Type Converse stream events from the :event-type header

[mattwebbio]

Problem

Follow-up to #14, which did not resolve the live failure it targeted: with extended/adaptive thinking enabled, a streamed Converse tool-call turn whose thinking block carries a signature but no text is rejected on the next round-trip of the same tool loop with Invalid \data` in `redacted_thinking` block`.

The reason #14 never engaged: real ConverseStream wire messages carry the event type in the eventstream :event-type header, and the JSON payload is the bare member struct. Captured from a live Bedrock ConverseStream response:

headers={":event-type"=>"contentBlockDelta", ...} payload={"contentBlockIndex":0,"delta":{"reasoningContent":{"text":"..."}},"p":"..."}
headers={":event-type"=>"messageStop", ...}       payload={"stopReason":"tool_use","p":"..."}

(This matches how the official AWS SDK parses event streams — aws-sdk-core/binary/event_parser.rb routes on the :event-type header.)

The streaming extractors instead dig by a nested event name ({'contentBlockDelta' => {...}}) — a shape only this repo's specs ever fabricated. Extractors with flat fallbacks (text/signature deltas via normalized_delta, finish_reason, tool-call start) worked by accident. The paths without fallbacks never matched a real event:

• track_thinking_delta's index lookup → thinking_state never populated
• contentBlockStop finalization → thinking.blocks never captured from any stream
• Finalize open thinking blocks on messageStop for Bedrock Converse streaming #14's event.key?('messageStop') safety net → dead code

So every streamed thinking turn persisted with blocks: nil, and a turn whose thinking had a signature but no text fell into format_single_thinking_block's redactedContent branch on replay — sending a signature where the API expects the opaque encrypted blob → 400.

Fix

1. decode_events re-nests each payload under its :event-type (or :exception-type) header, so the nested-shape paths engage on real traffic. Already-nested payloads (the old spec shape) pass through unchanged. Exception events now also match stream_error_event? by their proper name (e.g. throttlingException).
2. format_single_thinking_block replays a signature-only thinking turn as an empty-text reasoningText block, never redactedContent. A persisted signature is always a real reasoningText signature (redacted blobs are only ever captured into blocks), so this is also the correct path for legacy messages persisted before this fix.

Verification

• Live repro against Bedrock (us.anthropic.claude-sonnet-5, adaptive thinking + tool call, streaming): before this change thinking.blocks was nil on every turn; after, blocks are captured raw and the tool-loop replay succeeds end-to-end.
• New specs frame events exactly as the wire does (aws-eventstream Encoder, :event-type headers, flat payloads) and drive them through parse_stream_chunk — the previous specs bypassed decoding entirely, which is how Finalize open thinking blocks on messageStop for Bedrock Converse streaming #14 passed while real streams stayed broken.
• Full suite: 1641 examples, 0 failures. RuboCop clean on changed files.

🤖 Generated with Claude Code

[staging-supernova-dx-appf-io]

🤖 Supernova Code Review — View trace

[staging-supernova-dx-appf-io]

Reviewed the full diff and both changed files at the PR head. No actionable findings.

decode_events / nest_event_under_type — the re-nesting logic is correct. The idempotency guard (event.key?(type)) correctly passes through the old nested-shape events used by pre-existing unit tests that call build_chunk directly, while real wire events (flat payloads + :event-type header) get wrapped as the extractors expect. The rescue StandardError => nil in event_type_header is intentionally broad to stay safe in the face of malformed eventstream objects; the worst outcome is a silently-dropped un-nested event, not a crash.

format_single_thinking_block — the reasoning in the comment is sound: anything persisted into thinking.signature came from a reasoningContent signature delta, so it is always a real reasoningText signature. Using redactedContent there was always wrong; the new always-reasoningText path is both correct and backward-compatible for legacy-persisted messages.

finalize_thinking_block — correctly handles the signature-only case: text.empty? && signature.nil? is false when a signature is present, so an empty-string text field is preserved in the output block and compact leaves it intact. The spec at line 234 and the wire-framing spec at line 332 both cover this path.

New wire-framing specs — the wire_chunk / stream_to_message helpers go through the real aws-eventstream encoder and the full parse_stream_chunk → decode_events → nest_event_under_type → build_chunk path, which is exactly what was missing before. These tests are a meaningful improvement over the previous approach of calling build_chunk with hand-crafted nested hashes.

Nothing to request changes on. The fix is clean and well-tested.

---

Re-run this review

[mattwebbio]

The initial CI run failed on acts_as_model_spec.rb's registry-integration examples across most matrix jobs — a pre-existing order-dependent flake, not a regression from this change: the logger specs in ruby_llm_spec.rb nil out RubyLLM.@config and leave it nil, destroying the model_registry_source that ActsAs.included installs once at load time. Whenever test-queue schedules ruby_llm_spec.rb before acts_as_model_spec.rb in the same worker, those two examples fail (this PR's added specs merely reshuffled the scheduling). Deterministic repro:

rspec spec/ruby_llm_spec.rb spec/ruby_llm/active_record/acts_as_model_spec.rb --order defined

Fixed in 3a6c20d by saving/restoring the original config around each logger example.