convert_messages() emitted reasoning items with ``content`` as a plain
string whenever preserve_reasoning was enabled (the DeepSeek spec).
DeepSeek's Responses gateway rejects that shape with a serde error
("input: invalid type: string ..., expected a sequence"), which surfaced
only after token consolidation cleared provider_state and forced the
full-history conversion path; replayed server items already carry list
content, which is why normal multi-turn requests never failed. Serialize
reasoning content as a list of output_text parts, matching the OpenAI
Responses schema and DeepSeek's accepted wire shape (verified live against
api.deepseek.com/responses).
The serde fallback classifier introduced in the previous commit remains as
a last-resort safeguard for any remaining wire incompatibility.
Tests: extend test_preserves_deepseek_reasoning_content to the array shape;
add a full-history regression with the observed failing item, a
replay/consolidation regression covering both replayed and converted
reasoning items, and provider-level request fixtures for both paths.
Full suite: 5773 passed, 22 skipped (only the known local-only
channels/sms packaging failure remains).
DeepSeek's new Responses endpoint (deepseek-v4-flash) intermittently rejects valid request bodies with serde deserialization errors such as 'input: invalid type: string ..., expected a sequence'. These were not classified as compatibility errors, so affected conversations died instead of falling back to Chat Completions.
The wire format is correct (input serializes as a list), so this is a server-side Responses compatibility issue; Chat Completions is strictly more permissive, making fallback safe. Extend the fallback classifier to recognize serde body-parsing markers. Repeated failures still trip the existing circuit breaker.
When the Responses API fails repeatedly (3 consecutive compatibility
errors), skip it and fall back directly to Chat Completions. Unlike a
permanent disable, the circuit re-probes after 5 minutes so recovery
is automatic when the API comes back. Success resets the counter.
Keyed per (model, reasoning_effort) so a failure with one model does
not affect others.
