@inixiative/foundry-core 0.2.0 → 0.3.0

package/src/model-registry.ts

@@ -0,0 +1,1587 @@
+/**
+ * The model registry: the vocabulary of what models are worth using for, the
+ * table of providers and models, and the lookups over it.
+ *
+ * It lives in core rather than beside the adapters because consumers that
+ * depend on core alone — Oracle among them — must be able to reason about which
+ * model suits a piece of work without taking a dependency on Foundry's HTTP and
+ * process handling. Adapters stay in `packages/foundry`; this file is data and
+ * pure functions only, and adds no runtime dependency.
+ */
+
+export type ModelTier = 'fast' | 'standard' | 'powerful';
+export type CostTier = 'low' | 'medium' | 'high';
+export type ProviderType =
+  | 'anthropic'
+  | 'openai'
+  | 'openai-compatible'
+  | 'gemini'
+  | 'claude-code'
+  | 'codex'
+  | 'xai'
+  | 'typesafe'
+  | 'custom';
+export type RuntimeKind = 'api' | 'native-harness' | 'typed-decision';
+
+/**
+ * What a model is worth using for. Every model is a judgment client — that is
+ * not a provider category, it is a tag every entry carries. The other tags name
+ * work Foundry already distinguishes:
+ *
+ * - judgment        classifier/router/decider agent kinds and the decision
+ *                   provider: which archive, which middleware, which domain.
+ * - execution       the executor kind (Artificer): tools on, multi-turn.
+ * - domain-advising the `domain-advising` flow role: read an owned domain layer
+ *                   and advise/guard from it.
+ * - learning-review the background `learning.review` pass over finished work.
+ * - tool-use        CompletionOpts.tools / toolDefinitions.
+ * - reasoning       CompletionOpts.thinking (reasoning effort / think budget).
+ */
+export type ModelCapability =
+  | 'judgment'
+  | 'execution'
+  | 'domain-advising'
+  | 'learning-review'
+  | 'tool-use'
+  | 'reasoning';
+
+export const MODEL_CAPABILITIES: readonly ModelCapability[] = [
+  'judgment',
+  'execution',
+  'domain-advising',
+  'learning-review',
+  'tool-use',
+  'reasoning',
+] as const;
+
+/** How the provider is paid for. `envKey` is present exactly when this is "api-key". */
+export type ProviderCredential = 'api-key' | 'subscription' | 'local';
+
+/** Effort levels hosts publish. A superset of CompletionOpts.thinking. */
+export type ReasoningEffort = 'none' | 'low' | 'medium' | 'high' | 'xhigh' | 'max' | 'ultra';
+
+/**
+ * How one model takes its reasoning controls. This is the map that replaced a
+ * model-name regex: naming is not behaviour, and the next generation's names
+ * are not knowable in advance.
+ */
+export interface ModelReasoning {
+  /**
+   * Where the effort goes. Only the first two reach the OpenAI-compatible
+   * adapter; Anthropic and Gemini have their own adapters and their own fields,
+   * recorded here so the map describes every model, not just the compatible ones.
+   */
+  param: 'reasoning_effort' | 'reasoning.effort' | 'output_config.effort' | 'thinking_level';
+  /** Accepted levels. Without "none", thinking cannot be switched off. */
+  efforts: readonly ReasoningEffort[];
+  /** Sent when the caller asks for nothing, or asks for a level this model rejects. */
+  fallback: ReasoningEffort;
+  /** OpenAI reasoning models cap output with max_completion_tokens, not max_tokens. */
+  outputField?: 'max_completion_tokens';
+  /** xAI reasoning models reject `stop` and the penalty parameters. */
+  rejectsStop?: boolean;
+}
+
+export interface FoundryModelInfo {
+  id: string;
+  label: string;
+  tier: ModelTier;
+  costTier?: CostTier;
+  contextWindow?: number;
+  maxOutputTokens?: number;
+  runtimeKind: RuntimeKind;
+  nativeAlias?: boolean;
+  /** Always includes "judgment". Includes "reasoning" exactly when `reasoning` is set. */
+  capabilities: ModelCapability[];
+  /** Request shape for reasoning. Absent = plain chat completions with max_tokens. */
+  reasoning?: ModelReasoning;
+  /**
+   * The vendor trains on prompts and completions sent to this model, by its own
+   * documentation. Set only where the vendor says so. This is a term of use, not
+   * a capability: it says nothing about what the model is fit for, so it must not
+   * be a `ModelCapability` — `modelOptionsByCapability` selects models for work,
+   * and a commercial term has no business changing that selection. Every entry
+   * carrying it also says so in its `label`, because a `<select>` renders nothing
+   * but the label.
+   */
+  trainsOnInput?: boolean;
+  notes?: string;
+}
+
+/** OpenAI's API reasoning models: effort can be switched off, output caps differ. */
+const OPENAI_REASONING: ModelReasoning = {
+  param: 'reasoning_effort',
+  efforts: ['none', 'low', 'medium', 'high', 'xhigh', 'max'],
+  fallback: 'none',
+  outputField: 'max_completion_tokens',
+};
+/** Codex CLI catalogue (~/.codex/models_cache.json): no "none", medium by default. */
+const CODEX_REASONING: ModelReasoning = {
+  param: 'reasoning_effort',
+  efforts: ['low', 'medium', 'high', 'xhigh', 'max'],
+  fallback: 'medium',
+};
+const CODEX_REASONING_ULTRA: ModelReasoning = {
+  ...CODEX_REASONING,
+  efforts: ['low', 'medium', 'high', 'xhigh', 'max', 'ultra'],
+};
+/** Anthropic: adaptive thinking, steered by output_config.effort, never by reasoning_effort. */
+const CLAUDE_REASONING: ModelReasoning = {
+  param: 'output_config.effort',
+  efforts: ['low', 'medium', 'high', 'xhigh', 'max'],
+  fallback: 'high',
+};
+/** Grok nests effort and cannot be switched off; reasoning models reject stop. */
+const GROK_REASONING: ModelReasoning = {
+  param: 'reasoning.effort',
+  efforts: ['low', 'medium', 'high'],
+  fallback: 'high',
+  rejectsStop: true,
+};
+const GROK_REASONING_XHIGH: ModelReasoning = {
+  ...GROK_REASONING,
+  efforts: ['low', 'medium', 'high', 'xhigh'],
+};
+/** Compatible hosts whose thinking is a plain effort with no "none" and no output-field change. */
+const EFFORT_ONLY: ModelReasoning = {
+  param: 'reasoning_effort',
+  efforts: ['low', 'medium', 'high'],
+  fallback: 'medium',
+};
+/** Gemini takes thinking_level, and rejects "minimal". */
+const GEMINI_REASONING: ModelReasoning = {
+  param: 'thinking_level',
+  efforts: ['low', 'medium', 'high'],
+  fallback: 'medium',
+};
+/** DeepSeek publishes its own ladder, including "none" and "max". */
+const DEEPSEEK_REASONING: ModelReasoning = {
+  param: 'reasoning_effort',
+  efforts: ['none', 'low', 'high', 'max'],
+  fallback: 'none',
+};
+/**
+ * Meta's Model API. Muse Spark always reasons: `reasoning_effort: "none"` is a 400.
