@opengsd/gsd-pi 1.0.2-dev.5961fbf → 1.0.2-dev.5f7864c

package/packages/pi-ai/dist/types.js.map

@@ -1 +1 @@
-{"version":3,"file":"types.js","sourceRoot":"","sources":["../src/types.ts"],"names":[],"mappings":"","sourcesContent":["import type { AssistantMessageDiagnostic } from \"./utils/diagnostics.js\";\nimport type { AssistantMessageEventStream } from \"./utils/event-stream.js\";\n\nexport type { AssistantMessageEventStream } from \"./utils/event-stream.js\";\n\nexport type KnownApi =\n\t| \"openai-completions\"\n\t| \"mistral-conversations\"\n\t| \"openai-responses\"\n\t| \"azure-openai-responses\"\n\t| \"openai-codex-responses\"\n\t| \"anthropic-messages\"\n\t| \"bedrock-converse-stream\"\n\t| \"google-generative-ai\"\n\t| \"google-vertex\";\n\nexport type Api = KnownApi | (string & {});\n\nexport type KnownImagesApi = \"openrouter-images\";\n\nexport type ImagesApi = KnownImagesApi | (string & {});\n\nexport type KnownProvider =\n\t| \"amazon-bedrock\"\n\t| \"anthropic\"\n\t| \"google\"\n\t| \"google-vertex\"\n\t| \"openai\"\n\t| \"azure-openai-responses\"\n\t| \"openai-codex\"\n\t| \"deepseek\"\n\t| \"github-copilot\"\n\t| \"xai\"\n\t| \"groq\"\n\t| \"cerebras\"\n\t| \"openrouter\"\n\t| \"vercel-ai-gateway\"\n\t| \"zai\"\n\t| \"mistral\"\n\t| \"minimax\"\n\t| \"minimax-cn\"\n\t| \"moonshotai\"\n\t| \"moonshotai-cn\"\n\t| \"huggingface\"\n\t| \"fireworks\"\n\t| \"together\"\n\t| \"opencode\"\n\t| \"opencode-go\"\n\t| \"kimi-coding\"\n\t| \"cloudflare-workers-ai\"\n\t| \"cloudflare-ai-gateway\"\n\t| \"xiaomi\"\n\t| \"xiaomi-token-plan-cn\"\n\t| \"xiaomi-token-plan-ams\"\n\t| \"xiaomi-token-plan-sgp\";\nexport type Provider = KnownProvider | string;\n\nexport type KnownImagesProvider = \"openrouter\";\n\nexport type ImagesProvider = KnownImagesProvider | string;\n\nexport type ThinkingLevel = \"minimal\" | \"low\" | \"medium\" | \"high\" | \"xhigh\";\nexport type ModelThinkingLevel = \"off\" | ThinkingLevel;\nexport type ThinkingLevelMap = Partial<Record<ModelThinkingLevel, string | null>>;\n\n/** Token budgets for each thinking level (token-based providers only) */\nexport interface ThinkingBudgets {\n\tminimal?: number;\n\tlow?: number;\n\tmedium?: number;\n\thigh?: number;\n}\n\n// Base options all providers share\nexport type CacheRetention = \"none\" | \"short\" | \"long\";\n\nexport type Transport = \"sse\" | \"websocket\" | \"websocket-cached\" | \"auto\";\n\nexport interface ProviderResponse {\n\tstatus: number;\n\theaders: Record<string, string>;\n}\n\nexport interface StreamOptions {\n\ttemperature?: number;\n\tmaxTokens?: number;\n\tsignal?: AbortSignal;\n\tapiKey?: string;\n\t/**\n\t * Preferred transport for providers that support multiple transports.\n\t * Providers that do not support this option ignore it.\n\t */\n\ttransport?: Transport;\n\t/**\n\t * Prompt cache retention preference. Providers map this to their supported values.\n\t * Default: \"short\".\n\t */\n\tcacheRetention?: CacheRetention;\n\t/**\n\t * Optional session identifier for providers that support session-based caching.\n\t * Providers can use this to enable prompt caching, request routing, or other\n\t * session-aware features. Ignored by providers that don't support it.\n\t */\n\tsessionId?: string;\n\t/**\n\t * Optional callback for inspecting or replacing provider payloads before sending.\n\t * Return undefined to keep the payload unchanged.\n\t */\n\tonPayload?: (payload: unknown, model: Model<Api>) => unknown | undefined | Promise<unknown | undefined>;\n\t/**\n\t * Optional callback invoked after an HTTP response is received and before\n\t * its body stream is consumed.\n\t */\n\tonResponse?: (response: ProviderResponse, model: Model<Api>) => void | Promise<void>;\n\t/**\n\t * Optional custom HTTP headers to include in API requests.\n\t * Merged with provider defaults; can override default headers.\n\t * Not supported by all providers (e.g., AWS Bedrock uses SDK auth).\n\t */\n\theaders?: Record<string, string>;\n\t/**\n\t * HTTP request timeout in milliseconds for providers/SDKs that support it.\n\t * For example, OpenAI and Anthropic SDK clients default to 10 minutes.\n\t */\n\ttimeoutMs?: number;\n\t/**\n\t * Maximum retry attempts for providers/SDKs that support client-side retries.\n\t * For example, OpenAI and Anthropic SDK clients default to 2.\n\t */\n\tmaxRetries?: number;\n\t/**\n\t * Maximum delay in milliseconds to wait for a retry when the server requests a long wait.\n\t * If the server's requested delay exceeds this value, the request fails immediately\n\t * with an error containing the requested delay, allowing higher-level retry logic\n\t * to handle it with user visibility.\n\t * Default: 60000 (60 seconds). Set to 0 to disable the cap.\n\t */\n\tmaxRetryDelayMs?: number;\n\t/**\n\t * Optional metadata to include in API requests.\n\t * Providers extract the fields they understand and ignore the rest.\n\t * For example, Anthropic uses `user_id` for abuse tracking and rate limiting.\n\t */\n\tmetadata?: Record<string, unknown>;\n}\n\nexport type ProviderStreamOptions = StreamOptions & Record<string, unknown>;\n\nexport interface ImagesOptions {\n\tsignal?: AbortSignal;\n\tapiKey?: string;\n\t/**\n\t * Optional callback for inspecting or replacing provider payloads before sending.\n\t * Return undefined to keep the payload unchanged.\n\t */\n\tonPayload?: (payload: unknown, model: ImagesModel<ImagesApi>) => unknown | undefined | Promise<unknown | undefined>;\n\t/**\n\t * Optional callback invoked after an HTTP response is received.\n\t */\n\tonResponse?: (response: ProviderResponse, model: ImagesModel<ImagesApi>) => void | Promise<void>;\n\t/**\n\t * Optional custom HTTP headers to include in API requests.\n\t * Merged with provider defaults; can override default headers.\n\t */\n\theaders?: Record<string, string>;\n\t/**\n\t * HTTP request timeout in milliseconds for providers/SDKs that support it.\n\t */\n\ttimeoutMs?: number;\n\t/**\n\t * Maximum retry attempts for providers/SDKs that support client-side retries.\n\t */\n\tmaxRetries?: number;\n\t/**\n\t * Maximum delay in milliseconds to wait for a retry when the server requests a long wait.\n\t * If the server's requested delay exceeds this value, the request fails immediately\n\t * with an error containing the requested delay, allowing higher-level retry logic\n\t * to handle it with user visibility.\n\t * Default: 60000 (60 seconds). Set to 0 to disable the cap.\n\t */\n\tmaxRetryDelayMs?: number;\n\t/**\n\t * Optional metadata to include in API requests.\n\t * Providers extract the fields they understand and ignore the rest.\n\t */\n\tmetadata?: Record<string, unknown>;\n}\n\nexport type ProviderImagesOptions = ImagesOptions & Record<string, unknown>;\n\n// Unified options with reasoning passed to streamSimple() and completeSimple()\nexport interface SimpleStreamOptions extends StreamOptions {\n\t/** Workspace root for providers that spawn local processes. */\n\tcwd?: string;\n\treasoning?: ThinkingLevel;\n\t/** Custom token budgets for thinking levels (token-based providers only) */\n\tthinkingBudgets?: ThinkingBudgets;\n}\n\n// Generic StreamFunction with typed options.\n//\n// Contract:\n// - Must return an AssistantMessageEventStream.\n// - Once invoked, request/model/runtime failures should be encoded in the\n//   returned stream, not thrown.\n// - Error termination must produce an AssistantMessage with stopReason\n//   \"error\" or \"aborted\" and errorMessage, emitted via the stream protocol.\nexport type StreamFunction<TApi extends Api = Api, TOptions extends StreamOptions = StreamOptions> = (\n\tmodel: Model<TApi>,\n\tcontext: Context,\n\toptions?: TOptions,\n) => AssistantMessageEventStream;\n\nexport type ImagesFunction<TApi extends ImagesApi = ImagesApi, TOptions extends ImagesOptions = ImagesOptions> = (\n\tmodel: ImagesModel<TApi>,\n\tcontext: ImagesContext,\n\toptions?: TOptions,\n) => Promise<AssistantImages>;\n\nexport interface TextSignatureV1 {\n\tv: 1;\n\tid: string;\n\tphase?: \"commentary\" | \"final_answer\";\n}\n\nexport interface TextContent {\n\ttype: \"text\";\n\ttext: string;\n\ttextSignature?: string; // e.g., for OpenAI responses, message metadata (legacy id string or TextSignatureV1 JSON)\n}\n\nexport interface ThinkingContent {\n\ttype: \"thinking\";\n\tthinking: string;\n\tthinkingSignature?: string; // e.g., for OpenAI responses, the reasoning item ID\n\t/** When true, the thinking content was redacted by safety filters. The opaque\n\t *  encrypted payload is stored in `thinkingSignature` so it can be passed back\n\t *  to the API for multi-turn continuity. */\n\tredacted?: boolean;\n}\n\nexport interface ImageContent {\n\ttype: \"image\";\n\tdata: string; // base64 encoded image data\n\tmimeType: string; // e.g., \"image/jpeg\", \"image/png\"\n}\n\nexport interface ToolCall {\n\ttype: \"toolCall\";\n\tid: string;\n\tname: string;\n\targuments: Record<string, any>;\n\tthoughtSignature?: string; // Google-specific: opaque signature for reusing thought context\n\t/** Result from an externally executed tool (e.g. Claude Code SDK). Skips local dispatch. */\n\texternalResult?: {\n\t\tcontent?: Array<{ type: string; text?: string; data?: string; mimeType?: string }>;\n\t\tdetails?: Record<string, unknown>;\n\t\tisError?: boolean;\n\t};\n}\n\n/** GSD compat: server-side tool invocation block (e.g. web search). */\nexport interface ServerToolUse {\n\ttype: \"serverToolUse\";\n\tid: string;\n\tname: string;\n\tinput: unknown;\n\texternalResult?: {\n\t\tcontent?: Array<{ type: string; text?: string; data?: string; mimeType?: string }>;\n\t\tdetails?: Record<string, unknown>;\n\t\tisError?: boolean;\n\t};\n}\n\n/** GSD compat: server-side tool result block paired with serverToolUse. */\nexport interface WebSearchResult {\n\ttype: \"webSearchResult\";\n\ttoolUseId: string;\n\tcontent: unknown;\n}\n\n/** GSD compat alias for {@link ServerToolUse}. */\nexport type ServerToolUseContent = ServerToolUse;\n\n/** GSD compat alias for {@link WebSearchResult}. */\nexport type WebSearchResultContent = WebSearchResult;\n\nexport interface Usage {\n\tinput: number;\n\toutput: number;\n\tcacheRead: number;\n\tcacheWrite: number;\n\ttotalTokens: number;\n\tcost: {\n\t\tinput: number;\n\t\toutput: number;\n\t\tcacheRead: number;\n\t\tcacheWrite: number;\n\t\ttotal: number;\n\t};\n}\n\nexport type StopReason = \"stop\" | \"length\" | \"toolUse\" | \"error\" | \"aborted\";\n\nexport interface UserMessage {\n\trole: \"user\";\n\tcontent: string | (TextContent | ImageContent)[];\n\ttimestamp: number; // Unix timestamp in milliseconds\n}\n\nexport interface AssistantMessage {\n\trole: \"assistant\";\n\tcontent: (TextContent | ThinkingContent | ToolCall | ServerToolUse | WebSearchResult)[];\n\tapi: Api;\n\tprovider: Provider;\n\tmodel: string;\n\tresponseModel?: string; // Concrete `chunk.model` when different from the requested `model` (e.g. OpenRouter `auto` -> `anthropic/...