Skip to content

Releases

v2.2.1

Choose a tag to compare

@github-actions github-actions released this 16 Nov 05:22
d3c87fa

Changed

  • PDF observation is back: the 2.2.0 safety measure that treated
    every PDF as unobserved is lifted. 2.2.0 shipped it because the
    visible-text extraction of the day let the never-painted text leak
    into handoff/ (one sentence of history, measured in the 2.2
    start-up report); Sieve Lens 0.22.1 now excludes the characters a
    PDF does not paint (render mode 3, for example) from its visible
    text, so Gate observes PDFs again and reports the hidden bodies on
    the item. A PDF without pypdf keeps the closed engine_unavailable
    reason (no detail), and the closed vocabulary is unchanged.

Fixed

  • Invisible text no longer appears as visible text: the leak is fixed
    at the source in Sieve Lens 0.22.1 and the Gate PDF lane rides that
    fix - the invisible marker stays out of the whole handoff/ tree
    (measured by the PD tests, raw bytes, ascii and utf-16).
  • gate --self-check carries one more item: a hand-built PDF (one
    visible line + one render-mode-3 line) must come back observed with
    the hidden body reported and the marker nowhere in handoff/ -
    the item proves pypdf actually works inside the frozen apps on
    every release OS (the expected count is 13 + 1 = 14).

v2.2.0

Choose a tag to compare

@github-actions github-actions released this 15 Nov 02:27
7c8dde2

Added

  • File list and check code (manifest): list a folder into a
    handoff receipt - every regular file as path/bytes/sha256 in
    relative POSIX form sorted by UTF-8 name bytes (NFC never applied,
    a name that is not valid UTF-8 rides as path null + path_hex),
    duplicates confirmed by real byte comparison, and the whole-folder
    checksum as the hash tree of docs/manifest-format.md §1.
    --against diffs a folder into the closed five states
    (same/changed/removed/added/moved, moved pairs bound per sha group
    with one ambiguous_move per multi-candidate group). The receipt
    keeps the 2.1.1 frame (receipt_version 1) and is written last; a
    run without the key stays byte-identical in shape.
  • Check code (--key-file): gate/cohort/manifest/pack share the
    optional key file - the receipt gains the check_code section
    (key_id = the first 8 HMAC bytes in hex, check_code = HMAC over the
    whole-folder checksum and the canonical receipt body,
    docs/manifest-format.md §2). verify gains --key-file and
    --root, adds exactly two not_verifiable reasons
    (code_key_missing / code_key_mismatch) while the file-list
    check keeps running and its results still land in items, and keeps
    the 0/1/2 exit codes and the published vocabulary; a real
    difference still dominates not_verifiable, so the list results are
    never hidden.
  • Rename and audit-names: rename plans by default and never
    touches the source - --apply builds the renamed copy at
    /renamed (the rule fields are closed - seq / seq:width / stem
    / ext / hash8 / parent - and collisions are detected at the plan
    stage over exact + casefold + NFC against each other and the
    existing directories); audit-names audits files AND directories
    with the closed 8 symbols and reports path + symbol only. The
    handoff carries the rule, the counts and the mapping hash only -
    the names stay in local/rename-plan.json
    (docs/rename-procedure.md).
  • Pack and unpack: pack writes the deterministic zip (entries
    in bare-path byte order, 1980-01-01 00:00:00 stamps, ZIP_STORED by
    default, --deflate on file entries only) at -o itself next to a
    handoff/pack-receipt.json carrying zip_sha256 and the
    manifest-identical whole-folder checksum; unpack runs every
    safety check (absolute path, .. component, drive letter, and
    collision over exact/casefold/file-inside-file) BEFORE creating
    anything and refuses with zero extraction (exit 1, the receipt
    still written); the roundtrip is byte-identical including empty
    directories.
  • verify.html (--verify-page): manifest/pack can write the
    self-contained handoff/verify.html BEFORE the receipt and record
    its sha256 in the optional verify_page section - the file list is
    embedded as escaped canonical JSON (a file name cannot close the
    script block), the SHA-256 is a pure-JS core checked against 4
    known values, and the receiver picks a folder (webkitdirectory)
    and reads every file through Blob.slice in 1 MiB chunks; same /
    changed / missing / extra is shown on screen only in plain ja and
    en - no network, no external references, and the check-code
    verification stays with Sieve Tray's verify.
  • Draw (draw): a verifiable draw over a roster - key(x) =
    SHA-256(UTF-8(seed) || NUL || x), the items ordered by key bytes
    and dealt round-robin into k groups (--groups, or --size so the
    last group can be short); the result with the names lands in
    local/draw-result.json (written before the receipt) and the
    receipt carries the roster hash, the public seed, the parameters
    and the result hash - never the names. docs/draw-procedure.md
    spells the procedure byte by byte with two measured known values
    (KV-D1/KV-D2) and the seed warning sentence.
  • The PDF clause: the Gate document lane short-circuits a PDF to
    unobserved / engine_unavailable (detail
    pdf_visible_extraction_disabled) BEFORE any engine.observe call,
    regardless of pypdf presence - pypdf's visible extraction ignores
    the 3 Tr render mode, so observing a PDF is where the invisible
    text would start to leak.
  • GUI: the sidebar grows to twelve pages - the three new pages
    (en: Manifest & Check Code / Rename / Draw, ja: ファイル一覧と照合コード /
    名前の整理 / 抽選・班分け) run the real engines behind the same
    manners (a description written from the actual behavior, a tooltip
    on every part, a placeholder in the result field), the scan
    operation row shows on the scan pages only, and the settings ride
    the menu as a PreferencesRole action (macOS relocates it into the
    application menu, Windows/Linux keep it in the window menu). The
    new pages render uncut at the 880x560 minimum under ja/en and a
    dark palette, and last_page restores them too.
  • README and docs: the Gate/Cohort sections are rewritten for the
    general software-manual style (the old technical paragraphs moved
    verbatim into docs/receipt-reference.md with one link line in each
    README), the three new features get sections in the same four
    manners (what it does, up to 3 numbered steps, one command block,
    one docs link), and the screenshots are re-shot offscreen at
    1042x900 for both languages (8 shots) with the icon at width 96.
  • The frozen app runs the three Seal self-checks (SL-T20):
    release.yml's post-processing starts the built artifact with
    manifest --self-check / pack --self-check / draw --self-check
    on every OS leg after gate/cohort - each builds a small known input
    in a temp folder at run time (the published KV-2 / KV-D1 / KV-D2
    fixtures) and compares the expected values under the same 0/1/2
    exit contract, so a tagged bundle that cannot list, pack or draw
    fails the release.

