@dreb/ai 2.43.0 → 2.43.3

package/README.md

@@ -794,7 +794,7 @@ interface OpenAICompletionsCompat {
   requiresToolResultName?: boolean;  // Whether tool results require the `name` field (default: false)
   requiresAssistantAfterToolResult?: boolean; // Whether tool results must be followed by an assistant message (default: false)
   requiresThinkingAsText?: boolean;  // Whether thinking blocks must be converted to text (default: false)
-  thinkingFormat?: 'openai' | 'zai' | 'qwen'; // Format for reasoning param: 'openai' uses reasoning_effort, 'zai' uses thinking: { type: "enabled" }, 'qwen' uses enable_thinking: boolean (default: openai)
+  thinkingFormat?: 'openai' | 'openrouter' | 'zai' | 'qwen' | 'qwen-chat-template' | 'kimi'; // Format for reasoning param: 'openai' uses reasoning_effort, 'openrouter' uses reasoning: { effort }, 'zai' uses top-level enable_thinking: boolean, 'qwen' uses top-level enable_thinking: boolean, 'qwen-chat-template' uses chat_template_kwargs.enable_thinking, and 'kimi' uses nested thinking: { type, effort? } plus prompt_cache_key (default: openai)
   openRouterRouting?: OpenRouterRouting; // OpenRouter routing preferences (default: {})
   vercelGatewayRouting?: VercelGatewayRouting; // Vercel AI Gateway routing preferences (default: {})
 }
@@ -830,16 +830,27 @@ await streamAnthropic(claude, context, options);
 
 ## Cross-Provider Handoffs
 
-The library supports seamless handoffs between different LLM providers within the same conversation. This allows you to switch models mid-conversation while preserving context, including thinking blocks, tool calls, and tool results.
+The library can prepare one conversation for another model without changing the stored `Context`. User messages, visible assistant text, tool calls, and tool results remain available, while model-bound tool signatures and IDs are normalized for the destination. Provider-specific reasoning state is handled conservatively.
 
-### How It Works
+### Reasoning State Compatibility
 
-When messages from one provider are sent to a different provider, the library automatically transforms them for compatibility:
+- **Exact model:** signed, encrypted, and redacted reasoning state is replayed unchanged.
+- **Compatible model switch:** structured reasoning is preserved only when source and target share a provider, both use `openai-completions`, the destination accepts structured reasoning, and the source uses a recognized plain field: `reasoning_content`, `reasoning`, or `reasoning_text`.
+- **Other readable reasoning:** it is retained as labelled plaintext inside `<reformatted-pre-switch-reasoning>` markers, with incompatible protocol metadata stripped.
+- **Opaque state:** redacted or encrypted-only reasoning is omitted for incompatible targets.
 
-- **User and tool result messages** are passed through unchanged
-- **Assistant messages from the same provider/API** are preserved as-is
-- **Assistant messages from different providers** have their thinking blocks converted to text with `<thinking>` tags
-- **Tool calls and regular text** are preserved unchanged
+Compatibility depends on provider, API, and signature behavior. This includes custom models: two models at the same endpoint are not compatible merely because their IDs or URLs match; they must share the configured provider identity as well.
+
+The transformation applies only to the outbound request. It does not mutate `Context.messages`, so a later switch back to the original model can replay its original reasoning state unless that history has been compacted or pruned by the caller.
+
+### Examples
+
+| Source and target | Outbound reasoning state |
+|---|---|
+| The same model | Original signed, encrypted, or redacted state is replayed unchanged. |
+| Two models under the same custom provider, both using `openai-completions`, where the destination accepts the source's `reasoning_content`, `reasoning`, or `reasoning_text` field | Recognized plain structured reasoning is preserved. |
+| Different providers or APIs with readable reasoning | Reasoning is sent as labelled plaintext in `<reformatted-pre-switch-reasoning>` markers. |
+| An incompatible target with redacted or encrypted-only reasoning | Opaque reasoning state is omitted. |
 
 ### Example: Multi-Provider Conversation
 
@@ -858,32 +869,17 @@ const claudeResponse = await complete(claude, context, {
 });
 context.messages.push(claudeResponse);
 
-// Switch to GPT-5 - it will see Claude's thinking as <thinking> tagged text
+// Switch to GPT-5. Readable Claude reasoning is reformatted for this outbound request.
 const gpt5 = getModel('openai', 'gpt-5-mini');
 context.messages.push({ role: 'user', content: 'Is that calculation correct?' });
 const gptResponse = await complete(gpt5, context);
 context.messages.push(gptResponse);
 
-// Switch to Gemini
-const gemini = getModel('google', 'gemini-2.5-flash');
+// Switching back to Claude can use the original Claude state in context.
