better-dsh 0.2.4 → 0.2.5

package/docs/80_upstream-docs-and-skills/dsh-dev-skill/pi-and-omp/pi/models.md

@@ -0,0 +1,587 @@
+<!--
+source: https://pi.dev/docs/latest/models
+fetched: 2026-09-06
+-->
+
+# Custom Models
+
+Add custom providers and models (Ollama, vLLM, LM Studio, proxies) via `~/.pi/agent/models.json`.
+
+## Table of Contents
+
+- [Minimal Example](#minimal-example)
+- [Full Example](#full-example)
+- [Supported APIs](#supported-apis)
+- [Provider Configuration](#provider-configuration)
+- [Model Configuration](#model-configuration)
+- [Overriding Built-in Providers](#overriding-built-in-providers)
+- [Per-model Overrides](#per-model-overrides)
+- [Anthropic Messages Compatibility](#anthropic-messages-compatibility)
+- [OpenAI Compatibility](#openai-compatibility)
+
+## Minimal Example
+
+For local models (Ollama, LM Studio, vLLM), only `id` is required per model:
+
+```
+{
+  "providers": {
+    "ollama": {
+      "baseUrl": "http://localhost:11434/v1",
+      "api": "openai-completions",
+      "apiKey": "ollama",
+      "models": [
+        { "id": "llama3.1:8b" },
+        { "id": "qwen2.5-coder:7b" }
+      ]
+    }
+  }
+}
+```
+
+The `apiKey` value is a placeholder because Ollama ignores it. pi still treats models as requiring auth before they appear in `/model`, so keyless local servers should keep a dummy value, save a key for that provider with `/login`, or pass `--api-key` when selecting the model.
+
+Some OpenAI-compatible servers do not understand the `developer` role used for reasoning-capable models. For those providers, set `compat.supportsDeveloperRole` to `false` so pi sends the system prompt as a `system` message instead. If the server also does not support `reasoning_effort`, set `compat.supportsReasoningEffort` to `false` too.
+
+You can set `compat` at the provider level to apply to all models, or at the model level to override a specific model. This commonly applies to Ollama, vLLM, SGLang, and similar OpenAI-compatible servers.
+
+```
+{
+  "providers": {
+    "ollama": {
+      "baseUrl": "http://localhost:11434/v1",
+      "api": "openai-completions",
+      "apiKey": "ollama",
+      "compat": {
+        "supportsDeveloperRole": false,
+        "supportsReasoningEffort": false
+      },
+      "models": [
+        {
+          "id": "gpt-oss:20b",
+          "reasoning": true
+        }
+      ]
+    }
+  }
+}
+```
+
+## Full Example
+
+Override defaults when you need specific values:
+
+```
+{
+  "providers": {
+    "ollama": {
+      "baseUrl": "http://localhost:11434/v1",
+      "api": "openai-completions",
+      "apiKey": "ollama",
+      "models": [
+        {
+          "id": "llama3.1:8b",
+          "name": "Llama 3.1 8B (Local)",
+          "reasoning": false,
+          "input": ["text"],
+          "contextWindow": 128000,
+          "maxTokens": 32000,
+          "cost": { "input": 0, "output": 0, "cacheRead": 0, "cacheWrite": 0 }
+        }
+      ]
+    }
+  }
+}
+```
+
+The file reloads each time you open `/model`. Edit during session; no restart needed.
+
+## Google AI Studio Example
+
+Use `google-generative-ai` with a `baseUrl` to add models from Google AI Studio, including custom Gemma 4 entries:
+
+```
+{
+  "providers": {
+    "my-google": {
+      "baseUrl": "https://generativelanguage.googleapis.com/v1beta",
+      "api": "google-generative-ai",
+      "apiKey": "$GEMINI_API_KEY",
+      "models": [
+        {
+          "id": "gemma-4-31b-it",
+          "name": "Gemma 4 31B",
+          "input": ["text", "image"],
+          "contextWindow": 262144,
+          "reasoning": true
+        }
+      ]
+    }
+  }
+}
+```
+
+The `baseUrl` is required when adding custom models to the `google-generative-ai` API type.
