experimental/1.1.6

[jolovicdev]

Summary

This PR ships experimental/1.1.6 and includes the earlier multimodality foundation commit:

• 4d26a1daed951927f3d875f796d9f1bf427475aa (feat: add multimodal support and image generation tools)

On top of that base, this PR hardens streaming/tool parsing and multimodal/image payload behavior,
and bumps package metadata to 1.1.6.

What changed

• Enabled/kept tool-turn streaming support and hardened streamed tool-call assembly.
• Fixed streamed tool-call parsing edge cases:
• Prevented position-key collisions across parallel streamed tool calls.
• Restored message.tool_calls fallback when delta.tool_calls is empty.
• Fixed mixed delta+message argument duplication (no more concatenated JSON like }{ / duplicate
  object append).
• Added fallback to non-stream completion when adapter/provider does not support stream=True for
  tool-enabled turns.
• Kept structured schema path strict by default; preview streaming remains opt-in via
  structured_stream_mode="preview".
• Fixed multimodal input detection so generic typed dict lists (e.g. {"type":"event"}) are JSON-
  serialized instead of misclassified as multimodal blocks.
• Fixed image tool content safety:
• Redacts data:image/... URLs in ToolResult.content.
• Redacts b64_json in ToolResult.content.
• Preserves full payload in ToolResult.data.
• Updated docs (configuration, desk, worker, events, core-types) for multimodal + tool
  streaming + structured preview behavior.
• Version bump + lock refresh:
• pyproject.toml: 1.1.6
• uv.lock updated via uv sync

Compatibility

This is intended as a non-breaking hardening/QoL release from 1.1.5:

• No mandatory new params.
• Existing defaults preserved.
• Changes mainly correct previously buggy edge behavior (stream parsing, multimodal
  misclassification, oversized image tool content).

Verification

• Local checks: ruff, mypy, pytest coverage added for each regression path.
• Full suite reported green (177 passed).
• Live model probes executed against:
• deepseek/deepseek-chat
• openrouter/z-ai/glm-5
• openrouter/minimax/minimax-m2.5
• Live image roundtrip probe executed with:
• generation: openrouter/google/gemini-3-pro-image-preview
• understanding: openrouter/google/gemini-3-flash-preview

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 53e4472948

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Parse image URLs from list-based chat content

_extract_chat_image_response only extracts image data from message.images or a string message.content, so the chat-completion fallback returns {} when providers send images in OpenAI-style content blocks (for example, message.content as a list containing an image_url block). In that case generate_image/agenerate_image can report a successful tool call with an empty payload even though the model returned an image, which breaks image generation on compatible providers that use block-based content.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Include input_text blocks when building context summaries

message_summary_text only keeps block text when part.get("type") == "text", but this commit also accepts multimodal input_text blocks in render_input; when context-window summarization runs, those messages are reduced to [multimodal message] and lose the user’s actual text instructions. That can materially degrade retry behavior after context-limit errors because summaries omit key prompt content.

Useful? React with 👍 / 👎.