Testing

Testing

Two codebases, two doctrines that fit their subject. The any server is tested by running the real binary against a real network — unit tests for the parts, then end-to-end suites that boot one or two servers and drive every endpoint over HTTP. The anybao runtime is tested around one property: everything nondeterministic is an effect, so a recorded trace replays the whole system deterministically, and every test layer exists to keep that cheap to assert.

Testing against any

make test                                   # unit tests with the fts vector tags
go test ./internal/e2e -run TestE2E_FullFlow # boot the binary, drive every endpoint
go test ./internal/e2e -run Multipeer        # two peers, invite/join, CRDT convergence

• Unit tests live next to the code; the search index legs are behind build tags, so make test (not bare go test) runs the full suite.
• End-to-end tests build the binary, run any run against a temporary data dir on an ephemeral port, and talk JSON over real TCP. They need a staging node configuration and skip without one — a test must never join production.
• Multi-peer tests start two servers with distinct accounts (or the same mnemonic on two device keys) and prove what a single process cannot: invites, membership, cross-replica convergence, LAN discovery.

Testing against anybao

cargo test --manifest-path runtime/Cargo.toml   # runtime logic + replay determinism
uv run pytest                                   # guest programs under the real kernel, offline
make test-integration                           # opt-in, against a live any server

• The Rust runtime's policy — effect boundary, capabilities, trace and replay, deploy, triggers — is unit-tested with fakes.
• Guest programs run under the real guest kernel host-side, with only the effect boundary faked, so an import that the sandbox forbids fails in the test exactly as it would in the wasm guest.
• The runtime binary is driven end-to-end as a subprocess against a stdlib fake that plays both the any server and the LLM provider.
• Integration and live-eval layers are opt-in and point at a real server.

Why it matters. A local-first system's hardest bugs are convergence bugs — two replicas, one network partition, one merge. Faking the sync layer would hide exactly those, so the server suites pay for real peers. And an agent's hardest bugs are non-reproducible ones — so the runtime makes every run a trace that replays bit-for-bit.

What to write when