+
+## Supported APIs
+
+| API | Description |
+| --- | --- |
+| `openai-completions` | OpenAI Chat Completions (most compatible) |
+| `openai-responses` | OpenAI Responses API |
+| `anthropic-messages` | Anthropic Messages API |
+| `google-generative-ai` | Google Generative AI |
+
+Set `api` at provider level (default for all models) or model level (override per model).
+
+## Provider Configuration
+
+| Field | Description |
+| --- | --- |
+| `baseUrl` | API endpoint URL |
+| `api` | API type (see above) |
+| `apiKey` | Optional API key config (see value resolution below). Omit it when auth is provided by `/login`/`auth.json` or CLI `--api-key`. |
+| `oauth` | Dynamic OAuth provider type. Currently supports `"radius"`; requires the gateway `baseUrl`. |
+| `headers` | Custom headers (see value resolution below) |
+| `authHeader` | Set `true` to add `Authorization: Bearer <apiKey>` automatically |
+| `models` | Array of model configurations |
+| `modelOverrides` | Per-model overrides for built-in or extension-registered models on this provider |
+
+For providers with `models`, non-built-in provider configs need `baseUrl` and an `api` value at either provider or model level. `apiKey` is not required to load the file: models become available when auth is configured through `/login`/`auth.json`, CLI `--api-key`, or provider `apiKey`. If no auth is configured, the models load but stay unavailable in `/model` and `--list-models`.
+
+### Value Resolution
+
+The `apiKey` and `headers` fields support command execution, environment interpolation, and literals:
+
+- **Shell command:** `"!command"` at the start executes the whole value as a command and uses stdout
+
+  ```
+  "apiKey": "!security find-generic-password -ws 'anthropic'"
+  "apiKey": "!op read 'op://vault/item/credential'"
+  ```
+- **Environment interpolation:** `"$ENV_VAR"` or `"${ENV_VAR}"` uses the value of the named variable. Interpolation works inside larger literals.
+
+  ```
+  "apiKey": "$MY_API_KEY"
+  "apiKey": "${KEY_PREFIX}_${KEY_SUFFIX}"
+  ```
+
+  `$FOO_BAR` is the variable `FOO_BAR`; use `${FOO}_BAR` when `BAR` is literal text. Missing environment variables make the value unresolved.
+- **Escapes:** `"$$"` emits a literal `"$"`; `"$!"` emits a literal `"!"` without triggering command execution.
+
+  ```
+  "apiKey": "$$literal-dollar-prefix"
+  "apiKey": "$!literal-bang-prefix"
+  ```
+- **Literal value:** Used directly. Plain uppercase strings such as `MY_API_KEY` are literals; use `$MY_API_KEY` for environment variables.
+
+  ```
+  "apiKey": "sk-..."
+  ```
+
+For `models.json`, shell commands are resolved at request time. pi intentionally does not apply built-in TTL, stale reuse, or recovery logic for arbitrary commands. Different commands need different caching and failure strategies, and pi cannot infer the right one.
+
+If your command is slow, expensive, rate-limited, or should keep using a previous value on transient failures, wrap it in your own script or command that implements the caching or TTL behavior you want.
+
+`/model` availability checks use configured auth presence and do not execute shell commands.
+
+### Custom Headers
+
+```
+{
+  "providers": {
+    "custom-proxy": {
+      "baseUrl": "https://proxy.example.com/v1",
+      "apiKey": "$MY_API_KEY",
+      "api": "anthropic-messages",
+      "headers": {
+        "x-portkey-api-key": "$PORTKEY_API_KEY",
+        "x-secret": "!op read 'op://vault/item/secret'"
+      },
+      "models": [...]
