Summary

• replace the auto image-preprocess prompt in GatewayRunner._enrich_message_with_vision() with a concise summary-oriented prompt
• add an internal max_tokens override to vision_analyze_tool()
• make the gateway auto-preprocess path use max_tokens=500 without changing explicit vision_analyze call sites

Fixes #10809.

Why this shape

The issue is specifically about the automatic image pre-process path inflating latency and prompt size. Lowering the global vision_analyze_tool() default would also make explicit user-invoked vision analysis shorter, which is a separate product decision.

This PR keeps the explicit tool behavior intact and tightens only the automatic gateway path.

Validation

• pytest -q -o addopts= tests/tools/test_vision_tools.py -k "TestHandleVisionAnalyze or TestVisionAnalyzeTokenBudget"
  • 6 passed, 63 deselected
• pytest -q -o addopts= tests/gateway/test_vision_preprocess.py
  • 1 passed

Notes

• no docs updated because this only changes internal defaults for the eager gateway image-analysis path
• timeout behavior was left unchanged in this PR; that should stay a separate discussion