Bubblegum v0.0.2-alpha

Bubblegum v0.0.2-alpha

Highlights

• Added callable-only OCR backend support with opt-in privacy gating.
• Added VisionProvider abstraction and callable vision provider support.
• Added optional OpenAIVisionProvider backend without making OpenAI a mandatory dependency.
• Added public vision provider lifecycle APIs: configure_vision_provider(...) and clear_vision_provider().
• Added SDK-level screenshot-to-vision wiring with strict privacy and cost gates.
• Added sanitized OpenAI vision diagnostics.
• Added manual OpenAI vision provider usage example.

Vision and OCR posture

• OCR remains callable-only; no bundled real OCR dependency is included.
• Vision provider usage is optional and dependency-light.
• Provider-based screenshot vision requires:
  • grounding.enable_vision=true
  • privacy.send_screenshots=true
  • privacy.process_screenshots_for_vision=true
  • max_cost_level="high"
  • registered provider
  • screenshot bytes present
• Manual vision_candidates remain supported and are not cost-blocked.
• vision://... refs remain synthetic and non-executable.

Safety and privacy

• Raw screenshots, base64 payloads, API keys, and secrets are not persisted or exposed.
• OpenAI provider diagnostics are sanitized.
• Provider failures fail safe to empty candidates.

Validation

• Package validation passed.
• Strict validation passed.
• Build passed.
• Static benchmark: 12/12.
• Execution benchmark: 12/12.
• Targeted Phase 11J/11N/11R/11X tests passed.
• Public API, packaging extras, and package metadata tests passed.
• pytest --collect-only -q: 539 tests collected.

Known limitations

• PyPI/TestPyPI publishing is deferred.
• Playwright/Appium runtime smoke remains manual and non-CI-gated.
• Real OpenAI usage requires user-installed OpenAI SDK and OPENAI_API_KEY.
• vision://... refs are not adapter-executable yet.

Publishing status

• GitHub pre-release only.
• Package version: 0.0.2a0.