+    }
+  }
+}
+```
+
+## Model Configuration
+
+| Field | Required | Default | Description |
+| --- | --- | --- | --- |
+| `id` | Yes | — | Model identifier (passed to the API) |
+| `name` | No | `id` | Human-readable model label. Used for matching (`--model` patterns) and shown as secondary model detail text. |
+| `api` | No | provider's `api` | Override provider's API for this model |
+| `reasoning` | No | `false` | Supports extended thinking |
+| `thinkingLevelMap` | No | omitted | Maps pi thinking levels to provider values and marks unsupported levels (see below) |
+| `input` | No | `["text"]` | Input types: `["text"]` or `["text", "image"]` |
+| `contextWindow` | No | `128000` | Context window size in tokens |
+| `maxTokens` | No | `16384` | Maximum output tokens |
+| `samplingParams` | No | omitted | Sampling parameters merged verbatim into every request body (see below) |
+| `cost` | No | all zeros | Per-million-token rates with optional request-wide input pricing tiers |
+| `compat` | No | provider `compat` | Provider compatibility overrides. Merged with provider-level `compat` when both are set. |
+
+A cost tier supplies a complete alternate rate set and applies to the full request when total input usage (`input + cacheRead + cacheWrite`) exceeds `inputTokensAbove`. When multiple tiers match, the highest threshold wins.
+
+```
+{
+  "cost": {
+    "input": 5,
+    "output": 30,
+    "cacheRead": 0.5,
+    "cacheWrite": 6.25,
+    "tiers": [
+      {
+        "inputTokensAbove": 272000,
+        "input": 10,
+        "output": 45,
+        "cacheRead": 1,
+        "cacheWrite": 12.5
+      }
+    ]
+  }
+}
+```
+
+Current behavior:
+
+- `/model`, `--list-models`, and the interactive footer display entries by model `id`.
+- The configured `name` is used for model matching and secondary model detail text. It does not replace the footer/status-bar model id.
+
+### Sampling Parameters
+
+`samplingParams` is a free-form object merged verbatim into every request body for the model, after the fields pi sets itself, so its keys win. Use it to send sampling parameters pi does not model — including server-specific ones like llama.cpp's `min_p` or vLLM's `top_k`:
+
+```
+{
+  "id": "deepseek-v4-flash",
+  "samplingParams": {
+    "temperature": 1.0,
+    "top_p": 0.95,
+    "top_k": 0,
+    "min_p": 0.0
+  }
+}
+```
+
+Only OpenAI-compatible APIs apply it (`openai-completions`, `openai-responses`, `azure-openai-responses`); other APIs ignore it. Keys override pi's named request fields (for example a `temperature` key here beats the request-level temperature), so prefer it as the single source of sampling truth for a model. In `modelOverrides`, `samplingParams` merges per key with the base model's value.
+
+A constant thinking-token cap can go here too, but it will not follow `thinkingBudgets` or leave room for the answer. Prefer `compat.thinkingTokenBudgetField` (or the `supportsThinkingTokenBudget` alias) for that.
+
+### Thinking Level Map
+
+Use `thinkingLevelMap` on a model to describe model-specific thinking controls. Keys are pi thinking levels: `off`, `minimal`, `low`, `medium`, `high`, `xhigh`, `max`. Maps may contain holes; for example, a model can expose `high` and `max` without exposing `xhigh`.
+
+Values are tristate:
+
+| Value | Meaning |
+| --- | --- |
+| omitted | Standard levels through `high` use the provider's default mapping; extended `xhigh` and `max` levels are unsupported |
+| string | Level is supported and this value is sent to the provider |
+| `null` | Level is unsupported and hidden/skipped/clamped away |
+
+Example for a model that only supports off, high, and max reasoning:
+
+```
+{
+  "id": "deepseek-v4-pro",
+  "reasoning": true,
+  "thinkingLevelMap": {
+    "minimal": null,
+    "low": null,
+    "medium": null,
+    "high": "high",
+    "xhigh": null,
+    "max": "max"
+  }
+}
+```
+
+Example for a model where thinking cannot be disabled:
+
+```
+{
+  "id": "always-thinking-model",
+  "reasoning": true,
+  "thinkingLevelMap": {
+    "off": null
+  }
+}
+```
+
+Migration: older configs that used `compat.reasoningEffortMap` should move that mapping to model-level `thinkingLevelMap`. Use `null` for levels that should not appear in the UI.