`)\n\tresponseId?: string; // Provider-specific response/message identifier when the upstream API exposes one\n\tdiagnostics?: AssistantMessageDiagnostic[]; // Redacted provider/runtime diagnostics for failures and recoveries.\n\tusage: Usage;\n\tstopReason: StopReason;\n\terrorMessage?: string;\n\t/** Server-requested retry delay in milliseconds (from Retry-After or rate limit headers). */\n\tretryAfterMs?: number;\n\t/** Provider inference performance metrics (e.g. tokens/sec from local models). */\n\tinferenceMetrics?: InferenceMetrics;\n\ttimestamp: number; // Unix timestamp in milliseconds\n}\n\n/** Inference performance metrics reported by providers that support it (e.g. Ollama). */\nexport interface InferenceMetrics {\n\t/** Tokens generated per second during eval phase. */\n\ttokensPerSecond: number;\n\t/** Wall-clock duration of the full request in milliseconds. */\n\ttotalDurationMs: number;\n\t/** Duration of the eval (generation) phase in milliseconds. */\n\tevalDurationMs: number;\n\t/** Duration of the prompt eval phase in milliseconds. */\n\tpromptEvalDurationMs: number;\n}\n\nexport interface ToolResultMessage<TDetails = any> {\n\trole: \"toolResult\";\n\ttoolCallId: string;\n\ttoolName: string;\n\tcontent: (TextContent | ImageContent)[]; // Supports text and images\n\tdetails?: TDetails;\n\tisError: boolean;\n\ttimestamp: number; // Unix timestamp in milliseconds\n}\n\nexport type Message = UserMessage | AssistantMessage | ToolResultMessage;\n\nexport type ImagesInputContent = TextContent | ImageContent;\nexport type ImagesOutputContent = TextContent | ImageContent;\n\nexport interface ImagesContext {\n\tinput: ImagesInputContent[];\n}\n\nexport type ImagesStopReason = \"stop\" | \"error\" | \"aborted\";\n\nexport interface AssistantImages {\n\tapi: ImagesApi;\n\tprovider: ImagesProvider;\n\tmodel: string;\n\toutput: ImagesOutputContent[];\n\tresponseId?: string;\n\tusage?: Usage;\n\tstopReason: ImagesStopReason;\n\terrorMessage?: string;\n\ttimestamp: number; // Unix timestamp in milliseconds\n}\n\nimport type { TSchema } from \"typebox\";\n\nexport interface Tool<TParameters extends TSchema = TSchema> {\n\tname: string;\n\tdescription: string;\n\tparameters: TParameters;\n}\n\nexport interface Context {\n\tsystemPrompt?: string;\n\tmessages: Message[];\n\ttools?: Tool[];\n}\n\n/**\n * Event protocol for AssistantMessageEventStream.\n *\n * Streams should emit `start` before partial updates, then terminate with either:\n * - `done` carrying the final successful AssistantMessage, or\n * - `error` carrying the final AssistantMessage with stopReason \"error\" or \"aborted\"\n *   and errorMessage.\n */\nexport type AssistantMessageEvent =\n\t| { type: \"start\"; partial: AssistantMessage }\n\t| { type: \"text_start\"; contentIndex: number; partial: AssistantMessage }\n\t| { type: \"text_delta\"; contentIndex: number; delta: string; partial: AssistantMessage }\n\t| { type: \"text_end\"; contentIndex: number; content: string; partial: AssistantMessage }\n\t| { type: \"thinking_start\"; contentIndex: number; partial: AssistantMessage }\n\t| { type: \"thinking_delta\"; contentIndex: number; delta: string; partial: AssistantMessage }\n\t| { type: \"thinking_end\"; contentIndex: number; content: string; partial: AssistantMessage }\n\t| { type: \"toolcall_start\"; contentIndex: number; partial: AssistantMessage }\n\t| { type: \"toolcall_delta\"; contentIndex: number; delta: string; partial: AssistantMessage }\n\t| {\n\t\t\ttype: \"toolcall_end\";\n\t\t\tcontentIndex: number;\n\t\t\ttoolCall: ToolCall;\n\t\t\tpartial: AssistantMessage;\n\t\t\tmalformedArguments?: boolean;\n\t  }\n\t| { type: \"server_tool_use\"; contentIndex: number; partial: AssistantMessage }\n\t| { type: \"done\"; reason: Extract<StopReason, \"stop\" | \"length\" | \"toolUse\">; message: AssistantMessage }\n\t| { type: \"error\"; reason: Extract<StopReason, \"aborted\" | \"error\">; error: AssistantMessage };\n\n/**\n * Compatibility settings for OpenAI-compatible completions APIs.\n * Use this to override URL-based auto-detection for custom providers.\n */\nexport interface OpenAICompletionsCompat {\n\t/** Whether the provider supports the `store` field. Default: auto-detected from URL. */\n\tsupportsStore?: boolean;\n\t/** Whether the provider supports the `developer` role (vs `system`). Default: auto-detected from URL. */\n\tsupportsDeveloperRole?: boolean;\n\t/** Whether the provider supports `reasoning_effort`. Default: auto-detected from URL. */\n\tsupportsReasoningEffort?: boolean;\n\t/** Whether the provider supports `stream_options: { include_usage: true }` for token usage in streaming responses. Default: true. */\n\tsupportsUsageInStreaming?: boolean;\n\t/** Which field to use for max tokens. Default: auto-detected from URL. */\n\tmaxTokensField?: \"max_completion_tokens\" | \"max_tokens\";\n\t/** Whether tool results require the `name` field. Default: auto-detected from URL. */\n\trequiresToolResultName?: boolean;\n\t/** Whether a user message after tool results requires an assistant message in between. Default: auto-detected from URL. */\n\trequiresAssistantAfterToolResult?: boolean;\n\t/** Whether thinking blocks must be converted to text blocks with <thinking> delimiters. Default: auto-detected from URL. */\n\trequiresThinkingAsText?: boolean;\n\t/** Whether all replayed assistant messages must include an empty reasoning_content field when reasoning is enabled. Default: auto-detected from URL. */\n\trequiresReasoningContentOnAssistantMessages?: boolean;\n\t/** Format for reasoning/thinking parameter. \"openai\" uses reasoning_effort, \"openrouter\" uses reasoning: { effort }, \"deepseek\" uses thinking: { type } plus reasoning_effort, \"together\" uses reasoning: { enabled } plus reasoning_effort when supported, \"zai\" uses top-level enable_thinking: boolean, \"qwen\" uses top-level enable_thinking: boolean, and \"qwen-chat-template\" uses chat_template_kwargs.enable_thinking. Default: \"openai\". */\n\tthinkingFormat?: \"openai\" | \"openrouter\" | \"deepseek\" | \"together\" | \"zai\" | \"qwen\" | \"qwen-chat-template\";\n\t/** OpenRouter-specific routing preferences. Only used when baseUrl points to OpenRouter. */\n\topenRouterRouting?: OpenRouterRouting;\n\t/** Vercel AI Gateway routing preferences. Only used when baseUrl points to Vercel AI Gateway. */\n\tvercelGatewayRouting?: VercelGatewayRouting;\n\t/** Whether z.ai supports top-level `tool_stream: true` for streaming tool call deltas. Default: false. */\n\tzaiToolStream?: boolean;\n\t/** Whether the provider supports the `strict` field in tool definitions. Default: true. */\n\tsupportsStrictMode?: boolean;\n\t/** Cache control convention for prompt caching. \"anthropic\" applies Anthropic-style `cache_control` markers to the system prompt, last tool definition, and last user/assistant text content. */\n\tcacheControlFormat?: \"anthropic\";\n\t/** Whether to send known session-affinity headers (`session_id`, `x-client-request-id`, `x-session-affinity`) from `options.sessionId` when caching is enabled. Default: false. */\n\tsendSessionAffinityHeaders?: boolean;\n\t/** Whether the provider supports long prompt cache retention (`prompt_cache_retention: \"24h\"` or Anthropic-style `cache_control.ttl: \"1h\"`, depending on format). Default: true. */\n\tsupportsLongCacheRetention?: boolean;\n}\n\n/** Compatibility settings for OpenAI Responses APIs. */\nexport interface OpenAIResponsesCompat {\n\t/** Whether to send the OpenAI `session_id` cache-affinity header from `options.sessionId` when caching is enabled. Default: true. */\n\tsendSessionIdHeader?: boolean;\n\t/** Whether the provider supports `prompt_cache_retention: \"24h\"`. Default: true. */\n\tsupportsLongCacheRetention?: boolean;\n}\n\n/** Compatibility settings for Anthropic Messages-compatible APIs. */\nexport interface AnthropicMessagesCompat {\n\t/**\n\t * Whether the provider accepts per-tool `eager_input_streaming`.\n\t * When false, the Anthropic provider omits `tools[].eager_input_streaming`\n\t * and sends the legacy `fine-grained-tool-streaming-2025-05-14` beta header\n\t * for tool-enabled requests.\n\t * Default: true.\n\t */\n\tsupportsEagerToolInputStreaming?: boolean;\n\t/** Whether the provider supports Anthropic long cache retention (`cache_control.ttl: \"1h\"`). Default: true. */\n\tsupportsLongCacheRetention?: boolean;\n\t/**\n\t * Whether to send the `x-session-affinity` header from `options.sessionId`\n\t * when caching is enabled. Required for providers like Fireworks that use\n\t * session affinity for prompt cache routing (requests to the same replica\n\t * maximize cache hits).\n\t * Default: false.\n\t */\n\tsendSessionAffinityHeaders?: boolean;\n\t/**\n\t * Whether the provider supports Anthropic-style `cache_control` markers on\n\t * tool definitions. When false, `cache_control` is omitted from tool params.\n\t * Some Anthropic-compatible providers (e.g., Fireworks) do not support this\n\t * field on tools and may reject or ignore it.\n\t * Default: true.\n\t */\n\tsupportsCacheControlOnTools?: boolean;\n\t/**\n\t * Whether to force adaptive thinking (`thinking.type: \"adaptive\"` plus\n\t * `output_config.effort`) regardless of the model id. Built-in models that\n\t * require adaptive thinking set this in generated metadata. Custom\n\t * Anthropic-compatible providers can set this to `true` for any model whose\n\t * upstream requires the adaptive format. Set to `false` to\n\t * opt out on overridden built-in models.\n\t * Default: false.\n\t */\n\tforceAdaptiveThinking?: boolean;\n}\n\n/**\n * OpenRouter provider routing preferences.\n * Controls which upstream providers OpenRouter routes requests to.