@skydiveai/pi-server 0.1.0-beta.165 → 0.1.0-beta.2448

package/dist/index.d.mts

@@ -208,30 +208,7 @@ declare function createPrewarm({
  */
 declare const KNOWN_PI_PROVIDERS: readonly ["amazon-bedrock", "anthropic", "google", "google-vertex", "openai", "azure-openai-responses", "openai-codex", "deepseek", "github-copilot", "xai", "groq", "cerebras", "openrouter", "vercel-ai-gateway", "zai", "mistral", "minimax", "minimax-cn", "moonshotai", "moonshotai-cn", "huggingface", "fireworks", "opencode", "opencode-go", "kimi-coding", "cloudflare-workers-ai", "cloudflare-ai-gateway", "xiaomi", "xiaomi-token-plan-cn", "xiaomi-token-plan-ams", "xiaomi-token-plan-sgp"];
 declare const modelSpecSchema: z.ZodDiscriminatedUnion<"api", [z.ZodObject<{
-  api: z.ZodLiteral<"openai-completions">;
-  compat: z.ZodOptional<z.ZodObject<{
-    supportsReasoningEffort: z.ZodOptional<z.ZodBoolean>;
-    requiresThinkingAsText: z.ZodOptional<z.ZodBoolean>;
-    thinkingFormat: z.ZodOptional<z.ZodEnum<["openai", "openrouter", "deepseek", "zai", "qwen", "qwen-chat-template"]>>;
-    supportsStrictMode: z.ZodOptional<z.ZodBoolean>;
-    maxTokensField: z.ZodOptional<z.ZodEnum<["max_completion_tokens", "max_tokens"]>>;
-    cacheControlFormat: z.ZodOptional<z.ZodLiteral<"anthropic">>;
-  }, "strip", z.ZodTypeAny, {
-    supportsReasoningEffort?: boolean | undefined;
-    requiresThinkingAsText?: boolean | undefined;
-    thinkingFormat?: "openai" | "deepseek" | "openrouter" | "zai" | "qwen" | "qwen-chat-template" | undefined;
-    supportsStrictMode?: boolean | undefined;
-    maxTokensField?: "max_completion_tokens" | "max_tokens" | undefined;
-    cacheControlFormat?: "anthropic" | undefined;
-  }, {
-    supportsReasoningEffort?: boolean | undefined;
-    requiresThinkingAsText?: boolean | undefined;
-    thinkingFormat?: "openai" | "deepseek" | "openrouter" | "zai" | "qwen" | "qwen-chat-template" | undefined;
-    supportsStrictMode?: boolean | undefined;
-    maxTokensField?: "max_completion_tokens" | "max_tokens" | undefined;
-    cacheControlFormat?: "anthropic" | undefined;
-  }>>; /** Exact model id the serving endpoint expects. */
-  id: z.ZodString; /** Display name shown in pi surfaces and session logs. */
+  /** Exact model id the serving endpoint expects. */id: z.ZodString; /** Display name shown in pi surfaces and session logs. */
   name: z.ZodString; /** pi-ai provider key — also the auth-storage key for `apiKey`. */
   provider: z.ZodEnum<["amazon-bedrock", "anthropic", "google", "google-vertex", "openai", "azure-openai-responses", "openai-codex", "deepseek", "github-copilot", "xai", "groq", "cerebras", "openrouter", "vercel-ai-gateway", "zai", "mistral", "minimax", "minimax-cn", "moonshotai", "moonshotai-cn", "huggingface", "fireworks", "opencode", "opencode-go", "kimi-coding", "cloudflare-workers-ai", "cloudflare-ai-gateway", "xiaomi", "xiaomi-token-plan-cn", "xiaomi-token-plan-ams", "xiaomi-token-plan-sgp"]>; /** Provider API root, e.g. "https://openrouter.ai/api/v1". */
   baseUrl: z.ZodString;
@@ -244,17 +221,17 @@ declare const modelSpecSchema: z.ZodDiscriminatedUnion<"api", [z.ZodObject<{
     high: z.ZodOptional<z.ZodNullable<z.ZodString>>;
     xhigh: z.ZodOptional<z.ZodNullable<z.ZodString>>;
   }, "strip", z.ZodTypeAny, {
-    medium?: string | null | undefined;
     off?: string | null | undefined;
     minimal?: string | null | undefined;
     low?: string | null | undefined;
+    medium?: string | null | undefined;
     high?: string | null | undefined;
     xhigh?: string | null | undefined;
   }, {
-    medium?: string | null | undefined;
     off?: string | null | undefined;
     minimal?: string | null | undefined;
     low?: string | null | undefined;
+    medium?: string | null | undefined;
     high?: string | null | undefined;
     xhigh?: string | null | undefined;
   }>>;
@@ -268,73 +245,85 @@ declare const modelSpecSchema: z.ZodDiscriminatedUnion<"api", [z.ZodObject<{
    * placeholder that layer recognizes — the value is opaque to this server.
