feat: add orcarouter llm2 extension

[jinhaosong-source]

What this PR does

Adds an OrcaRouter LLM2 provider extension: orcarouter_llm2_python.

OrcaRouter is an OpenAI-compatible meta-router — a single endpoint over many upstream models that also layers gateway-level zero-trust security for AI agents (scoped keys, guardrails, tool allow-lists, audit trail) and a routing DSL for per-request upstream selection. Because it speaks the OpenAI Chat Completions protocol, the extension mirrors gemini_llm2_python / openai_llm2_python: it reuses the OpenAI SDK and only changes the defaults (base_url → https://api.orcarouter.ai/v1, model → orcarouter/auto).

Disclosure: I'm an engineer on the OrcaRouter team.

Included

• ai_agents/agents/ten_packages/extension/orcarouter_llm2_python/ — the extension (manifest.json, property.json, addon.py, extension.py, orcarouter.py, think_parser.py, __init__.py, pyproject.toml, requirements.txt, README.md).
• tests/ — pure-logic unit tests for the streaming <think> reasoning parser (run anywhere) and config-default tests (skipped where ten_ai_base is unavailable), plus the standard conftest.py harness and config fixtures.
• Registered in the voice-assistant example dependency list (additive — the default graph is unchanged).
• ORCAROUTER_* documented in ai_agents/.env.example.

Features

• Streaming, tool/function calling, and reasoning (<think> blocks and reasoning_content) work unchanged via the OpenAI-compatible protocol.
• model="orcarouter/auto" lets OrcaRouter pick a live model per request (strategy configured in the OrcaRouter console).
• Model-fallback chain via request parameters: {"models": ["openai/gpt-4o", "anthropic/claude-haiku-4.5"], "route": "fallback"}.
• custom_headers forwards HTTP-Referer / X-Title for console attribution.

How I verified

Beyond the unit tests, I ran the extension's exact client path (OpenAI SDK with base_url + default_headers) against the live OrcaRouter API:

• Streaming — model="openai/gpt-4o" streamed content deltas correctly.
• Tool calling — a get_weather tool was invoked as expected.
• Fallback — extra_body={"models": [...], "route": "fallback"} resolved to a live model and returned a completion.
• Reasoning params — confirmed OrcaRouter normalizes provider-specific parameter constraints server-side (sampling params like temperature are accepted even when a request is routed to a reasoning model), so the extension sends the standard OpenAI parameters unconditionally.

Notes

• No changes to framework internals; this is a self-contained extension plus additive example/env wiring.
• Conventional-commit title; commits authored under my GitHub account for the CLA.