fix: capture reasoning blocks from streaming responses

[sjawhar]

Capture thinking and redacted_thinking blocks when rebuilding a streamed response for conversation history.

_reconstruct_response_from_stream_events had branches for text and tool_use only; the line this replaces read # thinking blocks are intentionally excluded from history with no recorded rationale (its origin is not reachable — the line predates this clone's grafted root). A streamed call therefore recorded a complete-looking response holding no reasoning content, while the non-streaming path already stored the same blocks verbatim via an unfiltered model_dump(). History was inconsistent between the two transports depending only on whether the client streamed.

Change

• thinking blocks accumulate thinking_delta text and record the signature_delta value. The signature arrives as a separate event from the text deltas, which is the failure mode a half-implementation would hit.
• redacted_thinking blocks keep their opaque data payload — dropping it loses the only record that a reasoning block was present.
• Block ordering is untouched, so reasoning blocks still precede visible text.

Why this is not a new data shape

Nothing under history/, debug/, ui/ or static/ references thinking at all. The event's consumers (passthrough_materialize/payloads.py, debug/service.py) handle the streaming and non-streaming events symmetrically, and already receive these blocks today from non-streaming calls. This makes streaming produce what non-streaming already produces.

Deliberately out of scope

The streaming path's record_inbound_response / record_outbound_response pass no body=, so request_logs.response_body stays NULL for streamed calls. Filling it would duplicate every streamed body into a second table under the 8 MB truncation rule, for no gain here. Separate concern.

Tests

Three cases added to TestReconstructResponseFromStreamEvents, written first and confirmed failing for the right reason (blocks dropped entirely — len([]) == 0):

• test_captures_thinking_text_and_signature — multi-delta text accumulation plus the separately-arriving signature.
• test_preserves_thinking_before_text_in_block_order — index ordering across mixed block types.
• test_captures_redacted_thinking_payload — opaque payload preserved.

uv run pytest ...::TestReconstructResponseFromStreamEvents → 11/11 pass.

Coverage is unit-level only. The tests construct the SDK's own pydantic event models (RawContentBlockStartEvent, ThinkingDelta, SignatureDelta), so a wrong block shape fails construction rather than passing a hand-rolled dict. What they do not prove is the live streaming path reaching this function — that link is exercised by the existing text/tool_use cases. MockAnthropicServer emits only hardcoded text/tool_use blocks, so an end-to-end mock_e2e case needs a new mock response type; that is a follow-on rather than scope creep here.

Pre-existing failures, not from this change

scripts/dev_checks.sh reports three. They reproduce identically on unmodified main under xdist (uv run -m pytest -q --no-cov -n 4, which is what dev_checks.sh runs by default) and all pass single-process, so they are parallelism-sensitive and unrelated to this diff:

• tests/luthien_cli/test_onboard.py::test_find_docker_ports_respects_env_vars
• tests/luthien_proxy/unit_tests/test_config_registry.py::TestResolveDefaults::test_default_value_when_no_overrides
• tests/luthien_proxy/unit_tests/test_config_registry.py::TestResolvePriority::test_db_ignored_for_non_db_settable

Changelog fragment: changelog.d/capture-streaming-thinking.md.