Skip to content

CHANGELOG.md

Latest commit

 

History

History
203 lines (173 loc) · 10.7 KB

CHANGELOG.md

File metadata and controls

203 lines (173 loc) · 10.7 KB
Raw

Changelog

All notable changes to this project are documented here. The format follows Keep a Changelog; versioning is SemVer with a 0.x experimental prefix. Evidence links point to files in this repository (docs/) so the trail is auditable offline.

[0.2.1] — 2026-09-29

Added

  • Multi-source download fallback — fetch now tries the official raw.githubusercontent.com URL first, then the equivalent github.com raw link, and only then community mirrors of the same URLs (default gh-proxy.com/ghproxy.net, overridable via DLSS_COMBO_MIRRORS). All failures list every source tried. We still never redistribute upstream files (src/dlss_combo/fetch.py, tests/test_fetch_sources.py).
  • fetch --runtime all — one command downloads both DLSS-G runtimes (310.9 + 310.1) plus the Swapper portable (src/dlss_combo/cli.py, tests/test_fetch_all.py).
  • DOWNLOADS.md — manual download & placement guide on the repo page: direct upstream links for every file, why NVIDIA SDK terms forbid bundling them into our zips, mirror configuration, integrity notes.

0.2.0 — 2026-09-29

Added

  • Daily upstream-structure watch: the scheduled upstream-check workflow verifies the upstream shape this toolkit depends on (dlssg_for_sm86 main commit resolves, kit files still exist at that commit, generated INI keys stay accepted, DLSS5-Swapper still publishes a selectable portable asset) and opens/updates an auto-labeled issue on drift, closing it when green again. Evidence: .github/workflows/upstream-check.yml, scripts/upstream_check.py, selftest tests/test_upstream_check.py.
  • report command: prints a paste-ready, offline, read-only game-test report (GPU landscape with the selected card, manifest summary, routing verdict, and an explicit "does NOT contain" section) for the game-test issue template. Evidence: src/dlss_combo/report.py, tests/test_report.py.
  • --check-update: non-blocking check for a newer dlss-combo release before any command — prints one line and continues; network failures degrade to a skip notice. Evidence: src/dlss_combo/update.py, tests/test_update.py.
  • Multi-GPU enumeration: all NVIDIA cards are parsed and the first card with a supported architecture (RTX 20/30) is selected; doctor lists multi-card rigs and names the primary. --arch still overrides. Evidence: src/dlss_combo/gpu.py, tests/test_gpu_multi.py.
  • Recommended test-games list for RTX 20/30 testers, with native-FG evidence links and a verify-in-game disclaimer. Evidence: docs/TEST-GAMES.md.

Changed

  • CLI is now English-only — help text, diagnostics and error messages. Internal source comments are unchanged. Guarded by tests/test_english_cli.py (CJK scan over user-facing strings).
  • Doctor session diagnostics hardened (see Fixed).

Fixed

  • Unreadable route logs are reported as problems instead of crashing doctor; the session verdict is now the last route event of the newest log (a later failure is never hidden by an earlier success), and image-layer files count as evidence only, never as layer activation. Evidence: tests/test_doctor_last_event.py.
  • Manifest schema/path validation error messages are now English — translated wording only, failure behavior is unchanged — and a bare invocation prints usage again (guarded in tests/test_english_cli.py).
  • Proxy fallback order regression-locked to the upstream tool-class order (version → winmm → dbghelp → dinput8 before d3d12/dxgi). Evidence: tests/test_proxy_order.py.

0.1.3 — 2026-09-20

Fixed

  • Corrected release for v0.1.2's Linux measurement defect: the v0.1.2 workflow's embedded-library extraction silently failed (wrong CArchiveReader.extract call signature) and fell back to the outer-file string scan, so build-info.txt again reported the under-stated floor (2.14). The step now uses the bytes-returning extraction API and fails loudly when zero embedded ELF libraries parse (no silent fallback). The real floor (expected 2.35, from the embedded libpython3.13) is recorded per build. v0.1.2 assets are left untouched; prefer this release.