 context.messages.push({ role: 'user', content: 'What was the original question?' });
-const geminiResponse = await complete(gemini, context);
+const finalClaudeResponse = await complete(claude, context);
 ```
 
-### Provider Compatibility
-
-All providers can handle messages from other providers, including:
-- Text content
-- Tool calls and tool results (including images in tool results)
-- Thinking/reasoning blocks (transformed to tagged text for cross-provider compatibility)
-- Aborted messages with partial content
-
-This enables flexible workflows where you can:
-- Start with a fast model for initial responses
-- Switch to a more capable model for complex reasoning
-- Use specialized models for specific tasks
-- Maintain conversation continuity across provider outages
-
 ## Context Serialization
 
 The `Context` object can be easily serialized and deserialized using standard JSON methods, making it simple to persist conversations, implement chat history, or transfer contexts between services:

package/dist/models.d.ts

@@ -47,6 +47,7 @@ export declare function calculateCost<TApi extends Api>(model: Model<TApi>, usag
  * Supported today:
  * - GPT-5.2 through GPT-5.6 model families
  * - Opus 4.6+ models (xhigh maps to adaptive effort "max" on Anthropic-compatible providers)
+ * - Kimi Code K3 (xhigh maps to its advertised "max" effort)
  */
 export declare function supportsXhigh<TApi extends Api>(model: Model<TApi>): boolean;
 /**

package/dist/models.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"models.d.ts","sourceRoot":"","sources":["../src/models.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,MAAM,EAAE,MAAM,uBAAuB,CAAC;AAC/C,OAAO,KAAK,EAAE,GAAG,EAAE,aAAa,EAAE,KAAK,EAAE,KAAK,EAAE,MAAM,YAAY,CAAC;AAanE,KAAK,QAAQ,CACZ,SAAS,SAAS,aAAa,EAC/B,QAAQ,SAAS,MAAM,CAAC,OAAO,MAAM,CAAC,CAAC,SAAS,CAAC,IAC9C,CAAC,OAAO,MAAM,CAAC,CAAC,SAAS,CAAC,CAAC,QAAQ,CAAC,SAAS;IAAE,GAAG,EAAE,MAAM,IAAI,CAAA;CAAE,GAAG,CAAC,IAAI,SAAS,GAAG,GAAG,IAAI,GAAG,KAAK,CAAC,GAAG,KAAK,CAAC;AAEjH,wBAAgB,QAAQ,CAAC,SAAS,SAAS,aAAa,EAAE,QAAQ,SAAS,MAAM,CAAC,OAAO,MAAM,CAAC,CAAC,SAAS,CAAC,EAC1G,QAAQ,EAAE,SAAS,EACnB,OAAO,EAAE,QAAQ,GACf,KAAK,CAAC,QAAQ,CAAC,SAAS,EAAE,QAAQ,CAAC,CAAC,CAGtC;AAED;;;;;;GAMG;AACH,wBAAgB,YAAY,CAAC,EAAE,EAAE,MAAM,GAAG,OAAO,CAGhD;AAED;;;;;;;;;;;;;;;;GAgBG;AACH,wBAAgB,SAAS,CAAC,QAAQ,EAAE,MAAM,EAAE,OAAO,EAAE,MAAM,GAAG,KAAK,CAAC,GAAG,CAAC,GAAG,SAAS,CA8BnF;AAED;;;;;;GAMG;AACH,wBAAgB,eAAe,CAAC,OAAO,EAAE,MAAM,EAAE,MAAM,EAAE,KAAK,CAAC,GAAG,CAAC,EAAE,GAAG,KAAK,CAAC,GAAG,CAAC,GAAG,SAAS,CA4B7F;AAED,wBAAgB,YAAY,IAAI,aAAa,EAAE,CAE9C;AAED,wBAAgB,SAAS,CAAC,SAAS,SAAS,aAAa,EACxD,QAAQ,EAAE,SAAS,GACjB,KAAK,CAAC,QAAQ,CAAC,SAAS,EAAE,MAAM,CAAC,OAAO,MAAM,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,EAAE,CAGhE;AAED,wBAAgB,aAAa,CAAC,IAAI,SAAS,GAAG,EAAE,KAAK,EAAE,KAAK,CAAC,IAAI,CAAC,EAAE,KAAK,EAAE,KAAK,GAAG,KAAK,CAAC,MAAM,CAAC,CAO/F;AAED;;;;;;GAMG;AACH,wBAAgB,aAAa,CAAC,IAAI,SAAS,GAAG,EAAE,KAAK,EAAE,KAAK,CAAC,IAAI,CAAC,GAAG,OAAO,CAmB3E;AAED;;;GAGG;AACH,wBAAgB,wBAAwB,CAAC,IAAI,SAAS,GAAG,EAAE,KAAK,EAAE,KAAK,CAAC,IAAI,CAAC,GAAG,OAAO,CAGtF;AAED;;;GAGG;AACH,wBAAgB,cAAc,CAAC,IAAI,SAAS,GAAG,EAC9C,CAAC,EAAE,KAAK,CAAC,IAAI,CAAC,GAAG,IAAI,GAAG,SAAS,EACjC,CAAC,EAAE,KAAK,CAAC,IAAI,CAAC,GAAG,IAAI,GAAG,SAAS,GAC/B,OAAO,CAGT","sourcesContent":["import { MODELS } from \"./models.generated.js\";\nimport type { Api, KnownProvider, Model, Usage } from \"./types.js\";\n\nconst modelRegistry: Map<string, Map<string, Model<Api>>> = new Map();\n\n// Initialize registry from MODELS on module load\nfor (const [provider, models] of Object.entries(MODELS)) {\n\tconst providerModels = new Map<string, Model<Api>>();\n\tfor (const [id, model] of Object.entries(models)) {\n\t\tproviderModels.set(id, model as