    */
   apiKey: z.ZodOptional<z.ZodString>;
+  api: z.ZodLiteral<"openai-completions">;
+  compat: z.ZodOptional<z.ZodObject<{
+    supportsReasoningEffort: z.ZodOptional<z.ZodBoolean>;
+    requiresThinkingAsText: z.ZodOptional<z.ZodBoolean>;
+    thinkingFormat: z.ZodOptional<z.ZodEnum<["openai", "openrouter", "deepseek", "zai", "qwen", "qwen-chat-template"]>>;
+    supportsStrictMode: z.ZodOptional<z.ZodBoolean>;
+    maxTokensField: z.ZodOptional<z.ZodEnum<["max_completion_tokens", "max_tokens"]>>;
+    cacheControlFormat: z.ZodOptional<z.ZodLiteral<"anthropic">>;
+  }, "strip", z.ZodTypeAny, {
+    supportsReasoningEffort?: boolean | undefined;
+    requiresThinkingAsText?: boolean | undefined;
+    thinkingFormat?: "deepseek" | "openai" | "openrouter" | "qwen" | "qwen-chat-template" | "zai" | undefined;
+    supportsStrictMode?: boolean | undefined;
+    maxTokensField?: "max_completion_tokens" | "max_tokens" | undefined;
+    cacheControlFormat?: "anthropic" | undefined;
+  }, {
+    supportsReasoningEffort?: boolean | undefined;
+    requiresThinkingAsText?: boolean | undefined;
+    thinkingFormat?: "deepseek" | "openai" | "openrouter" | "qwen" | "qwen-chat-template" | "zai" | undefined;
+    supportsStrictMode?: boolean | undefined;
+    maxTokensField?: "max_completion_tokens" | "max_tokens" | undefined;
+    cacheControlFormat?: "anthropic" | undefined;
+  }>>;
 }, "strip", z.ZodTypeAny, {
-  api: "openai-completions";
   id: string;
   name: string;
-  provider: "anthropic" | "amazon-bedrock" | "google" | "google-vertex" | "openai" | "azure-openai-responses" | "openai-codex" | "deepseek" | "github-copilot" | "xai" | "groq" | "cerebras" | "openrouter" | "vercel-ai-gateway" | "zai" | "mistral" | "minimax" | "minimax-cn" | "moonshotai" | "moonshotai-cn" | "huggingface" | "fireworks" | "opencode" | "opencode-go" | "kimi-coding" | "cloudflare-workers-ai" | "cloudflare-ai-gateway" | "xiaomi" | "xiaomi-token-plan-cn" | "xiaomi-token-plan-ams" | "xiaomi-token-plan-sgp";
+  provider: "amazon-bedrock" | "anthropic" | "azure-openai-responses" | "cerebras" | "cloudflare-ai-gateway" | "cloudflare-workers-ai" | "deepseek" | "fireworks" | "github-copilot" | "google" | "google-vertex" | "groq" | "huggingface" | "kimi-coding" | "minimax" | "minimax-cn" | "mistral" | "moonshotai" | "moonshotai-cn" | "openai" | "openai-codex" | "opencode" | "opencode-go" | "openrouter" | "vercel-ai-gateway" | "xai" | "xiaomi" | "xiaomi-token-plan-ams" | "xiaomi-token-plan-cn" | "xiaomi-token-plan-sgp" | "zai";
   baseUrl: string;
   reasoning: boolean;
-  input: ["text" | "image", ...("text" | "image")[]];
+  thinkingLevelMap?: {
+    off?: string | null | undefined;
+    minimal?: string | null | undefined;
+    low?: string | null | undefined;
+    medium?: string | null | undefined;
+    high?: string | null | undefined;
+    xhigh?: string | null | undefined;
+  } | undefined;
+  input: ["image" | "text", ...("image" | "text")[]];
   contextWindow: number;
   maxTokens: number;
+  apiKey?: string | undefined;
+  api: "openai-completions";
   compat?: {
     supportsReasoningEffort?: boolean | undefined;
     requiresThinkingAsText?: boolean | undefined;
-    thinkingFormat?: "openai" | "deepseek" | "openrouter" | "zai" | "qwen" | "qwen-chat-template" | undefined;
+    thinkingFormat?: "deepseek" | "openai" | "openrouter" | "qwen" | "qwen-chat-template" | "zai" | undefined;