+ * The docs also accept "minimal", which this vocabulary does not name, so it is not offered.
+ * Output is capped with max_completion_tokens; max_tokens survives as a deprecated alias.
+ */
+const MUSE_REASONING: ModelReasoning = {
+  param: 'reasoning_effort',
+  efforts: ['low', 'medium', 'high', 'xhigh'],
+  fallback: 'medium',
+  outputField: 'max_completion_tokens',
+};
+
+export interface FoundryProviderInfo {
+  id: string;
+  type: ProviderType;
+  label: string;
+  description: string;
+  envKey?: string;
+  credential: ProviderCredential;
+  /**
+   * Versioned API root for OpenAI-compatible hosts, used verbatim: requests go
+   * to `${apiRoot}/chat/completions`. Hosts disagree about the version segment
+   * (/v1, /api/paas/v4, /compatible-mode/v1, none at all), so this is not
+   * normalized. A user's `providers[id].baseUrl` override is.
+   */
+  apiRoot?: string;
+  enabledByDefault: boolean;
+  models: FoundryModelInfo[];
+}
+
+export interface ModelSweepOption {
+  provider: string;
+  model: string;
+  label: string;
+}
+
+export const MODEL_REGISTRY_UPDATED_AT = '2026-09-24';
+
+/**
+ * Decision default: the cheapest current model tagged `judgment`. Held here
+ * rather than in an adapter so one map answers "what runs decisions".
+ */
+export const DECISION_PROVIDER = 'openai';
+export const DECISION_MODEL = 'gpt-6-luna';
+
+/** Tag rule: every model judges. Cheap models review and route; capable models execute and advise. */
+const FAST = ['judgment', 'learning-review', 'tool-use'] as ModelCapability[];
+const FAST_REASONING = [...FAST, 'reasoning'] as ModelCapability[];
+const STANDARD = [
+  'judgment',
+  'learning-review',
+  'execution',
+  'domain-advising',
+  'tool-use',
+] as ModelCapability[];
+const STANDARD_REASONING = [...STANDARD, 'reasoning'] as ModelCapability[];
+const POWERFUL = ['judgment', 'execution', 'domain-advising', 'tool-use'] as ModelCapability[];
+const POWERFUL_REASONING = [...POWERFUL, 'reasoning'] as ModelCapability[];
+
+export const MODEL_REGISTRY: Record<string, FoundryProviderInfo> = {
+  'claude-code': {
+    id: 'claude-code',
+    type: 'claude-code',
+    label: 'Claude Code (CLI subscription)',
+    description: "Native Claude Code harness using the user's authenticated subscription.",
+    credential: 'subscription',
+    enabledByDefault: true,
+    models: [
+      {
+        id: 'fable',
+        label: 'Fable 5.1',
+        tier: 'powerful',
+        costTier: 'high',
+        contextWindow: 1_000_000,
+        maxOutputTokens: 128_000,
+        runtimeKind: 'native-harness',
+        nativeAlias: true,
+        capabilities: POWERFUL_REASONING,
+        reasoning: CLAUDE_REASONING,
+        notes: 'Claude Code alias for claude-fable-5-1.',
+      },
+      {
+        id: 'opus',
+        label: 'Opus 5',
+        tier: 'powerful',
+        costTier: 'high',
+        contextWindow: 1_000_000,
+        runtimeKind: 'native-harness',
+        nativeAlias: true,
+        capabilities: POWERFUL_REASONING,
+        reasoning: CLAUDE_REASONING,
+      },
+      {
+        id: 'sonnet',
+        label: 'Sonnet 5',
+        tier: 'standard',
+        costTier: 'medium',
+        contextWindow: 1_000_000,
+        runtimeKind: 'native-harness',
+        nativeAlias: true,
+        capabilities: STANDARD_REASONING,
+        reasoning: CLAUDE_REASONING,
+      },
+      {
+        id: 'haiku',
+        label: 'Haiku 4.5',
+        tier: 'fast',
+        costTier: 'low',
+        contextWindow: 200_000,
+        runtimeKind: 'native-harness',
+        nativeAlias: true,
+        capabilities: FAST_REASONING,
+        reasoning: { ...CLAUDE_REASONING, efforts: ['low', 'medium', 'high'] },
+      },
+    ],
+  },
+  anthropic: {
+    id: 'anthropic',
+    type: 'anthropic',
+    label: 'Anthropic (API key)',
+    description: 'Direct Anthropic API access without the Claude Code native harness.',
+    envKey: 'ANTHROPIC_API_KEY',
+    credential: 'api-key',
+    enabledByDefault: true,
+    models: [
+      {
+        id: 'claude-fable-5-1',
+        label: 'Claude Fable 5.1',
+        tier: 'powerful',
+        costTier: 'high',
+        contextWindow: 1_000_000,
+        maxOutputTokens: 128_000,
+        runtimeKind: 'api',
+        capabilities: POWERFUL_REASONING,
+        reasoning: CLAUDE_REASONING,
+        notes: 'Thinking is always on; steer with effort.',
+      },
+      {
+        id: 'claude-opus-5',
+        label: 'Claude Opus 5',
+        tier: 'powerful',
+        costTier: 'high',
+        contextWindow: 1_000_000,
+        maxOutputTokens: 128_000,
+        runtimeKind: 'api',
+        capabilities: POWERFUL_REASONING,
+        reasoning: CLAUDE_REASONING,
+      },
+      {
+        id: 'claude-sonnet-5',
+        label: 'Claude Sonnet 5',
+        tier: 'standard',
+        costTier: 'medium',
+        contextWindow: 1_000_000,
+        maxOutputTokens: 128_000,
+        runtimeKind: 'api',
+        capabilities: STANDARD_REASONING,
+        reasoning: CLAUDE_REASONING,
+      },
+      {
+        id: 'claude-haiku-4-5',
+        label: 'Claude Haiku 4.5',
+        tier: 'fast',
+        costTier: 'low',
+        contextWindow: 200_000,
+        runtimeKind: 'api',
+        capabilities: FAST_REASONING,
+        reasoning: { ...CLAUDE_REASONING, efforts: ['low', 'medium', 'high'] },
+        notes: 'Takes a thinking budget, not an effort ladder.',
+      },
+      // Previous generation, still served.