+
+## Overriding Built-in Providers
+
+Route a built-in provider through a proxy without redefining models:
+
+```
+{
+  "providers": {
+    "anthropic": {
+      "baseUrl": "https://my-proxy.example.com/v1"
+    }
+  }
+}
+```
+
+All built-in Anthropic models remain available. Existing OAuth or API key auth continues to work.
+
+To merge custom models into a built-in provider, include the `models` array:
+
+```
+{
+  "providers": {
+    "anthropic": {
+      "baseUrl": "https://my-proxy.example.com/v1",
+      "apiKey": "$ANTHROPIC_API_KEY",
+      "api": "anthropic-messages",
+      "models": [...]
+    }
+  }
+}
+```
+
+Merge semantics:
+
+- Built-in models are kept.
+- Custom models are upserted by `id` within the provider.
+- If a custom model `id` matches a built-in model `id`, the custom model replaces that built-in model.
+- If a custom model `id` is new, it is added alongside built-in models.
+
+## Per-model Overrides
+
+Use `modelOverrides` to customize built-in models and matching extension-registered models without replacing the provider's full model list.
+
+```
+{
+  "providers": {
+    "openrouter": {
+      "modelOverrides": {
+        "anthropic/claude-sonnet-4": {
+          "name": "Claude Sonnet 4 (Bedrock Route)",
+          "compat": {
+            "openRouterRouting": {
+              "only": ["amazon-bedrock"]
+            }
+          }
+        }
+      }
+    }
+  }
+}
+```
+
+`modelOverrides` supports these fields per model: `name`, `reasoning`, `thinkingLevelMap`, `input`, `cost` (partial), `contextWindow`, `maxTokens`, `samplingParams` (merged per key), `headers`, `compat`.
+
+Direct OpenAI GPT-5.6 Sol, Terra, and Luna default to a `272000` context window so requests remain within OpenAI's short-context pricing tier. To opt into OpenAI's 1.05M context window, increase it for each model you use:
+
+```
+{
+  "providers": {
+    "openai": {
+      "modelOverrides": {
+        "gpt-5.6-sol": {
+          "contextWindow": 1050000
+        }
+      }
+    }
+  }
+}
+```
+
+The override preserves the built-in pricing metadata. Requests with more than 272K total input tokens use GPT-5.6's long-context rates for the entire request. Apply the same override to `gpt-5.6-terra` or `gpt-5.6-luna` when needed.
+
+Behavior notes:
+
+- `modelOverrides` are applied to built-in provider models and matching extension-registered provider models.
+- Unknown model IDs are ignored.
+- You can combine provider-level `baseUrl`/`headers` with `modelOverrides`.
+- Overriding `name` changes model matching and secondary detail text only; the footer and primary model lists continue to show the model `id`.
+- If `models` is also defined for a provider, custom models are merged after built-in overrides. A custom model with the same `id` replaces the overridden built-in model entry.
+
+## Anthropic Messages Compatibility
+
+For providers or proxies using `api: "anthropic-messages"`, use `compat` to control Anthropic-specific request compatibility.
+
+By default pi sends per-tool `eager_input_streaming: true`. If a proxy or Anthropic-compatible backend rejects that field, set `supportsEagerToolInputStreaming` to `false`. Pi will omit `tools[].eager_input_streaming` and send the legacy `fine-grained-tool-streaming-2025-05-14` beta header for tool-enabled requests instead.
