[rudycelekli]

What this fixes

AgentCore's RecallPolicy(mode="reflect") receives the SDK's ReflectResponse, whose response field is text. The adapter reads answer instead. A successful real reflection therefore raises AttributeError, is caught by the existing graceful-degradation handler, and produces empty memory context.

Read resp.text and preserve empty-response behavior. Replace the existing fabricated MagicMock.answer test with the actual generated SDK model. Add Unicode/empty-text controls and a loopback HTTP regression through the real Hindsight Python SDK, asserting both the returned context and reflection request fields.

Validation

• Current-main reproduction with final tests: 4 failed, 42 passed, 1 skipped.
• Full AgentCore integration suite: 46 passed, 1 skipped on Python 3.11.15 and 3.13.5. The existing manual cloud test remains skipped.
• Native tests exercise actual generated ReflectResponse; the HTTP test calls real Hindsight.areflect against a local server and closes the owned SDK session.
• Focused Ruff checks, Ruff format checks and git diff --check pass.
• Required ./scripts/hooks/lint.sh invoked offline stops before lint jobs because the sparse checkout lacks hindsight-all; full repository lint is not claimed green.
• No AWS deployment, live Hindsight server, LLM, full API/system suite or cloud credential was used. This fixes the existing integration response contract; it adds no API field, SDK capability or reflection-prompt behavior.

AI-assisted implementation and tests; source and regression evidence independently reviewed before publication.

[nicoloboschi]

Thanks! Pushed one small follow-up: the HTTP reflect test now lets the adapter build its own SDK client from the URL (instead of swapping in a hand-built one), so the real client path is covered too.