Model<Api>);\n\t}\n\tmodelRegistry.set(provider, providerModels);\n}\n\ntype ModelApi<\n\tTProvider extends KnownProvider,\n\tTModelId extends keyof (typeof MODELS)[TProvider],\n> = (typeof MODELS)[TProvider][TModelId] extends { api: infer TApi } ? (TApi extends Api ? TApi : never) : never;\n\nexport function getModel<TProvider extends KnownProvider, TModelId extends keyof (typeof MODELS)[TProvider]>(\n\tprovider: TProvider,\n\tmodelId: TModelId,\n): Model<ModelApi<TProvider, TModelId>> {\n\tconst providerModels = modelRegistry.get(provider);\n\treturn providerModels?.get(modelId as string) as Model<ModelApi<TProvider, TModelId>>;\n}\n\n/**\n * Check if a model ID looks like an alias (no date suffix).\n * Aliases are preferred over dated versions when fuzzy matching.\n *\n * IDs ending with `-latest` are treated as aliases.\n * IDs ending with a date pattern (`-YYYYMMDD`) are treated as dated versions.\n */\nexport function isModelAlias(id: string): boolean {\n\tif (id.endsWith(\"-latest\")) return true;\n\treturn !/-\\d{8}$/.test(id);\n}\n\n/**\n * Find a model by fuzzy matching against the provider's registered models.\n *\n * Resolution order:\n * 1. Exact match by provider + model ID (via registry Map.get)\n * 2. Case-insensitive substring match against model ID and display name\n * 3. Among matches, prefer aliases (non-dated IDs) over dated versions\n * 4. Among ties, pick the lexicographically highest (latest) ID\n *\n * This is the same matching logic used by the CLI, subagent model resolution,\n * and interactive mode — centralised here so tests can exercise the real path.\n *\n * @example\n * findModel(\"anthropic\", \"sonnet\")  // → latest claude-sonnet alias\n * findModel(\"anthropic\", \"haiku\")   // → latest claude-haiku alias\n * findModel(\"openai\", \"gpt-5\")     // → latest gpt-5 alias\n */\nexport function findModel(provider: string, pattern: string): Model<Api> | undefined {\n\tconst providerModels = modelRegistry.get(provider);\n\tif (!providerModels) return undefined;\n\n\t// Try exact match first\n\tconst exact = providerModels.get(pattern);\n\tif (exact) return exact;\n\n\t// Substring match (case-insensitive)\n\tconst normalizedPattern = pattern.toLowerCase();\n\tconst matches = Array.from(providerModels.values()).filter(\n\t\t(m) => m.id.toLowerCase().includes(normalizedPattern) || m.name?.toLowerCase().includes(normalizedPattern),\n\t);\n\n\tif (matches.length === 0) return undefined;\n\tif (matches.length === 1) return matches[0];\n\n\t// Multiple matches — separate into aliases and dated versions\n\tconst aliases = matches.filter((m) => isModelAlias(m.id));\n\tconst datedVersions = matches.filter((m) => !isModelAlias(m.id));\n\n\tif (aliases.length > 0) {\n\t\t// Prefer alias — if multiple, pick the lexicographically highest\n\t\taliases.sort((a, b) => b.id.localeCompare(a.id));\n\t\treturn aliases[0];\n\t}\n\n\t// All dated — prefer the latest\n\tdatedVersions.sort((a, b) => b.id.localeCompare(a.id));\n\treturn datedVersions[0];\n}\n\n/**\n * Find a model by fuzzy matching against a flat array of models.\n * Same algorithm as findModel() but operates on an arbitrary model list\n * instead of the built-in registry.\n *\n * Used by model-resolver.ts and other code that manages its own model lists.