v0.3.5

Changed

• The complete Simplified Chinese guide is now integrated into the root README.md with a language
  switch. README.zh-CN.md remains as a compatibility pointer instead of a second source of truth.
• npm is now the primary installation path throughout the project and package READMEs; GitHub release
  tarballs remain documented as offline fallbacks.
• The README status and package links reflect the published 0.3.4 registry packages instead of the
  earlier pending-verification state.

v0.3.4

Fixed

• The release workflow is rerunnable. An existing GitHub release is edited instead of recreated, and
  an npm package version that is already present is skipped. This lets maintainers add NPM_TOKEN
  and rerun the release after GitHub artifacts have already been published.

v0.3.3

Fixed

• SHA256SUMS now contains artifact basenames instead of build-directory paths, so downloading the
  three release assets into one directory is enough to run sha256sum -c SHA256SUMS.

v0.3.2

Changed

• The CLI is published as @nowyousee1/agentloom: the unscoped agentloom name on npm belongs to an
  unrelated package. The installed command is still agentloom, and dsh-agentloom stays unscoped.

Added

• Every GitHub release now carries stable agentloom-cli.tgz, dsh-agentloom.tgz and
  SHA256SUMS assets. Installation no longer depends on npm registry availability or credentials.
• scripts/verify.mjs rejects a release workflow that stops attaching those installable artifacts.

v0.3.1

Added

• Release engineering: scripts/verify.mjs (versions, syntax, packed contents, required files,
  documentation links) and scripts/version.mjs (bump every package.json plus
  CODEX_PLUGIN_VERSION in one step).
• .github/workflows/release.yml: tag v* → verify → test → npm publish --provenance for both
  packages → GitHub release notes taken from this file.
• Community files: CODE_OF_CONDUCT.md, issue forms, PR template, Dependabot, .editorconfig,
  .gitattributes, .nvmrc.
• Docs: docs/FAQ.md and docs/PUBLISHING.md; "Where to find it" tables in both READMEs.
• Tests for the Codex marketplace output and for config portability/migration (34 tests total).

Changed

• defaultConfig() now writes memory.dshBudgetBytes: "auto" and portablePaths() normalises path
  separators, so one hub.json round-trips unchanged on every platform.
• migrateLegacyHub(legacyRoot, hubRoot) takes explicit roots, which is what makes the migration
  path testable; the CLI call site is unchanged.
• Marketplace and plugin display names are Agentloom (they still said Agent Hub).

Fixed

• scripts/verify.mjs no longer triggers Node's DEP0190 deprecation warning when invoking
  npm pack on Windows.
• LICENSE is shipped inside both published packages instead of only at the repository root.
• Link handling on macOS: isLink() compared realpath() with the lexical path, so every real
  directory under /var/... (which resolves to /private/var/...) looked like a link. The probe is
  now lstat()-first everywhere and the realpath fallback is Windows-only; ownership, pruning and
  linkStatus() compare paths through new samePath() / isUnder() helpers.

v0.3.0

First public release under the name agentloom (formerly agent-hub).

Added

• Monorepo layout: packages/cli (the agentloom CLI) and packages/dsh-plugin
  (the DeepSeek Harness bridge plugin).
• Shared, both-sides writable memory store at ~/.agentloom/memory/shared.md
  with metadata (kind, scope, tags, timestamps) and update-by-title semantics.
• agentloom skill import --from=codex to pull existing Codex skills into the single
  source of truth.
• agentloom doctor: checks Node version, hub config, DSH link state, Codex skill
  inventory, MCP bridge, Codex plugin version/cache, the AGENTS.md mirror block,
  index freshness, legacy directories and DSH compatibility patches.
• MCP server (packages/cli/mcp/server.mjs) exposing loom_memory_search,
  loom_memory_add, loom_memory_list, loom_skill_list, loom_skill_show,
  loom_status and loom_sync, with readOnlyHint annotations so Codex can call
  the read-only tools without an approval prompt.
• AGENTS.md mirror block (<!-- agentloom:shared:start --> … end) so Codex
  automatically sees shared memories instead of having to call a tool.
• Windows/Unix link handling: junctions on Windows, directory symlinks elsewhere;
  dangling links are detected through the link target, not through fs.exists.
• Pruning of managed DSH links whose skill was renamed or deleted.
• Unit tests for the edit engine, skill rendering, shared memory and link lifecycle.
• GitHub Actions CI on Linux, Windows and macOS with Node 22 and 24.
• English and Chinese documentation, architecture notes and a compatibility matrix.

Changed

• Hub directory moved from ~/.agent-hub to ~/.agentloom (automatic one-time
  migration on first run, copy-then-remove so it also works across volumes).
• Default paths are platform-neutral (~/.dsh/..., ~/.codex/...) instead of
  ${USERPROFILE}/....
• The DSH injection budget defaults to "auto" and follows the harness's
  agent-instructions maxBytes instead of a hard-coded 64 KiB.
• Codex plugin content is versioned (CODEX_PLUGIN_VERSION) so the per-version
  cache in ~/.codex/plugins/cache/... actually misses after a layout change.
• The Codex plugin's .mcp.json now ships only the agentloom bridge, never the
  user's other MCP servers (their headers and tokens stay out of the plugin).

Fixed

• settings-compat patch script was not idempotent: a second run could duplicate an
  import line and break the plugin with a duplicate ESM binding.
• @deepseek-ai/dsh-settings compatibility shims for plugins built against a newer
  host (settingsNamespace, installSettingsSection, typert create() codecs).
• Invalid skill names and empty shared-memory titles are rejected instead of written.