\n * Sent as the `provider` field in the OpenRouter API request body.\n * @see https://openrouter.ai/docs/guides/routing/provider-selection\n */\nexport interface OpenRouterRouting {\n\t/** Whether to allow backup providers to serve requests. Default: true. */\n\tallow_fallbacks?: boolean;\n\t/** Whether to filter providers to only those that support all parameters in the request. Default: false. */\n\trequire_parameters?: boolean;\n\t/** Data collection setting. \"allow\" (default): allow providers that may store/train on data. \"deny\": only use providers that don't collect user data. */\n\tdata_collection?: \"deny\" | \"allow\";\n\t/** Whether to restrict routing to only ZDR (Zero Data Retention) endpoints. */\n\tzdr?: boolean;\n\t/** Whether to restrict routing to only models that allow text distillation. */\n\tenforce_distillable_text?: boolean;\n\t/** An ordered list of provider names/slugs to try in sequence, falling back to the next if unavailable. */\n\torder?: string[];\n\t/** List of provider names/slugs to exclusively allow for this request. */\n\tonly?: string[];\n\t/** List of provider names/slugs to skip for this request. */\n\tignore?: string[];\n\t/** A list of quantization levels to filter providers by (e.g., [\"fp16\", \"bf16\", \"fp8\", \"fp6\", \"int8\", \"int4\", \"fp4\", \"fp32\"]). */\n\tquantizations?: string[];\n\t/** Sorting strategy. Can be a string (e.g., \"price\", \"throughput\", \"latency\") or an object with `by` and `partition`. */\n\tsort?:\n\t\t| string\n\t\t| {\n\t\t\t\t/** The sorting metric: \"price\", \"throughput\", \"latency\". */\n\t\t\t\tby?: string;\n\t\t\t\t/** Partitioning strategy: \"model\" (default) or \"none\". */\n\t\t\t\tpartition?: string | null;\n\t\t  };\n\t/** Maximum price per million tokens (USD). */\n\tmax_price?: {\n\t\t/** Price per million prompt tokens. */\n\t\tprompt?: number | string;\n\t\t/** Price per million completion tokens. */\n\t\tcompletion?: number | string;\n\t\t/** Price per image. */\n\t\timage?: number | string;\n\t\t/** Price per audio unit. */\n\t\taudio?: number | string;\n\t\t/** Price per request. */\n\t\trequest?: number | string;\n\t};\n\t/** Preferred minimum throughput (tokens/second). Can be a number (applies to p50) or an object with percentile-specific cutoffs. */\n\tpreferred_min_throughput?:\n\t\t| number\n\t\t| {\n\t\t\t\t/** Minimum tokens/second at the 50th percentile. */\n\t\t\t\tp50?: number;\n\t\t\t\t/** Minimum tokens/second at the 75th percentile. */\n\t\t\t\tp75?: number;\n\t\t\t\t/** Minimum tokens/second at the 90th percentile. */\n\t\t\t\tp90?: number;\n\t\t\t\t/** Minimum tokens/second at the 99th percentile. */\n\t\t\t\tp99?: number;\n\t\t  };\n\t/** Preferred maximum latency (seconds). Can be a number (applies to p50) or an object with percentile-specific cutoffs. */\n\tpreferred_max_latency?:\n\t\t| number\n\t\t| {\n\t\t\t\t/** Maximum latency in seconds at the 50th percentile. */\n\t\t\t\tp50?: number;\n\t\t\t\t/** Maximum latency in seconds at the 75th percentile. */\n\t\t\t\tp75?: number;\n\t\t\t\t/** Maximum latency in seconds at the 90th percentile. */\n\t\t\t\tp90?: number;\n\t\t\t\t/** Maximum latency in seconds at the 99th percentile. */\n\t\t\t\tp99?: number;\n\t\t  };\n}\n\n/**\n * Vercel AI Gateway routing preferences.\n * Controls which upstream providers the gateway routes requests to.\n * @see https://vercel.com/docs/ai-gateway/models-and-providers/provider-options\n */\nexport interface VercelGatewayRouting {\n\t/** List of provider slugs to exclusively use for this request (e.g., [\"bedrock\", \"anthropic\"]). */\n\tonly?: string[];\n\t/** List of provider slugs to try in order (e.g., [\"anthropic\", \"openai\"]). */\n\torder?: string[];\n}\n\n// Model interface for the unified model system\nexport interface Model<TApi extends Api> {\n\tid: string;\n\tname: string;\n\tapi: TApi;\n\tprovider: Provider;\n\tbaseUrl: string;\n\treasoning: boolean;\n\t/**\n\t * Maps pi thinking levels to provider/model-specific values.\n\t * Missing keys use provider defaults. null marks a level as unsupported.\n\t */\n\tthinkingLevelMap?: ThinkingLevelMap;\n\tinput: (\"text\" | \"image\")[];\n\tcost: {\n\t\tinput: number; // $/million tokens\n\t\toutput: number; // $/million tokens\n\t\tcacheRead: number; // $/million tokens\n\t\tcacheWrite: number; // $/million tokens\n\t};\n\tcontextWindow: number;\n\tmaxTokens: number;\n\theaders?: Record<string, string>;\n\t/** Compatibility overrides for OpenAI-compatible APIs. If not set, auto-detected from baseUrl. */\n\tcompat?: TApi extends \"openai-completions\"\n\t\t? OpenAICompletionsCompat\n\t\t: TApi extends \"openai-responses\"\n\t\t\t? OpenAIResponsesCompat\n\t\t\t: TApi extends \"anthropic-messages\"\n\t\t\t\t? AnthropicMessagesCompat\n\t\t\t\t: never;\n\t/** Provider-specific options passed through to stream handlers. */\n\tproviderOptions?: Record<string, unknown>;\n}\n\nexport interface ImagesModel<TApi extends ImagesApi>\n\textends Omit<Model<Api>, \"api\" | \"provider\" | \"reasoning\" | \"contextWindow\" | \"maxTokens\" | \"compat\"> {\n\tapi: TApi;\n\tprovider: ImagesProvider;\n\toutput: (\"text\" | \"image\")[];\n}\n"]}
+{"version":3,"file":"types.js","sourceRoot":"","sources":["../src/types.ts"],"names":[],"mappings":"","sourcesContent":["import type { AssistantMessageDiagnostic } from \"./utils/diagnostics.js\";\nimport type { AssistantMessageEventStream } from \"./utils/event-stream.js\";\n\nexport type { AssistantMessageEventStream } from \"./utils/event-stream.js\";\n\nexport type KnownApi =\n\t| \"openai-completions\"\n\t| \"mistral-conversations\"\n\t| \"openai-responses\"\n\t| \"azure-openai-responses\"\n\t| \"openai-codex-responses\"\n\t| \"anthropic-messages\"\n\t| \"bedrock-converse-stream\"\n\t| \"google-generative-ai\"\n\t| \"google-vertex\";\n\nexport type Api = KnownApi | (string & {});\n\nexport type KnownImagesApi = \"openrouter-images\";\n\nexport type ImagesApi = KnownImagesApi | (string & {});\n\nexport type KnownProvider =\n\t| \"amazon-bedrock\"\n\t| \"anthropic\"\n\t| \"google\"\n\t| \"google-vertex\"\n\t| \"openai\"\n\t| \"azure-openai-responses\"\n\t| \"openai-codex\"\n\t| \"deepseek\"\n\t| \"github-copilot\"\n\t| \"xai\"\n\t| \"groq\"\n\t| \"cerebras\"\n\t| \"openrouter\"\n\t| \"vercel-ai-gateway\"\n\t| \"zai\"\n\t| \"mistral\"\n\t| \"minimax\"\n\t| \"minimax-cn\"\n\t| \"moonshotai\"\n\t| \"moonshotai-cn\"\n\t| \"huggingface\"\n\t| \"fireworks\"\n\t| \"together\"\n\t| \"opencode\"\n\t| \"opencode-go\"\n\t| \"kimi-coding\"\n\t| \"cloudflare-workers-ai\"\n\t| \"cloudflare-ai-gateway\"\n\t| \"xiaomi\"\n\t| \"xiaomi-token-plan-cn\"\n\t| \"xiaomi-token-plan-ams\"\n\t| \"xiaomi-token-plan-sgp\";\nexport type Provider = KnownProvider | string;\n\nexport type KnownImagesProvider = \"openrouter\";\n\nexport type ImagesProvider = KnownImagesProvider | string;\n\nexport type ThinkingLevel = \"minimal\" | \"low\" | \"medium\" | \"high\" | \"xhigh\";\nexport type ModelThinkingLevel = \"off\" | ThinkingLevel;\nexport type ThinkingLevelMap = Partial<Record<ModelThinkingLevel, string | null>>;\n\n/** Token budgets for each thinking level (token-based providers only) */\nexport interface ThinkingBudgets {\n\tminimal?: number;\n\tlow?: number;\n\tmedium?: number;\n\thigh?: number;\n}\n\n// Base options all providers share\nexport type CacheRetention = \"none\" | \"short\" | \"long\";\n\nexport type Transport = \"sse\" | \"websocket\" | \"websocket-cached\" | \"auto\";\n\nexport interface ProviderResponse {\n\tstatus: number;\n\theaders: Record<string, string>;\n}\n\nexport interface StreamOptions {\n\ttemperature?: number;\n\tmaxTokens?: number;\n\tsignal?: AbortSignal;\n\tapiKey?: string;\n\t/**\n\t * Preferred transport for providers that support multiple transports.\n\t * Providers that do not support this option ignore it.\n\t */\n\ttransport?: Transport;\n\t/**\n\t * Prompt cache retention preference. Providers map this to their supported values.\n\t * Default: \"short\".\n\t */\n\tcacheRetention?: CacheRetention;\n\t/**\n\t * Optional session identifier for providers that support session-based caching.\n\t * Providers can use this to enable prompt caching, request routing, or other\n\t * session-aware features. Ignored by providers that don't support it.\n\t */\n\tsessionId?: string;\n\t/**\n\t * Optional callback for inspecting or replacing provider payloads before sending.\n\t * Return undefined to keep the payload unchanged.\n\t */\n\tonPayload?: (payload: unknown, model: Model<Api>) => unknown | undefined | Promise<unknown | undefined>;\n\t/**\n\t * Optional callback invoked after an HTTP response is received and before\n\t * its body stream is consumed.\n\t */\n\tonResponse?: (response: ProviderResponse, model: Model<Api>) => void | Promise<void>;\n\t/**\n\t * Optional custom HTTP headers to include in API requests.\n\t * Merged with provider defaults; can override default headers.\n\t * Not supported by all providers (e.g., AWS Bedrock uses SDK auth).\n\t */\n\theaders?: Record<string, string>;\n\t/**\n\t * HTTP request timeout in milliseconds for providers/SDKs that support it.\n\t * For example, OpenAI and Anthropic SDK clients default to 10 minutes.\n\t */\n\ttimeoutMs?: number;\n\t/**\n\t * Maximum retry attempts for providers/SDKs that support client-side retries.\n\t * For example, OpenAI and Anthropic SDK clients default to 2.\n\t */\n\tmaxRetries?: number;\n\t/**\n\t * Maximum delay in milliseconds to wait for a retry when the server requests a long wait.\n\t * If the server's requested delay exceeds this value, the request fails immediately\n\t * with an error containing the requested delay, allowing higher-level retry logic\n\t * to handle it with user visibility.\n\t * Default: 60000 (60 seconds). Set to 0 to disable the cap.\n\t */\n\tmaxRetryDelayMs?: number;\n\t/**\n\t * Optional metadata to include in API requests.\n\t * Providers extract the fields they understand and ignore the rest.\n\t * For example, Anthropic uses `user_id` for abuse tracking and rate limiting.\n\t */\n\tmetadata?: Record<string, unknown>;\n}\n\nexport type ProviderStreamOptions = StreamOptions & Record<string, unknown>;\n\nexport interface ImagesOptions {\n\tsignal?: AbortSignal;\n\tapiKey?: string;\n\t/**\n\t * Optional callback for inspecting or replacing provider payloads before sending.