Responses streams report zero input usage and mask context overflow

[LPFchan]

Duplicate Check / 重复检查

• I have searched existing issues and this is not a duplicate / 我已搜索现有 issue，确认非重复

Claudex Version

0.2.4

Operating System

Linux (x86_64)

Description / 描述

The OpenAI Responses streaming translator always starts Anthropic usage at input_tokens: 0 and only forwards output_tokens at the end. Claude Code consequently records every translated assistant turn with zero input tokens, so its context accounting cannot trigger proactive auto-compaction reliably.

When the Responses backend eventually emits response.failed with error.code: context_length_exceeded, upstream Claudex silently finalizes the stream as a normal end_turn. Newer error-aware variants that classify it as a generic api_error can instead make Claude Code enter its non-streaming fallback; ChatGPT's Codex Responses endpoint then rejects that fallback with Stream must be set to true, masking the original context overflow.

This is distinct from #5 / #6: those cover the proxy's HTTP request-body limit, while this report concerns translated token usage and terminal SSE error semantics.

Steps to Reproduce / 复现步骤

1. Configure an OpenAIResponses profile, such as ChatGPT Codex OAuth.
2. Run a long Claude Code session through Claudex until its upstream context window is full.
3. Inspect Claude Code's transcript usage: translated assistant turns report input_tokens: 0.
4. Observe the upstream response.failed event with context_length_exceeded; the client does not receive a recognizable Anthropic invalid_request_error and does not compact normally.

Expected Behavior / 预期行为

• Terminal Responses usage is propagated through the final Anthropic message_delta, including input and cached-input accounting.
• context_length_exceeded becomes one terminal Anthropic invalid_request_error with no fabricated message_stop, allowing Claude Code's context-limit recovery to run.

Actual Behavior / 实际行为

• Input usage remains zero for the entire session.
• Context overflow is ignored or masked by a secondary non-streaming error.

Configuration / 配置

[[profiles]]
provider_type = "OpenAIResponses"
auth_type = "oauth"
oauth_provider = "chatgpt"

Proxy Logs / 代理日志

streaming=true model=<model>
Responses stream terminated with an error upstream_code=context_length_exceeded
streaming=false model=<model>
status=400 Bad Request body={"detail":"Stream must be set to true"}

Additional Context / 补充信息

Anthropic's TypeScript SDK merges input_tokens, cache_creation_input_tokens, and cache_read_input_tokens supplied in message_delta.usage, so the terminal Responses usage can be forwarded without buffering the stream.