Changed

  • Lens extra: the dependency is now
    sieve-lens[html-css-full]>=0.22.0 (the floor unchanged) - measured
    2026-11-14 in a fresh venv: a bare install cannot observe the
    hidden HTML text (the item is reported engine_unavailable) while
    the same input is observed and flagged once tinycss2 + cssselect2
    are present; pypdf stays out because the Gate PDF lane
    short-circuits by design with or without it. Both READMEs name the
    extra packages in the install note.
  • Screen vocabulary: the screen strings and the new docs follow
    the M2 vocabulary ruling (キーファイル / 処理記録); the ja wording
    drops the legacy words and docs/manifest-format.md says 順序付け
    (sort key).

v2.1.1

Choose a tag to compare

@github-actions github-actions released this 13 Nov 13:40
c4f8b44

Fixed

  • The windows leg of the three-OS Test matrix (run #84: 18 failed,
    five root classes)
    : every path string entering a ScanResult, a
    report and the history takes the canonical POSIX-slash form (the
    gate-walk idiom, and _posix_path_str at the storage boundary) while
    the engine still opens the native path; the two sample generators pin
    ZipInfo.create_system = 3 and write the cohort MANIFEST through
    write_bytes; the mold fixtures are written with fixed LF
    (newline=""); the POSIX-only observation scenarios (the unreadable
    chmod lane, the mkfifo lane, the keygen 0o600 pin) ride the v1.4.0
    os.name == "posix" idiom; the GUI startup-geometry pins accept the
    platform layout minimum (max(decision, minimumSizeHint())); and the
    single-image sanitize history pin compares the canonical form the
    boundary certifies. The repository pins its own checkout
    normalization (.gitattributes: * text=auto eol=lf), so every text
    file checks out in the canonical LF form the byte-pins certify on all
    three runners.
  • gate --self-check rides the 0/1/2 exit codes (CH-T21): 0 = the
    expected observations matched, 1 = a difference, 2 = could not run -
    a difference now exits 1 and fails CI, the same contract the cohort
    leg already had.
  • GUI wording: the Gate result gains its placeholder (GU-T06), the
    About credits Sieve Order and the tagline names the export (GU-T08),
    the base picker asks for files (GU-T18), the history header fits
    (GU-T19).