+      {
+        id: 'claude-fable-5',
+        label: 'Claude Fable 5',
+        tier: 'powerful',
+        costTier: 'high',
+        contextWindow: 1_000_000,
+        maxOutputTokens: 128_000,
+        runtimeKind: 'api',
+        capabilities: POWERFUL_REASONING,
+        reasoning: CLAUDE_REASONING,
+      },
+      {
+        id: 'claude-opus-4-8',
+        label: 'Claude Opus 4.8',
+        tier: 'powerful',
+        costTier: 'high',
+        contextWindow: 1_000_000,
+        maxOutputTokens: 128_000,
+        runtimeKind: 'api',
+        capabilities: POWERFUL_REASONING,
+        reasoning: CLAUDE_REASONING,
+      },
+      {
+        id: 'claude-opus-4-7',
+        label: 'Claude Opus 4.7',
+        tier: 'powerful',
+        costTier: 'high',
+        contextWindow: 1_000_000,
+        maxOutputTokens: 128_000,
+        runtimeKind: 'api',
+        capabilities: POWERFUL_REASONING,
+        reasoning: CLAUDE_REASONING,
+      },
+      {
+        id: 'claude-opus-4-6',
+        label: 'Claude Opus 4.6',
+        tier: 'powerful',
+        costTier: 'high',
+        contextWindow: 1_000_000,
+        maxOutputTokens: 128_000,
+        runtimeKind: 'api',
+        capabilities: POWERFUL_REASONING,
+        reasoning: { ...CLAUDE_REASONING, efforts: ['low', 'medium', 'high', 'max'] },
+      },
+      {
+        id: 'claude-sonnet-4-6',
+        label: 'Claude Sonnet 4.6',
+        tier: 'standard',
+        costTier: 'medium',
+        contextWindow: 1_000_000,
+        maxOutputTokens: 128_000,
+        runtimeKind: 'api',
+        capabilities: STANDARD_REASONING,
+        reasoning: { ...CLAUDE_REASONING, efforts: ['low', 'medium', 'high', 'max'] },
+      },
+    ],
+  },
+  codex: {
+    id: 'codex',
+    type: 'codex',
+    label: 'Codex CLI (OpenAI subscription)',
+    description: "Native Codex harness using the user's authenticated OpenAI/Codex subscription.",
+    credential: 'subscription',
+    enabledByDefault: true,
+    models: [
+      // Slugs, effort ladders and the 272k session window come from the local
+      // Codex catalogue (~/.codex/models_cache.json); the same models carry a
+      // 1,050,000 window on the platform API, which the openai entry records.
+      {
+        id: 'gpt-6-astra',
+        label: 'GPT-6 Astra',
+        tier: 'powerful',
+        costTier: 'high',
+        contextWindow: 272_000,
+        runtimeKind: 'native-harness',
+        capabilities: POWERFUL_REASONING,
+        reasoning: CODEX_REASONING_ULTRA,
+        notes: 'The Codex CLI default.',
+      },
+      {
+        id: 'gpt-6-sol',
+        label: 'GPT-6 Sol',
+        tier: 'powerful',
+        costTier: 'high',
+        contextWindow: 272_000,
+        runtimeKind: 'native-harness',
+        capabilities: POWERFUL_REASONING,
+        reasoning: CODEX_REASONING_ULTRA,
+        notes: 'Workhorse for coding and everyday work.',
+      },
+      {
+        id: 'gpt-6-luna',
+        label: 'GPT-6 Luna',
+        tier: 'fast',
+        costTier: 'low',
+        contextWindow: 272_000,
+        runtimeKind: 'native-harness',
+        capabilities: FAST_REASONING,
+        reasoning: CODEX_REASONING,
+        notes: "Foundry's decision model.",
+      },
+      {
+        id: 'gpt-5.6-sol',
+        label: 'GPT-5.6 Sol',
+        tier: 'powerful',
+        costTier: 'high',
+        contextWindow: 272_000,
+        runtimeKind: 'native-harness',
+        capabilities: POWERFUL_REASONING,
+        reasoning: CODEX_REASONING_ULTRA,
+      },
+      {
+        id: 'gpt-5.6-terra',
+        label: 'GPT-5.6 Terra',
+        tier: 'standard',
+        costTier: 'medium',
+        contextWindow: 272_000,
+        runtimeKind: 'native-harness',
+        capabilities: STANDARD_REASONING,
+        reasoning: CODEX_REASONING_ULTRA,
+      },
+      {
+        id: 'gpt-5.6-luna',
+        label: 'GPT-5.6 Luna',
+        tier: 'fast',
+        costTier: 'low',
+        contextWindow: 272_000,
+        runtimeKind: 'native-harness',
+        capabilities: FAST_REASONING,
+        reasoning: CODEX_REASONING,
+      },
+      {
+        id: 'gpt-5.5',
+        label: 'GPT-5.5',
+        tier: 'standard',
+        costTier: 'medium',
+        contextWindow: 272_000,
+        runtimeKind: 'native-harness',
+        capabilities: STANDARD_REASONING,
+        reasoning: {
+          ...CODEX_REASONING,
+          efforts: ['low', 'medium', 'high', 'xhigh'],
+          fallback: 'xhigh',
+        },
+      },
+    ],
+  },
+  openai: {
+    id: 'openai',
+    type: 'openai',
+    label: 'OpenAI',
+    description: 'OpenAI API models for classification, routing, and execution.',
+    envKey: 'OPENAI_API_KEY',
+    credential: 'api-key',
+    apiRoot: 'https://api.openai.com/v1',
+    enabledByDefault: true,
+    models: [
+      {
+        id: 'gpt-6-astra',
+        label: 'GPT-6 Astra',
+        tier: 'powerful',
+        costTier: 'high',
+        contextWindow: 1_050_000,
+        maxOutputTokens: 128_000,
+        runtimeKind: 'api',
+        capabilities: POWERFUL_REASONING,
+        reasoning: {
+          ...OPENAI_REASONING,
+          efforts: ['low', 'medium', 'high', 'xhigh', 'max'],
+          fallback: 'low',
+        },
+        notes: 'Rejects effort "none" with a 400; tool calling needs the Responses API.',
+      },
+      {
+        id: 'gpt-6-sol',
+        label: 'GPT-6 Sol',
+        tier: 'powerful',
+        costTier: 'high',
+        contextWindow: 1_050_000,
+        maxOutputTokens: 128_000,
+        runtimeKind: 'api',
+        capabilities: POWERFUL_REASONING,
+        reasoning: OPENAI_REASONING,
+      },
+      {
+        id: 'gpt-6-luna',
+        label: 'GPT-6 Luna',
+        tier: 'fast',
+        costTier: 'low',
+        contextWindow: 1_050_000,
+        maxOutputTokens: 128_000,
+        runtimeKind: 'api',
+        capabilities: FAST_REASONING,
+        reasoning: OPENAI_REASONING,
+        notes: "Foundry's decision model.",
+      },
+      {
+        id: 'gpt-5.6-sol',
+        label: 'GPT-5.6 Sol',
+        tier: 'powerful',
+        costTier: 'high',
+        contextWindow: 1_050_000,
+        maxOutputTokens: 128_000,
+        runtimeKind: 'api',
+        capabilities: POWERFUL_REASONING,
+        reasoning: OPENAI_REASONING,
+      },
+      {
+        id: 'gpt-5.6-terra',
+        label: 'GPT-5.6 Terra',
+        tier: 'standard',
+        costTier: 'medium',
+        contextWindow: 1_050_000,
+        maxOutputTokens: 128_000,
+        runtimeKind: 'api',
+        capabilities: STANDARD_REASONING,
+        reasoning: OPENAI_REASONING,
+      },
+      {
+        id: 'gpt-5.6-luna',
+        label: 'GPT-5.6 Luna',
+        tier: 'fast',
+        costTier: 'low',
+        contextWindow: 1_050_000,
+        maxOutputTokens: 128_000,
+        runtimeKind: 'api',
+        capabilities: FAST_REASONING,
+        reasoning: OPENAI_REASONING,
+      },
+      {
+        id: 'gpt-5.5',
+        label: 'GPT-5.5',
+        tier: 'standard',
+        costTier: 'medium',
+        contextWindow: 1_050_000,
+        maxOutputTokens: 128_000,
+        runtimeKind: 'api',
+        capabilities: STANDARD_REASONING,
+        reasoning: { ...OPENAI_REASONING, efforts: ['none', 'low', 'medium', 'high', 'xhigh'] },
+      },
+    ],
+  },
+  gemini: {
+    id: 'gemini',
+    type: 'gemini',
+    label: 'Google Gemini',
+    description: 'Gemini API models for high-volume agent work.',
+    envKey: 'GEMINI_API_KEY',
+    credential: 'api-key',
+    enabledByDefault: true,
+    models: [
+      // The Pro and Flash lines carry different version numbers; Pro has no GA
+      // successor to 3.1 yet. thinking_level takes low/medium/high, not minimal.
