v0.4.0

Added

• Release workflow publishes the wheels and sdist to PyPI (trusted publishing).

Fixed

• On Linux/macOS, item lock files are removed on normal release, including error and cancellation unwinding.
  Waiters validate the locked inode and reopen an unlinked file before entering; Windows keeps persistent locks.
  A crash or failed unlink may still leave a reusable lock file. Stop older clients before upgrading: versions
  without inode validation must not download to the same destination concurrently with this version.

v0.3.0

Fixed

• Failed updates preserve the previous item even when a file becomes a directory or a whole-file hash fails after
  other files were verified. Items are staged separately, then installed with rollback on ordinary rename errors.
• Cancellation during the last in-flight chunk, an up-to-date item's progress callback or a legacy request is checked
  before installation. Legacy progress callback failures preserve the old file too.
• Concurrent downloads of the same destination wait for a persistent .<item id>.lock with the owner's PID and a
  kernel lock. Waiting supports cancellation; process exit releases the lock and the next owner replaces stale PID data.

Added

• examples/: programs using the whole C++, Python and asyncio API.
• pxsteamdl::Version(), the PXSTEAMDL_VERSION_* macros (<pxsteamdl/version.hpp>), pxsteamdl --version and
  pxsteamdl.__version__.
• cmake --install and find_package(PxSteamDL): headers, the CLI, and one static archive that already contains
  curl, mbedTLS, zlib, liblzma and zstd.
• Transient failures are retried up to four times with exponential backoff (0.5 s, 1 s, 2 s): item details, the CDN
  server list, depot keys, manifest request codes, legacy files and the CM server list. After a transient failure
  (HTTP 429/5xx, a network error) CDN downloads pause before they try the next host. Pauses end when the download is
  cancelled.
• Steam's result codes are spelled out: Steam rejected the item: not found (EResult 9) instead of Steam result 9.
• ErrorKind (kNone, kCancelled, kNotFound, kRejected, kNetwork, kData, kFilesystem, kOther) in
  Result::error_kind and ItemInfo::error_kind, so callers need not parse messages; pxsteamdl::Error (a
  std::runtime_error with kind()) is what Client's constructor throws. Python: ErrorKind, Error,
  Result.error_kind, ItemInfo.error_kind.
• Result::ok(), Result::cancelled(), ItemInfo::ok(), ItemInfo::cancelled() (Python properties ok and
  cancelled; Result.ok existed).
• ClientOptions (C++) and Client(proxy=, connect_timeout=, stall_timeout=) (Python): a proxy, a connect timeout and
  a stall timeout for all connections.
• Result::downloaded_bytes and unpacked_bytes: what a run fetched for the item, over the network and on disk.
• Client can be moved. Python: Result, ItemInfo and Progress can be constructed (keyword arguments), for tests.
• pxsteamdl -h/--help (exit code 0) and -- (the rest are item IDs).
• Unit tests (GoogleTest, -DPXSTEAMDL_BUILD_TESTS=ON), including end-to-end tests of Client against a pretend
  Steam; the network smoke tests run in CI again, on all three platforms.

Changed

• Breaking: Progress::bytes_done and bytes_total are replaced by downloaded_bytes/downloaded_total (what came
  over the network: encrypted, compressed chunks) and unpacked_bytes/unpacked_total (what was decrypted,
  decompressed and written). The old pair counted the unpacked bytes, so unpacked_* is the drop-in replacement. Same
  in Python.
• Behavior change: download() returns one Result per entry of the ID list, also for a repeated ID: the item is
  downloaded once and every occurrence gets the same result. Before, a repeated ID made two jobs write the same
  directory at once, and one of them could fail.
• Behavior change: download() throws std::invalid_argument if parallel_items or threads_per_item is 0
  (the CLI refuses -j 0 and -t 0 with exit code 2); before, 0 was taken as 1. Python already raised ValueError.
• Result::path is documented: it is root/<id> for every item, failed ones included. An exception thrown by
  on_progress fails its item (kind kOther), one thrown by on_resolved propagates out of download(); both are
  now documented.
• Item details are looked up in batches while earlier items already download, instead of all before the first byte.
  on_resolved is still called once per item, in order, on the thread that called download(), but progress of
  earlier items may now arrive before it.
• Behavior change: a failed item-details request no longer throws out of download(). The items of the failed
  batch carry the error in Result.error (item details: ...), and the other items are downloaded. Items that were
  not looked up before a stop request report cancelled.
• A depot key reply that succeeds without a key now says so, instead of reporting EResult 1.
• The source is split into modules, one header each, and follows the style described in
  CONTRIBUTING.md; nothing changes for users.

v0.2.0

Full Changelog: v0.1.1...v0.2.0

v0.1.1

What's Changed

• Review fixes: fair item scheduling, CM reconnect, fd limits, case-insensitive paths by @PyXiion in #2

Full Changelog: v0.1.0...v0.1.1

v0.1.0

What's Changed

• Cross-platform support (Linux, macOS, Windows), CPM vendored dependencies, and Python wheels by @PyXiion in #1

New Contributors

• @PyXiion made their first contribution in #1

Full Changelog: https://github.com/PyXiion/PxSteamDL/commits/v0.1.0