lsm / HyperNeo / 31901879932

79%

push

github

test(daemon): enforce loopback-only fetch in unit tests (#2496)

* test(daemon): enforce loopback-only fetch in unit tests

CLAUDE.md documents that daemon unit tests 'never call real APIs', but
the setup only enforced that by clearing provider env keys. Availability
probes funnel through the global fetch with per-probe timeouts
(src/lib/providers/shared/credential-probe.ts), so any provider whose
availability check passes without an env key still burns real network
round-trips inside the per-test budget — the root of the model-service
flake family registered in flaky-tests.json and of yesterday's
model-service-provider-routing CI timeout.

Replace global fetch with a guard that allows loopback hosts (suites
that spin up local bridge servers keep working) and rejects everything
else with a pointing error. The original fetch is stashed as
__originalFetch, mirroring the existing __originalConsole stash.

* test(daemon): scope unit-fetch guard to test-spawned server ports

Address review feedback: a blanket loopback allowance is still
environment-dependent (a developer machine running Ollama on
localhost:11434 satisfies OllamaProvider.isAvailable() and loads real
local models), and native redirect-following would silently bypass the
host check on the second hop.

Patch net.Server.prototype.listen to record ports bound by servers the
tests spawn; fetch admits loopback only on those ports. Redirects are
followed manually so every hop is re-validated against the guard.

* test(daemon): address fetch-guard review round 2

- import node:net as an ES module instead of CommonJS require
- normalize bracketed IPv6 hostnames ('[::1]' vs '::1') before compare
- pass the original Request/input on the first hop so its method,
  headers, and body survive the guard
- honor caller-specified redirect:'manual'/'error' semantics; only the
  default follow mode uses the guard's hop-by-hop re-validation
- make the Server.listen prototype patch idempotent across per-file
  se... (continued)

42619 of 57973 branches covered (73.52%)

Branch coverage included in aggregate %.

58083 of 69922 relevant lines covered (83.07%)

3868.01 hits per line