+      {
+        id: 'gemini-3.1-pro-preview',
+        label: 'Gemini 3.1 Pro',
+        tier: 'powerful',
+        costTier: 'high',
+        contextWindow: 1_048_576,
+        maxOutputTokens: 65_536,
+        runtimeKind: 'api',
+        capabilities: POWERFUL_REASONING,
+        reasoning: GEMINI_REASONING,
+      },
+      {
+        id: 'gemini-3.8-flash',
+        label: 'Gemini 3.8 Flash',
+        tier: 'standard',
+        costTier: 'medium',
+        contextWindow: 1_048_576,
+        maxOutputTokens: 65_536,
+        runtimeKind: 'api',
+        capabilities: STANDARD_REASONING,
+        reasoning: GEMINI_REASONING,
+      },
+      {
+        id: 'gemini-3.7-flash',
+        label: 'Gemini 3.7 Flash',
+        tier: 'standard',
+        costTier: 'medium',
+        contextWindow: 1_048_576,
+        maxOutputTokens: 65_536,
+        runtimeKind: 'api',
+        capabilities: STANDARD_REASONING,
+        reasoning: GEMINI_REASONING,
+      },
+      {
+        id: 'gemini-3.5-flash-lite',
+        label: 'Gemini 3.5 Flash Lite',
+        tier: 'fast',
+        costTier: 'low',
+        runtimeKind: 'api',
+        capabilities: FAST_REASONING,
+        reasoning: GEMINI_REASONING,
+      },
+      {
+        id: 'gemini-3.1-flash-lite',
+        label: 'Gemini 3.1 Flash Lite',
+        tier: 'fast',
+        costTier: 'low',
+        contextWindow: 1_048_576,
+        maxOutputTokens: 65_536,
+        runtimeKind: 'api',
+        capabilities: FAST_REASONING,
+        reasoning: GEMINI_REASONING,
+      },
+    ],
+  },
+  meta: {
+    id: 'meta',
+    type: 'openai-compatible',
+    // Llama is not Meta's frontier line any more; Muse Spark replaced it in April 2026
+    // and there is no "Llama 5". The open-weight Llama entries under openrouter,
+    // together, fireworks and ollama are the older line, and stay where they are.
+    label: 'Meta (Muse Spark)',
+    description:
+      "Meta's Model API. OpenAI-compatible; it also serves an Anthropic-compatible /v1/messages. The documented key variable is the unprefixed MODEL_API_KEY.",
+    envKey: 'MODEL_API_KEY',
+    credential: 'api-key',
+    apiRoot: 'https://api.meta.ai/v1',
+    enabledByDefault: true,
+    models: [
+      // No max output is documented for any Muse Spark model, so none is recorded;
+      // a guess here would be sent as max_completion_tokens and truncate real replies.
+      // The -contributor ids are the same models at $0.10/$0.20 per M instead of
+      // $1.25/$4.25, bought with permission to train on the traffic. That is priced
+      // into the label as well as `trainsOnInput`, so no surface can show one without the other.
+      {
+        id: 'muse-spark-1.3',
+        label: 'Muse Spark 1.3',
+        tier: 'powerful',
+        costTier: 'high',
+        contextWindow: 1_048_576,
+        runtimeKind: 'api',
+        capabilities: POWERFUL_REASONING,
+        reasoning: MUSE_REASONING,
+      },
+      {
+        id: 'muse-spark-1.3-contributor',
+        label: 'Muse Spark 1.3 (Contributor — Meta trains on your data)',
+        tier: 'powerful',
+        costTier: 'low',
+        contextWindow: 1_048_576,
+        runtimeKind: 'api',
+        capabilities: POWERFUL_REASONING,
+        reasoning: MUSE_REASONING,
+        trainsOnInput: true,
+        notes:
+          '$0.10/$0.20 per M instead of $1.25/$4.25, in exchange for Meta training future models on your prompts and completions.',
+      },
+      {
+        id: 'muse-spark-1.2',
+        label: 'Muse Spark 1.2',
+        tier: 'powerful',
+        costTier: 'high',
+        contextWindow: 1_048_576,
+        runtimeKind: 'api',
+        capabilities: POWERFUL_REASONING,
+        reasoning: MUSE_REASONING,
+      },
+      {
+        id: 'muse-spark-1.2-contributor',
+        label: 'Muse Spark 1.2 (Contributor — Meta trains on your data)',
+        tier: 'powerful',
+        costTier: 'low',
+        contextWindow: 1_048_576,
+        runtimeKind: 'api',
+        capabilities: POWERFUL_REASONING,
+        reasoning: MUSE_REASONING,
+        trainsOnInput: true,
+        notes:
+          '$0.10/$0.20 per M instead of $1.25/$4.25, in exchange for Meta training future models on your prompts and completions.',
+      },
+      {
+        id: 'muse-spark-1.1',
+        label: 'Muse Spark 1.1',
+        tier: 'powerful',
+        costTier: 'high',
+        contextWindow: 1_048_576,
+        runtimeKind: 'api',
+        capabilities: POWERFUL_REASONING,
+        reasoning: MUSE_REASONING,
+        notes: 'No contributor tier is published for 1.1.',
+      },
+      // Not registered: muse-image, muse-voice-transcribe-1.0 and sam-3.1 are not
+      // chat models and do not answer /v1/chat/completions.
