[guernica0131]

Release PR for 0.2.0. Folds in #1 (security and conduct contacts at klv.ai), which is in this branch, so merging this marks it merged.

CI was failing

pytest -q died with an INTERNALERROR on every Python version before a test ran: pytest-asyncio 0.24+ warns when asyncio_default_fixture_loop_scope is unset, and filterwarnings = ["error"] made that fatal. Set explicitly. The workflows also move to the Node 24 majors of checkout and setup-python.

The live job has never actually run: there is no EXPERTS_TEST_KEY secret, so it skips and reports success.

Knowledge never worked

Every call except list_collections was against the wrong field or route, in 0.1 against 1.0.49 as much as against 1.1:

Call	0.1	0.2
upload	file + collection_uid → 422	files + collection on /documents/job-queue-upload; returns the UploadJob
wait_for_processing	polled a document row that has no "indexed" field → always timed out	polls the job; returns an UploadStatus (truthy when searchable; per-file uid or error)
search	conversation-context route, wrong body → 422	/similarity_search/; returns SearchHit passages
list_documents	collection_uid ignored; names/types empty	collection, and doc_name/doc_type/summary mapped

New: upload_status(), delete_document(), list_documents(skip=, search=).

1.1 behaviour

• Read-only licence window: LicenseReadOnlyError (subclass of LicenseError, read_only=True) for license_expired_readonly. Before, it came through as a plain permission error.
• Refused experts: /chat now 403s an expert the caller may not use. Raised as ExpertNotAvailableError (its {"detail": {...}} body used to read as "Request failed with status 403"). send() settles the placeholder, so the conversation does not show it typing forever.
• Guest visibility also requires no org-profile scope (Expert.scope_uids), matching what minting a session refuses.
• Mesh: QueuedEvent while a busy member's model has the request in line; served_by, mesh_fallback, mesh_unavailable and model on the result; stream_timeout client option.
• data_artifacts, generated_images; split_uid, start_ms, end_ms on source documents.

Fixes

• A failed streamed request (a refused chat turn, an upload) crashed with httpx.ResponseNotRead instead of raising the server's error. The unit tests' mocked responses were pre-read, so they never saw it.
• A timeout or dropped connection mid-answer raises StreamError, not httpx's own error.
• The knowledge service's 401 "insufficient permissions" is a permission error, not an auth error.
• cancel() returns whether the server stopped anything.

Breaking (0.x minor)

• upload() returns UploadJob, not KnowledgeDocument.
• wait_for_processing() takes the job and returns UploadStatus. Its truthiness keeps if wait_for_processing(...): working.
• search() returns SearchHits.

None of the three worked before.

Spec

spec/openapi.json is regenerated at 1.1.0. It adds the knowledge routes the SDK now calls and the 1.1 errors. The generator change is klv-ai/experts-knowledge sdk/openapi-1.1, and that file is byte-identical to this one.

Tested

• ruff, mypy strict, 100 unit tests. New tests/test_knowledge.py uses payloads copied from a 1.1 install.
• pytest tests/live: 10 passed against a Proxmox VM on 1.1.0-rc.127 (gateway rc.37, conversations rc.30, embedding rc.23). That includes a new upload → index → search → delete round trip.
• Not covered live: the read-only licence window, expert_not_available (the test key is an admin) and mesh events. Those are unit-tested from the services' code.

After merge

Tag v0.2.0 on main to publish to PyPI (release.yml checks the tag against both version strings).