\n\t * Return undefined to keep the payload unchanged.\n\t */\n\tonPayload?: (payload: unknown, model: ImagesModel<ImagesApi>) => unknown | undefined | Promise<unknown | undefined>;\n\t/**\n\t * Optional callback invoked after an HTTP response is received.\n\t */\n\tonResponse?: (response: ProviderResponse, model: ImagesModel<ImagesApi>) => void | Promise<void>;\n\t/**\n\t * Optional custom HTTP headers to include in API requests.\n\t * Merged with provider defaults; can override default headers.\n\t */\n\theaders?: Record<string, string>;\n\t/**\n\t * HTTP request timeout in milliseconds for providers/SDKs that support it.\n\t */\n\ttimeoutMs?: number;\n\t/**\n\t * Maximum retry attempts for providers/SDKs that support client-side retries.\n\t */\n\tmaxRetries?: number;\n\t/**\n\t * Maximum delay in milliseconds to wait for a retry when the server requests a long wait.\n\t * If the server's requested delay exceeds this value, the request fails immediately\n\t * with an error containing the requested delay, allowing higher-level retry logic\n\t * to handle it with user visibility.\n\t * Default: 60000 (60 seconds). Set to 0 to disable the cap.\n\t */\n\tmaxRetryDelayMs?: number;\n\t/**\n\t * Optional metadata to include in API requests.\n\t * Providers extract the fields they understand and ignore the rest.\n\t */\n\tmetadata?: Record<string, unknown>;\n}\n\nexport type ProviderImagesOptions = ImagesOptions & Record<string, unknown>;\n\n// Unified options with reasoning passed to streamSimple() and completeSimple()\nexport interface SimpleStreamOptions extends StreamOptions {\n\t/** Workspace root for providers that spawn local processes. */\n\tcwd?: string;\n\treasoning?: ThinkingLevel;\n\t/** Custom token budgets for thinking levels (token-based providers only) */\n\tthinkingBudgets?: ThinkingBudgets;\n}\n\n// Generic StreamFunction with typed options.\n//\n// Contract:\n// - Must return an AssistantMessageEventStream.\n// - Once invoked, request/model/runtime failures should be encoded in the\n//   returned stream, not thrown.\n// - Error termination must produce an AssistantMessage with stopReason\n//   \"error\" or \"aborted\" and errorMessage, emitted via the stream protocol.\nexport type StreamFunction<TApi extends Api = Api, TOptions extends StreamOptions = StreamOptions> = (\n\tmodel: Model<TApi>,\n\tcontext: Context,\n\toptions?: TOptions,\n) => AssistantMessageEventStream;\n\nexport type ImagesFunction<TApi extends ImagesApi = ImagesApi, TOptions extends ImagesOptions = ImagesOptions> = (\n\tmodel: ImagesModel<TApi>,\n\tcontext: ImagesContext,\n\toptions?: TOptions,\n) => Promise<AssistantImages>;\n\nexport interface TextSignatureV1 {\n\tv: 1;\n\tid: string;\n\tphase?: \"commentary\" | \"final_answer\";\n}\n\nexport interface TextContent {\n\ttype: \"text\";\n\ttext: string;\n\ttextSignature?: string; // e.g., for OpenAI responses, message metadata (legacy id string or TextSignatureV1 JSON)\n}\n\nexport interface ThinkingContent {\n\ttype: \"thinking\";\n\tthinking: string;\n\tthinkingSignature?: string; // e.g., for OpenAI responses, the reasoning item ID\n\t/** When true, the thinking content was redacted by safety filters. The opaque\n\t *  encrypted payload is stored in `thinkingSignature` so it can be passed back\n\t *  to the API for multi-turn continuity. */\n\tredacted?: boolean;\n}\n\nexport interface ImageContent {\n\ttype: \"image\";\n\tdata: string; // base64 encoded image data\n\tmimeType: string; // e.g., \"image/jpeg\", \"image/png\"\n}\n\nexport interface ToolCall {\n\ttype: \"toolCall\";\n\tid: string;\n\tname: string;\n\targuments: Record<string, any>;\n\tthoughtSignature?: string; // Google-specific: opaque signature for reusing thought context\n\t/** Result from an externally executed tool (e.g. Claude Code SDK). Skips local dispatch. */\n\texternalResult?: {\n\t\tcontent?: Array<{ type: string; text?: string; data?: string; mimeType?: string }>;\n\t\tdetails?: Record<string, unknown>;\n\t\tisError?: boolean;\n\t};\n}\n\n/** GSD compat: server-side tool invocation block (e.g. web search). */\nexport interface ServerToolUse {\n\ttype: \"serverToolUse\";\n\tid: string;\n\tname: string;\n\tinput: unknown;\n\tcaller?: unknown;\n\texternalResult?: {\n\t\tcontent?: Array<{ type: string; text?: string; data?: string; mimeType?: string }>;\n\t\tdetails?: Record<string, unknown>;\n\t\tisError?: boolean;\n\t};\n}\n\n/** GSD compat: server-side tool result block paired with serverToolUse. */\nexport interface WebSearchResult {\n\ttype: \"webSearchResult\";\n\ttoolUseId: string;\n\tcontent: unknown;\n\tcaller?: unknown;\n}\n\n/** GSD compat alias for {@link ServerToolUse}. */\nexport type ServerToolUseContent = ServerToolUse;\n\n/** GSD compat alias for {@link WebSearchResult}. */\nexport type WebSearchResultContent = WebSearchResult;\n\nexport interface Usage {\n\tinput: number;\n\toutput: number;\n\tcacheRead: number;\n\tcacheWrite: number;\n\ttotalTokens: number;\n\tcost: {\n\t\tinput: number;\n\t\toutput: number;\n\t\tcacheRead: number;\n\t\tcacheWrite: number;\n\t\ttotal: number;\n\t};\n}\n\nexport type StopReason = \"stop\" | \"length\" | \"toolUse\" | \"error\" | \"aborted\";\n\nexport interface UserMessage {\n\trole: \"user\";\n\tcontent: string | (TextContent | ImageContent)[];\n\ttimestamp: number; // Unix timestamp in milliseconds\n}\n\nexport interface AssistantMessage {\n\trole: \"assistant\";\n\tcontent: (TextContent | ThinkingContent | ToolCall | ServerToolUse | WebSearchResult)[];\n\tapi: Api;\n\tprovider: Provider;\n\tmodel: string;\n\tresponseModel?: string; // Concrete `chunk.model` when different from the requested `model` (e.g. OpenRouter `auto` -> `anthropic/...`)\n\tresponseId?: string; // Provider-specific response/message identifier when the upstream API exposes one\n\tdiagnostics?: AssistantMessageDiagnostic[]; // Redacted provider/runtime diagnostics for failures and recoveries.\n\tusage: Usage;\n\tstopReason: StopReason;\n\terrorMessage?: string;\n\t/** Server-requested retry delay in milliseconds (from Retry-After or rate limit headers). */\n\tretryAfterMs?: number;\n\t/** Provider inference performance metrics (e.g. tokens/sec from local models). */\n\tinferenceMetrics?: InferenceMetrics;\n\ttimestamp: number; // Unix timestamp in milliseconds\n}\n\n/** Inference performance metrics reported by providers that support it (e.g. Ollama). */\nexport interface InferenceMetrics {\n\t/** Tokens generated per second during eval phase. */\n\ttokensPerSecond: number;\n\t/** Wall-clock duration of the full request in milliseconds. */\n\ttotalDurationMs: number;\n\t/** Duration of the eval (generation) phase in milliseconds. */\n\tevalDurationMs: number;\n\t/** Duration of the prompt eval phase in milliseconds. */\n\tpromptEvalDurationMs: number;\n}\n\nexport interface ToolResultMessage<TDetails = any> {\n\trole: \"toolResult\";\n\ttoolCallId: string;\n\ttoolName: string;\n\tcontent: (TextContent | ImageContent)[]; // Supports text and images\n\tdetails?: TDetails;\n\tisError: boolean;\n\ttimestamp: number; // Unix timestamp in milliseconds\n}\n\nexport type Message = UserMessage | AssistantMessage | ToolResultMessage;\n\nexport type ImagesInputContent = TextContent | ImageContent;\nexport type ImagesOutputContent = TextContent | ImageContent;\n\nexport interface ImagesContext {\n\tinput: ImagesInputContent[];\n}\n\nexport type ImagesStopReason = \"stop\" | \"error\" | \"aborted\";\n\nexport interface AssistantImages {\n\tapi: ImagesApi;\n\tprovider: ImagesProvider;\n\tmodel: string;\n\toutput: ImagesOutputContent[];\n\tresponseId?: string;\n\tusage?: Usage;\n\tstopReason: ImagesStopReason;\n\terrorMessage?: string;\n\ttimestamp: number; // Unix timestamp in milliseconds\n}\n\nimport type { TSchema } from \"typebox\";\n\nexport interface Tool<TParameters extends TSchema = TSchema> {\n\tname: string;\n\tdescription: string;\n\tparameters: TParameters;\n}\n\nexport interface Context {\n\tsystemPrompt?: string;\n\tmessages: Message[];\n\ttools?: Tool[];\n}\n\n/**\n * Event protocol for AssistantMessageEventStream.\n *\n * Streams should emit `start` before partial updates, then terminate with either:\n * - `done` carrying the final successful AssistantMessage, or\n * - `error` carrying the final AssistantMessage with stopReason \"error\" or \"aborted\"\n *   and errorMessage.\n */\nexport type AssistantMessageEvent =\n\t| { type: \"start\"; partial: AssistantMessage }\n\t| { type: \"text_start\"; contentIndex: number; partial: AssistantMessage }\n\t| { type: \"text_delta\"; contentIndex: number; delta: string; partial: AssistantMessage }\n\t| { type: \"text_end\"; contentIndex: number; content: string; partial: AssistantMessage }\n\t| { type: \"thinking_start\"; contentIndex: number; partial: AssistantMessage }\n\t| { type: \"thinking_delta\"; contentIndex: number; delta: string; partial: AssistantMessage }\n\t| { type: \"thinking_end\"; contentIndex: number; content: string; partial: AssistantMessage }\n\t| { type: \"toolcall_start\"; contentIndex: number; partial: AssistantMessage }\n\t| { type: \"toolcall_delta\"; contentIndex: number; delta: string; partial: AssistantMessage }\n\t| {\n\t\t\ttype: \"toolcall_end\";\n\t\t\tcontentIndex: number;\n\t\t\ttoolCall: ToolCall;\n\t\t\tpartial: AssistantMessage;\n\t\t\tmalformedArguments?: boolean;\n\t  }\n\t| { type: \"server_tool_use\"; contentIndex: number; partial: AssistantMessage }\n\t| { type: \"done\"; reason: Extract<StopReason, \"stop\" | \"length\" | \"toolUse\">; message: AssistantMessage }\n\t| { type: \"error\"; reason: Extract<StopReason, \"aborted\" | \"error\">; error: AssistantMessage };\n\n/**\n * Compatibility settings for OpenAI-compatible completions APIs.\n * Use this to override URL-based auto-detection for custom providers.