\n */\nexport function findModelInList(pattern: string, models: Model<Api>[]): Model<Api> | undefined {\n\tif (models.length === 0) return undefined;\n\n\tconst normalizedPattern = pattern.toLowerCase();\n\n\t// Exact ID match (case-insensitive)\n\tconst exactById = models.find((m) => m.id.toLowerCase() === normalizedPattern);\n\tif (exactById) return exactById;\n\n\t// Substring match (case-insensitive)\n\tconst matches = models.filter(\n\t\t(m) => m.id.toLowerCase().includes(normalizedPattern) || m.name?.toLowerCase().includes(normalizedPattern),\n\t);\n\n\tif (matches.length === 0) return undefined;\n\tif (matches.length === 1) return matches[0];\n\n\t// Multiple matches — separate into aliases and dated versions\n\tconst aliases = matches.filter((m) => isModelAlias(m.id));\n\tconst datedVersions = matches.filter((m) => !isModelAlias(m.id));\n\n\tif (aliases.length > 0) {\n\t\taliases.sort((a, b) => b.id.localeCompare(a.id));\n\t\treturn aliases[0];\n\t}\n\n\tdatedVersions.sort((a, b) => b.id.localeCompare(a.id));\n\treturn datedVersions[0];\n}\n\nexport function getProviders(): KnownProvider[] {\n\treturn Array.from(modelRegistry.keys()) as KnownProvider[];\n}\n\nexport function getModels<TProvider extends KnownProvider>(\n\tprovider: TProvider,\n): Model<ModelApi<TProvider, keyof (typeof MODELS)[TProvider]>>[] {\n\tconst models = modelRegistry.get(provider);\n\treturn models ? (Array.from(models.values()) as Model<ModelApi<TProvider, keyof (typeof MODELS)[TProvider]>>[]) : [];\n}\n\nexport function calculateCost<TApi extends Api>(model: Model<TApi>, usage: Usage): Usage[\"cost\"] {\n\tusage.cost.input = (model.cost.input / 1000000) * usage.input;\n\tusage.cost.output = (model.cost.output / 1000000) * usage.output;\n\tusage.cost.cacheRead = (model.cost.cacheRead / 1000000) * usage.cacheRead;\n\tusage.cost.cacheWrite = (model.cost.cacheWrite / 1000000) * usage.cacheWrite;\n\tusage.cost.total = usage.cost.input + usage.cost.output + usage.cost.cacheRead + usage.cost.cacheWrite;\n\treturn usage.cost;\n}\n\n/**\n * Check if a model supports xhigh thinking level.\n *\n * Supported today:\n * - GPT-5.2 through GPT-5.6 model families\n * - Opus 4.6+ models (xhigh maps to adaptive effort \"max\" on Anthropic-compatible providers)\n */\nexport function supportsXhigh<TApi extends Api>(model: Model<TApi>): boolean {\n\tif (\n\t\tmodel.id.includes(\"gpt-5.2\") ||\n\t\tmodel.id.includes(\"gpt-5.3\") ||\n\t\tmodel.id.includes(\"gpt-5.4\") ||\n\t\tmodel.id.includes(\"gpt-5.5\") ||\n\t\tmodel.id.includes(\"gpt-5.6\")\n\t) {\n\t\treturn true;\n\t}\n\n\t// Opus 4.6+ supports xhigh (adaptive effort \"max\").\n\t// Match any opus-4-N or opus-4.N where N >= 6 (1-2 digit minor version, not date suffixes).\n\tconst opusMatch = model.id.match(/opus-4[.-](\\d{1,2})(?!\\d)/);\n\tif (opusMatch && Number.parseInt(opusMatch[1], 10) >= 6) {\n\t\treturn true;\n\t}\n\n\treturn false;\n}\n\n/**\n * Check if a model uses adaptive thinking (Opus 4.6+, Sonnet 4.6+), where the\n * `thinkingDisplay` option is honored. Mirrors the per-provider internal checks.\n */\nexport function supportsAdaptiveThinking<TApi extends Api>(model: Model<TApi>): boolean {\n\tconst m = model.id.match(/(opus|sonnet)-4[.-](\\d{1,2})(?!\\d)/);\n\treturn m != null && Number.parseInt(m[2], 10) >= 6;\n}\n\n/**\n * Check if two models are equal by comparing both their id and provider.\n * Returns false if either model is null or undefined.\n */\nexport function modelsAreEqual<TApi extends Api>(\n\ta: Model<TApi> | null | undefined,\n\tb: Model<TApi> | null | undefined,\n): boolean {\n\tif (!a || !b) return false;\n\treturn a.id === b.id && a.provider === b.provider;\n}\n"]}
