Skip to content

ADR-016: Relicense from MIT to the proprietary Sigil License 1.0

  • Status: Accepted
  • Date: 2026-09-14
  • Decision driver: Sigil is about to ship v0.1.0-alpha, its first public artifact. Under MIT, the moment that artifact exists so does the right for anyone to fork the engine, rebrand it, and ship a competitor — and the engine is the product. The license had to be decided before the tag, not after, because a license change cannot reach copies already distributed.
  • Scope: replaces LICENSE; changes the package license metadata, the copyright holder recorded in every source file, the contribution policy, and every document that described Sigil as open source. No src/ behavior changes. Supersedes nothing; ADR-016 is the first licensing ADR.

Decision (TL;DR)

Relicense Sigil, effective 2026-09-14, from MIT to the Sigil License 1.0 — a proprietary, source-available license that grants unlimited use of the tools (including commercial use) and an explicit royalty-free right to redistribute the artifacts the tools generate, while withholding every right to copy, modify, or reuse the source code itself. Record Yevhen Khudoliiv as the sole copyright holder in LICENSE, in package metadata, and in a per-file header on every .cs and .axaml file, enforced by IDE0073 as a build error. Close the repository to external code contributions.

Context

git shortlog -sne --all reports 364 commits from one human across two addresses:

   300  Yevhen Khudoliiv <eugenkhudoliiv@gmail.com>
    64  Yevhen Khudoliiv <yevhen.khudoliiv@helixleisure.com>

There is no third-party contributor whose permission would be needed, and no CLA to reconstruct after the fact. This is the single fact that makes the whole decision cheap; it is also the fact the contribution policy below exists to preserve, because it stops being true the first time an outside pull request merges.

The ~77 commits carrying Co-Authored-By: Claude trailers do not affect this. A model is not an author under copyright law, and those trailers assert provenance, not ownership.

The dependencies do not obstruct it

Every direct NuGet dependency is permissively licensed (MIT, BSD-3-Clause, or Apache-2.0), and THIRD-PARTY-NOTICES.md already enumerates the native components that carry binary-redistribution attribution requirements — Skia, HarfBuzz, ANGLE and their transitive pieces. The one copyleft exposure in the tree, SkiaSharp's eCos/LGPL-2.1 dual-licensed HTTP-server component, is not used. Nothing in the dependency graph requires Sigil's own code to stay open.

What those terms do require is attribution on binary redistribution, and that obligation passes through to every vendor who ships a Setup.exe. The license must say so rather than quietly inherit it.

A generic "no redistribution" license would break the product

This is the constraint that disqualified the obvious answer. Sigil's output is not a document describing an installer; it is Sigil's own binary. WrapperRuntimeLocator.cs resolves the Native-AOT-published SigilBuild.Installer.Host.exe for each target RID, and ExeWrapperPackager stamps that executable with the vendor's step blob, payload, icon, and the bundled native dependencies as a SIGIL_RUNTIME_V1 resource. The vendor then ships the result to their own users.

So the ordinary use of Sigil is: a third party reproduces and distributes the licensor's compiled code, to an unbounded audience, as a component of their own product. An off-the-shelf source-available license that forbids redistribution — PolyForm Strict 1.0.0, for instance — would make every correct user of Sigil an infringer for using it as designed, and would make their end users infringers for running the installer they received. The license needs a redistribution grant carved out by construction, not by exception.

The MIT grant already made cannot be recalled

The repository has been public under MIT since 2026-04-30. Anyone who cloned it in the intervening four and a half months holds MIT rights to that snapshot permanently, including the right to fork and redistribute it. Relicensing binds future versions only.

The exposure is nevertheless small and worth naming precisely: no v* tag exists, no GitHub release has been published, and the only NuGet package is SigilBuild 0.0.0-reserved (published 2026-05-05), a metadata-only placeholder containing no code. What is at risk is a source snapshot of a pre-MVP tree, not a shipped product with users.

Decision in detail

  1. LICENSE becomes the Sigil License 1.0, drafted in PolyForm's register — plain English, defined terms, short numbered sections — because PolyForm is explicitly designed as a family assembled from reusable blocks, and a bespoke grant was needed anyway (see below).
  2. Use is unrestricted. Any purpose including commercial, any number of machines, any number of personnel. A distribution tool nobody may run is not a product.
  3. Generated Artifacts carry an explicit redistribution grant. Royalty-free, perpetual, to unlimited recipients, conditioned on the Redistributable Components being unmodified, being shipped only inside an artifact that installs the vendor's own product, and third-party notices being preserved. Recipients get a direct right to run what they received, and that right survives termination of the vendor's own license — otherwise a vendor's breach would retroactively strand their customers.
  4. Source reuse is withheld entirely — copying, modifying, deriving, redistributing the software itself, extracting the Redistributable Components from an artifact, and building a competing product from what the source teaches. Reverse engineering is carved back where applicable law forbids the restriction, rather than asserting an unenforceable prohibition.
  5. Third-party notice obligations are passed through explicitly, since the licensor has no power to waive them and a vendor who strips THIRD-PARTY-NOTICES.md from their installer breaches Skia's and HarfBuzz's terms, not only Sigil's.
  6. Per-file authorship headers on all sources, naming the holder and the license. Volume at the time of writing: 466 .cs files under src/ and tests/, plus 16 .axaml.
  7. Contributions are closed. CONTRIBUTING.md states that pull requests are not accepted and that issues and bug reports are. Soliciting code for a codebase nobody may derive from is incoherent, and each merged outside contribution would fracture the sole ownership that section one rests on.

