feat(models): add OrcaRouter provider

models/orcarouter/.difyignore

@@ -0,0 +1,7 @@
+.venv/
+__pycache__/
+*.py[cod]
+.pytest_cache/
+.ruff_cache/
+.env
+*.difypkg

models/orcarouter/.env.example

@@ -0,0 +1,3 @@
+INSTALL_METHOD=remote
+REMOTE_INSTALL_URL=debug-plugin.dify.dev:5003
+REMOTE_INSTALL_KEY=********-****-****-****-************

models/orcarouter/.python-version

@@ -0,0 +1 @@
+3.12

models/orcarouter/README.md

@@ -0,0 +1,54 @@
+# Overview
+
+OrcaRouter is an **OpenAI-compatible LLM gateway** that routes requests across 40+ upstream providers (OpenAI, Anthropic, Google, DeepSeek, Qwen, Grok, and more) to the cheapest or fastest path that can serve the model you asked for. Users pay **below list price**, and the dashboard tracks how much they save in real time.
+
+# Configuration
+
+After installation, sign up at [orcarouter.ai](https://www.orcarouter.ai), grab an API key from the [console](https://www.orcarouter.ai/console), and set it up under **Settings → Model Provider** in Dify.
+
+# Adaptive routing — `orcarouter/auto`
+
+The `orcarouter/auto` model is a virtual router that picks the best upstream per request. Configure its strategy from the [routing console](https://www.orcarouter.ai/console/routing). Available strategies:
+
+| Strategy | Behavior |
+|---|---|
+| `cheapest` | Lowest-priced upstream that can serve the request (default) |
+| `balanced` | Trades off price vs latency vs quality |
+| `quality` | Highest-quality upstream |
+| `adaptive` | Linear contextual bandit picks among candidates based on per-request features (prompt length, code/math/JSON density, declared `max_tokens` budget tier, MinHash-LSH similarity to recent traffic) |
+| `gated_adaptive` | Layers a task-difficulty score on top of `adaptive` — mundane prompts restricted to a "weak" model pool, hard prompts to a "strong" pool |
+
+**Why this matters**
+
+- **Self-tuning** — adaptive strategies learn from your own traffic; performance shifts when workload changes
+- **Microsecond overhead** — feature extraction and bandit are closed-form math; routing adds no measurable latency
+- **Workload-aware in one call** — same `orcarouter/auto` endpoint serves cheap summarization and premium code-refactor requests with no client-side dispatch logic
+- **Cost & reliability guardrails by construction** — reward explicitly penalizes cost, latency, rate-limit, and format failures
+- **Admin-tunable without redeploys** — strategies, pools, thresholds, and reward weights are changed from console, not client code
+
+# Fallback routing (`extra_body`)
+
+OrcaRouter supports an OpenAI-compatible extension to specify per-request fallback models. Use the **`Fallback models`** and **`Routing mode`** parameters in the model node:
+
+| Parameter | Example | Effect |
+|---|---|---|
+| `Fallback models` (string, JSON array) | `["openai/gpt-4o-mini", "openai/gpt-4o"]` | If the primary upstream fails, try these in order |
+| `Routing mode` | `fallback` | Activate the fallback list above |
+
+These are translated to the request body's `extra_body: {models, route}` key — see the [API reference](https://docs.orcarouter.ai).
+
+# Reasoning models
+
+Some models (OpenAI `o1`/`o3`/`gpt-5`, Anthropic `claude-opus-4.7`, DeepSeek `deepseek-reasoner`/`deepseek-r1`) expose reasoning controls:
+
+- **OpenAI o-style**: `reasoning_effort` (high / medium / low / minimal), `verbosity`, `exclude_reasoning_tokens`
+- **Anthropic thinking**: `enable_thinking`, `reasoning_budget` (token budget), `exclude_reasoning_tokens`
+- **DeepSeek r-style**: `exclude_reasoning_tokens` (the model reasons by default)
+
+These map onto the upstream provider's native reasoning protocol — OrcaRouter handles the translation.
+
+Reasoning models do not accept `temperature`. The Dify UI will show `temperature` only for non-reasoning models.
+
+# Pricing
+
+All prices are configured at the per-model level in this plugin and reflect the **effective price you pay through OrcaRouter** (which may be below the upstream's list price). See the live [models page](https://www.orcarouter.ai/models) or `GET https://www.orcarouter.ai/api/pricing` for the authoritative source.