+{"version":3,"file":"models.d.ts","sourceRoot":"","sources":["../src/models.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,MAAM,EAAE,MAAM,uBAAuB,CAAC;AAC/C,OAAO,KAAK,EAAE,GAAG,EAAE,aAAa,EAAE,KAAK,EAAE,KAAK,EAAE,MAAM,YAAY,CAAC;AAanE,KAAK,QAAQ,CACZ,SAAS,SAAS,aAAa,EAC/B,QAAQ,SAAS,MAAM,CAAC,OAAO,MAAM,CAAC,CAAC,SAAS,CAAC,IAC9C,CAAC,OAAO,MAAM,CAAC,CAAC,SAAS,CAAC,CAAC,QAAQ,CAAC,SAAS;IAAE,GAAG,EAAE,MAAM,IAAI,CAAA;CAAE,GAAG,CAAC,IAAI,SAAS,GAAG,GAAG,IAAI,GAAG,KAAK,CAAC,GAAG,KAAK,CAAC;AAEjH,wBAAgB,QAAQ,CAAC,SAAS,SAAS,aAAa,EAAE,QAAQ,SAAS,MAAM,CAAC,OAAO,MAAM,CAAC,CAAC,SAAS,CAAC,EAC1G,QAAQ,EAAE,SAAS,EACnB,OAAO,EAAE,QAAQ,GACf,KAAK,CAAC,QAAQ,CAAC,SAAS,EAAE,QAAQ,CAAC,CAAC,CAGtC;AAED;;;;;;GAMG;AACH,wBAAgB,YAAY,CAAC,EAAE,EAAE,MAAM,GAAG,OAAO,CAGhD;AAED;;;;;;;;;;;;;;;;GAgBG;AACH,wBAAgB,SAAS,CAAC,QAAQ,EAAE,MAAM,EAAE,OAAO,EAAE,MAAM,GAAG,KAAK,CAAC,GAAG,CAAC,GAAG,SAAS,CA8BnF;AAED;;;;;;GAMG;AACH,wBAAgB,eAAe,CAAC,OAAO,EAAE,MAAM,EAAE,MAAM,EAAE,KAAK,CAAC,GAAG,CAAC,EAAE,GAAG,KAAK,CAAC,GAAG,CAAC,GAAG,SAAS,CA4B7F;AAED,wBAAgB,YAAY,IAAI,aAAa,EAAE,CAE9C;AAED,wBAAgB,SAAS,CAAC,SAAS,SAAS,aAAa,EACxD,QAAQ,EAAE,SAAS,GACjB,KAAK,CAAC,QAAQ,CAAC,SAAS,EAAE,MAAM,CAAC,OAAO,MAAM,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,EAAE,CAGhE;AAED,wBAAgB,aAAa,CAAC,IAAI,SAAS,GAAG,EAAE,KAAK,EAAE,KAAK,CAAC,IAAI,CAAC,EAAE,KAAK,EAAE,KAAK,GAAG,KAAK,CAAC,MAAM,CAAC,CAO/F;AAED;;;;;;;GAOG;AACH,wBAAgB,aAAa,CAAC,IAAI,SAAS,GAAG,EAAE,KAAK,EAAE,KAAK,CAAC,IAAI,CAAC,GAAG,OAAO,CAoB3E;AAED;;;GAGG;AACH,wBAAgB,wBAAwB,CAAC,IAAI,SAAS,GAAG,EAAE,KAAK,EAAE,KAAK,CAAC,IAAI,CAAC,GAAG,OAAO,CAGtF;AAED;;;GAGG;AACH,wBAAgB,cAAc,CAAC,IAAI,SAAS,GAAG,EAC9C,CAAC,EAAE,KAAK,CAAC,IAAI,CAAC,GAAG,IAAI,GAAG,SAAS,EACjC,CAAC,EAAE,KAAK,CAAC,IAAI,CAAC,GAAG,IAAI,GAAG,SAAS,GAC/B,OAAO,CAGT","sourcesContent":["import { MODELS } from \"./models.generated.js\";\nimport type { Api, KnownProvider, Model, Usage } from \"./types.js\";\n\nconst modelRegistry: Map<string, Map<string, Model<Api>>> = new Map();\n\n// Initialize registry from MODELS on module load\nfor (const [provider, models] of Object.entries(MODELS)) {\n\tconst providerModels = new Map<string, Model<Api>>();\n\tfor (const [id, model] of Object.entries(models)) {\n\t\tproviderModels.set(id, model as Model<Api>);\n\t}\n\tmodelRegistry.set(provider, providerModels);\n}\n\ntype ModelApi<\n\tTProvider extends KnownProvider,\n\tTModelId extends keyof (typeof MODELS)[TProvider],\n> = (typeof MODELS)[TProvider][TModelId] extends { api: infer TApi } ? (TApi extends Api ? TApi : never) : never;\n\nexport function getModel<TProvider extends KnownProvider, TModelId extends keyof (typeof MODELS)[TProvider]>(\n\tprovider: TProvider,\n\tmodelId: TModelId,\n): Model<ModelApi<TProvider, TModelId>> {\n\tconst providerModels = modelRegistry.get(provider);\n\treturn providerModels?.get(modelId as string) as Model<ModelApi<TProvider, TModelId>>;\n}\n\n/**\n * Check if a model ID looks like an alias (no date suffix).\n * Aliases are preferred over dated versions when fuzzy matching.\n *\n * IDs ending with `-latest` are treated as aliases.\n * IDs ending with a date pattern (`-YYYYMMDD`) are treated as dated versions.\n */\nexport function isModelAlias(id: string): boolean {\n\tif (id.endsWith(\"-latest\")) return true;\n\treturn !/-\\d{8}$/.test(id);\n}\n\n/**\n * Find a model by fuzzy matching against the provider's registered models.\n *\n * Resolution order:\n * 1. Exact match by provider + model ID (via registry Map.get)\n * 2. Case-insensitive substring match against model ID and display name\n * 3. Among matches, prefer aliases (non-dated IDs) over dated versions\n * 4. Among ties, pick the lexicographically highest (latest) ID\n *\n * This is the same matching logic used by the CLI, subagent model resolution,\n * and interactive mode — centralised here so tests can exercise the real path.