+    ],
+  },
+  xai: {
+    id: 'xai',
+    type: 'xai',
+    label: 'xAI (Grok)',
+    description:
+      'Grok models. Near-OpenAI-compatible; reasoning effort is a nested object and cannot be turned off.',
+    envKey: 'XAI_API_KEY',
+    credential: 'api-key',
+    apiRoot: 'https://api.x.ai/v1',
+    enabledByDefault: true,
+    models: [
+      {
+        id: 'grok-4.7',
+        label: 'Grok 4.7',
+        tier: 'powerful',
+        costTier: 'high',
+        contextWindow: 500_000,
+        runtimeKind: 'api',
+        capabilities: POWERFUL_REASONING,
+        reasoning: GROK_REASONING_XHIGH,
+        notes: 'Reasoning effort low/medium/high/xhigh; cannot be disabled.',
+      },
+      {
+        id: 'grok-4.6',
+        label: 'Grok 4.6',
+        tier: 'powerful',
+        costTier: 'high',
+        contextWindow: 500_000,
+        runtimeKind: 'api',
+        capabilities: POWERFUL_REASONING,
+        reasoning: GROK_REASONING_XHIGH,
+      },
+      {
+        id: 'grok-4.5',
+        label: 'Grok 4.5',
+        tier: 'standard',
+        costTier: 'medium',
+        contextWindow: 500_000,
+        runtimeKind: 'api',
+        capabilities: STANDARD_REASONING,
+        reasoning: GROK_REASONING,
+        notes: 'xhigh effort is served as high.',
+      },
+      {
+        id: 'grok-4.3',
+        label: 'Grok 4.3',
+        tier: 'standard',
+        costTier: 'medium',
+        contextWindow: 1_000_000,
+        runtimeKind: 'api',
+        capabilities: STANDARD,
+      },
+      {
+        id: 'grok-4.20-0309-reasoning',
+        label: 'Grok 4.20 (reasoning)',
+        tier: 'powerful',
+        costTier: 'high',
+        contextWindow: 1_000_000,
+        runtimeKind: 'api',
+        capabilities: POWERFUL_REASONING,
+        reasoning: GROK_REASONING,
+      },
+      {
+        id: 'grok-4.20-0309-non-reasoning',
+        label: 'Grok 4.20 (non-reasoning)',
+        tier: 'fast',
+        costTier: 'low',
+        contextWindow: 1_000_000,
+        runtimeKind: 'api',
+        capabilities: FAST,
+      },
+      {
+        id: 'grok-build-0.1',
+        label: 'Grok Build 0.1',
+        tier: 'fast',
+        costTier: 'low',
+        contextWindow: 256_000,
+        runtimeKind: 'api',
+        capabilities: FAST,
+      },
+    ],
+  },
+  deepseek: {
+    id: 'deepseek',
+    type: 'openai-compatible',
+    label: 'DeepSeek',
+    description:
+      'DeepSeek API. OpenAI-compatible; penalties are silently ignored and output defaults differ.',
+    envKey: 'DEEPSEEK_API_KEY',
+    credential: 'api-key',
+    apiRoot: 'https://api.deepseek.com',
+    enabledByDefault: true,
+    models: [
+      {
+        id: 'deepseek-flash',
+        label: 'DeepSeek Flash (V4.1)',
+        tier: 'fast',
+        costTier: 'low',
+        contextWindow: 1_000_000,
+        maxOutputTokens: 393_216,
+        runtimeKind: 'api',
+        capabilities: FAST_REASONING,
+        reasoning: { ...DEEPSEEK_REASONING, fallback: 'low' },
+        notes: 'Thinking on by default; the legacy deepseek-v4-flash ids route here.',
+      },
+      {
+        id: 'deepseek-v4-pro',
+        label: 'DeepSeek V4 Pro',
+        tier: 'powerful',
+        costTier: 'high',
+        contextWindow: 1_000_000,
+        maxOutputTokens: 393_216,
+        runtimeKind: 'api',
+        capabilities: POWERFUL_REASONING,
+        reasoning: DEEPSEEK_REASONING,
+      },
+    ],
+  },
+  kimi: {
+    id: 'kimi',
+    type: 'openai-compatible',
+    label: 'Kimi (Moonshot AI)',
+    description: 'Moonshot Kimi models. OpenAI-compatible; temperature range is 0-1, not 0-2.',
+    envKey: 'MOONSHOT_API_KEY',
+    credential: 'api-key',
+    apiRoot: 'https://api.moonshot.ai/v1',
+    enabledByDefault: true,
+    models: [
+      {
+        id: 'kimi-k3',
+        label: 'Kimi K3',
+        tier: 'powerful',
+        costTier: 'high',
+        contextWindow: 1_048_576,
+        maxOutputTokens: 131_072,
+        runtimeKind: 'api',
+        capabilities: POWERFUL_REASONING,
+        reasoning: EFFORT_ONLY,
+        notes: 'Thinking is always on and cannot be disabled.',
+      },
+      {
+        id: 'kimi-k2.6',
+        label: 'Kimi K2.6',
+        tier: 'standard',
+        costTier: 'medium',
+        contextWindow: 262_144,
+        runtimeKind: 'api',
+        capabilities: STANDARD_REASONING,
+        reasoning: EFFORT_ONLY,
+        notes: 'Thinking requires temperature 1.0; non-thinking requires 0.6.',
+      },
+      {
+        id: 'kimi-k2.7-code',
+        label: 'Kimi K2.7 Code',
+        tier: 'standard',
+        costTier: 'medium',
+        contextWindow: 262_144,
+        runtimeKind: 'api',
+        capabilities: STANDARD,
+        notes: 'tool_choice required is unsupported.',
+      },
+      {
+        id: 'kimi-k2.7-code-highspeed',
+        label: 'Kimi K2.7 Code (high speed)',
+        tier: 'fast',
+        costTier: 'medium',
+        contextWindow: 262_144,
+        runtimeKind: 'api',
+        capabilities: FAST,
+      },
+    ],
+  },
+  glm: {
+    id: 'glm',
+    type: 'openai-compatible',
+    label: 'GLM (Z.ai / Zhipu)',
+    description:
+      "GLM over Z.ai's OpenAI-compatible v4 endpoint. Mainland Zhipu is https://open.bigmodel.cn/api/paas/v4.",
+    envKey: 'ZAI_API_KEY',
+    credential: 'api-key',
+    apiRoot: 'https://api.z.ai/api/paas/v4',
+    enabledByDefault: true,
+    models: [
+      {
+        id: 'glm-4.6',
+        label: 'GLM-4.6',
+        tier: 'standard',
+        costTier: 'medium',
+        contextWindow: 204_800,
+        maxOutputTokens: 128_000,
+        runtimeKind: 'api',
+        capabilities: STANDARD_REASONING,
+        reasoning: EFFORT_ONLY,
+        notes: 'Thinking is opt-in via the thinking parameter.',
+      },
+      {
+        id: 'glm-4.7',
+        label: 'GLM-4.7',
+        tier: 'standard',
+        costTier: 'medium',
+        contextWindow: 200_000,
+        maxOutputTokens: 128_000,
+        runtimeKind: 'api',
+        capabilities: STANDARD_REASONING,
+        reasoning: EFFORT_ONLY,
+      },
+      {
+        id: 'glm-4.7-flash',
+        label: 'GLM-4.7 Flash',
+        tier: 'fast',
+        costTier: 'low',
+        contextWindow: 200_000,
+        maxOutputTokens: 128_000,
+        runtimeKind: 'api',
+        capabilities: FAST_REASONING,
+        reasoning: EFFORT_ONLY,
+        notes: "Free on Z.ai's pricing page.",
+      },
+      {
+        id: 'glm-5.3',
+        label: 'GLM-5.3',
+        tier: 'powerful',
+        costTier: 'high',
+        contextWindow: 1_000_000,
+        maxOutputTokens: 128_000,
+        runtimeKind: 'api',
+        capabilities: POWERFUL_REASONING,
+        reasoning: EFFORT_ONLY,
+      },
+      {
+        id: 'glm-5.3-flash',
+        label: 'GLM-5.3 Flash',
+        tier: 'fast',
+        costTier: 'low',
+        contextWindow: 1_000_000,
+        maxOutputTokens: 128_000,
+        runtimeKind: 'api',
+        capabilities: FAST_REASONING,
+        reasoning: EFFORT_ONLY,
+      },
+      // The 4.x and 5.x lines run in parallel; neither is sunset.