\n */\nexport interface OpenAICompletionsCompat {\n\t/** Whether the provider supports the `store` field. Default: auto-detected from URL. */\n\tsupportsStore?: boolean;\n\t/** Whether the provider supports the `developer` role (vs `system`). Default: auto-detected from URL. */\n\tsupportsDeveloperRole?: boolean;\n\t/** Whether the provider supports `reasoning_effort`. Default: auto-detected from URL. */\n\tsupportsReasoningEffort?: boolean;\n\t/** Whether the provider supports `stream_options: { include_usage: true }` for token usage in streaming responses. Default: true. */\n\tsupportsUsageInStreaming?: boolean;\n\t/** Which field to use for max tokens. Default: auto-detected from URL. */\n\tmaxTokensField?: \"max_completion_tokens\" | \"max_tokens\";\n\t/** Whether tool results require the `name` field. Default: auto-detected from URL. */\n\trequiresToolResultName?: boolean;\n\t/** Whether a user message after tool results requires an assistant message in between. Default: auto-detected from URL. */\n\trequiresAssistantAfterToolResult?: boolean;\n\t/** Whether thinking blocks must be converted to text blocks with <thinking> delimiters. Default: auto-detected from URL. */\n\trequiresThinkingAsText?: boolean;\n\t/** Whether all replayed assistant messages must include an empty reasoning_content field when reasoning is enabled. Default: auto-detected from URL. */\n\trequiresReasoningContentOnAssistantMessages?: boolean;\n\t/** Format for reasoning/thinking parameter. \"openai\" uses reasoning_effort, \"openrouter\" uses reasoning: { effort }, \"deepseek\" uses thinking: { type } plus reasoning_effort, \"together\" uses reasoning: { enabled } plus reasoning_effort when supported, \"zai\" uses top-level enable_thinking: boolean, \"qwen\" uses top-level enable_thinking: boolean, and \"qwen-chat-template\" uses chat_template_kwargs.enable_thinking. Default: \"openai\". */\n\tthinkingFormat?: \"openai\" | \"openrouter\" | \"deepseek\" | \"together\" | \"zai\" | \"qwen\" | \"qwen-chat-template\";\n\t/** OpenRouter-specific routing preferences. Only used when baseUrl points to OpenRouter. */\n\topenRouterRouting?: OpenRouterRouting;\n\t/** Vercel AI Gateway routing preferences. Only used when baseUrl points to Vercel AI Gateway. */\n\tvercelGatewayRouting?: VercelGatewayRouting;\n\t/** Whether z.ai supports top-level `tool_stream: true` for streaming tool call deltas. Default: false. */\n\tzaiToolStream?: boolean;\n\t/** Whether the provider supports the `strict` field in tool definitions. Default: true. */\n\tsupportsStrictMode?: boolean;\n\t/** Cache control convention for prompt caching. \"anthropic\" applies Anthropic-style `cache_control` markers to the system prompt, last tool definition, and last user/assistant text content. */\n\tcacheControlFormat?: \"anthropic\";\n\t/** Whether to send known session-affinity headers (`session_id`, `x-client-request-id`, `x-session-affinity`) from `options.sessionId` when caching is enabled. Default: false. */\n\tsendSessionAffinityHeaders?: boolean;\n\t/** Whether the provider supports long prompt cache retention (`prompt_cache_retention: \"24h\"` or Anthropic-style `cache_control.ttl: \"1h\"`, depending on format). Default: true. */\n\tsupportsLongCacheRetention?: boolean;\n}\n\n/** Compatibility settings for OpenAI Responses APIs. */\nexport interface OpenAIResponsesCompat {\n\t/** Whether to send the OpenAI `session_id` cache-affinity header from `options.sessionId` when caching is enabled. Default: true. */\n\tsendSessionIdHeader?: boolean;\n\t/** Whether the provider supports `prompt_cache_retention: \"24h\"`. Default: true. */\n\tsupportsLongCacheRetention?: boolean;\n}\n\n/** Compatibility settings for Anthropic Messages-compatible APIs. */\nexport interface AnthropicMessagesCompat {\n\t/**\n\t * Whether the provider accepts per-tool `eager_input_streaming`.\n\t * When false, the Anthropic provider omits `tools[].eager_input_streaming`\n\t * and sends the legacy `fine-grained-tool-streaming-2025-05-14` beta header\n\t * for tool-enabled requests.\n\t * Default: true.\n\t */\n\tsupportsEagerToolInputStreaming?: boolean;\n\t/** Whether the provider supports Anthropic long cache retention (`cache_control.ttl: \"1h\"`). Default: true. */\n\tsupportsLongCacheRetention?: boolean;\n\t/**\n\t * Whether to send the `x-session-affinity` header from `options.sessionId`\n\t * when caching is enabled. Required for providers like Fireworks that use\n\t * session affinity for prompt cache routing (requests to the same replica\n\t * maximize cache hits).\n\t * Default: false.\n\t */\n\tsendSessionAffinityHeaders?: boolean;\n\t/**\n\t * Whether the provider supports Anthropic-style `cache_control` markers on\n\t * tool definitions. When false, `cache_control` is omitted from tool params.\n\t * Some Anthropic-compatible providers (e.g., Fireworks) do not support this\n\t * field on tools and may reject or ignore it.\n\t * Default: true.\n\t */\n\tsupportsCacheControlOnTools?: boolean;\n\t/**\n\t * Whether to force adaptive thinking (`thinking.type: \"adaptive\"` plus\n\t * `output_config.effort`) regardless of the model id. Built-in models that\n\t * require adaptive thinking set this in generated metadata. Custom\n\t * Anthropic-compatible providers can set this to `true` for any model whose\n\t * upstream requires the adaptive format. Set to `false` to\n\t * opt out on overridden built-in models.\n\t * Default: false.\n\t */\n\tforceAdaptiveThinking?: boolean;\n}\n\n/**\n * OpenRouter provider routing preferences.\n * Controls which upstream providers OpenRouter routes requests to.\n * Sent as the `provider` field in the OpenRouter API request body.\n * @see https://openrouter.ai/docs/guides/routing/provider-selection\n */\nexport interface OpenRouterRouting {\n\t/** Whether to allow backup providers to serve requests. Default: true. */\n\tallow_fallbacks?: boolean;\n\t/** Whether to filter providers to only those that support all parameters in the request. Default: false. */\n\trequire_parameters?: boolean;\n\t/** Data collection setting. \"allow\" (default): allow providers that may store/train on data. \"deny\": only use providers that don't collect user data. */\n\tdata_collection?: \"deny\" | \"allow\";\n\t/** Whether to restrict routing to only ZDR (Zero Data Retention) endpoints. */\n\tzdr?: boolean;\n\t/** Whether to restrict routing to only models that allow text distillation. */\n\tenforce_distillable_text?: boolean;\n\t/** An ordered list of provider names/slugs to try in sequence, falling back to the next if unavailable. */\n\torder?: string[];\n\t/** List of provider names/slugs to exclusively allow for this request. */\n\tonly?: string[];\n\t/** List of provider names/slugs to skip for this request. */\n\tignore?: string[];\n\t/** A list of quantization levels to filter providers by (e.g., [\"fp16\", \"bf16\", \"fp8\", \"fp6\", \"int8\", \"int4\", \"fp4\", \"fp32\"]). */\n\tquantizations?: string[];\n\t/** Sorting strategy. Can be a string (e.g., \"price\", \"throughput\", \"latency\") or an object with `by` and `partition`. */\n\tsort?:\n\t\t| string\n\t\t| {\n\t\t\t\t/** The sorting metric: \"price\", \"throughput\", \"latency\". */\n\t\t\t\tby?: string;\n\t\t\t\t/** Partitioning strategy: \"model\" (default) or \"none\". */\n\t\t\t\tpartition?: string | null;\n\t\t  };\n\t/** Maximum price per million tokens (USD). */\n\tmax_price?: {\n\t\t/** Price per million prompt tokens. */\n\t\tprompt?: number | string;\n\t\t/** Price per million completion tokens. */\n\t\tcompletion?: number | string;\n\t\t/** Price per image. */\n\t\timage?: number | string;\n\t\t/** Price per audio unit. */\n\t\taudio?: number | string;\n\t\t/** Price per request. */\n\t\trequest?: number | string;\n\t};\n\t/** Preferred minimum throughput (tokens/second). Can be a number (applies to p50) or an object with percentile-specific cutoffs. */\n\tpreferred_min_throughput?:\n\t\t| number\n\t\t| {\n\t\t\t\t/** Minimum tokens/second at the 50th percentile. */\n\t\t\t\tp50?: number;\n\t\t\t\t/** Minimum tokens/second at the 75th percentile. */\n\t\t\t\tp75?: number;\n\t\t\t\t/** Minimum tokens/second at the 90th percentile. */\n\t\t\t\tp90?: number;\n\t\t\t\t/** Minimum tokens/second at the 99th percentile. */\n\t\t\t\tp99?: number;\n\t\t  };\n\t/** Preferred maximum latency (seconds). Can be a number (applies to p50) or an object with percentile-specific cutoffs. */\n\tpreferred_max_latency?:\n\t\t| number\n\t\t| {\n\t\t\t\t/** Maximum latency in seconds at the 50th percentile. */\n\t\t\t\tp50?: number;\n\t\t\t\t/** Maximum latency in seconds at the 75th percentile. */\n\t\t\t\tp75?: number;\n\t\t\t\t/** Maximum latency in seconds at the 90th percentile. */\n\t\t\t\tp90?: number;\n\t\t\t\t/** Maximum latency in seconds at the 99th percentile. */\n\t\t\t\tp99?: number;\n\t\t  };\n}\n\n/**\n * Vercel AI Gateway routing preferences.\n * Controls which upstream providers the gateway routes requests to.\n * @see https://vercel.com/docs/ai-gateway/models-and-providers/provider-options\n */\nexport interface VercelGatewayRouting {\n\t/** List of provider slugs to exclusively use for this request (e.g., [\"bedrock\", \"anthropic\"]). */\n\tonly?: string[];\n\t/** List of provider slugs to try in order (e.g., [\"anthropic\", \"openai\"]). */\n\torder?: string[];\n}\n\n// Model interface for the unified model system\nexport interface Model<TApi extends Api> {\n\tid: string;\n\tname: string;\n\tapi: TApi;\n\tprovider: Provider;\n\tbaseUrl: string;\n\treasoning: boolean;\n\t/**\n\t * Maps pi thinking levels to provider/model-specific values.\n\t * Missing keys use provider defaults. null marks a level as unsupported.\n\t */\n\tthinkingLevelMap?: ThinkingLevelMap;\n\tinput: (\"text\" | \"image\")[];\n\tcost: {\n\t\tinput: number; // $/million tokens\n\t\toutput: number; // $/million tokens\n\t\tcacheRead: number; // $/million tokens\n\t\tcacheWrite: number; // $/million tokens\n\t};\n\tcontextWindow: number;\n\tmaxTokens: number;\n\theaders?: Record<string, string>;\n\t/** Compatibility overrides for OpenAI-compatible APIs. If not set, auto-detected from baseUrl. */\n\tcompat?: TApi extends \"openai-completions\"\n\t\t? OpenAICompletionsCompat\n\t\t: TApi extends \"openai-responses\"\n\t\t\t? OpenAIResponsesCompat\n\t\t\t: TApi extends \"anthropic-messages\"\n\t\t\t\t? AnthropicMessagesCompat\n\t\t\t\t: never;\n\t/** Provider-specific options passed through to stream handlers. */\n\tproviderOptions?: Record<string, unknown>;\n}\n\nexport interface ImagesModel<TApi extends ImagesApi>\n\textends Omit<Model<Api>, \"api\" | \"provider\" | \"reasoning\" | \"contextWindow\" | \"maxTokens\" | \"compat\"> {\n\tapi: TApi;\n\tprovider: ImagesProvider;\n\toutput: (\"text\" | \"image\")[];\n}\n"]}