     supportsStrictMode?: boolean | undefined;
     maxTokensField?: "max_completion_tokens" | "max_tokens" | undefined;
     cacheControlFormat?: "anthropic" | undefined;
   } | undefined;
+}, {
+  id: string;
+  name: string;
+  provider: "amazon-bedrock" | "anthropic" | "azure-openai-responses" | "cerebras" | "cloudflare-ai-gateway" | "cloudflare-workers-ai" | "deepseek" | "fireworks" | "github-copilot" | "google" | "google-vertex" | "groq" | "huggingface" | "kimi-coding" | "minimax" | "minimax-cn" | "mistral" | "moonshotai" | "moonshotai-cn" | "openai" | "openai-codex" | "opencode" | "opencode-go" | "openrouter" | "vercel-ai-gateway" | "xai" | "xiaomi" | "xiaomi-token-plan-ams" | "xiaomi-token-plan-cn" | "xiaomi-token-plan-sgp" | "zai";
+  baseUrl: string;
+  reasoning: boolean;
   thinkingLevelMap?: {
-    medium?: string | null | undefined;
     off?: string | null | undefined;
     minimal?: string | null | undefined;
     low?: string | null | undefined;
+    medium?: string | null | undefined;
     high?: string | null | undefined;
     xhigh?: string | null | undefined;
   } | undefined;
-  apiKey?: string | undefined;
-}, {
-  api: "openai-completions";
-  id: string;
-  name: string;
-  provider: "anthropic" | "amazon-bedrock" | "google" | "google-vertex" | "openai" | "azure-openai-responses" | "openai-codex" | "deepseek" | "github-copilot" | "xai" | "groq" | "cerebras" | "openrouter" | "vercel-ai-gateway" | "zai" | "mistral" | "minimax" | "minimax-cn" | "moonshotai" | "moonshotai-cn" | "huggingface" | "fireworks" | "opencode" | "opencode-go" | "kimi-coding" | "cloudflare-workers-ai" | "cloudflare-ai-gateway" | "xiaomi" | "xiaomi-token-plan-cn" | "xiaomi-token-plan-ams" | "xiaomi-token-plan-sgp";
-  baseUrl: string;
-  reasoning: boolean;
-  input: ["text" | "image", ...("text" | "image")[]];
+  input: ["image" | "text", ...("image" | "text")[]];
   contextWindow: number;
   maxTokens: number;
+  apiKey?: string | undefined;
+  api: "openai-completions";
   compat?: {
     supportsReasoningEffort?: boolean | undefined;
     requiresThinkingAsText?: boolean | undefined;
-    thinkingFormat?: "openai" | "deepseek" | "openrouter" | "zai" | "qwen" | "qwen-chat-template" | undefined;
+    thinkingFormat?: "deepseek" | "openai" | "openrouter" | "qwen" | "qwen-chat-template" | "zai" | undefined;
     supportsStrictMode?: boolean | undefined;
     maxTokensField?: "max_completion_tokens" | "max_tokens" | undefined;
     cacheControlFormat?: "anthropic" | undefined;
   } | undefined;
-  thinkingLevelMap?: {
-    medium?: string | null | undefined;
-    off?: string | null | undefined;
-    minimal?: string | null | undefined;
-    low?: string | null | undefined;
-    high?: string | null | undefined;
-    xhigh?: string | null | undefined;
-  } | undefined;
-  apiKey?: string | undefined;
 }>, z.ZodObject<{
-  api: z.ZodLiteral<"anthropic-messages">;
-  compat: z.ZodOptional<z.ZodObject<{
-    supportsEagerToolInputStreaming: z.ZodOptional<z.ZodBoolean>;
-    supportsLongCacheRetention: z.ZodOptional<z.ZodBoolean>;
-  }, "strip", z.ZodTypeAny, {
-    supportsEagerToolInputStreaming?: boolean | undefined;
-    supportsLongCacheRetention?: boolean | undefined;
-  }, {
-    supportsEagerToolInputStreaming?: boolean | undefined;
-    supportsLongCacheRetention?: boolean | undefined;
-  }>>; /** Exact model id the serving endpoint expects. */
