For your consideration some testing fake seams risks RunWield found during init

[gandazgul]

Question

when RunWield (https://github.com/gandazgul/runwield) initializes on a new repo it scans its tests for fake seams, places where the test should be using product owned machinery instead of a fake because you risk a green suite hiding a real problem. I havent examined the code to confirm these hence an issue and not a PR. Let me know if you would like me to dig deeper or if these are false positives.

Possible test-seam risks

1. crates/openab-core/src/dispatch.rs — MockDispatchTarget

   • Replaceable behavior: real AdapterRouter, SessionPool, session creation/reset, and ACP turn streaming are replaced by a recorder.
   • Why it may be product-owned: dispatch buffering and batching are OpenAB-owned orchestration. The real session pool and router coupling could break while these tests
     stay green.
   • Fixture environment: use a real AdapterRouter and SessionPool with a small local ACP stub process; keep only the platform ChatAdapter fake because the platform is an
     external boundary.
   • Confidence: medium.
   • Uncertain facts: whether existing higher-level tests already cover this exact dispatcher/router/session path.

2. src/acp_tunnel_source.rs — FakeTunnel

   • Replaceable behavior: real RootAcpTunnel, gateway tunnel registry lookup, detach/reattach behavior, and tool-list routing are simulated.
   • Why it may be product-owned: tunnel registry resolution and routing are OpenAB-owned, even though the remote MCP server is external.
   • Fixture environment: use the real tunnel registry with controlled in-process tunnel handles, then exercise discovery and call routing through AcpTunnelSource.
   • Confidence: low to medium.
   • Uncertain facts: how practical it is to construct real TunnelHandle fixtures without a live WebSocket.

3. crates/openab-core/src/acp/pool.rs — CountingRegistrar

   • Replaceable behavior: real FacadeRegistrar and SessionTokens are replaced by a counter.
   • Why it may be product-owned: session token mint/revoke lifecycle is OpenAB-owned security machinery.
   • Fixture environment: use real SessionTokens through FacadeRegistrar, then assert token resolution and revocation behavior.
   • Confidence: medium.
   • Uncertain facts: existing SessionTokens unit tests may already cover enough of the real registry behavior.