trimDanglingAssistantToolCalls in the provider router held the same
consecutive-tool-message assumption as the replay projector: when a
model emits function_call items before explanation text within one
response (e.g. gpt-5.3-codex-spark), the sanitized sequence becomes
assistant[tool_call] -> assistant[text] -> tool[result] and responded
calls were stripped at the provider boundary even though the projector
already replayed them correctly. The orphan tool result then forced the
Responses adapter to synthesize a placeholder function_call with empty
arguments, degrading the arguments the model sees on in-loop requests.
Widen the response collection window to skip interleaved plain
assistant text messages and drop orphan tool results, mirroring the
projector fix.
上玄