Fix a test that had been failing for two weeks, and run the suite in CI

[michalakmarcinth]

Follow-up to #26. Adding that docs job was the first time anything in this repo ran a test — release.yml only publishes. Pointing it at the whole suite turned up four failures on a clean checkout, and they weren't the same kind of thing.

Split out: the follow-on work that un-skips the framework integration tests (and fixes three more stale ones it uncovered) is #28, stacked on this branch.

Three were missing optional extras

test_span_tree.py and test_integrations.py import agentx.integrations.google_adk at call time. That module raises ImportError by design when google-adk is absent, so the tests failed instead of skipping — while their crewai, litellm, llamaindex and autogen neighbours in the same files guard the identical situation with pytest.importorskip.

They now do too. With the extra installed all three run and pass, so this skips what genuinely isn't installed rather than hiding anything.

The fourth was real — and stale by two weeks

test_trace_tool_call_emits_real_child_span asserted the root trace carries no flat tool_calls list when a tool call became a real child span:

assert "tool_calls" not in root or root.get("tool_calls") in (None, [])

That was true when the test was written on 2026-08-07. Four days later, 40c6f6e deliberately added the dual-write — the code says why:

The child span above is only for the trace detail's span tree — the engine's built-in "Tool failure" check and the dashboard's Tool quality column read the ROOT trace's flat tool_calls list, so a summary lands there too... the reason a failed trace_tool_call() used to be invisible to both surfaces.

The behaviour is right; the assertion was stale. It had been failing since 2026-08-11 — unnoticed, because no CI ran it. "Fixing" the code to satisfy it would have re-broken the Tool failure check and the Tool quality column.

So the assertion is corrected, and corrected upward:

assert [tc["name"] for tc in root["tool_calls"]] == ["policy_lookup"]
assert root["tool_calls"][0]["output"] == "digital purchases are final"

It now pins the dual-write positively and pins it at exactly one entry. Verified against a modified tree: removing the dual-write fails it, and writing a third copy fails it too — neither of which the old assertion could catch, since it only checked for absence.

Widening the CI

docs.yml becomes tests.yml and runs the whole suite. The old name stopped being true once it covered more than the docs.

Install	Result
base (.[langchain])	73 passed, 28 skipped, 0 failed
with google-adk	76 passed, 25 skipped, 0 failed

Optional-extra tests skip themselves, so the base install is green — a skip is honest about not being covered, where a failure would have drowned out real ones. #28 takes the next step and actually runs them. test_integration.py talks to the hosted API and skips itself without a key; no secret is passed on purpose, so fork PRs behave the same.

Verification

• Exact CI command run on both paths — with and without google-adk
• The 3 adk tests confirmed to run and pass with the extra, and to skip with a clear reason without it (not silently vanish)
• The corrected assertion confirmed to fail in both directions it's meant to catch
• tests.yml validated as YAML
• Eval-skill audit re-run against the freshly published 0.6.37 — clean, 29 breaks still caught, no regression from the release

Tests and CI only; no library code changed, so no version bump is warranted by the diff.

🤖 Generated with Claude Code

https://claude.ai/code/session_01WYi6akpDa8qfiojRed2cAE