh-fs-archive; it mirrors @deepseek-ai/dsh-fs-archive in the hy-sde deepseek-harness fork.

Format detection sniffs content signatures (sniffArchiveFormat) and falls back to extension inference (archiveFormatFromPath); member reads are bounded by ArchiveLimits (entry count, index size, in-memory size, member size, path bytes, link depth) so attacker-controlled archives cannot drive unbounded allocation.

Port differences

No Bun runtime: Bun.hash.crc32 is table-driven CRC-32, Bun.CryptoHasher is node:crypto SHA-256, and Bun.file / Bun.write are node:fs / node:fs/promises equivalents.
All relative imports carry the .ts extension (repo rewriteRelativeImportExtensions).
formatBytes, the small LRUCache, and the public index.ts surface replace the upstream package-level re-exports; tests load fixtures from one bundled tar.gz (the same layout upstream uses to keep checkout state quiet).
The optional ./invariant companion registers the package into a Cordis host's ctx.invariants service (no-op); it needs @deepseek-ai/cordis + @deepseek-ai/dsh-invariants peers only when you use that entry.

Otherwise the port is algorithmically 1:1 with the original source. Keep the two sources in sync when the fork backports upstream archive-engine fixes.

Known Limitations and Deferred Work

The harness read tool member reads go through ctx.fs.readBytes with a bounded readMaxArchiveBytes (default 256 MiB): the archive file is loaded into memory for indexing. No streaming/local fileByteSource path is wired into the tool yet, so very large archives stay beyond the cap even though the engine itself supports lazy source reads.
The harness read tool serves member text and listings only; archive writing is not exposed through the tool (members are immutable there), though this package's writeArchive API supports zip/tar/tar.gz/asar writing for programmatic use.
Detection needs either content sniffing or an archive extension on the path; a member named with a supported archive extension inside a container is interpreted by longest-prefix rules from parseArchivePathCandidates.

Package API:

import {
  openArchive,
  parseArchivePathCandidates,
  sniffArchiveFormat,
  formatArchiveEntryLines,
} from '@hy-sde-org/dsh-fs-archive'

openArchive(source, options) — open a file path or { bytes, format } into an ArchiveReader; getNode resolves members, listDirectory lists, readFile reads one member's bytes, indexEntries enumerates.
parseArchivePathCandidates(filePath) — split archive.ext:member/path into resolution candidates (longest archive prefix first).
sniffArchiveFormat(bytes) / archiveFormatFromPath(path) — format rules.
formatArchiveEntryLines(entries) — one line per member, name/ for directories and name (size) for files.
See src/index.ts for the full export surface.

@hy-sde-org/dsh-fs-archive

Install

README