0.1.2 — 2026-09-20

Fixed (independent v0.1.1 review, R1–R7)

  • R1: first-install failures now restore the user's original INI at every failure point (DLL copy, manifest write, backup write; reinstall failures restore the previous working state). Covered by fault-injection tests in tests/test_review4.py.
  • R2: uninstall never overwrites any existing file or symlink when saving a restore copy — copies are created exclusively (O_CREAT|O_EXCL) with numbered fallback names (*.pre-existing.1, …).
  • R3: temporary files use random, exclusively-created names (tempfile.mkstemp, same directory) — predictable-name symlink attacks are inert — and cleanup touches only paths registered by the current operation.
  • R4: the Linux glibc floor is now measured by extracting the embedded ELF libraries from the PyInstaller CArchive and parsing their .gnu.version_r requirements (outer-file string scans under-report); docs state the real floor (glibc >= 2.35, from the embedded libpython3.13) and the release notes read the measured value per build.
  • R5: a missing or corrupted pre-existing backup aborts uninstall before any deletion (backups now record a SHA256 that is verified when present).
  • R6: doctor treats the newest log as the diagnostic scope; a newer session without route events reports "未验证/unverified" while older route events are labelled historical.
  • R7: kit refreshes download into a staging directory and switch only after every file is fetched — an interrupted refresh leaves the previous kit fully valid (offline fallback preserved).

The eight reproduced scenarios from docs/reviews/2026-09-20-v0.1.1-independent-repro.py all assert fixed, and are migrated into formal regressions (tests/test_review4.py). Docs consistency: STATUS links/counts, README quick-start no longer forces --arch (auto-detection first), Linux requirements updated.

0.1.1 — 2026-09-20

Fixed

  • Preserve user-modified dlssg_sm86.ini across reinstall and uninstall (release-acceptance review P1). Reinstall now keeps an externally modified INI instead of overwriting it after warning (the manifest records an ini_user_managed flag that persists across upgrades); uninstall no longer overwrites an in-place user file with the pre-existing backup — the original config is saved beside it as dlssg_sm86.ini.pre-existing. Regression tests: tests/test_review3.py.
  • Single-source package version (pyproject.toml now reads dlss_combo.__version__); --version previously could disagree with the built artifact.

Changed

  • Release process gate: the release workflow now runs the offline test suite and the review repro audit and a functional smoke test of the packaged binary (install → doctor → uninstall against a fabricated kit), on both Windows and Linux — no longer a --version-only check.
  • Linux binary glibc floor: built on ubuntu-22.04 (was ubuntu-latest); the whole-binary GLIBC symbol floor is measured per build and ships in the zip's build-info.txt (v0.1.1 measured GLIBC_2.14, down from 2.38 in v0.1.0). Other distros remain unverified; glibc >= 2.17 is the conservative recommendation.
  • Release page: marked pre-release (matches the developer-preview stage); notes are English-first with Windows as the main download and Linux labelled experimental; each zip now contains QUICKSTART.txt, LICENSE and build-info.txt (source commit, build time, measured glibc floor).
  • Linux/Proton guidance (README + release notes): locate the game via Steam → Browse local files instead of assuming a fixed prefix layout; WINEDLLOVERRIDES/PROTON_* launch options are labelled community convention, not verified with this toolkit, with a pointer to Valve's Proton config docs.
  • Unsigned-binary wording is neutral: verify SHA256, decide about security warnings yourself; no instruction to bypass protections.

0.1.0 — 2026-09-20

Added

  • Initial public toolkit: fetch / install / uninstall / doctor CLI coordinating DLSS5-Swapper (DLSS 5 image layer) with dlssg_for_sm86 (frame generation on RTX 20/30); SHA256-verified kit downloads pinned to an upstream commit; manifest-driven install/rollback/uninstall; bilingual docs.
  • First release binaries for Windows x64 and Linux x64 via GitHub Actions.

Security / file-safety (from the 2026-09-20 project review, all fixed here)

  • Manifest validation before any destructive step: known schema version, allow-listed filenames only, relative paths, no ../absolute/symlink escapes, backups confined to .dlss-combo/ (review A).
  • Hash-based ownership: files whose current hash no longer matches the install record are treated as third-party and never touched (review B).
  • Pre-existing user INI backed up (pre-existing, distinct from tool-history backups) and restored on uninstall (review C).
  • Every required kit file must carry a well-formed matching SHA256; cache hits re-verify with automatic refetch; Swapper binary hash-checked before launch and refused if a ZIP (review D, J).
  • Atomic disk writes everywhere (temp file + replace), rollback restores deleted files and cleans partial temp files (review E).
  • Kit fetch is transactional per file; per-runtime commit records; Swapper metadata survives refreshes (review F).
  • Doctor accuracy: newest-log/newest-event route parsing, strict booleans, empty ReShade folders no longer reported as an installed image layer, non-zero exit on detected problems (review G).
  • Proxy order follows the upstream tool-class recommendation (version → winmm → dbghelp → dinput8 before the riskier d3d12/dxgi), --proxy pins a name explicitly (review H).
  • --arch no longer skips driver checks; 310.1 + 6x rejected per upstream limits (review I).

The nine reproduced problem behaviors are pinned by docs/reviews/2026-09-20-repro.py (all report fixed) and migrated into regression tests (tests/test_review2.py).