跳转到主要内容

ADR 0077 - XTL 1.0 ships with a single maintainer

  • Status: accepted (process gate)
  • Date: 2026-07-28
  • Spec target: XTL 1.0
  • Affects: GOVERNANCE.md ("Current state", "Roles", "How this document evolves"), ROADMAP gate G16

Context​

ROADMAP gate G16 asks for ≥ 2 people with accept/reject rights on ADRs and impl PRs before 1.0, with the fallback: "explicit accept of single-maintainer 1.0 governance shape via amendment to GOVERNANCE."

GOVERNANCE.md already describes the project as single-maintainer, but describes it as a phase — "xl3 is in its formative phase", "the maintainer set grows when external contributors sustainably review and land changes". Read as a gate, that is ambiguous in the way that matters: someone deciding whether to depend on xl3 cannot tell whether 1.0 is blocked on recruiting a second maintainer, or whether 1.0 ships as-is.

This ADR removes the ambiguity by taking the fallback explicitly rather than leaving G16 open until a second maintainer happens to appear.

Decision​

XTL 1.0 ships with one maintainer. G16 is closed by its fallback, not by satisfying its primary criterion.

The clause in GOVERNANCE.md about the maintainer set growing stays live — this decision is about the 1.0 cut, not a statement that the project will always have one maintainer.

Why this is defensible here​

The risk G16 guards against is real: a spec with a single accepter can drift with that person, and a project with a bus factor of one can stop. What makes it acceptable for this project at 1.0 is that the contract does not live in the maintainer's head.

The specification is executable and independent. The corpus is the normative definition of behavior — see conformance/DASHBOARD.md for the live figures — and conformance/AUTHORING.md forbids capturing expected outputs from the reference implementation precisely so the corpus does not become "whatever the JS impl does". Every accepted behavior has an ADR recording its reasoning.

A second implementation already exists and passes. xl3-py clears Stage 1 (see the dashboard). That is the strongest available evidence that the corpus is sufficient to reimplement XTL from, by someone who is not the maintainer. A spec you can build a second implementation from survives its author in a way a single codebase does not.

Both artifacts are permissively licensed. Code MIT, spec CC-BY-4.0. A stalled project can be forked and continued without permission.

The surface cannot churn quietly. G3 and G6 freeze the error catalog and the public API against breaking change, and additions must be logged to spec/STABILITY.md. Those gates constrain the maintainer as much as anyone.

What this does not fix​

Stated plainly, because the mitigations above do not cover it:

  • Releases stop if the maintainer stops. Nobody else can publish to npm or cut a tag. A fork can continue the spec and the code; it cannot continue the package name.
  • Nobody else can accept an ADR. An external contributor can author one (that is G14), but acceptance is still one person.
  • Review latency is one person's availability. There is no second reviewer to route around an absence.

A user for whom any of those is disqualifying should treat xl3 as a fork-if-needed dependency rather than a managed one. That is the honest framing, and it is why this is recorded as a decision with consequences rather than a formality.

Revisit trigger​

Named so this does not sit open-ended:

  • Any external contributor who lands changes sustainably — the existing GOVERNANCE.md trigger, unchanged.
  • G14 closing is the natural moment to look again: someone who has authored an accepted ADR has demonstrated the judgment the role needs.
  • A second production user appearing (beyond G15's reference case) raises the cost of a bus factor of one enough to revisit regardless of contributor flow.

Consequences​

GOVERNANCE.md gains a short statement that 1.0 ships in this shape, pointing here for the reasoning. ROADMAP G16 closes via its fallback with the fallback named, so the table does not read as though the primary criterion was met.

No spec behavior changes; no fixture is affected. This is a process decision recorded the same way ADR-0043 and ADR-0048 record theirs.