- Keep real model streaming checks in the manual workflow
.github/workflows/live-provider-streaming.yml. - Do not move these provider-backed checks into the always-on
ci.ymlpath unless the user explicitly asks for spendier default coverage. - The manual workflow is the canonical way to prove token streaming against real providers.
- OpenAI-compatible:
- repo variable
OPENAI_COMPAT_MODEL - repo variable
OPENAI_COMPAT_BASE_URL - repo secret
OPENAI_COMPAT_API_KEY
- repo variable
- Anthropic-compatible:
- repo variable
ANTHROPIC_COMPAT_MODEL - repo variable
ANTHROPIC_COMPAT_BASE_URL - repo secret
ANTHROPIC_COMPAT_API_KEY
- repo variable
workflow_dispatchinputs may override the model and base URL for a single run.- API keys stay in GitHub secrets only; do not add dispatch inputs for secrets.
- The workflow runs
tests/integration/test_live_provider_streaming.pyagainst manifest-registered graphs inexamples/live_provider_graphs/manifest.json. - The proof target is SSE
message_chunkevents from real provider-backed graphs, not just a final successful response.