Shrink the Windows MSI to runtime files

[leoisadev1]

Public feedback on X: the Windows MSI is about 500 MB while the Mac and Linux downloads are about 50 MB. That is almost all packaging, not the app.

What was in the v0.2.0 MSI (537 MiB)

Extracted with 7-Zip from convt-0.2.0-windows-x86_64.msi. The cabinets already use LZX:18 (high compression). The File table uses hashed names; sizes and magic identify the members.

Cabinet	Compressed	Uncompressed	What it is
cab2	443.8 MiB	445.8 MiB	documents.tar.gz — the optional LibreOffice pack, already gzipped so CAB barely helps
cab3	61.2 MiB	194.1 MiB	ffmpeg.exe (97.2 MiB) and ffprobe.exe (97.0 MiB), Gyan essentials static
cab1	18.0 MiB	57.7 MiB	convt-app.exe (35.7 MiB), convt.exe (15.0 MiB), aom.dll, CRT, LICENSE.txt
cab4	13.3 MiB	67.5 MiB	2,316 files: codec source trees under licenses/native-source/ plus the remaining runtime DLLs (heif.dll, libde265, x265, pdfium, more CRT)
MSI	537 MiB		Mac DMG 47.7 MiB, Linux AppImage 55.8 MiB

WiX harvested packaging/out/windows/payload/**. That directory held the document pack (product docs already call this a separate opt-in on Mac and Linux) and the full x265 / libde265 / aom / libheif source trees.

No PDBs and no uncompressed CAB. The leftover bulk after this fix is the two static FFmpeg binaries (~61 MiB compressed), which the app needs at runtime. Expected MSI after the change: about 80–90 MiB (cab1 + cab3 + the runtime DLLs from cab4). CI fails the job if the MSI is 120 MiB or larger.

How document support works

Mac and Linux already download the LibreOffice pack on demand. Selecting a Word/Excel/PowerPoint file shows Quick convert's Download card; nothing hits the network until the user clicks Install (or runs convt pack install documents). The engines then fetch CONVT_DOCUMENT_PACK_URL, verify the compiled SHA-256, and install under the per-user data directory. A silent convert never downloads.

Those platforms do not ship the pack inside the AppImage/DMG, and they do not yet publish a hosted archive, so a released Mac/Linux build only converts documents if LibreOffice is already on the machine.

Windows used to embed the 446 MiB pack in the MSI (bundle:documents.tar.gz) so Install worked offline. That is what made the installer 537 MiB. Windows now matches the Mac/Linux installer code:

• The MSI stays runtime-only.
• The Windows job publishes convt-<version>-windows-x86_64-documents.tar.gz and .sha256 on the same windows-release-review artifact (and then on the GitHub release).
• The app is compiled with CONVT_DOCUMENT_PACK_URL pointing at that release asset, plus the digest and sizes. Quick convert shows the existing Download card (size, progress, checksum failure, retry). convt pack install documents fetches and verifies it.

The website update manifest does not list the pack as a Windows app download. Only the compiled URL fetches it.

What changed

• build.ps1 writes the pack beside the payload, names it convt-<version>-windows-x86_64-documents.tar.gz, writes the checksum, and compiles the GitHub release URL.
• build-native.ps1 copies codec LICENSE/COPYING/NOTICE files only, not the source trees.
• stage_payload.py is the harvest allowlist: executables, DLLs, and notices. It refuses the document pack, native-source, PDBs, .lib files, and headers.
• assemble.py requires the pack and checksum whenever the Windows MSI is included.
• WiX MediaTemplate sets CompressionLevel="high". MSI artifact name stays windows-release-review / convt-<version>-windows-x86_64.msi.
• release-windows.yml still runs on windows-latest. After packaging, smoke.ps1 extracts the MSI, converts a PNG and an MP4, installs the sibling document pack, converts a Word file to PDF, and fails if convt-app.exe exits or does not start.
• installer.ps1 retries by removing only the owned msi-payload harvest directory.
• Changeset on @convt/desktop (and the fixed CLI/web group) so this ships in the next release.