package/packages/pi-ai/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@gsd/pi-ai",
-  "version": "1.0.2-dev.5961fbf",
+  "version": "1.0.2-dev.5f7864c",
   "description": "Unified LLM API with automatic model discovery and provider configuration",
   "type": "module",
   "gsd": {

package/packages/pi-coding-agent/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@gsd/pi-coding-agent",
-  "version": "1.0.2-dev.5961fbf",
+  "version": "1.0.2-dev.5f7864c",
   "description": "Coding agent CLI (vendored from earendil-works/pi)",
   "type": "module",
   "gsd": {
@@ -33,7 +33,7 @@
     "copy-assets": "node scripts/copy-assets.cjs"
   },
   "dependencies": {
-    "@opengsd/contracts": "^1.0.2-dev.5961fbf",
+    "@opengsd/contracts": "^1.0.2-dev.5f7864c",
     "@mariozechner/jiti": "^2.6.2",
     "@silvia-odwyer/photon-node": "0.3.4",
     "chalk": "5.6.2",
@@ -53,11 +53,11 @@
     "typebox": "1.1.38",
     "undici": "7.26.0",
     "yaml": "2.9.0",
-    "@gsd/agent-core": "^1.0.2-dev.5961fbf",
-    "@gsd/native": "^1.0.2-dev.5961fbf",
-    "@gsd/pi-agent-core": "^1.0.2-dev.5961fbf",
-    "@gsd/pi-ai": "^1.0.2-dev.5961fbf",
-    "@gsd/pi-tui": "^1.0.2-dev.5961fbf",
+    "@gsd/agent-core": "^1.0.2-dev.5f7864c",
+    "@gsd/native": "^1.0.2-dev.5f7864c",
+    "@gsd/pi-agent-core": "^1.0.2-dev.5f7864c",
+    "@gsd/pi-ai": "^1.0.2-dev.5f7864c",
+    "@gsd/pi-tui": "^1.0.2-dev.5f7864c",
     "@sinclair/typebox": "^0.34.41"
   },
   "devDependencies": {

