400: reasoning.summary not stripped for gpt-5.3-codex-spark

[salty-flower]

Problem

When Claude Code sends a request with reasoning.summary parameter to the proxy, and the target model is gpt-5.3-codex-spark via a ChatGPT/Codex OAuth profile, the parameter is forwarded as-is to the Codex Responses endpoint, which rejects it:

API Error: 400 Unsupported parameter: 'reasoning.summary' is not supported with the 'gpt-5.3-codex-spark' model.

Expected behavior

claudex should strip reasoning.summary (and potentially other Anthropic-specific thinking/reasoning parameters) before forwarding to the Codex Responses API, similar to how strip_params = "auto" handles temperature/top_p/top_k for chatgpt.com endpoints.

Setup

• claudex v0.9.49
• Profile: codex-sub (OpenAIResponses, ChatGPT OAuth)
• Claude Code v2.1.168 with per-model endpoint routing (custom patch)
• ANTHROPIC_MODEL_BASE_URL_gpt_5_3_codex_spark pointed at claudex proxy
• Claude Code effort level: high (default)

Reproduction

claudex-config proxy start -p 51456 &

curl -s "http://127.0.0.1:51456/proxy/codex-sub/v1/messages" \
  -H "content-type: application/json" \
  -H "x-api-key: passthrough" \
  -H "anthropic-version: 2023-06-01" \
  -d '{
    "model": "gpt-5.3-codex-spark",
    "max_tokens": 100,
    "stream": true,
    "reasoning": {"summary": "auto"},
    "messages": [{"role": "user", "content": "hi"}]
  }'

[salty-flower]

Closing — curl reproduction works fine (claudex strips reasoning.summary for small payloads). The 400 only occurs with large Claude Code requests (~76KB body). Will reopen with a better reproduction once I can capture the exact request body that triggers the bug.

[salty-flower]

Reopening with actual evidence from the error dump.

claudex does capture the translated request in ~/Library/Caches/claudex/errors/. The dump shows the Responses API request body still contains:

"reasoning": {
  "effort": "high",
  "summary": "detailed"
}

The summary field is what Codex rejects. My earlier curl test passed because I used "summary": "auto" — claudex may strip "auto" but not "detailed".

The original Anthropic request from Claude Code uses thinking with budget_tokens at effort level "high". The reasoning.effort mapping looks correct, but reasoning.summary should be stripped for models that don't support it (like gpt-5.3-codex-spark).

Error dump path: ~/Library/Caches/claudex/errors/20260609-195314.938-39869-codex-sub-openai_error.json

[salty-flower]

Minimal reproduction:

curl -s "http://127.0.0.1:51456/proxy/codex-sub/v1/messages" \
  -H "content-type: application/json" \
  -H "x-api-key: passthrough" \
  -H "anthropic-version: 2023-06-01" \
  -d '{
    "model": "gpt-5.3-codex-spark",
    "max_tokens": 100,
    "stream": true,
    "thinking": {
      "type": "enabled",
      "budget_tokens": 10000
    },
    "messages": [{"role": "user", "content": "hi"}]
  }'

Response:

{"error":{"message":"Unsupported parameter: 'reasoning.summary' is not supported with the 'gpt-5.3-codex-spark' model.","type":"invalid_request_error"},"type":"error"}

The Anthropic thinking block gets translated to reasoning: {effort: "high", summary: "detailed"}. The summary field needs to be stripped for models that don't support it.