+
+Some Anthropic models require adaptive thinking (`thinking.type: "adaptive"` plus `output_config.effort`) instead of the legacy budget-based thinking payload. Built-in models set this automatically. For custom providers or aliases that route to those models, set `forceAdaptiveThinking` to `true`.
+
+Claude models with per-turn effort support use `supportsMidConvoEffort`. Pi then persists each response's provider effort, reconstructs effort-only system messages on later requests, and sends thinking binding controls with `prefix_mismatch_behavior: "drop_block"` to avoid stale signed-thinking prefixes causing persistent 400 responses. Set this only for the exact supported Claude model on a faithful Anthropic Messages transport; do not enable it for APIs that merely imitate the Messages shape.
+
+Some Anthropic-compatible providers emit thinking blocks with empty signatures and still expect them on replay. Set `allowEmptySignature` to `true` only for those providers; real Anthropic rejects empty thinking signatures.
+
+Built-in Anthropic models enable `supportsStrictTools` in their model metadata. Custom Anthropic-compatible models must set it to `true` when their endpoint accepts strict JSON-schema tool definitions.
+
+```
+{
+  "providers": {
+    "anthropic-proxy": {
+      "baseUrl": "https://proxy.example.com",
+      "api": "anthropic-messages",
+      "apiKey": "$ANTHROPIC_PROXY_KEY",
+      "compat": {
+        "supportsEagerToolInputStreaming": false,
+        "supportsLongCacheRetention": true,
+        "forceAdaptiveThinking": true,
+        "allowEmptySignature": true
+      },
+      "models": [
+        {
+          "id": "claude-opus-4-7",
+          "reasoning": true,
+          "input": ["text", "image"]
+        }
+      ]
+    }
+  }
+}
+```
+
+| Field | Description |
+| --- | --- |
+| `supportsEagerToolInputStreaming` | Whether the provider accepts per-tool `eager_input_streaming`. Default: `true`. Set to `false` to omit that field and use the legacy fine-grained tool streaming beta header on tool-enabled requests. |
+| `supportsLongCacheRetention` | Whether the provider accepts Anthropic long cache retention (`cache_control.ttl: "1h"`) when cache retention is `long`. Default: `true`. |
+| `sendSessionAffinityHeaders` | Whether to send `x-session-affinity` from the session id when caching is enabled. Default: auto-detected for known providers. |
+| `supportsCacheControlOnTools` | Whether the provider accepts Anthropic-style `cache_control` markers on tool definitions. Default: `true`. |
+| `forceAdaptiveThinking` | Whether to send adaptive thinking (`thinking.type: "adaptive"` plus `output_config.effort`) for this model. Built-in adaptive models set this automatically. Default: `false`. |
+| `supportsMidConvoEffort` | Whether the exact Claude model transport supports per-turn effort system messages and thinking binding controls. Pi persists native effort levels and always sends `drop_block` when enabled. Default: `false`. |
+| `allowEmptySignature` | Whether to replay empty thinking signatures as `signature: ""` instead of converting thinking to text. Default: `false`. |
+| `supportsStrictTools` | Whether the provider accepts strict JSON-schema tool definitions. Default: `false`; built-in Anthropic models enable it in generated metadata. |
+
+## OpenAI Compatibility
+
+For providers with partial OpenAI compatibility, use the `compat` field.
+
+- Provider-level `compat` applies defaults to all models under that provider.
+- Model-level `compat` overrides provider-level values for that model.
+
+```
+{
+  "providers": {
+    "local-llm": {
+      "baseUrl": "http://localhost:8080/v1",
+      "api": "openai-completions",
+      "compat": {
+        "supportsUsageInStreaming": false,
+        "maxTokensField": "max_tokens"
+      },
+      "models": [...]