-  id: z.ZodString; /** Display name shown in pi surfaces and session logs. */
+  /** Exact model id the serving endpoint expects. */id: z.ZodString; /** Display name shown in pi surfaces and session logs. */
   name: z.ZodString; /** pi-ai provider key — also the auth-storage key for `apiKey`. */
   provider: z.ZodEnum<["amazon-bedrock", "anthropic", "google", "google-vertex", "openai", "azure-openai-responses", "openai-codex", "deepseek", "github-copilot", "xai", "groq", "cerebras", "openrouter", "vercel-ai-gateway", "zai", "mistral", "minimax", "minimax-cn", "moonshotai", "moonshotai-cn", "huggingface", "fireworks", "opencode", "opencode-go", "kimi-coding", "cloudflare-workers-ai", "cloudflare-ai-gateway", "xiaomi", "xiaomi-token-plan-cn", "xiaomi-token-plan-ams", "xiaomi-token-plan-sgp"]>; /** Provider API root, e.g. "https://openrouter.ai/api/v1". */
   baseUrl: z.ZodString;
@@ -347,17 +336,17 @@ declare const modelSpecSchema: z.ZodDiscriminatedUnion<"api", [z.ZodObject<{
     high: z.ZodOptional<z.ZodNullable<z.ZodString>>;
     xhigh: z.ZodOptional<z.ZodNullable<z.ZodString>>;
   }, "strip", z.ZodTypeAny, {
-    medium?: string | null | undefined;
     off?: string | null | undefined;
     minimal?: string | null | undefined;
     low?: string | null | undefined;
+    medium?: string | null | undefined;
     high?: string | null | undefined;
     xhigh?: string | null | undefined;
   }, {
-    medium?: string | null | undefined;
     off?: string | null | undefined;
     minimal?: string | null | undefined;
     low?: string | null | undefined;
+    medium?: string | null | undefined;
     high?: string | null | undefined;
     xhigh?: string | null | undefined;
   }>>;
@@ -371,52 +360,63 @@ declare const modelSpecSchema: z.ZodDiscriminatedUnion<"api", [z.ZodObject<{
    * placeholder that layer recognizes — the value is opaque to this server.
    */
   apiKey: z.ZodOptional<z.ZodString>;
+  api: z.ZodLiteral<"anthropic-messages">;
+  compat: z.ZodOptional<z.ZodObject<{
+    supportsEagerToolInputStreaming: z.ZodOptional<z.ZodBoolean>;
+    supportsLongCacheRetention: z.ZodOptional<z.ZodBoolean>;
+  }, "strip", z.ZodTypeAny, {
+    supportsEagerToolInputStreaming?: boolean | undefined;
+    supportsLongCacheRetention?: boolean | undefined;
+  }, {
+    supportsEagerToolInputStreaming?: boolean | undefined;
+    supportsLongCacheRetention?: boolean | undefined;
+  }>>;
 }, "strip", z.ZodTypeAny, {
-  api: "anthropic-messages";
   id: string;
   name: string;
-  provider: "anthropic" | "amazon-bedrock" | "google" | "google-vertex" | "openai" | "azure-openai-responses" | "openai-codex" | "deepseek" | "github-copilot" | "xai" | "groq" | "cerebras" | "openrouter" | "vercel-ai-gateway" | "zai" | "mistral" | "minimax" | "minimax-cn" | "moonshotai" | "moonshotai-cn" | "huggingface" | "fireworks" | "opencode" | "opencode-go" | "kimi-coding" | "cloudflare-workers-ai" | "cloudflare-ai-gateway" | "xiaomi" | "xiaomi-token-plan-cn" | "xiaomi-token-plan-ams" | "xiaomi-token-plan-sgp";
+  provider: "amazon-bedrock" | "anthropic" | "azure-openai-responses" | "cerebras" | "cloudflare-ai-gateway" | "cloudflare-workers-ai" | "deepseek" | "fireworks" | "github-copilot" | "google" | "google-vertex" | "groq" | "huggingface" | "kimi-coding" | "minimax" | "minimax-cn" | "mistral" | "moonshotai" | "moonshotai-cn" | "openai" | "openai-codex" | "opencode" | "opencode-go" | "openrouter" | "vercel-ai-gateway" | "xai" | "xiaomi" | "xiaomi-token-plan-ams" | "xiaomi-token-plan-cn" | "xiaomi-token-plan-sgp" | "zai";