models/orcarouter/main.py

@@ -0,0 +1,6 @@
+from dify_plugin import Plugin, DifyPluginEnv
+
+plugin = Plugin(DifyPluginEnv(MAX_REQUEST_TIMEOUT=120))
+
+if __name__ == "__main__":
+    plugin.run()

models/orcarouter/manifest.yaml

@@ -0,0 +1,30 @@
+meta:
+  arch:
+    - amd64
+    - arm64
+  runner:
+    entrypoint: main
+    language: python
+    version: "3.12"
+  version: 0.0.1
+name: orcarouter
+author: langgenius
+icon: orcarouter_square.svg
+description:
+  en_US: OrcaRouter is an OpenAI-compatible LLM gateway that routes requests across 40+ upstream providers (OpenAI, Anthropic, Google, DeepSeek, Qwen, Grok, and more) to the cheapest or fastest path, with adaptive routing, fallback chains, and pay-below-list-price billing.
+  zh_Hans: OrcaRouter 是兼容 OpenAI 协议的 LLM 网关，跨 40+ 上游 provider（OpenAI、Anthropic、Google、DeepSeek、Qwen、Grok 等）按最便宜或最快路径分发请求，内置自适应路由、降级链与低于官方价计费。
+  ja_JP: OrcaRouter は OpenAI 互換の LLM ゲートウェイで、40 以上の上流プロバイダー（OpenAI、Anthropic、Google、DeepSeek、Qwen、Grok など）への最安または最速のパスへリクエストをルーティングします。アダプティブルーティング、フォールバックチェーン、表示価格以下の課金を内蔵。
+  pt_BR: OrcaRouter é um gateway LLM compatível com OpenAI que roteia solicitações entre mais de 40 provedores upstream (OpenAI, Anthropic, Google, DeepSeek, Qwen, Grok e outros) pelo caminho mais barato ou mais rápido, com roteamento adaptativo, cadeias de fallback e preços abaixo da tabela.
+label:
+  en_US: OrcaRouter
+resource:
+  memory: 268435456
+  permission:
+    model:
+      enabled: false
+type: plugin
+plugins:
+  models:
+    - provider/orcarouter.yaml
+version: 0.0.1
+created_at: 2026-05-14T19:20:00.000000000-04:00

models/orcarouter/models/_endpoint_utils.py

@@ -0,0 +1,13 @@
+DEFAULT_ENDPOINT_URL = "https://api.orcarouter.ai/v1"
+
+
+def normalize_endpoint_url(credentials: dict) -> str:
+    """Resolve and normalize the API endpoint URL from credentials.
+
+    Strips leading/trailing whitespace and trailing slashes from the URL.
+    Falls back to DEFAULT_ENDPOINT_URL when the value is missing or blank.
+    """
+    endpoint_url = (
+        (credentials.get("endpoint_url") or DEFAULT_ENDPOINT_URL).strip().rstrip("/")
+    )
+    return endpoint_url or DEFAULT_ENDPOINT_URL