+    }
+  }
+}
+```
+
+| Field | Description |
+| --- | --- |
+| `supportsStore` | Provider supports `store` field |
+| `supportsDeveloperRole` | Use `developer` vs `system` role |
+| `supportsReasoningEffort` | Support for `reasoning_effort` parameter |
+| `supportsUsageInStreaming` | Supports `stream_options: { include_usage: true }` (default: `true`) |
+| `supportsFinishReason` | Whether streamed responses include `finish_reason`. When `false`, pi infers `stop` or `toolUse` when the stream ends. Default: `true`. |
+| `maxTokensField` | Use `max_completion_tokens` or `max_tokens` |
+| `requiresToolResultName` | Include `name` on tool result messages |
+| `requiresAssistantAfterToolResult` | Insert an assistant message before a user message after tool results |
+| `requiresThinkingAsText` | Convert thinking blocks to plain text |
+| `requiresReasoningContentOnAssistantMessages` | Include empty `reasoning_content` on all replayed assistant messages when reasoning is enabled |
+| `thinkingFormat` | Use `reasoning_effort`, `openrouter`, `deepseek`, `together`, `baseten`, `zai`, `qwen`, `chat-template`, or `qwen-chat-template` thinking parameters |
+| `chatTemplateKwargs` | `chat_template_kwargs` values for `thinkingFormat: "chat-template"`; use `{ "$var": "thinking.enabled" }`, `{ "$var": "thinking.effort" }`, or `{ "$var": "thinking.budget" }` for pi-controlled thinking values |
+| `chatTemplateArgs` | `chat_template_args` values for `thinkingFormat: "baseten"`; use `{ "$var": "thinking.enabled" }`, `{ "$var": "thinking.effort" }`, or `{ "$var": "thinking.budget" }` for pi-controlled thinking values |
+| `thinkingTokenBudgetField` | Top-level request field used to cap reasoning tokens from `thinkingBudgets`, clamped so at least 1024 tokens remain for the answer. `"thinking_token_budget"` (vLLM), `"thinking_budget"` (Qwen/DashScope/SGLang), `"thinking_budget_tokens"` (llama.cpp). Off by default; not set on the generated catalog. |
+| `supportsThinkingTokenBudget` | Alias for `thinkingTokenBudgetField: "thinking_token_budget"` (vLLM). Prefer `thinkingTokenBudgetField`. Default: `false`. |
+| `cacheControlFormat` | Use Anthropic-style `cache_control` markers on the system prompt, last tool definition, and last user, assistant, or tool-result text content. Currently only `anthropic` is supported. |
+| `sendSessionAffinityHeaders` | For `openai-completions`, send session-affinity headers from the session id when caching is enabled. Default: `false`. |
+| `sessionAffinityFormat` | For `openai-completions` and `openai-responses`, the session-affinity header format: `openai` sends `session_id`/`x-client-request-id` (completions also `x-session-affinity`), `openai-nosession` omits the underscore-containing `session_id` header, `openrouter` sends `x-session-id`. Does not affect the `prompt_cache_key` body param. Default: auto-detected. |
+| `supportsStrictMode` | Whether the provider accepts strict JSON-schema function tool definitions. Defaults depend on the API; built-in OpenAI models carry explicit capability metadata. |
+| `supportsOpenAIGrammarTools` | Whether OpenAI-compatible APIs emit custom Lark/regex grammar tools. When `false`, grammar-constrained tools fall back to normal function tools. Default: `false`; the built-in model catalog enables it for GPT-5+ models on OpenAI, OpenAI Codex, Azure OpenAI, GitHub Copilot, opencode, and Cloudflare AI Gateway. |
+| `deferredToolsMode` | Use provider-specific deferred tool serialization. Currently only `"kimi"` is supported for Kimi's OpenAI-compatible Chat Completions format. |
+| `supportsLongCacheRetention` | Whether the provider accepts long cache retention when cache retention is `long`: `prompt_cache_options.ttl: "30m"` for GPT-5.6+ Responses models, `prompt_cache_retention: "24h"` for earlier OpenAI models, or `cache_control.ttl: "1h"` when `cacheControlFormat` is `anthropic`. Default: `true`. |