   baseUrl: string;
   reasoning: boolean;
-  input: ["text" | "image", ...("text" | "image")[]];
-  contextWindow: number;
-  maxTokens: number;
-  compat?: {
-    supportsEagerToolInputStreaming?: boolean | undefined;
-    supportsLongCacheRetention?: boolean | undefined;
-  } | undefined;
   thinkingLevelMap?: {
-    medium?: string | null | undefined;
     off?: string | null | undefined;
     minimal?: string | null | undefined;
     low?: string | null | undefined;
+    medium?: string | null | undefined;
     high?: string | null | undefined;
     xhigh?: string | null | undefined;
   } | undefined;
-  apiKey?: string | undefined;
-}, {
-  api: "anthropic-messages";
-  id: string;
-  name: string;
-  provider: "anthropic" | "amazon-bedrock" | "google" | "google-vertex" | "openai" | "azure-openai-responses" | "openai-codex" | "deepseek" | "github-copilot" | "xai" | "groq" | "cerebras" | "openrouter" | "vercel-ai-gateway" | "zai" | "mistral" | "minimax" | "minimax-cn" | "moonshotai" | "moonshotai-cn" | "huggingface" | "fireworks" | "opencode" | "opencode-go" | "kimi-coding" | "cloudflare-workers-ai" | "cloudflare-ai-gateway" | "xiaomi" | "xiaomi-token-plan-cn" | "xiaomi-token-plan-ams" | "xiaomi-token-plan-sgp";
-  baseUrl: string;
-  reasoning: boolean;
-  input: ["text" | "image", ...("text" | "image")[]];
+  input: ["image" | "text", ...("image" | "text")[]];
   contextWindow: number;
   maxTokens: number;
+  apiKey?: string | undefined;
+  api: "anthropic-messages";
   compat?: {
     supportsEagerToolInputStreaming?: boolean | undefined;
     supportsLongCacheRetention?: boolean | undefined;
   } | undefined;
+}, {
+  id: string;
+  name: string;
+  provider: "amazon-bedrock" | "anthropic" | "azure-openai-responses" | "cerebras" | "cloudflare-ai-gateway" | "cloudflare-workers-ai" | "deepseek" | "fireworks" | "github-copilot" | "google" | "google-vertex" | "groq" | "huggingface" | "kimi-coding" | "minimax" | "minimax-cn" | "mistral" | "moonshotai" | "moonshotai-cn" | "openai" | "openai-codex" | "opencode" | "opencode-go" | "openrouter" | "vercel-ai-gateway" | "xai" | "xiaomi" | "xiaomi-token-plan-ams" | "xiaomi-token-plan-cn" | "xiaomi-token-plan-sgp" | "zai";
+  baseUrl: string;
+  reasoning: boolean;
   thinkingLevelMap?: {
-    medium?: string | null | undefined;
     off?: string | null | undefined;
     minimal?: string | null | undefined;
     low?: string | null | undefined;
+    medium?: string | null | undefined;
     high?: string | null | undefined;
     xhigh?: string | null | undefined;
   } | undefined;
+  input: ["image" | "text", ...("image" | "text")[]];
+  contextWindow: number;
+  maxTokens: number;
   apiKey?: string | undefined;
+  api: "anthropic-messages";
+  compat?: {
+    supportsEagerToolInputStreaming?: boolean | undefined;
+    supportsLongCacheRetention?: boolean | undefined;
+  } | undefined;
 }>]>;
 type ModelSpec = z.infer<typeof modelSpecSchema>;
 /**
@@ -426,16 +426,20 @@ type ModelSpec = z.infer<typeof modelSpecSchema>;
  */
 declare function parseModelSpec(input: unknown, log: Logger): ModelSpec | null;
 declare function buildModelFromSpec(spec: ModelSpec): {
-  compat?: {
-    supportsReasoningEffort?: boolean | undefined;