models/orcarouter/models/llm/_position.yaml

@@ -0,0 +1,107 @@
+- orcarouter/auto
+- openai/gpt-5.5-2026-04-23
+- openai/gpt-3.5-turbo-1106
+- openai/gpt-4.1-mini-2025-04-14
+- qwen/qwen3-vl-235b-a22b-thinking
+- qwen/qwen3.6-plus
+- z-ai/glm-5
+- openai/gpt-3.5-turbo-0125
+- openai/gpt-4o-mini-search-preview-2025-03-11
+- openai/gpt-5-chat-latest
+- openai/gpt-5.4-mini-2026-03-17
+- z-ai/glm-4.5
+- kimi/kimi-k2.6
+- openai/gpt-5.2-2025-12-11
+- qwen/qwen3-vl-235b-a22b-instruct
+- openai/gpt-4-0613
+- openai/gpt-5.4-mini
+- openai/gpt-5.4-nano-2026-03-17
+- z-ai/glm-4.6
+- minimax/minimax-m2.7-highspeed
+- openai/gpt-4-turbo-2024-04-09
+- openai/gpt-4.1
+- google/gemini-3-pro-preview
+- openai/gpt-4o-2024-05-13
+- openai/gpt-5.2-chat-latest
+- qwen/qwen3.5-27b
+- anthropic/claude-opus-4.7
+- openai/gpt-4.1-nano
+- openai/gpt-5.1
+- google/gemini-2.5-flash-lite
+- google/gemini-2.5-pro
+- google/gemini-flash-latest
+- openai/gpt-3.5-turbo
+- openai/gpt-5-nano
+- qwen/qwen3.6-flash
+- openai/gpt-4o-2024-11-20
+- openai/gpt-5
+- google/gemini-3.1-flash-lite-preview
+- minimax/minimax-m2.7
+- anthropic/claude-opus-4.5
+- google/gemma-4-31b-it
+- openai/gpt-4o-mini-search-preview
+- openai/gpt-5-2025-08-07
+- openai/gpt-5.1-chat-latest
+- qwen/qwen3.5-plus-2026-02-15
+- google/gemini-robotics-er-1.6-preview
+- openai/gpt-3.5-turbo-16k
+- openai/gpt-4o-mini-2024-07-18
+- openai/gpt-5.4-nano
+- openai/gpt-4o-2024-08-06
+- openai/gpt-5.2
+- qwen/qwen3-max
+- anthropic/claude-opus-4
+- google/gemini-pro-latest
+- openai/gpt-4-turbo
+- qwen/qwen3-vl-8b-instruct
+- qwen/qwen3.5-flash
+- z-ai/glm-5.1
+- openai/gpt-4.1-2025-04-14
+- openai/gpt-4.1-mini
+- openai/gpt-5-mini
+- qwen/qwen3-max-preview
+- anthropic/claude-opus-4.1
+- anthropic/claude-sonnet-4.5
+- deepseek/deepseek-v4-flash
+- qwen/qwen3.6-plus-2026-04-02
+- anthropic/claude-sonnet-4.6
+- minimax/minimax-m2.5
+- openai/gpt-5-search-api-2025-10-14
+- qwen/qwen3.5-plus
+- qwen/qwen3.6-35b-a3b
+- deepseek/deepseek-chat
+- google/gemini-flash-lite-latest
+- google/gemini-2.5-flash
+- google/gemini-3-flash-preview
+- openai/gpt-4.1-nano-2025-04-14
+- qwen/qwen3.6-flash-2026-04-16
+- minimax/minimax-m2.5-highspeed
+- qwen/qwen3-vl-8b-thinking
+- anthropic/claude-haiku-4.5
+- kimi/kimi-k2.5
+- openai/gpt-5.5
+- qwen/qwen3.5-397b-a17b
+- anthropic/claude-sonnet-4
+- google/gemini-3.1-pro-preview-customtools
+- openai/gpt-5.4-2026-03-05
+- deepseek/deepseek-reasoner
+- deepseek/deepseek-v4-pro
+- openai/gpt-4o-search-preview
+- openai/gpt-5.4
+- z-ai/glm-4.7
+- openai/gpt-4o-search-preview-2025-03-11
+- qwen/qwen3.5-122b-a10b
+- google/gemma-4-26b-a4b-it
+- qwen/qwen3.5-35b-a3b
+- z-ai/glm-4.5-air
+- google/gemini-3.1-pro-preview
+- openai/gpt-5.1-2025-11-13
+- anthropic/claude-opus-4.6
+- openai/gpt-4
+- openai/gpt-5-mini-2025-08-07
+- openai/gpt-5-nano-2025-08-07
+- openai/gpt-5-search-api
+- openai/gpt-5.3-chat-latest
+- openai/gpt-4o-mini
+- qwen/qwen3.5-flash-2026-02-23
+- openai/gpt-4o

