The agentlink domain.

agentlink gives every AI coding agent in a repository the same brain, without copying a single file. It does that by exploiting a fact about the 2026 ecosystem: for the resources that matter most, the format has already converged and only the location differs.

Instructions converged on AGENTS.md (Linux Foundation, December 2025).
Skills converged on Agent Skills / SKILL.md (open specification, December 2025), read by 30+ tools.

So each pairing of an agent with a resource resolves to one of four verdicts:

Verdict	Meaning	Cost
`Native`	the tool already reads the canonical path	zero
`Link`	same format, different path	one link
`Import`	link impossible; write a one-line stub	one small file
[`plan::Blocked`]	needs a human decision	nothing is written

The payoff is that bidirectional propagation is not a feature anyone has to implement. For a Native or Link verdict there is exactly one inode, so editing, renaming, moving or deleting through any agent's path is immediately visible to every other agent — with no daemon, no watcher and no reconciliation pass.

Layering

agentlink-cli    composition root, human-facing output
      ↓
agentlink-domain   this crate: pure domain, no `std::fs`
      ↑
agentlink-fs     adapter implementing [`Workspace`] against a real directory

Everything here is deterministic and testable without a filesystem: planning is a pure function of observable state, and [testing::FakeWorkspace] simulates hosts — including Windows without symlink privileges — that a given CI runner cannot otherwise reproduce.

agentlink-domain 0.0.1

Layering