lsm / neokai / 25591566667 / 15
82%
dev: 82%

DEFAULT BRANCH: dev
Ran
Files 278
Run time 7s
Badge
Embed ▾
README BADGES
x

If you need to use a raster PNG badge, change the '.svg' to '.png' in the link

Markdown

Textile

RDoc

HTML

Rst
coverage: 22.698% (-0.03%) from 22.729%
25591566667.15

push

github

web-flow 
fix: forward image attachments through OpenAI Responses bridge (#1831)

* fix: forward image attachments through OpenAI Responses bridge

The OpenAI (Codex) provider was silently dropping image content blocks
sent by the SDK. The appendUserBlocks function only handled text and
tool_result blocks, ignoring image blocks entirely.

Changes:
- Add AnthropicContentBlockImage to the translator types
- Extend ResponsesInputItem to support input_image content blocks
- Convert image blocks to OpenAI Responses API input_image format
  (data URL: data:image/jpeg;base64,...)
- Update token estimation to account for image content
- Add unit tests for image translation and end-to-end forwarding

* fix: address review feedback on image attachment bridge

- Preserve source-order of interleaved text/image blocks in user messages
- Support URL-based image sources (source.type: 'url') in addition to base64
- Replace base64-length token heuristic with fixed per-image estimate (100)
  that aligns with OpenAI's vision tokenization (~85 for low detail)
- Set ProviderCapabilities.vision = true in AnthropicToCodexBridgeProvider

* fix: handle non-text tool_result blocks, validate image sources, reject unsupported blocks

- toolResultText now preserves non-text blocks (images) as placeholders
  instead of producing undefined via block.text
- imageBlockToInputImage validates source.type and throws for unknown
  types (e.g. 'file') instead of falling through to invalid data URLs
- appendUserBlocks throws on unsupported user block types (document,
  etc.) instead of silently dropping them
- Increase ESTIMATED_IMAGE_TOKENS from 100 to 300 to cover the common
  auto/high-detail case
- Update AnthropicContentBlockToolResult type to allow image blocks in
  tool_result.content, and fix downstream consumers (gemini, ollama,
  token-estimator) to handle mixed content gracefully

* fix: catch translation errors, preserve multimodal tool outputs, return 400s

- Wrap buildResponsesRequest ... (continued)

16576 of 73027 relevant lines covered (22.7%)

11.8 hits per line

Source Files on job daemon-online-rewind-1 - 25591566667.15