package/packages/pi-tui/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@gsd/pi-tui",
-  "version": "1.0.2-dev.5961fbf",
+  "version": "1.0.2-dev.5f7864c",
   "description": "Terminal UI library (vendored from earendil-works/pi)",
   "type": "module",
   "gsd": {

package/packages/rpc-client/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@opengsd/rpc-client",
-  "version": "1.0.2-dev.5961fbf",
+  "version": "1.0.2-dev.5f7864c",
   "description": "Standalone RPC client SDK for GSD — zero internal dependencies",
   "license": "MIT",
   "gsd": {
@@ -34,7 +34,7 @@
     "test": "node --test dist/rpc-client.test.js"
   },
   "dependencies": {
-    "@opengsd/contracts": "^1.0.2-dev.5961fbf"
+    "@opengsd/contracts": "^1.0.2-dev.5f7864c"
   },
   "engines": {
     "node": ">=22.0.0"

package/pkg/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@glittercowboy/gsd",
-  "version": "1.0.2-dev.5961fbf",
+  "version": "1.0.2-dev.5f7864c",
   "piConfig": {
     "name": "gsd",
     "configDir": ".gsd"

package/scripts/install/detect-existing.js

@@ -5,11 +5,20 @@
 import { createRequire } from 'module'
 import { dirname, join } from 'path'
 import { fileURLToPath } from 'url'
+import { detectPackageManager } from './npm-global.js'
 
 const require = createRequire(import.meta.url)
 const { GSD_PI_BRAND } = require(join(dirname(fileURLToPath(import.meta.url)), '..', 'lib', 'logo.cjs'))
 
 export function parseInstalledVersion(npmListJson) {
+  if (Array.isArray(npmListJson)) {
+    for (const entry of npmListJson) {
+      const found = parseInstalledVersion(entry)
+      if (found) return found
+    }
+    return null
+  }
+
   if (!npmListJson || typeof npmListJson !== 'object') return null
 
   const direct = npmListJson.dependencies?.['@opengsd/gsd-pi']?.version
@@ -30,6 +39,10 @@ export function parseInstalledVersion(npmListJson) {
   return walk(npmListJson)
 }
 
+function getPackageManagerBin(packageManager) {
+  return process.platform === 'win32' ? `${packageManager}.cmd` : packageManager
+}
+
 /**
  * Decide install action for non-interactive or pre-prompt routing.
  */
@@ -39,11 +52,12 @@ export function compareActions({ installed, yesMode }) {
   return 'prompt'
 }
 
-export async function detectInstalledVersion() {
+export async function detectInstalledVersion(options = {}) {
   const { execFileSync } = await import('child_process')
-  const npm = process.platform === 'win32' ? 'npm.cmd' : 'npm'
+  const packageManager = options.packageManager ?? detectPackageManager()
+  const bin = getPackageManagerBin(packageManager)
   try {
-    const raw = execFileSync(npm, ['list', '-g', '@opengsd/gsd-pi', '--json'], {
+    const raw = execFileSync(bin, ['list', '-g', '@opengsd/gsd-pi', '--json'], {
       encoding: 'utf-8',
       stdio: ['ignore', 'pipe', 'pipe'],
       timeout: 30_000,

package/scripts/install/npm-global.js

@@ -1,14 +1,61 @@
 import { execFileSync, spawn } from 'child_process'
-import { join } from 'path'
+import { homedir } from 'os'
+import { join, resolve as resolvePath, sep } from 'path'
 
-function getNpm() {
-  return process.platform === 'win32' ? 'npm.cmd' : 'npm'
+const PACKAGE_MANAGERS = new Set(['npm', 'pnpm'])
+
+function normalizePackageManager(value) {
+  return PACKAGE_MANAGERS.has(value) ? value : 'npm'
+}
+
+function hasPnpmPath(value = '') {
+  const normalized = value.replace(/\\/g, '/').toLowerCase()
+  return (
+    normalized.includes('/.pnpm/') ||
+    normalized.endsWith('/pnpm') ||
+    normalized.endsWith('/pnpm.cjs') ||
+    normalized.endsWith('/pnpm.js')
+  )
+}
+
+function pathStartsWith(pathValue = '', dir) {
+  if (!pathValue) return false
+  const resolvedPath = resolvePath(pathValue)
+  const resolvedDir = resolvePath(dir)
+  return resolvedPath === resolvedDir || resolvedPath.startsWith(resolvedDir + sep)
+}
+
+function hasPnpmBinPath(env, argv1) {
+  const pnpmBinDirs = []
+  if (env.PNPM_HOME) pnpmBinDirs.push(env.PNPM_HOME)
+  pnpmBinDirs.push(join(homedir(), 'Library', 'pnpm'))
+  pnpmBinDirs.push(join(homedir(), '.local', 'share', 'pnpm'))
+
+  return pnpmBinDirs.some((dir) => pathStartsWith(argv1, dir) || pathStartsWith(env.npm_execpath, dir))
+}
+
+export function detectPackageManager(env = process.env, argv1 = process.argv[1]) {
+  const userAgent = env.npm_config_user_agent || ''
+  if (userAgent.startsWith('pnpm/')) return 'pnpm'
+  // Installer runs under npm during npm install; keep that context authoritative.
+  if (userAgent.startsWith('npm/')) return 'npm'
+
+  if (hasPnpmPath(env.npm_execpath || '')) return 'pnpm'
+  if (hasPnpmPath(argv1 || '')) return 'pnpm'
+  if (hasPnpmBinPath(env, argv1)) return 'pnpm'
+
+  return 'npm'
+}
+
+function getPackageManagerBin(packageManager) {
+  const bin = normalizePackageManager(packageManager)
+  return process.platform === 'win32' ? `${bin}.cmd` : bin
 }
 
-const NPM_OUTPUT_LIMIT = 64 * 1024
+const PM_OUTPUT_LIMIT = 64 * 1024
 
-function runNpm(args) {
-  return execFileSync(getNpm(), args, {
+function runPackageManager(packageManager, args) {
+  return execFileSync(getPackageManagerBin(packageManager), args, {
     encoding: 'utf-8',
     stdio: ['ignore', 'pipe', 'pipe'],
     timeout: 120_000,
@@ -16,30 +63,30 @@ function runNpm(args) {
   }).trim()
 }
 
-function formatNpmFailure(result) {
+function formatPackageManagerFailure(packageManager, result) {
   const output = `${result.stderr}\n${result.stdout}`.trim()
   const meaningful = output
     .split('\n')
-    .filter((line) => !line.includes('npm warn') && !line.includes('npm WARN') && line.trim())
+    .filter((line) => line.trim() && !/^(?:npm\s+warn|warn)\b/i.test(line.trim()))
     .slice(-3)
     .join('; ')
-  return meaningful || result.error?.message || 'npm install failed'
+  return meaningful || result.error?.message || `${packageManager} install failed`
 }
 
 function appendLimited(value, chunk) {
-  if (value.length >= NPM_OUTPUT_LIMIT) return value
-  return value + chunk.slice(0, NPM_OUTPUT_LIMIT - value.length)
+  if (value.length >= PM_OUTPUT_LIMIT) return value
+  return value + chunk.slice(0, PM_OUTPUT_LIMIT - value.length)
 }
 
-function runNpmAsync(args, {
+function runPackageManagerAsync(packageManager, args, {
   captureStdout = false,
   cwd,
   timeout = 300_000,
 } = {}) {
-  const npm = getNpm()
+  const bin = getPackageManagerBin(packageManager)
 
   return new Promise((resolve) => {
-    const child = spawn(npm, args, {
+    const child = spawn(bin, args, {
       cwd,
       shell: process.platform === 'win32',
       stdio: ['ignore', captureStdout ? 'pipe' : 'ignore', 'pipe'],
@@ -85,7 +132,7 @@ function runNpmAsync(args, {
           ok: false,
           stdout,
           stderr,
-          error: new Error(`npm ${args.join(' ')} timed out after ${timeout}ms`),
+          error: new Error(`${packageManager} ${args.join(' ')} timed out after ${timeout}ms`),
         })
         return
       }
@@ -100,20 +147,35 @@ function runNpmAsync(args, {
         ok: false,
         stdout,
         stderr,
-        error: new Error(`npm ${args.join(' ')} failed with ${reason}`),
+        error: new Error(`${packageManager} ${args.join(' ')} failed with ${reason}`),
       })
     })
   })
 }
 
-export function getGlobalPaths() {
-  const prefix = runNpm(['prefix', '-g'])
-  const root = runNpm(['root', '-g'])
+export function getGlobalPaths(options = {}) {
+  const packageManager = normalizePackageManager(options.packageManager ?? detectPackageManager())
+
+  if (packageManager === 'pnpm') {
+    const binDir = runPackageManager(packageManager, ['bin', '-g'])
+    const root = runPackageManager(packageManager, ['root', '-g'])
+    return {
+      prefix: binDir,
+      root,
+      binDir,
+      packageRoot: join(root, '@opengsd', 'gsd-pi'),
+      packageManager,
+    }
+  }
+
+  const prefix = runPackageManager(packageManager, ['prefix', '-g'])
+  const root = runPackageManager(packageManager, ['root', '-g'])
   return {
     prefix,
     root,
     binDir: process.platform === 'win32' ? prefix : join(prefix, 'bin'),
     packageRoot: join(root, '@opengsd', 'gsd-pi'),
+    packageManager,
   }
 }
 
@@ -121,33 +183,41 @@ export function getLocalPackageRoot(cwd = process.cwd()) {
   return join(cwd, 'node_modules', '@opengsd', 'gsd-pi')
 }
 