+      {
+        id: 'glm-5.2',
+        label: 'GLM-5.2',
+        tier: 'powerful',
+        costTier: 'high',
+        contextWindow: 1_000_000,
+        maxOutputTokens: 128_000,
+        runtimeKind: 'api',
+        capabilities: POWERFUL_REASONING,
+        reasoning: EFFORT_ONLY,
+      },
+      {
+        id: 'glm-5.1',
+        label: 'GLM-5.1',
+        tier: 'standard',
+        costTier: 'medium',
+        contextWindow: 200_000,
+        maxOutputTokens: 128_000,
+        runtimeKind: 'api',
+        capabilities: STANDARD_REASONING,
+        reasoning: EFFORT_ONLY,
+      },
+      {
+        id: 'glm-5',
+        label: 'GLM-5',
+        tier: 'standard',
+        costTier: 'medium',
+        contextWindow: 200_000,
+        maxOutputTokens: 128_000,
+        runtimeKind: 'api',
+        capabilities: STANDARD_REASONING,
+        reasoning: EFFORT_ONLY,
+      },
+      {
+        id: 'glm-4.5',
+        label: 'GLM-4.5',
+        tier: 'standard',
+        costTier: 'medium',
+        contextWindow: 128_000,
+        maxOutputTokens: 96_000,
+        runtimeKind: 'api',
+        capabilities: STANDARD_REASONING,
+        reasoning: EFFORT_ONLY,
+      },
+      {
+        id: 'glm-4.5-air',
+        label: 'GLM-4.5 Air',
+        tier: 'fast',
+        costTier: 'low',
+        contextWindow: 128_000,
+        maxOutputTokens: 96_000,
+        runtimeKind: 'api',
+        capabilities: FAST_REASONING,
+        reasoning: EFFORT_ONLY,
+      },
+      {
+        id: 'glm-4.5-flash',
+        label: 'GLM-4.5 Flash',
+        tier: 'fast',
+        costTier: 'low',
+        contextWindow: 128_000,
+        maxOutputTokens: 96_000,
+        runtimeKind: 'api',
+        capabilities: FAST_REASONING,
+        reasoning: EFFORT_ONLY,
+      },
+    ],
+  },
+  qwen: {
+    id: 'qwen',
+    type: 'openai-compatible',
+    label: 'Qwen (Alibaba Model Studio)',
+    // The legacy static domain still works; Alibaba now recommends the per-workspace
+    // domain (https://{WorkspaceId}.ap-southeast-1.maas.aliyuncs.com/compatible-mode/v1),
+    // which has no static form. Keys are region-bound: a wrong-region key reads as a bad key.
+    description:
+      'DashScope compatible mode, Singapore. Set baseUrl to your workspace or region domain; mainland China is https://dashscope.aliyuncs.com/compatible-mode/v1.',
+    envKey: 'DASHSCOPE_API_KEY',
+    credential: 'api-key',
+    apiRoot: 'https://dashscope-intl.aliyuncs.com/compatible-mode/v1',
+    enabledByDefault: true,
+    models: [
+      // Alibaba's own grouping: 3.8 and 3.7-plus/flash are "recommended"; 3.7-max
+      // and the 3.6 line are "legacy"; the undotted names are older still.
+      {
+        id: 'qwen3.8-max',
+        label: 'Qwen3.8 Max',
+        tier: 'powerful',
+        costTier: 'high',
+        contextWindow: 1_000_000,
+        maxOutputTokens: 131_072,
+        runtimeKind: 'api',
+        capabilities: POWERFUL_REASONING,
+        reasoning: EFFORT_ONLY,
+      },
+      {
+        id: 'qwen3.8-flash',
+        label: 'Qwen3.8 Flash',
+        tier: 'fast',
+        costTier: 'low',
+        contextWindow: 1_000_000,
+        maxOutputTokens: 131_072,
+        runtimeKind: 'api',
+        capabilities: FAST_REASONING,
+        reasoning: EFFORT_ONLY,
+      },
+      {
+        id: 'qwen3.7-plus',
+        label: 'Qwen3.7 Plus',
+        tier: 'standard',
+        costTier: 'medium',
+        runtimeKind: 'api',
+        capabilities: STANDARD_REASONING,
+        reasoning: EFFORT_ONLY,
+      },
+      {
+        id: 'qwen3.7-flash',
+        label: 'Qwen3.7 Flash',
+        tier: 'fast',
+        costTier: 'low',
+        runtimeKind: 'api',
+        capabilities: FAST_REASONING,
+        reasoning: EFFORT_ONLY,
+      },
+      {
+        id: 'qwen3.7-max',
+        label: 'Qwen3.7 Max',
+        tier: 'powerful',
+        costTier: 'high',
+        contextWindow: 991_808,
+        maxOutputTokens: 131_072,
+        runtimeKind: 'api',
+        capabilities: POWERFUL_REASONING,
+        reasoning: EFFORT_ONLY,
+      },
+      {
+        id: 'qwen3.6-plus',
+        label: 'Qwen3.6 Plus',
+        tier: 'standard',
+        costTier: 'medium',
+        contextWindow: 1_000_000,
+        maxOutputTokens: 65_536,
+        runtimeKind: 'api',
+        capabilities: STANDARD_REASONING,
+        reasoning: EFFORT_ONLY,
+      },
+      {
+        id: 'qwen3.6-flash',
+        label: 'Qwen3.6 Flash',
+        tier: 'fast',
+        costTier: 'low',
+        contextWindow: 1_000_000,
+        maxOutputTokens: 65_536,
+        runtimeKind: 'api',
+        capabilities: FAST_REASONING,
+        reasoning: EFFORT_ONLY,
+      },
+      {
+        id: 'qwen-plus',
+        label: 'Qwen Plus (legacy)',
+        tier: 'standard',
+        costTier: 'low',
+        contextWindow: 1_000_000,
+        maxOutputTokens: 32_768,
+        runtimeKind: 'api',
+        capabilities: STANDARD,
+      },
+      {
+        id: 'qwen-flash',
+        label: 'Qwen Flash (legacy)',
+        tier: 'fast',
+        costTier: 'low',
+        contextWindow: 1_000_000,
+        maxOutputTokens: 32_768,
+        runtimeKind: 'api',
+        capabilities: FAST,
+      },
+    ],
+  },
+  openrouter: {
+    id: 'openrouter',
+    type: 'openai-compatible',
+    label: 'OpenRouter',
+    description:
+      'One key, many vendors. Model ids are namespaced vendor/model; a :free suffix marks free models.',
+    envKey: 'OPENROUTER_API_KEY',
+    credential: 'api-key',
+    apiRoot: 'https://openrouter.ai/api/v1',
+    enabledByDefault: true,
+    models: [
+      {
+        id: 'meta-llama/llama-3.1-8b-instruct',
+        label: 'Llama 3.1 8B Instruct',
+        tier: 'fast',
+        costTier: 'low',
+        contextWindow: 131_072,
+        runtimeKind: 'api',
+        capabilities: FAST,
+        notes: 'Cheapest Llama on OpenRouter.',
+      },
+      {
+        id: 'meta-llama/llama-3.3-70b-instruct',
+        label: 'Llama 3.3 70B Instruct',
+        tier: 'standard',
+        costTier: 'medium',
+        contextWindow: 131_072,
+        runtimeKind: 'api',
+        capabilities: STANDARD,
+      },
+      {
+        id: 'meta-llama/llama-4-scout',
+        label: 'Llama 4 Scout',