-    requiresThinkingAsText?: boolean | undefined;
-    thinkingFormat?: "openai" | "deepseek" | "openrouter" | "zai" | "qwen" | "qwen-chat-template" | undefined;
-    supportsStrictMode?: boolean | undefined;
-    maxTokensField?: "max_completion_tokens" | "max_tokens" | undefined;
-    cacheControlFormat?: "anthropic" | undefined;
+  id: string;
+  name: string;
+  provider: "amazon-bedrock" | "anthropic" | "azure-openai-responses" | "cerebras" | "cloudflare-ai-gateway" | "cloudflare-workers-ai" | "deepseek" | "fireworks" | "github-copilot" | "google" | "google-vertex" | "groq" | "huggingface" | "kimi-coding" | "minimax" | "minimax-cn" | "mistral" | "moonshotai" | "moonshotai-cn" | "openai" | "openai-codex" | "opencode" | "opencode-go" | "openrouter" | "vercel-ai-gateway" | "xai" | "xiaomi" | "xiaomi-token-plan-ams" | "xiaomi-token-plan-cn" | "xiaomi-token-plan-sgp" | "zai";
+  baseUrl: string;
+  reasoning: boolean;
+  thinkingLevelMap?: {
+    off?: string | null | undefined;
+    minimal?: string | null | undefined;
+    low?: string | null | undefined;
+    medium?: string | null | undefined;
+    high?: string | null | undefined;
+    xhigh?: string | null | undefined;
   } | undefined;
-  api: "openai-completions";
-  input: ["text" | "image", ...("text" | "image")[]];
+  input: ["image" | "text", ...("image" | "text")[]];
   cost: {
     input: number;
     output: number;
@@ -444,26 +448,30 @@ declare function buildModelFromSpec(spec: ModelSpec): {
   };
   contextWindow: number;
   maxTokens: number;
+  api: "openai-completions";
+  compat?: {
+    supportsReasoningEffort?: boolean | undefined;
+    requiresThinkingAsText?: boolean | undefined;
+    thinkingFormat?: "deepseek" | "openai" | "openrouter" | "qwen" | "qwen-chat-template" | "zai" | undefined;
+    supportsStrictMode?: boolean | undefined;
+    maxTokensField?: "max_completion_tokens" | "max_tokens" | undefined;
+    cacheControlFormat?: "anthropic" | undefined;
+  } | undefined;
+} | {
+  id: string;
+  name: string;
+  provider: "amazon-bedrock" | "anthropic" | "azure-openai-responses" | "cerebras" | "cloudflare-ai-gateway" | "cloudflare-workers-ai" | "deepseek" | "fireworks" | "github-copilot" | "google" | "google-vertex" | "groq" | "huggingface" | "kimi-coding" | "minimax" | "minimax-cn" | "mistral" | "moonshotai" | "moonshotai-cn" | "openai" | "openai-codex" | "opencode" | "opencode-go" | "openrouter" | "vercel-ai-gateway" | "xai" | "xiaomi" | "xiaomi-token-plan-ams" | "xiaomi-token-plan-cn" | "xiaomi-token-plan-sgp" | "zai";
+  baseUrl: string;
+  reasoning: boolean;
   thinkingLevelMap?: {
-    medium?: string | null | undefined;
     off?: string | null | undefined;
     minimal?: string | null | undefined;
     low?: string | null | undefined;
+    medium?: string | null | undefined;
     high?: string | null | undefined;
     xhigh?: string | null | undefined;
   } | undefined;
-  id: string;
-  name: string;
-  provider: "anthropic" | "amazon-bedrock" | "google" | "google-vertex" | "openai" | "azure-openai-responses" | "openai-codex" | "deepseek" | "github-copilot" | "xai" | "groq" | "cerebras" | "openrouter" | "vercel-ai-gateway" | "zai" | "mistral" | "minimax" | "minimax-cn" | "moonshotai" | "moonshotai-cn" | "huggingface" | "fireworks" | "opencode" | "opencode-go" | "kimi-coding" | "cloudflare-workers-ai" | "cloudflare-ai-gateway" | "xiaomi" | "xiaomi-token-plan-cn" | "xiaomi-token-plan-ams" | "xiaomi-token-plan-sgp";
