You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
CI run https://github.com/micro/go-micro/actions/runs/28645508619 failed only in the Harness (E2E) workflow's Provider harnesses (live LLM conformance) job on master at c645c5faa7c0f6be53e7643db00451fed5349659; the mock harness job passed.
The live run configured only atlascloud (ATLASCLOUD_MODEL=minimaxai/minimax-m3); anthropic/openai/gemini/groq/mistral/together were skipped because their API keys were unset.
atlascloud passed agent, agent-flow, and a2a-stream-fallback, but failed universe and plan-delegate.
In internal/harness/universe, the checkout flow recovered from the intentional payment failure and resumed correctly through inventory, charge, and order confirmation. The failure happened at the concierge notification step: the notify service logged ignored non-buyer notification to=buyer-of-order-1 "Your order order-1 has been confirmed. Thank you for your purchase!", so the buyer notification count remained zero.
Because the notify step returned concierge completed without notifying buyer: notify count stayed at 0, the resumed flow stayed pending and the final persisted flow assertions failed (no pending runs and flow run ... marked done). This is a harness/code conformance defect in the concierge -> notify argument/recipient path, not a provider outage, rate limit, or network timeout.
Where:
internal/harness/universe checkout/concierge notify path and buyer-recipient validation.
The atlascloud/minimax universe harness reliably resumes checkout after the intentional payment retry and sends exactly one accepted buyer notification through the concierge agent.
The notify recipient accepted by the harness matches the expected buyer identity for the order; attempts with malformed/non-buyer recipients remain rejected.
A failed concierge notify must leave actionable diagnostics showing the expected recipient and actual tool arguments.
Add or update deterministic coverage around the resumed checkout notify step so a completed agent wrapper call cannot be mistaken for a successful buyer notification when the notify service rejects the recipient.
A follow-up Harness (E2E) provider-conformance run passes atlascloud / universe and leaves no pending checkout flow run.
Root cause:
Harness (E2E)workflow'sProvider harnesses (live LLM conformance)job onmasteratc645c5faa7c0f6be53e7643db00451fed5349659; the mock harness job passed.ATLASCLOUD_MODEL=minimaxai/minimax-m3); anthropic/openai/gemini/groq/mistral/together were skipped because their API keys were unset.agent,agent-flow, anda2a-stream-fallback, but faileduniverseandplan-delegate.internal/harness/universe, the checkout flow recovered from the intentional payment failure and resumed correctly through inventory, charge, and order confirmation. The failure happened at the concierge notification step: the notify service loggedignored non-buyer notification to=buyer-of-order-1 "Your order order-1 has been confirmed. Thank you for your purchase!", so the buyer notification count remained zero.concierge completed without notifying buyer: notify count stayed at 0, the resumed flow stayed pending and the final persisted flow assertions failed (no pending runsandflow run ... marked done). This is a harness/code conformance defect in the concierge -> notify argument/recipient path, not a provider outage, rate limit, or network timeout.Where:
internal/harness/universecheckout/concierge notify path and buyer-recipient validation.Acceptance criteria:
Harness (E2E)provider-conformance run passesatlascloud / universeand leaves no pending checkout flow run.