PR #73 is also editing release CI. This only adds the smoke step, the extra Windows review files, a short release-notes line, and unit tests on the existing ubuntu-24.04 release-scripts job. Runners are unchanged.

Proof so far

• python3 -m unittest discover -s packaging/windows (payload allowlist, document-pack URL/checksum, smoke script contract).
• python3 -m unittest discover -s scripts/release (assemble now requires the Windows pack and checksum).
• App UI for Download / progress / checksum already has tests in crates/convt-app/src/ui/tests.rs (a_build_without_a_pinned_pack_offers_no_download vs the configured Download card).

How to dry-run the real MSI

A reviewer can:

1. Actions → Build release windows → Run workflow.
2. Use branch cursor/windows-msi-size-8c4c.
3. Inputs: version=0.2.0 (must match this commit), source_date_epoch = any UTC unix time (for example 1791331200), dry_run=true.
4. The job uploads windows-release-review: the MSI, convt-0.2.0-windows-x86_64-documents.tar.gz, and .sha256.
5. The Smoke-test the extracted MSI step prints the MSI byte size, converts a PNG and an MP4, installs the document pack from that archive, converts a Word file to PDF, and fails if convt-app.exe does not stay running. The job fails if the MSI is 120 MiB or larger.

Do not merge; a reviewer will merge it.

  

---

[devin-ai-integration]

Devin Review found 2 potential issues.

[devin-ai-integration]

🟡 Missing runtime files pass staging

When convt.exe is missing from the payload root, validate accepts a namesake under licenses. stage leaves that namesake under licenses, so the MSI lacks its CLI.

Suggested change

	names = {path.name for path in files}
	for required in REQUIRED:
	if required not in names:
	errors.append(f"missing required runtime file: {required}")
	names = {path.name for path in files if path.parent == root}
	for required in REQUIRED:
	if required not in names:
	errors.append(f"missing required runtime file: {required}")

---

Was this helpful? React with 👍 or 👎 to provide feedback.

[devin-ai-integration]

🔍 PDFium discovery remains untested

Setting CONVT_PDFIUM_DIR bypasses normal PDFium discovery. The smoke test also skips PDF conversion, so it cannot validate PDF support in an ordinary install.

---

Was this helpful? React with 👍 or 👎 to provide feedback.

[greptile-apps]

Confidence Score: 5/5

[High risk] Restructures Windows installer build and packaging.

The PR appears safe to merge; both numbered previous findings are fixed and no new blocking issue was found.

What we checked:

• Rebuilds use the current URL: publish() now derives the URL from the current version rather than the environment. The build restores the caller’s URL after compilation.
• Smoke runs use separate folders: Each run gets a new GUID for its extraction, work, and profile folders. Cleanup uses those run-specific paths.

Summary

The Windows MSI now contains runtime files and notices instead of the document pack and codec source trees.

• The document pack and checksum are published beside the MSI.
• Packaging and smoke checks enforce the 120 MiB limit.
• The latest changes fix stale build URLs and shared smoke folders.
• No new actionable issues were found. The Windows MSI and smoke run were not executed in this Linux environment.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart TD
  Build[Windows build] --> Runtime[Executables, DLLs and notices]
  Runtime --> Stage[Runtime-only harvest]
  Stage --> MSI[Windows MSI]
  Build --> Pack[Versioned document pack and checksum]
  Pack --> Release[GitHub release assets]
  Build --> Pin[Compile release URL and digest]
  Release --> Install[User clicks Install]
  Pin --> Install
  Install --> Verify[Verify digest before installing]

Loading

_{Reviews (7) · Last reviewed commit: "fix(windows): isolate smoke dirs and sto..." · Reviewed by Greptile}