[WH-2099]

Summary

Closes #4033.

Blocking SDK aggregation loses the OpenAI Responses snapshot, so the next tool turn omits encrypted reasoning, message phase, item IDs, and ordering.
Adopt the SDK fix and cover both invocation modes through the real OpenAI client and SDK Session.
The OpenAI-compatible Dify endpoint extracts visible text from content blocks for its public response format.

Release Notes

• Preserve complete Responses output through model invocations and subsequent Agent tool turns.
• Retain encrypted reasoning, phase, item IDs, call IDs, and ordering.
• Return visible text from content blocks through the compatible endpoint.
• Use the SDK collector that retains complete tool calls and final usage snapshots.

Change Type

• LLM plugin
• Non-LLM plugin (endpoint extension)

LLM Plugin Checklist

• Message flow
• Tool interaction flow
• Reasoning replay
• Token consumption metrics

Versions and release gates

OpenAI is 1.0.9 and the endpoint extension is 0.0.13.
Both declare SDK >=0.11,<0.12 and lock published review commit 0fa060a11fd774fe8ed70c915d7881e20a765b64 from SDK #390.

Coordinate with provider/Agent #4034, Dify #44502, its Graphon candidate, and daemon #828.
Replace Git sources with the selected published SDK release and set minimum Dify versions before production rollout.
Keep this PR in draft until those release requirements are met.

Testing

• Fresh Python 3.12 environments installed frozen dependencies from the committed locks.
• OpenAI: 319 local tests passed; live tests were excluded.
• Endpoint extension: 2 tests passed.
• Controlled Vertex/OpenAI provider-to-Agent checks also covered Function Calling and Self Refine in streaming and blocking modes, verifying two tools, original replay data, and final usage.

Real vendor signature acceptance and production deployment remain untested.