Changed

  • The Test workflow rides the three release OS legs (ubuntu-24.04 /
    windows-latest / macos-15 on Python 3.12, fail-fast: false): the
    distributed Windows/macOS builds are tested on their own runners, and
    the vendored sieve_redact verification is checkout-normalization-proof.
  • Release trials without publishing: release.yml gained
    workflow_dispatch - the three-OS build and the two frozen-artifact
    self-checks run headlessly (the frozen GUI entry dispatches
    gate|cohort --self-check to the CLI before QApplication, exit 1/2
    fails the release) while every Release attach sits behind the push
    guard.
  • Release assets take the tag, and the body comes from the
    CHANGELOG
    (the 2026-11-13 ruling): the asset names ride
    github.ref_name (SieveTray-v2.1.1-windows.zip and siblings), the
    SHA-256 list and the provenance attestation follow the built zips,
    and the release body is the CHANGELOG section of the tag version,
    extracted by tools/extract_changelog.py - a tag without its CHANGELOG
    section fails the release structurally.
  • Version notation is unified (the 2026-11-13 ruling): the README
    titles carry the release version as their single version literal,
    downloads point at releases/latest with the asset names described
    generically, and one test pins pyproject (dynamic attr) ==
    version == the first CHANGELOG section == both README titles;
    the six version observation pins follow the bump.

Note

  • 2.1.0 was published to PyPI only - no GitHub tag/Release was cut for
    it (the 2026-11-11 ruling: the pre-fix workflow must not turn a
    v2.1.0 tag into a Release), and PyPI 2.1.0 stays as published. The
    first GitHub tag/Release is v2.1.1.

v2.0.0

Choose a tag to compare

@github-actions github-actions released this 09 Nov 03:34
fix: give CI and the bundled GUI the Lens engine extras

Found by the 2.0.0 acceptance/release run (2026-11-09):

- test.yml installed .[gui] only, so the Lens extras lane
  (pypdf / tinycss2 / cssselect2) was missing and the Gate engine
  tests ran as engine_unavailable (t03, t08, t17).
- test.yml invoked plain pytest, which does not put the repo root on
  sys.path; the site-packages copy installed by .[gui] was imported
  instead of the tree, so REPO_ROOT pointed away from samples/gate
  and the self-check tests got the sample_pack_not_available error
  dict (t14, UI button). The acceptance contract (tests/conftest.py)
  is 'python -m pytest'.
- release.yml built the GUI bundle without the engine: the
  sieve_lens_ext extensions load via importlib (invisible to
  PyInstaller analysis), so the distributed app silently degraded
  every HTML observation to engine_unavailable even with
  samples/gate bundled. --collect-all sieve_lens_ext plus the extras
  in the build env bundles the extensions and their libraries
  (sandbox verified: the PYZ carries tinycss2/cssselect2/pypdf and
  _internal carries sieve_lens_ext + chartable_v1.json).

v1.4.0

Choose a tag to compare

@github-actions github-actions released this 06 Nov 18:25
fix: make the three v1.4.0 GUI probes hold on a real Qt (sort order, …

v1.0.0

Choose a tag to compare

@github-actions github-actions released this 30 Oct 22:57
chore: bump version to 1.0.0

v0.10.1

Choose a tag to compare

@github-actions github-actions released this 29 Oct 05:11

What's Changed

  • README (EN/JA) now documents the two bundled engines: Sieve Redact (behind the Sanitize actions) and Sieve Crowd (behind the Anonymize tab).
  • Series table refreshed to 8 packages — sieve-order and sieve-crowd added, role descriptions rewritten.
  • About dialog lists all five bundled engines: Lens, Scope, Referee, Redact, Crowd.
  • Tab renamed: "Crowd" → "Anonymize" (EN) / 「匿名化」(JA).
  • 130 tests (About-dialog regression test added).

v0.10.0

Choose a tag to compare

@github-actions github-actions released this 29 Oct 03:17

First release with the bundled anonymization engine: the Anonymize tab runs vendored Sieve Crowd (k-anonymity for tables), and crowd run history is included in exported reports. 129 tests.

v0.9.1

Choose a tag to compare

@github-actions github-actions released this 26 Oct 21:48
chore: bump version to 0.9.1

v0.9.0

Choose a tag to compare

@github-actions github-actions released this 26 Oct 20:13
feat: v0.9.0 - file selection, history clear, Markdown export, signal…