\n *\n * @example\n * findModel(\"anthropic\", \"sonnet\")  // → latest claude-sonnet alias\n * findModel(\"anthropic\", \"haiku\")   // → latest claude-haiku alias\n * findModel(\"openai\", \"gpt-5\")     // → latest gpt-5 alias\n */\nexport function findModel(provider: string, pattern: string): Model<Api> | undefined {\n\tconst providerModels = modelRegistry.get(provider);\n\tif (!providerModels) return undefined;\n\n\t// Try exact match first\n\tconst exact = providerModels.get(pattern);\n\tif (exact) return exact;\n\n\t// Substring match (case-insensitive)\n\tconst normalizedPattern = pattern.toLowerCase();\n\tconst matches = Array.from(providerModels.values()).filter(\n\t\t(m) => m.id.toLowerCase().includes(normalizedPattern) || m.name?.toLowerCase().includes(normalizedPattern),\n\t);\n\n\tif (matches.length === 0) return undefined;\n\tif (matches.length === 1) return matches[0];\n\n\t// Multiple matches — separate into aliases and dated versions\n\tconst aliases = matches.filter((m) => isModelAlias(m.id));\n\tconst datedVersions = matches.filter((m) => !isModelAlias(m.id));\n\n\tif (aliases.length > 0) {\n\t\t// Prefer alias — if multiple, pick the lexicographically highest\n\t\taliases.sort((a, b) => b.id.localeCompare(a.id));\n\t\treturn aliases[0];\n\t}\n\n\t// All dated — prefer the latest\n\tdatedVersions.sort((a, b) => b.id.localeCompare(a.id));\n\treturn datedVersions[0];\n}\n\n/**\n * Find a model by fuzzy matching against a flat array of models.\n * Same algorithm as findModel() but operates on an arbitrary model list\n * instead of the built-in registry.\n *\n * Used by model-resolver.ts and other code that manages its own model lists.\n */\nexport function findModelInList(pattern: string, models: Model<Api>[]): Model<Api> | undefined {\n\tif (models.length === 0) return undefined;\n\n\tconst normalizedPattern = pattern.toLowerCase();\n\n\t// Exact ID match (case-insensitive)\n\tconst exactById = models.find((m) => m.id.toLowerCase() === normalizedPattern);\n\tif (exactById) return exactById;\n\n\t// Substring match (case-insensitive)\n\tconst matches = models.filter(\n\t\t(m) => m.id.toLowerCase().includes(normalizedPattern) || m.name?.toLowerCase().includes(normalizedPattern),\n\t);\n\n\tif (matches.length === 0) return undefined;\n\tif (matches.length === 1) return matches[0];\n\n\t// Multiple matches — separate into aliases and dated versions\n\tconst aliases = matches.filter((m) => isModelAlias(m.id));\n\tconst datedVersions = matches.filter((m) => !isModelAlias(m.id));\n\n\tif (aliases.length > 0) {\n\t\taliases.sort((a, b) => b.id.localeCompare(a.id));\n\t\treturn aliases[0];\n\t}\n\n\tdatedVersions.sort((a, b) => b.id.localeCompare(a.id));\n\treturn datedVersions[0];\n}\n\nexport function getProviders(): KnownProvider[] {\n\treturn Array.from(modelRegistry.keys()) as KnownProvider[];\n}\n\nexport function getModels<TProvider extends KnownProvider>(\n\tprovider: TProvider,\n): Model<ModelApi<TProvider, keyof (typeof MODELS)[TProvider]>>[] {\n\tconst models = modelRegistry.get(provider);\n\treturn models ? (Array.from(models.values()) as Model<ModelApi<TProvider, keyof (typeof MODELS)[TProvider]>>[]) : [];\n}\n\nexport function calculateCost<TApi extends Api>(model: Model<TApi>, usage: Usage): Usage[\"cost\"] {\n\tusage.cost.input = (model.cost.input / 1000000) * usage.input;\n\tusage.cost.output = (model.cost.output / 1000000) * usage.output;\n\tusage.cost.cacheRead = (model.cost.cacheRead / 1000000) * usage.cacheRead;\n\tusage.cost.cacheWrite = (model.cost.cacheWrite / 1000000) * usage.cacheWrite;\n\tusage.cost.total = usage.cost.input + usage.cost.output + usage.cost.cacheRead + usage.cost.cacheWrite;\n\treturn usage.cost;\n}\n\n/**\n * Check if a model supports xhigh thinking level.\n *\n * Supported today:\n * - GPT-5.2 through GPT-5.6 model families\n * - Opus 4.6+ models (xhigh maps to adaptive effort \"max\" on Anthropic-compatible providers)\n * - Kimi Code K3 (xhigh maps to its advertised \"max\" effort)\n */\nexport function supportsXhigh<TApi extends Api>(model: Model<TApi>): boolean {\n\tif (model.provider === \"kimi-coding-oauth\" && model.id === \"k3\") return true;\n\tif (\n\t\tmodel.id.includes(\"gpt-5.2\") ||\n\t\tmodel.id.includes(\"gpt-5.3\") ||\n\t\tmodel.id.includes(\"gpt-5.4\") ||\n\t\tmodel.id.includes(\"gpt-5.5\") ||\n\t\tmodel.id.includes(\"gpt-5.6\")\n\t) {\n\t\treturn true;\n\t}\n\n\t// Opus 4.6+ supports xhigh (adaptive effort \"max\").