7.0.0

Added

• Tests for the MLX → canonical tool call bridge, which had none. They pin the payload contract:
  ToolCall.arguments is always a valid JSON object, {} for a call with no arguments rather
  than empty bytes, and every JSONValue shape MLX can hand over — nested objects, arrays, nulls,
  quotes, escapes, non-ASCII — survives the round trip unchanged.

6.0.0

Removed

• BREAKING — the calls that answer "what is on disk" and "how much memory is there" now throw
  instead of guessing. LocalModelInventory, ModelRegistry, AdapterRegistry, MLXBackend,
  DestinationHubDownloader and LLMLocalService gained throwing initializers;
  MemoryProvider.availableMemoryBytes() throws; DownloadState.paused carries Data?;
  DownloadedModel.sizeInBytes is no longer optional.
• StubDownloadDelegate and StubAdapterNetworkDelegate are gone. They were the injected
  defaults, and they reported a completed 1 MB download of a model that was never fetched and
  recorded a 13-byte text file as a cached adapter. A caller who forgot to wire a delegate was told
  the work succeeded. The delegates are now optional, and the two calls that actually need bytes
  moved — downloadWithProgress and a remote AdapterRegistry.resolve — fail without one. The uses
  that never needed a delegate keep working.

Fixed

• An unreadable model directory reported as "not downloaded". isDownloaded used
  try? contentsOfDirectory, so a permission error looked like an absent model — and
  DestinationHubDownloader responded by re-fetching a multi-gigabyte model already on disk.
  snapshotState(at:) now distinguishes complete, incomplete and absent, and throws when it cannot
  tell.
• Storage size understated without saying so. The directory walk used errorHandler: nil and
  try? per file, so an unreadable subtree silently contributed zero — producing a plausible short
  number that a user acts on when deciding what to delete.
• The backup-exclusion flag dropped silently. try? on setResourceValues meant multi-gigabyte
  weights could end up in iCloud backup with no signal, contradicting the doc comment directly
  above it.
• A fabricated memory measurement. host_page_size's return code was discarded, and a failed
  host_statistics64 returned physicalMemory / 2 — a made-up number indistinguishable from a
  real one.
• A download that could not be cancelled reported as stopped. try? on the delegate call, then
  bookkeeping torn down regardless, so the transfer kept running while isDownloading said false.
  A pause with no resume data stored Data(), so hasResumeData said true and resume restarted
  from byte 0.
• Eviction reported storage freed when the files were still there. deleteCache dropped the entry
  whether or not the delete succeeded — and that entry was the last thing pointing at the
  gigabytes, so nothing could find them to try again. The entry is now kept and the failure
  reported; clearAllCache saves what it did manage to free.
• Application Support resolution had two opposite policies in three files — .first! in two
  (a crash the caller cannot catch) and ?? temporaryDirectory in the third (a purgeable directory,
  so "downloaded" quietly stops being true). One throwing helper now serves all three.

5.0.0

Changed

• BREAKING — ModelRegistry and AdapterRegistry no longer report a registry file they could
  not read as an empty registry. A file that exists but will not read or decode now throws
  LLMLocalError.registryUnreadable(reason:); only a registry that was never written still reads
  as empty. ModelRegistry.cachedModels() and isCached(_:), and AdapterRegistry's
  cachedAdapters(), isCached(_:) and isUpdateAvailable(for:latestTag:), gained throws to
  carry it.

  The two states called for opposite responses and were being given the same answer. Downwards,
  every mutating method is a load-mutate-save over the whole file, so treating an unreadable
  registry as empty wrote that emptiness back over a file that was still recoverable by hand and
  orphaned the model directories the lost entries pointed at — deleteCache(for:) could never free
  them again. Upwards, a caller that heard "nothing is downloaded" re-fetched gigabytes the device
  was already holding, and AdapterRegistry.resolve(_:) downloaded over the adapter files already
  sitting at those paths. Nothing is cached from a failed read, so the next call reads again and
  succeeds once the file is repaired or removed.

• Raised the swift-persistence pin to 3.0.0, whose RegistryStore.load() throws rather than
  returning a default. This package had been swallowing that distinction one layer further up.

Added

• LLMLocalError.registryUnreadable(reason:).

4.0.0

Removed

• BREAKING — LocalAgentClient.generateWithUsage takes a GenerationOptions value instead of
  separate systemPrompt, temperature and maxTokens arguments, following swift-llm-client
  4.0.0. The requirement's signature now differs from the convenience method beside it, which is
  what stops a conformance that forgets the requirement from calling the convenience back into
  itself forever. Call sites that pass the arguments individually keep working — llm-client still
  ships those as defaulted convenience overloads.

Changed

• Raised the swift-llm-client pin to 4.0.0.

Nothing.

3.0.0

Full Changelog: 2.2.6...3.0.0

1.5.0

chore: swift-llm-client を URL 参照に変更

1.4.0

chore: Examples・specs を削除、Package.resolved を gitignore に追加

1.3.0

feat: LocalAgentClient 追加、swift-persistence 導入、Manager → Registry リネーム