models/orcarouter/models/llm/anthropic-claude-haiku-4.5.yaml

@@ -0,0 +1,97 @@
+# https://www.orcarouter.ai/models/anthropic/claude-haiku-4.5
+model: anthropic/claude-haiku-4.5
+label:
+  en_US: claude-haiku-4.5
+model_type: llm
+features:
+  - agent-thought
+  - tool-call
+  - multi-tool-call
+  - stream-tool-call
+  - vision
+  - document
+model_properties:
+  mode: chat
+  context_size: 200000
+parameter_rules:
+  - name: temperature
+    use_template: temperature
+  - name: top_p
+    use_template: top_p
+  - name: top_k
+    label:
+      zh_Hans: 取样数量
+      en_US: Top k
+    type: int
+    required: false
+    help:
+      zh_Hans: 仅从每个后续标记的前 K 个选项中采样。
+      en_US: Only sample from the top K options for each subsequent token.
+  - name: max_tokens
+    use_template: max_tokens
+    default: 1024
+    min: 1
+    max: 64000
+  - name: response_format
+    label:
+      zh_Hans: 回复格式
+      en_US: Response Format
+    type: string
+    required: false
+    options:
+      - text
+      - json_object
+  - name: enable_thinking
+    type: boolean
+    required: false
+    default: false
+    label:
+      zh_Hans: 推理模式
+      en_US: Thinking Mode
+    help:
+      zh_Hans: 启用 Anthropic native thinking 模式；启用时输出会包含思考过程。
+      en_US: Enable Anthropic native thinking. When enabled, the output includes reasoning content.
+  - name: reasoning_budget
+    type: int
+    required: false
+    default: 2000
+    min: 1024
+    max: 64000
+    label:
+      en_US: Thinking Budget
+      zh_Hans: 推理预算
+    help:
+      en_US: Token budget for thinking (min 1024, must be less than max_tokens). Only valid when thinking is enabled.
+      zh_Hans: 推理 token 预算（最小 1024，必须小于 max_tokens）。仅 thinking 模式生效。
+  - name: exclude_reasoning_tokens
+    type: boolean
+    required: false
+    default: false
+    label:
+      zh_Hans: 隐藏思考过程
+      en_US: Hide the thought process
+  - name: orcarouter_fallback_models
+    type: string
+    required: false
+    label:
+      en_US: Fallback models (JSON array)
+      zh_Hans: 降级模型（JSON 数组）
+    help:
+      en_US: 'Upstream models to try if the primary fails. Example: ["openai/gpt-4o-mini","openai/gpt-4o"]'
+      zh_Hans: 主上游失败时按顺序尝试这些模型。例：["openai/gpt-4o-mini","openai/gpt-4o"]
+  - name: orcarouter_route
+    type: string
+    required: false
+    options:
+      - fallback
+    label:
+      en_US: Routing mode
+      zh_Hans: 路由模式
+    help:
+      en_US: Activate fallback list. Set to 'fallback' to use orcarouter_fallback_models in order.
+      zh_Hans: 启用降级列表。设为 'fallback' 后按 orcarouter_fallback_models 顺序尝试。
+pricing:
+  input: "1.0"
+  output: "5.0"
+  unit: "0.000001"
+  currency: USD