\n\t// Match any opus-4-N or opus-4.N where N >= 6 (1-2 digit minor version, not date suffixes).\n\tconst opusMatch = model.id.match(/opus-4[.-](\\d{1,2})(?!\\d)/);\n\tif (opusMatch && Number.parseInt(opusMatch[1], 10) >= 6) {\n\t\treturn true;\n\t}\n\n\treturn false;\n}\n\n/**\n * Check if a model uses adaptive thinking (Opus 4.6+, Sonnet 4.6+), where the\n * `thinkingDisplay` option is honored. Mirrors the per-provider internal checks.\n */\nexport function supportsAdaptiveThinking<TApi extends Api>(model: Model<TApi>): boolean {\n\tconst m = model.id.match(/(opus|sonnet)-4[.-](\\d{1,2})(?!\\d)/);\n\treturn m != null && Number.parseInt(m[2], 10) >= 6;\n}\n\n/**\n * Check if two models are equal by comparing both their id and provider.\n * Returns false if either model is null or undefined.\n */\nexport function modelsAreEqual<TApi extends Api>(\n\ta: Model<TApi> | null | undefined,\n\tb: Model<TApi> | null | undefined,\n): boolean {\n\tif (!a || !b) return false;\n\treturn a.id === b.id && a.provider === b.provider;\n}\n"]}

package/dist/models.generated.d.ts

@@ -5552,6 +5552,34 @@ export declare const MODELS: {
         };
     };
     readonly "kimi-coding-oauth": {
+        readonly k3: {
+            id: string;
+            name: string;
+            api: "openai-completions";
+            provider: string;
+            baseUrl: string;
+            compat: {
+                thinkingFormat: "kimi";
+                supportsDeveloperRole: false;
+                reasoningEffortMap: {
+                    minimal: string;
+                    low: string;
+                    medium: string;
+                    high: string;
+                    xhigh: string;
+                };
+            };
+            reasoning: true;
+            input: ("image" | "text")[];
+            cost: {
+                input: number;
+                output: number;
+                cacheRead: number;
+                cacheWrite: number;
+            };
+            contextWindow: number;
+            maxTokens: number;
+        };
         readonly "kimi-for-coding": {
             id: string;
             name: string;
@@ -5561,6 +5589,41 @@ export declare const MODELS: {
             compat: {
                 thinkingFormat: "kimi";
                 supportsDeveloperRole: false;
+                reasoningEffortMap: {
+                    minimal: string;
+                    low: string;
+                    medium: string;
+                    high: string;
+                    xhigh: string;
+                };
+            };
+            reasoning: true;
+            input: ("image" | "text")[];
+            cost: {
+                input: number;
+                output: number;
+                cacheRead: number;
+                cacheWrite: number;
+            };
+            contextWindow: number;
+            maxTokens: number;
+        };
+        readonly "kimi-for-coding-highspeed": {
+            id: string;
+            name: string;
+            api: "openai-completions";
+            provider: string;
+            baseUrl: string;
+            compat: {
+                thinkingFormat: "kimi";
+                supportsDeveloperRole: false;
+                reasoningEffortMap: {
+                    minimal: string;
+                    low: string;
+                    medium: string;
+                    high: string;
+                    xhigh: string;
+                };
