ADR 0001 - Clean-room derivation from permitted sources

Status

Accepted. This decision predates any code and constrains all of it.

Context

The contracts Relay implements are already public. The SDKs that speak the executor protocol are MIT licensed and contain the whole client side of it. The HTTP API is published as an OpenAPI document which, on the evidence of the public documentation, is served without authentication. The command-line client that consumes it is open source, though its exact licence is an open question until someone reads the file. Both of those are confirmed during discovery before anything depends on them.

The server that sits between them is proprietary, and its licence appears to prohibit reverse engineering, to prohibit using confidential information to design similar software, and to treat benchmark results as confidential. Free test and development use appears to be available, which is what makes running it tempting as an oracle. These readings come from the public licence page and have not been reviewed by a lawyer; the decision below does not depend on which of them is exactly right.

Taking that path would be the single fastest way to make this project indefensible. The value of a compatible control plane is that people can deploy it without worrying about where it came from, and that value does not survive an ambiguous provenance story.

Decision

Relay is derived only from permitted sources. The canonical list of what is permitted and what is forbidden lives in the clean-room rules, which is the single copy in the project.

The proprietary server and console images are never downloaded, run, observed, or benchmarked, at any point, for any reason, including testing.

Two rules follow from that and are enforced rather than trusted:

• Cite or test. Every protocol or API fact reaching code or docs names its source, or is proven by a test against a real SDK. An uncited fact is a guess and is marked as one.
• Re-derive, do not copy. Non-trivial code is not copied from the SDKs even though MIT permits it. The protocol is written down as a specification first, and the implementation follows the specification. Small constants such as message type strings and field names are expected. Wire-type declarations (like those in messages.go) are transliterations from the MIT-licensed SDK, and this is permitted as long as no functional logic or behaviour is copied.

When a fact is only obtainable from a forbidden source, the gap is surfaced and a fallback is designed. Relay accepts a superset or degrades gracefully rather than guessing.

Consequences

• Development is slower. There is no oracle to diff against, so the protocol specification and the conformance suite carry weight that a reference implementation would otherwise carry.
• The specification is a first-class deliverable, not documentation of the code. It is written so that a reviewer can implement a test executor from it without opening an SDK, which is also what makes it reviewable.
• Testing is done against real SDK sample applications and a test double, both of which are permitted inputs.
• The rules are written into AGENTS.md and the clean-room rules because they bind agent sessions as much as people, and an agent will otherwise reach for the fastest source.
• The hooks cannot enforce this one. A pattern match cannot tell where an idea came from, so provenance is checked by the reviewer on every change, and the provenance ledger is the only durable evidence.
• This is engineering guidance. It is not legal advice, and independent legal review is a blocker on the first public release.