+        tier: 'standard',
+        costTier: 'low',
+        contextWindow: 1_310_720,
+        runtimeKind: 'api',
+        capabilities: STANDARD,
+      },
+      {
+        id: 'meta-llama/llama-4-maverick',
+        label: 'Llama 4 Maverick',
+        tier: 'powerful',
+        costTier: 'medium',
+        contextWindow: 1_048_576,
+        runtimeKind: 'api',
+        capabilities: POWERFUL,
+      },
+      {
+        id: 'mistralai/mistral-nemo',
+        label: 'Mistral Nemo',
+        tier: 'fast',
+        costTier: 'low',
+        runtimeKind: 'api',
+        capabilities: FAST,
+        notes: 'Cheapest model in the OpenRouter catalogue.',
+      },
+      {
+        id: 'z-ai/glm-5.2:free',
+        label: 'GLM-5.2 (free)',
+        tier: 'standard',
+        costTier: 'low',
+        runtimeKind: 'api',
+        capabilities: STANDARD,
+        notes: 'Free tier; the cheapest hosted judgment model that needs no local runtime.',
+      },
+      {
+        id: 'qwen/qwen3.8-27b:free',
+        label: 'Qwen3.8 27B (free)',
+        tier: 'standard',
+        costTier: 'low',
+        runtimeKind: 'api',
+        capabilities: STANDARD,
+        notes: 'Free tier.',
+      },
+    ],
+  },
+  groq: {
+    id: 'groq',
+    type: 'openai-compatible',
+    label: 'Groq',
+    description:
+      'Groq LPU inference. Meta Llama chat models are enterprise-contract only here; GPT-OSS is the self-serve open-weight line.',
+    envKey: 'GROQ_API_KEY',
+    credential: 'api-key',
+    apiRoot: 'https://api.groq.com/openai/v1',
+    enabledByDefault: true,
+    models: [
+      {
+        id: 'openai/gpt-oss-20b',
+        label: 'GPT-OSS 20B',
+        tier: 'fast',
+        costTier: 'low',
+        contextWindow: 131_072,
+        runtimeKind: 'api',
+        capabilities: FAST_REASONING,
+        reasoning: EFFORT_ONLY,
+      },
+      {
+        id: 'openai/gpt-oss-120b',
+        label: 'GPT-OSS 120B',
+        tier: 'standard',
+        costTier: 'low',
+        contextWindow: 131_072,
+        runtimeKind: 'api',
+        capabilities: STANDARD_REASONING,
+        reasoning: EFFORT_ONLY,
+      },
+    ],
+  },
+  together: {
+    id: 'together',
+    type: 'openai-compatible',
+    label: 'Together AI',
+    description:
+      'Hosted open-weight models. Llama ids are not uniform across versions; check each one.',
+    envKey: 'TOGETHER_API_KEY',
+    credential: 'api-key',
+    apiRoot: 'https://api.together.ai/v1',
+    enabledByDefault: true,
+    models: [
+      {
+        id: 'meta-llama/Llama-3.3-70B-Instruct-Turbo',
+        label: 'Llama 3.3 70B Instruct Turbo',
+        tier: 'standard',
+        costTier: 'medium',
+        contextWindow: 131_072,
+        runtimeKind: 'api',
+        capabilities: STANDARD,
+      },
+      {
+        id: 'meta-llama/Meta-Llama-3.1-405B-Instruct-Turbo',
+        label: 'Llama 3.1 405B Instruct Turbo',
+        tier: 'powerful',
+        costTier: 'high',
+        runtimeKind: 'api',
+        capabilities: POWERFUL,
+        notes: 'The 3.1 line keeps the Meta- prefix; 3.3 and 4 drop it.',
+      },
+      {
+        id: 'meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8',
+        label: 'Llama 4 Maverick 17B 128E FP8',
+        tier: 'powerful',
+        costTier: 'medium',
+        runtimeKind: 'api',
+        capabilities: POWERFUL,
+      },
+    ],
+  },
+  fireworks: {
+    id: 'fireworks',
+    type: 'openai-compatible',
+    label: 'Fireworks AI',
+    description: 'Hosted open-weight models. Model ids are full account paths, not bare names.',
+    envKey: 'FIREWORKS_API_KEY',
+    credential: 'api-key',
+    apiRoot: 'https://api.fireworks.ai/inference/v1',
+    enabledByDefault: true,
+    models: [
+      {
+        id: 'accounts/fireworks/models/llama-v3p1-8b-instruct',
+        label: 'Llama 3.1 8B Instruct',
+        tier: 'fast',
+        costTier: 'low',
+        runtimeKind: 'api',
+        capabilities: FAST,
+      },
+      {
+        id: 'accounts/fireworks/models/llama-v3p3-70b-instruct',
+        label: 'Llama 3.3 70B Instruct',
+        tier: 'standard',
+        costTier: 'medium',
+        runtimeKind: 'api',
+        capabilities: STANDARD,
+      },
+      {
+        id: 'accounts/fireworks/models/llama-v3p1-405b-instruct',
+        label: 'Llama 3.1 405B Instruct',
+        tier: 'powerful',
+        costTier: 'high',
+        runtimeKind: 'api',
+        capabilities: POWERFUL,
+      },
+      {
+        id: 'accounts/fireworks/models/llama4-scout-instruct-basic',
+        label: 'Llama 4 Scout',
+        tier: 'standard',
+        costTier: 'low',
+        runtimeKind: 'api',
+        capabilities: STANDARD,
+      },
+      {
+        id: 'accounts/fireworks/models/llama4-maverick-instruct-basic',
+        label: 'Llama 4 Maverick',
+        tier: 'powerful',
+        costTier: 'medium',
+        runtimeKind: 'api',
+        capabilities: POWERFUL,
+      },
+    ],
+  },
+  ollama: {
+    id: 'ollama',
+    type: 'openai-compatible',
+    label: 'Ollama (local)',
+    description:
+      'Local Ollama server. No credential: clients must send a bearer, the server ignores it.',
+    credential: 'local',
+    apiRoot: 'http://localhost:11434/v1',
+    enabledByDefault: true,
+    models: [
+      {
+        id: 'llama3.2:3b',
+        label: 'Llama 3.2 3B',
+        tier: 'fast',
+        costTier: 'low',
+        contextWindow: 128_000,
+        runtimeKind: 'api',
+        capabilities: FAST,
+        notes: 'ollama pull llama3.2:3b — the cheapest working judgment model.',
+      },
+      {
+        id: 'llama3.1:8b',
+        label: 'Llama 3.1 8B',
+        tier: 'fast',
+        costTier: 'low',
+        contextWindow: 128_000,
+        runtimeKind: 'api',
+        capabilities: FAST,
+      },
+      {
+        id: 'llama3.3:70b',
+        label: 'Llama 3.3 70B',
+        tier: 'standard',
+        costTier: 'low',
+        contextWindow: 128_000,
+        runtimeKind: 'api',
+        capabilities: STANDARD,
+      },
+      {
+        id: 'llama4:16x17b',
+        label: 'Llama 4 Scout',
+        tier: 'standard',
+        costTier: 'low',
+        contextWindow: 10_000_000,
+        runtimeKind: 'api',
+        capabilities: STANDARD,
+      },
+      {
