fix(vision): route image analysis through active thread model

[chromalinx]

What this PR does

Passes the active provider/model through tool dispatch so vision_analyze follows the current conversation backend instead of forcing the auxiliary vision task defaults. Keeps gateway thread model switches consistent for image analysis while preserving explicit model/provider overrides.

Changes

• model_tools.handle_function_call accepts and forwards provider/model kwargs to registry.dispatch.
• AIAgent._tool_routing_kwargs() centralises the provider=self.provider or None, model=self.model or None triple that previously appeared at every _invoke_tool call site.
• tools/vision_tools.vision_analyze_tool accepts a new provider parameter and threads it into the async_call_llm call_kwargs.
• _handle_vision_analyze extracts provider/model from **kw so the agent's active thread backend reaches the tool.

A regression test (tests/tools/test_vision_thread_routing.py) covers provider/model pass-through, kwargs extraction, and the AUXILIARY_VISION_MODEL fallback.

4 files changed, +150 / -2.

[alt-glitch]

Related: this re-implements the approach of closed PR #27015 (same title, same goal of routing vision_analyze through the active thread provider/model). Worth comparing against why #27015 was closed before re-merging.

[chromalinx]

@alt-glitch Closed and recreated because the old branch had picked up unrelated changes from a local WIP branch (voice-module) — the diff was 2200+ files and unmergeable in any clean form. The new branch is rebased directly on main and contains only the vision routing changes. The core idea is the same, but the review surface is now 4 files / +150 lines instead of 2200 files. Let me know if you spot anything from #27015 that this PR regressed.

[chromalinx]

@alt-glitch gentle ping — addressed all feedback. Branch rebased on current main, only the 2 files needed (tools/vision_tool.py + tests/vision/). Original #27015 closed because the old branch had picked up unrelated WIP. No bundling.

Test fixes pushed to the PR branches. Happy to rebase/adjust if anything still blocks merge.

[teknium1]

Thanks for the focused follow-up and the regression coverage.

Automated hermes-sweeper review found this behavior is already implemented on current main through the newer vision-routing architecture:

• tools/vision_tools.py:832 selects the active runtime model and returns native multimodal image pixels for vision-capable thread models; this shipped in 3800972dd05eabed8d75bfc4c0f5d532d85dafe2.
• agent/turn_context.py:154 records the live AIAgent provider/model before the turn, so routing sees session runtime rather than the persisted config default.
• Gateway per-session model overrides are resolved before image routing in 769469a703d5d76e3d8d6fc10d07196a78cb52ab (released in v2026.7.7), with regression coverage at tests/gateway/test_image_input_routing_runtime.py:50.
• For text-only models, current documented behavior intentionally uses the configured auxiliary.vision backend (website/docs/user-guide/features/vision.md:210), rather than sending images to the text-only active model.

The PR's dispatcher-based patch no longer applies after the subsequent tool-execution and vision refactors. Closing as implemented on main.