Alternatives considered

PolyForm Strict 1.0.0, unmodified. The closest off-the-shelf fit and the default recommendation for "you may look, you may not take". Rejected because it grants no redistribution right, and Sigil's entire function is to hand the user a redistributable binary built from Sigil's own compiled code. Adopting it would have criminalised the product's happy path. Adding a carve-out as an appendix to someone else's license text was considered and rejected too: a license read as "document A except where document B disagrees" is worse to reason about than one coherent page.

Two documents — proprietary source license plus a separate Runtime Redistribution License. This is how the commercial installer vendors (InstallShield and peers) structure it, and it is the most defensible split legally, because the grant that flows to third parties is physically separate from the grant that binds the licensee. Rejected on proportionality: it makes every evaluating vendor read two agreements to answer one question, at a stage where Sigil has no customers and needs adoption friction to be zero. Revisit if the redistribution grant ever needs to differ by customer tier.

PolyForm Noncommercial 1.0.0. Free for noncommercial use, paid otherwise. Rejected because Sigil's users are, by definition, people shipping commercial Windows software — the noncommercial carve-out would cover almost nobody, so the practical effect is "commercial license required to evaluate", and organic adoption goes to zero.

BUSL-1.1 with a four-year Apache-2.0 conversion. The HashiCorp/Sentry compromise: protection now, open source eventually. Rejected because BUSL is written around hosted services, and its pivotal term — "production use" — is close to meaningless for a build-time CLI. A vendor could not tell whether packaging a nightly build counted. A license whose central condition is ambiguous for the actual product is a support burden, not a protection.

All rights reserved, no grant at all. Maximally protective and three lines long. Rejected because it would make v0.1.0-alpha unusable by anyone who downloaded it, converting the release from a product into a display case.

Staying on MIT. Rejected on the forcing problem in Context: the engine is the product, and MIT invites a rebranded fork the day the first artifact ships.

Consequences

What gets easier. The license question is settled before the first tag rather than after, which is the only cheap moment it exists in. Package metadata, the NuGet listing, and the repository's own documentation can be made to agree on one answer.

What gets harder.

  • GitHub will classify the repository as "Other" instead of showing an MIT badge, and license-scanning tools in downstream vendors' pipelines will flag Sigil for manual review. That review is a conversation the licensor now has to be ready to have; it is the cost of the grant being bespoke.
  • Every new source file needs a header. Enforced rather than remembered — see Verification — but it is friction on file creation, and any code emitted by SigilBuild.Localization.Generator must carry the header from the emitter, because IDE0073 cannot see generator output.
  • External contributions stop. Bug reports still arrive; fixes do not.
  • Pre-2026-09-14 snapshots stay MIT forever. Nothing recovers them. Future discussions of "can we stop X from forking" must start from that fact rather than relitigate it.

Size budgets: unaffected. The change adds a two-line comment to 466 source files and touches no code path; the sigil.exe 15 MB and installer-host 45 MB gates are unmoved. Header comments do not survive compilation.

New lockstep surface. The license identity now appears in LICENSE, README.md, Directory.Build.props (Copyright, PackageLicenseFile, PackageRequireLicenseAcceptance), THIRD-PARTY-NOTICES.md, CONTRIBUTING.md, AGENTS.md, docs/architecture-overview.md, and the .editorconfig header template. Changing the license means changing all of them; this row belongs in AGENTS.md's lockstep table.

NuGet: PackageLicenseExpression accepts SPDX identifiers only, so a bespoke license must ship as PackageLicenseFile with PackageRequireLicenseAcceptance set. The existing SigilBuild 0.0.0-reserved listing records MIT in its metadata permanently — NuGet packages cannot be deleted, only unlisted — but it contains no code, and it should be unlisted once a real version ships.

Verification

  • IDE0073 is set to error with file_header_template in .editorconfig. Because Directory.Build.props sets TreatWarningsAsErrors=true and EnforceCodeStyleInBuild=true, a source file without the header fails the build. dotnet format applies the header automatically, and the dotnet format CI job is a required status check, so the header cannot rot on a merge path.
  • The .axaml headers and the localization generator's emitted header have no analyzer and are guarded by review against this ADR's Consequences section.
  • AGENTS.md carries the lockstep row, so an agent changing one license surface is told about the other eight.
  • No test asserts the license text, deliberately: a test that pins a legal document's wording would be edited by whoever changes the document, and proves nothing a reviewer would not catch.