Multi-provider sync (fan-out)

[github-actions]

Multi-provider sync (fan-out)

specsync is tracker-agnostic, but Sync projects to exactly one provider per
run. A project that wants the same OpenSpec change visible in two places at once —
GitHub for humans and Beads for an agent task graph, say — must run specsync
twice and can never reconcile inbound state from both in one pass. Fan-out makes
"the same change, projected to every tracker you use" a single operation.

This is the tracker-agnostic principle taken to its natural end: not "pick a
tracker," but "keep all your trackers aligned with the spec at once."

Principle: one source of truth, many spokes (a star, never a mesh)

This is the line that keeps fan-out a sync problem and not a distributed-database
problem:

• OpenSpec is the single source of truth. Every provider is a projection of
  it.
• Every edge is OpenSpec↔provider. There is never a provider↔provider edge.
• Reconcile is N-inbound-to-1: done-state read from any provider merges into
  tasks.md by monotonic union; OpenSpec owns task wording/order.

Because each provider reconciles independently against the one authority —
exactly as today, just looped — fan-out needs no conflict-resolution model, no
stored base state, and no runtime. The instant state were allowed to flow
provider→provider (a mesh), specsync would need vector clocks / 3-way merge and a
running reconciler, forfeiting the invoke-and-exit property that makes it
adoptable anywhere. That door stays shut (see Non-goals).

What

• Sync accepts a set of providers instead of one.
• Provider-set selection via a repeatable -provider flag (and/or config), e.g.
  -provider github -provider beads; default remains github alone.
• Per change: reconcile inbound from every provider (union into tasks.md), then
  push the rendered item to every provider. .specsync/refs.json already keys by
  provider name, so each provider's ref coexists per change.
• Result reporting aggregates per-provider outcomes (created/updated, reconcile
  flips) per change.

Scope

• Repeatable -provider flag + config resolution.
• Sync loop over the provider set; aggregate Result across providers.
• Per-provider dry-run output.
• Failure isolation: one provider erroring must not corrupt another provider's
  refs or abort the rest of the run unceremoniously.

Non-goals

• No second source of truth. Providers receive wording/scope and contribute
  only done-state inbound; none becomes authoritative over OpenSpec.
• No spoke-to-spoke sync. State never flows provider→provider directly; it
  always passes through OpenSpec. No mesh, no CRDTs, no cross-tracker conflict
  resolution.
• No runtime. Still invoke-and-exit — no daemon, no watch loop, no persisted
  base state.
• Not a control plane and not a memory tool (inherited stance).

Tasks

Tasks: Multi-provider sync (fan-out)

• Make -provider repeatable (and config-resolvable); default to a single github when unset
• Change Sync to take a set of providers; loop projection + reconcile over each
• Reconcile inbound from every provider into tasks.md via the shared monotonic union (no provider↔provider path)
• Aggregate Result across providers (per-change, per-provider created/updated + reconcile flips)
• Per-provider dry-run output
• Failure isolation: one provider's error must not corrupt another's refs or abort the whole run silently
• Tests: a change fans out to two fake providers; inbound union from both; refs coexist in refs.json; one provider failing leaves the other's ref intact
• Docs: the fan-out model + the star-not-mesh / single-source-of-truth invariant