+        id: 'llama4:128x17b',
+        label: 'Llama 4 Maverick',
+        tier: 'powerful',
+        costTier: 'low',
+        contextWindow: 1_000_000,
+        runtimeKind: 'api',
+        capabilities: POWERFUL,
+      },
+      {
+        id: 'qwen3:8b',
+        label: 'Qwen3 8B',
+        tier: 'fast',
+        costTier: 'low',
+        runtimeKind: 'api',
+        capabilities: FAST,
+      },
+      {
+        id: 'gpt-oss:20b',
+        label: 'GPT-OSS 20B',
+        tier: 'fast',
+        costTier: 'low',
+        runtimeKind: 'api',
+        capabilities: FAST_REASONING,
+        reasoning: EFFORT_ONLY,
+      },
+    ],
+  },
+  vllm: {
+    id: 'vllm',
+    type: 'openai-compatible',
+    label: 'vLLM (local)',
+    description:
+      'Local vLLM server on port 8000. Keyless unless started with --api-key; the model id is whatever --served-model-name says.',
+    credential: 'local',
+    apiRoot: 'http://localhost:8000/v1',
+    enabledByDefault: true,
+    models: [
+      {
+        id: 'NousResearch/Meta-Llama-3-8B-Instruct',
+        label: 'vLLM served model',
+        tier: 'standard',
+        costTier: 'low',
+        runtimeKind: 'api',
+        capabilities: STANDARD,
+        notes:
+          "vLLM's own quickstart example. Replace with this server's --served-model-name; vLLM has no fixed catalogue.",
+      },
+    ],
+  },
+  mistral: {
+    id: 'mistral',
+    type: 'openai-compatible',
+    label: 'Mistral AI',
+    description:
+      'Mistral API. Wire-compatible with OpenAI; JSON mode needs the word JSON in the prompt.',
+    envKey: 'MISTRAL_API_KEY',
+    credential: 'api-key',
+    apiRoot: 'https://api.mistral.ai/v1',
+    enabledByDefault: true,
+    models: [
+      {
+        id: 'ministral-3b-2512',
+        label: 'Ministral 3 3B',
+        tier: 'fast',
+        costTier: 'low',
+        runtimeKind: 'api',
+        capabilities: FAST,
+      },
+      {
+        id: 'ministral-8b-2512',
+        label: 'Ministral 3 8B',
+        tier: 'fast',
+        costTier: 'low',
+        runtimeKind: 'api',
+        capabilities: FAST,
+      },
+      {
+        id: 'ministral-14b-2512',
+        label: 'Ministral 3 14B',
+        tier: 'fast',
+        costTier: 'low',
+        runtimeKind: 'api',
+        capabilities: FAST,
+      },
+      {
+        id: 'mistral-small-2603',
+        label: 'Mistral Small 4',
+        tier: 'standard',
+        costTier: 'low',
+        contextWindow: 256_000,
+        runtimeKind: 'api',
+        capabilities: STANDARD_REASONING,
+        reasoning: EFFORT_ONLY,
+        notes: 'Unifies instruct, reasoning and coding.',
+      },
+      {
+        id: 'mistral-medium-3-5',
+        label: 'Mistral Medium 3.5',
+        tier: 'powerful',
+        costTier: 'medium',
+        contextWindow: 256_000,
+        runtimeKind: 'api',
+        capabilities: POWERFUL_REASONING,
+        reasoning: EFFORT_ONLY,
+        notes: 'Takes reasoning_effort.',
+      },
+      {
+        id: 'mistral-large-2512',
+        label: 'Mistral Large 3',
+        tier: 'powerful',
+        costTier: 'high',
+        contextWindow: 256_000,
+        runtimeKind: 'api',
+        capabilities: POWERFUL,
+      },
+    ],
+  },
+  typesafe: {
+    id: 'typesafe',
+    type: 'typesafe',
+    label: 'TypeSafe (Jev)',
+    description:
+      'Typed decisions — choice, score and noul with calibrated confidence. Judgment only: it answers typed questions, it does not complete chat.',
+    envKey: 'TYPESAFE_API_KEY',
+    credential: 'api-key',
+    apiRoot: 'https://api.typesafe.ai/v1',
+    enabledByDefault: false,
+    models: [
+      {
+        id: 'jev-latest',
+        label: 'Jev (latest)',
+        tier: 'fast',
+        costTier: 'low',
+        runtimeKind: 'typed-decision',
+        capabilities: ['judgment'],
+        notes:
+          'Not an LLMProvider: TypeSafeDecisionClient.evaluate takes typed questions, so it cannot back a chat agent.',
+      },
+    ],
+  },
+};
+
+/** Every registered model carrying all of the requested capabilities. */
+export function modelOptionsByCapability(...capabilities: ModelCapability[]): ModelSweepOption[] {
+  return Object.values(MODEL_REGISTRY).flatMap((provider) =>
+    provider.models
+      .filter((model) =>
+        capabilities.every((capability) => model.capabilities.includes(capability)),
+      )
+      .map((model) => ({ provider: provider.id, model: model.id, label: model.id })),
+  );
+}
+
+export function registryModel(providerId: string, modelId: string): FoundryModelInfo | undefined {
+  return MODEL_REGISTRY[providerId]?.models.find((model) => model.id === modelId);
+}
+
+/** Unknown provider/model pairs report no capabilities rather than throwing. */
+export function modelCapabilities(providerId: string, modelId: string): ModelCapability[] {
+  return [...(registryModel(providerId, modelId)?.capabilities ?? [])];
+}
+
+export function modelHasCapability(
+  providerId: string,
+  modelId: string,
+  capability: ModelCapability,
+): boolean {
+  return modelCapabilities(providerId, modelId).includes(capability);
+}
+
+/** Providers with at least one model carrying the capability. */
+export function providersWithCapability(capability: ModelCapability): FoundryProviderInfo[] {
+  return Object.values(MODEL_REGISTRY).filter((provider) =>
+    provider.models.some((model) => model.capabilities.includes(capability)),
+  );
+}
+
+/** The part of a Foundry configuration `resolveDecisionModel` reads. */
+export interface DecisionModelDefaults {
+  defaults: {
+    classifierProvider?: string;
+    classifierModel?: string;
+  };
+}
+
+/** What the API-token path would construct, before any credential is read. */
+export function resolveDecisionModel(config: DecisionModelDefaults): {
+  provider: string;
+  model: string;
+} {
+  const configured = config.defaults.classifierProvider;
+  // A classifier provider outside the registry — the subscription decision
+  // profile — belongs to the native path, which constructs its own client.
+  const provider = configured && MODEL_REGISTRY[configured] ? configured : DECISION_PROVIDER;
+  const model = (provider === configured && config.defaults.classifierModel) || DECISION_MODEL;
+  return { provider, model };
+}