-export async function installGlobalPackage(version) {
-  const result = await runNpmAsync([
-    'install',
-    '-g',
-    '--ignore-scripts',
-    `@opengsd/gsd-pi@${version}`,
-  ])
+export async function installGlobalPackage(version, options = {}) {
+  const packageManager = normalizePackageManager(options.packageManager ?? detectPackageManager())
+  const packageSpec = `@opengsd/gsd-pi@${version}`
+  const installArgs = packageManager === 'pnpm'
+    ? ['add', '-g', '--ignore-scripts', packageSpec]
+    : ['install', '-g', '--ignore-scripts', packageSpec]
+
+  const result = await runPackageManagerAsync(packageManager, installArgs)
   if (!result.ok) {
-    throw new Error(formatNpmFailure(result))
+    throw new Error(formatPackageManagerFailure(packageManager, result))
   }
-  const rootResult = await runNpmAsync(['root', '-g'], {
+  const rootResult = await runPackageManagerAsync(packageManager, ['root', '-g'], {
     captureStdout: true,
     timeout: 120_000,
   })
   if (!rootResult.ok) {
-    throw new Error(formatNpmFailure(rootResult))
+    throw new Error(formatPackageManagerFailure(packageManager, rootResult))
   }
   return join(rootResult.stdout.trim(), '@opengsd', 'gsd-pi')
 }
 
-export async function installLocalPackage(version, cwd = process.cwd()) {
-  const result = await runNpmAsync(
-    ['install', '--ignore-scripts', `@opengsd/gsd-pi@${version}`],
+export async function installLocalPackage(version, cwd = process.cwd(), options = {}) {
+  const packageManager = normalizePackageManager(options.packageManager ?? detectPackageManager())
+  const packageSpec = `@opengsd/gsd-pi@${version}`
+  const installArgs = packageManager === 'pnpm'
+    ? ['add', '--ignore-scripts', packageSpec]
+    : ['install', '--ignore-scripts', packageSpec]
+
+  const result = await runPackageManagerAsync(
+    packageManager,
+    installArgs,
     { cwd },
   )
   if (!result.ok) {
-    throw new Error(formatNpmFailure(result))
+    throw new Error(formatPackageManagerFailure(packageManager, result))
   }
   return getLocalPackageRoot(cwd)
 }

package/scripts/install.js

@@ -71,6 +71,7 @@ if (HAS_HELP) {
 
   ${colors.yellow}Usage:${colors.reset}
     npx @opengsd/gsd-pi@latest              Install GSD-Pi globally (recommended)
+    pnpm dlx @opengsd/gsd-pi@latest         Install GSD-Pi globally with pnpm
     npx @opengsd/gsd-pi@latest --local      Install to current project (advanced)
 
   ${colors.yellow}Options:${colors.reset}

package/src/resources/extensions/context7/index.ts

@@ -132,6 +132,19 @@ function formatLibraryList(libs: C7Library[], query: string): string {
 	return lines.join("\n");
 }
 
+function getFirstTextContent(result: { content?: unknown }): string | undefined {
+	const content = result.content;
+	if (!Array.isArray(content)) return undefined;
+	const textBlock = content.find(
+		(block: unknown) =>
+			typeof block === "object" &&
+			block !== null &&
+			(block as { type?: unknown }).type === "text" &&
+			typeof (block as { text?: unknown }).text === "string",
+	) as { text: string } | undefined;
+	return textBlock?.text.trim() || undefined;
+}
+
 // ─── Tool details types ───────────────────────────────────────────────────────
 
 interface ResolveDetails {
@@ -234,7 +247,7 @@ export default function (pi: ExtensionAPI) {
 			const d = result.details as ResolveDetails | undefined;
 			if (isPartial) return new Text(theme.fg("warning", "Searching Context7..."), 0, 0);
 			if ((result as any).isError || d?.error) {
-				return new Text(theme.fg("error", `Error: ${d?.error ?? "unknown"}`), 0, 0);
+				return new Text(theme.fg("error", `Error: ${d?.error ?? getFirstTextContent(result) ?? "unknown"}`), 0, 0);
 			}
 			let text = theme.fg("success", `${d?.resultCount ?? 0} ${d?.resultCount === 1 ? "library" : "libraries"} found`);
 			if (d?.cached) text += theme.fg("dim", " (cached)");
@@ -389,7 +402,7 @@ export default function (pi: ExtensionAPI) {
 
 			if (isPartial) return new Text(theme.fg("warning", "Fetching documentation..."), 0, 0);
 			if ((result as any).isError || d?.error) {
-				return new Text(theme.fg("error", `Error: ${d?.error ?? "unknown"}`), 0, 0);
+				return new Text(theme.fg("error", `Error: ${d?.error ?? getFirstTextContent(result) ?? "unknown"}`), 0, 0);
 			}
 
 			let text = theme.fg("success", `${(d?.charCount ?? 0).toLocaleString()} chars`);

package/src/resources/extensions/get-secrets-from-user.ts

@@ -13,6 +13,7 @@ import { resolve } from "node:path";
 import type { ExtensionAPI, Theme } from "@gsd/pi-coding-agent";
 import { Editor, type EditorTheme, Key, matchesKey, Text, truncateToWidth, wrapTextWithAnsi } from "@gsd/pi-tui";
 import { Type } from "@sinclair/typebox";
+import { renderSharedDialogFrame } from "./shared/dialog-frame.js";
 import { makeUI } from "./shared/tui.js";
 import { maskEditorLine, type ProgressStatus } from "./shared/mod.js";
 import { parseSecretsManifest, formatSecretsManifest } from "./gsd/files.js";
@@ -43,6 +44,10 @@ function maskPreview(value: string): string {
 	return `${value.slice(0, 4)}${"*".repeat(Math.max(4, value.length - 8))}${value.slice(-4)}`;
 }
 
+function dialogContentWidth(width: number): number {
+	return width < 4 ? Math.max(1, width) : Math.max(1, width - 4);
+}
+
 function shellEscapeSingle(value: string): string {
 	return `'${value.replace(/'/g, `'\\''`)}'`;
 }
@@ -168,10 +173,10 @@ async function collectOneSecret(
 
 		function render(width: number): string[] {
 			if (cachedLines) return cachedLines;
+			const contentWidth = dialogContentWidth(width);
 			const lines: string[] = [];
-			const add = (s: string) => lines.push(truncateToWidth(s, width));
+			const add = (s: string) => lines.push(truncateToWidth(s, contentWidth));
 
-			add(theme.fg("accent", "─".repeat(width)));
 			add(theme.fg("dim", ` Page ${pageIndex + 1}/${totalPages} · Secure Env Setup`));
 			lines.push("");
 
@@ -187,7 +192,7 @@ async function collectOneSecret(
 				for (let g = 0; g < guidance.length; g++) {
 					const prefix = `  ${g + 1}. `;
 					const step = guidance[g] as string;
-					const wrappedLines = wrapTextWithAnsi(step, width - 4);
+					const wrappedLines = wrapTextWithAnsi(step, Math.max(1, contentWidth - 4));
 					for (let w = 0; w < wrappedLines.length; w++) {
 						const indent = w === 0 ? prefix : " ".repeat(prefix.length);
 						lines.push(theme.fg("dim", `${indent}${wrappedLines[w]}`));
@@ -205,16 +210,15 @@ async function collectOneSecret(
 
 			// Editor
 			add(theme.fg("muted", " Enter value:"));
-			for (const line of editor.render(width - 2)) {
+			for (const line of editor.render(Math.max(1, contentWidth - 2))) {
 				add(theme.fg("text", maskEditorLine(line)));
 			}
 
 			lines.push("");
-			add(theme.fg("dim", ` enter to confirm  |  ctrl+s or esc to skip  |  esc cancels`));
-			add(theme.fg("accent", "─".repeat(width)));
+			const footer = theme.fg("dim", " enter to confirm  |  ctrl+s or esc to skip  |  esc cancels");
 
-			cachedLines = lines;
-			return lines;
+			cachedLines = renderSharedDialogFrame(theme, "Secure Env Setup", lines, width, { footer });
+			return cachedLines;
 		}
 
 		return {
@@ -286,13 +290,11 @@ export async function showSecretsSummary(
 		function render(width: number): string[] {
 			if (cachedLines) return cachedLines;
 
-			const ui = makeUI(theme, width);
+			const contentWidth = dialogContentWidth(width);
+			const ui = makeUI(theme, contentWidth);
 			const lines: string[] = [];
 			const push = (...rows: string[][]) => { for (const r of rows) lines.push(...r); };
 
-			push(ui.bar());
-			push(ui.blank());
-			push(ui.header("  Secrets Summary"));
 			push(ui.blank());
 
 			for (const entry of entries) {
@@ -314,11 +316,10 @@ export async function showSecretsSummary(
 			}
 
 			push(ui.blank());
-			push(ui.hints(["any key to continue"]));
-			push(ui.bar());
+			const footer = ui.hints(["any key to continue"])[0] ?? "";
 
-			cachedLines = lines;
-			return lines;
+			cachedLines = renderSharedDialogFrame(theme, "Secrets Summary", lines, width, { footer });
+			return cachedLines;
 		}
 
 		return {

package/src/resources/extensions/google-cli/index.ts

@@ -0,0 +1,34 @@
+/**
+ * Google local CLI providers.
+ *
+ * These deliberately use authMode "externalCli": GSD never owns the browser
+ * OAuth flow or cached tokens. Users authenticate with the official CLI, then
+ * /login activates the provider once the local binary is available.
+ */
+
+import type { ExtensionAPI } from "@gsd/pi-coding-agent";
+import { GOOGLE_ANTIGRAVITY_MODELS, GOOGLE_GEMINI_CLI_MODELS } from "./models.js";
+import { isAntigravityCliReady, isGeminiCliReady } from "./readiness.js";
+import { streamViaGoogleCli } from "./stream-adapter.js";
+
+export default function googleCli(pi: ExtensionAPI) {
+	pi.registerProvider("google-gemini-cli", {
+		name: "Google Gemini CLI",
+		authMode: "externalCli",
+		api: "google-gemini-cli",
+		baseUrl: "local://google-gemini-cli",
+		isReady: isGeminiCliReady,
+		streamSimple: streamViaGoogleCli,
+		models: GOOGLE_GEMINI_CLI_MODELS,
+	});
+
+	pi.registerProvider("google-antigravity", {
+		name: "Google Antigravity",
+		authMode: "externalCli",
+		api: "google-antigravity",
+		baseUrl: "local://google-antigravity",
+		isReady: isAntigravityCliReady,
+		streamSimple: streamViaGoogleCli,
+		models: GOOGLE_ANTIGRAVITY_MODELS,
+	});
+}