ci: run the gateway integration tests as their own slow-tier check

[AlexS-scale]

What and why

CI did not run tst/integration/env/gateway/gateway_test.py: both integration jobs used --ignore on it. CONTRIBUTING.md gave the reason as "needs an x86 Chromium build that the hosted runners do not have". The hosted runners are x86, and the tests pass on them.

Without CI, two tests had become stale:

• test_snapshot_cli looked for the clean snapshot without the image fingerprint. Snapshots are keyed on that fingerprint, so a lookup without it finds only legacy rows, and it found nothing. The test now does the same lookup as MultiEnv (multi_env.py:215). Its dependent, test_load_from_snapshot, now runs too.
• test_env_validator_task deploys and prompts the default A2A agent. It now skips with the capability reason when there is no model endpoint or no default agent, as the same tests in task_steps_test.py do.

Also, test_gateway_triggers_nl_executor skipped each provider except modal_vm, with a free-form reason that the skip policy refuses. It is now parametrized on modal_vm only, behind the remote_sandbox gate.

The gateway tests take about as long as the remainder of the slow tier. Thus they run in parallel with it, as the new check integration-local-gateway. One job definition gives both checks through a matrix with name. Thus integration-local-slow keeps its name, and the Docker set-up is not copied. Each check keeps the slow tier's 40-minute job limit. On the reviewed run, integration-local-gateway took 16 min 48 s and integration-local-slow took 17 min 33 s, so CI takes approximately the same time as before.

Make integration-local-gateway a required check of the main ruleset directly after this pull request merges. Do not do it before the merge: other open pull requests do not report this check, so GitHub would block them.

How it was tested

On the reviewed CI run (run):

• integration-local-gateway: 24 passed, 441 skipped. Each skip has a capability reason: 376 remote_sandbox, 64 mcp_server_sources and 1 default_a2a_agent.
• integration-local-slow: 36 passed, 48 skipped.
• installer fails because of uv 0.13.1, on main too. test(installer): compare the tool's files before uv rebuilds it #199 fixes that.

On the current head (c2ca3a6), integration-local-gateway took 18 min 6 s and integration-local-slow took 18 min 54 s. Each check passes except installer, for the same reason (run).

Before the test fixes, the first run had the 2 failures above (run).

Confidence Score: 5/5

The PR appears safe to merge; no actionable issues were found.

Summary

Gateway integration tests now run beside the remaining slow tests as integration-local-gateway.

• CI runs gateway integration tests in a separate slow-tier check.
• Gateway tests use current snapshot keys and capability-based skips.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart TD
  A[Local backends workflow] --> B[integration-local: fast tests]
  A --> C[Shared slow-job definition]
  C --> D[integration-local-slow: other slow tests]
  C --> E[integration-local-gateway: gateway tests]
  D --> F[Check declared skips]
  E --> G[Check declared skips]

Loading

_{Reviews (3) · Last reviewed commit: "Merge remote-tracking branch 'origin/mai..." · Reviewed by Greptile}