+| `openRouterRouting` | OpenRouter provider routing preferences. This object is sent as-is in the `provider` field of the [OpenRouter API request](https://openrouter.ai/docs/guides/routing/provider-selection). |
+| `vercelGatewayRouting` | Vercel AI Gateway routing config for provider selection (`only`, `order`) |
+
+`openrouter` uses `reasoning: { effort }`. `together` uses `reasoning: { enabled }` and also `reasoning_effort` when `supportsReasoningEffort` is enabled. `qwen` uses top-level `enable_thinking`. Use `qwen-chat-template` for local Qwen-compatible servers that require `chat_template_kwargs.enable_thinking` and `preserve_thinking`. Use `chat-template` for vLLM/Hugging Face chat templates that need configurable `chat_template_kwargs`, such as `chatTemplateKwargs: { "thinking": { "$var": "thinking.enabled" } }` for DeepSeek V3.x templates. Use `thinkingFormat: "baseten"` with `chatTemplateArgs` for providers that expose toggle controls through `chat_template_args` and optionally support top-level `reasoning_effort`.
+
+`thinkingTokenBudgetField` is independent of `thinkingFormat`. Do not enable it on the generated Qwen catalog: those models already send `reasoning_effort`, and DashScope rejects `thinking_budget` together with `reasoning_effort`.
+
+`cacheControlFormat: "anthropic"` is for OpenAI-compatible providers that expose Anthropic-style prompt caching through `cache_control` markers on text content and tool definitions.
+
+Example:
+
+```
+{
+  "providers": {
+    "openrouter": {
+      "baseUrl": "https://openrouter.ai/api/v1",
+      "apiKey": "$OPENROUTER_API_KEY",
+      "api": "openai-completions",
+      "models": [
+        {
+          "id": "openrouter/anthropic/claude-3.5-sonnet",
+          "name": "OpenRouter Claude 3.5 Sonnet",
+          "compat": {
+            "openRouterRouting": {
+              "allow_fallbacks": true,
+              "require_parameters": false,
+              "data_collection": "deny",
+              "zdr": true,
+              "enforce_distillable_text": false,
+              "order": ["anthropic", "amazon-bedrock", "google-vertex"],
+              "only": ["anthropic", "amazon-bedrock"],
+              "ignore": ["gmicloud", "friendli"],
+              "quantizations": ["fp16", "bf16"],
+              "sort": {
+                "by": "price",
+                "partition": "model"
+              },
+              "max_price": {
+                "prompt": 10,
+                "completion": 20
+              },
+              "preferred_min_throughput": {
+                "p50": 100,
+                "p90": 50
+              },
+              "preferred_max_latency": {
+                "p50": 1,
+                "p90": 3,
+                "p99": 5
+              }
+            }
+          }
+        }
+      ]
+    }
+  }
+}
+```
+
+Vercel AI Gateway example:
+
+```
+{
+  "providers": {
+    "vercel-ai-gateway": {
+      "baseUrl": "https://ai-gateway.vercel.sh/v1",
+      "apiKey": "$AI_GATEWAY_API_KEY",
+      "api": "openai-completions",
+      "models": [
+        {
+          "id": "moonshotai/kimi-k2.5",
+          "name": "Kimi K2.5 (Fireworks via Vercel)",
+          "reasoning": true,
+          "input": ["text", "image"],
+          "cost": { "input": 0.6, "output": 3, "cacheRead": 0, "cacheWrite": 0 },
+          "contextWindow": 262144,
+          "maxTokens": 262144,
+          "compat": {
+            "vercelGatewayRouting": {
+              "only": ["fireworks", "novita"],
+              "order": ["fireworks", "novita"]
+            }
+          }
+        }
+      ]
+    }
+  }
+}
+```