Summary

This PR adds optional end-to-end OpenAI usage and cost tracking for presentation generation, including LLM and image
generation calls.

What’s included

• Added optional usage tracking flag in generate request:
  • track_openai_usage in GeneratePresentationRequest
• Extended generate response with usage summary:
  • openai_usage in PresentationPathAndEditPath
• Implemented context-based tracker:
  • servers/fastapi/services/openai_usage_tracker.py
• Added usage/cost response models:
  • servers/fastapi/models/openai_usage_cost.py
• Wired tracking into generation pipeline:
  • wraps generate_presentation_handler execution in tracker context
• Captures OpenAI usage from:
  • chat.completions (regular + structured + streaming with include_usage)
  • responses API (web search path)
  • images.generate
• Added pricing support:
  • built-in defaults for gpt-5.2 and gpt-image-1.5
  • env override via OPENAI_PRICING_JSON
  • unpriced model warnings and exclusion from total
• Added env helpers and docs:
  • TRACK_OPENAI_USAGE
  • OPENAI_PRICING_JSON
• Added tests:
  • servers/fastapi/tests/test_openai_usage_tracker.py
  • validates cached tokens, pricing override, unpriced models, image modal pricing

Additional fix

• Installed zstd in both Docker images:
  • Dockerfile
  • Dockerfile.dev
• Purpose: avoid runtime/container issues where zstd support is required by dependencies/tooling.

Validation

• uv run pytest -q tests/test_openai_usage_tracker.py → passed
• compile checks for modified FastAPI files → passed
• manual API generation call with track_openai_usage=true returned populated openai_usage with total USD cost and
  per-model breakdown.

Backward compatibility

• Tracking is optional and disabled by default unless enabled by:
  • request field track_openai_usage=true, or
  • env TRACK_OPENAI_USAGE=true.
• Existing clients remain compatible; new response field is optional.