             };
             reasoning: true;
             input: ("image" | "text")[];
@@ -7815,7 +7878,7 @@ export declare const MODELS: {
         readonly "grok-4.5": {
             id: string;
             name: string;
-            api: "openai-completions";
+            api: "openai-responses";
             provider: string;
             baseUrl: string;
             reasoning: true;
@@ -8123,7 +8186,7 @@ export declare const MODELS: {
         readonly "grok-4.5": {
             id: string;
             name: string;
-            api: "openai-completions";
+            api: "openai-responses";
             provider: string;
             baseUrl: string;
             reasoning: true;
@@ -9550,23 +9613,6 @@ export declare const MODELS: {
             contextWindow: number;
             maxTokens: number;
         };
-        readonly "meta-llama/llama-3.3-70b-instruct:free": {
-            id: string;
-            name: string;
-            api: "openai-completions";
-            provider: string;
-            baseUrl: string;
-            reasoning: false;
-            input: "text"[];
-            cost: {
-                input: number;
-                output: number;
-                cacheRead: number;
-                cacheWrite: number;
-            };
-            contextWindow: number;
-            maxTokens: number;
-        };
         readonly "meta-llama/llama-4-maverick": {
             id: string;
             name: string;
@@ -11760,23 +11806,6 @@ export declare const MODELS: {
             contextWindow: number;
             maxTokens: number;
         };
-        readonly "qwen/qwen3-coder:free": {
-            id: string;
-            name: string;
-            api: "openai-completions";
-            provider: string;
-            baseUrl: string;
-            reasoning: false;
-            input: "text"[];
-            cost: {
-                input: number;
-                output: number;
-                cacheRead: number;
-                cacheWrite: number;
-            };
-            contextWindow: number;
-            maxTokens: number;
-        };
         readonly "qwen/qwen3-max": {
             id: string;
             name: string;
@@ -11828,23 +11857,6 @@ export declare const MODELS: {
             contextWindow: number;
             maxTokens: number;
         };
-        readonly "qwen/qwen3-next-80b-a3b-instruct:free": {
-            id: string;
-            name: string;
-            api: "openai-completions";
-            provider: string;
-            baseUrl: string;
-            reasoning: false;
-            input: "text"[];
-            cost: {
-                input: number;
-                output: number;
-                cacheRead: number;
-                cacheWrite: number;
-            };
-            contextWindow: number;
-            maxTokens: number;
-        };
         readonly "qwen/qwen3-next-80b-a3b-thinking": {
             id: string;
             name: string;
@@ -12406,6 +12418,23 @@ export declare const MODELS: {
             contextWindow: number;
             maxTokens: number;
         };
+        readonly "thinkingmachines/inkling": {
+            id: string;
+            name: string;
+            api: "openai-completions";
+            provider: string;
+            baseUrl: string;
+            reasoning: true;
+            input: ("image" | "text")[];
+            cost: {
+                input: number;
+                output: number;
+                cacheRead: number;
+                cacheWrite: number;
+            };
+            contextWindow: number;
+            maxTokens: number;
+        };
         readonly "upstage/solar-pro-3": {
             id: string;
             name: string;