-  baseUrl: string;
-  reasoning: boolean;
-} | {
-  compat?: {
-    supportsEagerToolInputStreaming?: boolean | undefined;
-    supportsLongCacheRetention?: boolean | undefined;
-  } | undefined;
-  api: "anthropic-messages";
-  input: ["text" | "image", ...("text" | "image")[]];
+  input: ["image" | "text", ...("image" | "text")[]];
   cost: {
     input: number;
     output: number;
@@ -472,19 +480,11 @@ declare function buildModelFromSpec(spec: ModelSpec): {
   };
   contextWindow: number;
   maxTokens: number;
-  thinkingLevelMap?: {
-    medium?: string | null | undefined;
-    off?: string | null | undefined;
-    minimal?: string | null | undefined;
-    low?: string | null | undefined;
-    high?: string | null | undefined;
-    xhigh?: string | null | undefined;
+  api: "anthropic-messages";
+  compat?: {
+    supportsEagerToolInputStreaming?: boolean | undefined;
+    supportsLongCacheRetention?: boolean | undefined;
   } | undefined;
-  id: string;
-  name: string;
-  provider: "anthropic" | "amazon-bedrock" | "google" | "google-vertex" | "openai" | "azure-openai-responses" | "openai-codex" | "deepseek" | "github-copilot" | "xai" | "groq" | "cerebras" | "openrouter" | "vercel-ai-gateway" | "zai" | "mistral" | "minimax" | "minimax-cn" | "moonshotai" | "moonshotai-cn" | "huggingface" | "fireworks" | "opencode" | "opencode-go" | "kimi-coding" | "cloudflare-workers-ai" | "cloudflare-ai-gateway" | "xiaomi" | "xiaomi-token-plan-cn" | "xiaomi-token-plan-ams" | "xiaomi-token-plan-sgp";
-  baseUrl: string;
-  reasoning: boolean;
 };
 /**
  * Resolve the model a request should run on: a valid `x_model` wins,
@@ -542,4 +542,27 @@ declare function getCurrentTraceparent(): string | null;
 /** Extract the trace id (second segment) from a `traceparent` value. */
 declare function parseTraceId(traceparent: string | null): string | null;
 //#endregion
-export { type A2AOptions, type AnthropicMessagesOptions, type ChatCompletionsOptions, type CreateSessionArgs, DEFAULT_MODEL, DEFAULT_PROVIDER, DEFAULT_THINKING_LEVEL, KNOWN_PI_PROVIDERS, type Middleware, type ModelSpec, PI_AGENT_DIR, type PostPromptCallback, type ProtocolHandler, type ProtocolsOptions, type ResponsesOptions, type SessionFactory, type SessionSetupCallback, type TraceContext, VALID_THINKING_LEVELS, type WebHandler, buildAgentCard, buildModelFromSpec, chainMiddleware, composeHandlers, createAgentExecutor, createMapSessionRegistry, createPrewarm, createProtocolHandlers, createProtocols, deriveTraceContext, getCurrentTraceparent, logger, modelSpecSchema, mountAt, newTraceContext, parseModelSpec, parseTraceId, requestHeaders, requestLogger, requestUrl, resolveRequestModel, runInTraceContext, setCurrentTraceparent, webHandlerToMiddleware, withSpecApiKey };
+//#region src/billing-blocked-provider-signal.d.ts
+/**
+ * Duplicated from the billing gate's `BlockReason` (apps/anyone/billing
+ * gate/block-reason.ts): pi-server runs inside the sandbox and must not
+ * depend on platform packages. A reason the proxy sends that predates this
+ * build fails the enum and degrades to the untyped `billing_blocked`
+ * handling — never a crash. Exported so agent-lifecycle's chat.test.ts can
+ * pin this copy against its own (which billing's block-reason.test.ts in
+ * turn pins against the gate), keeping every copy drift-checked.
+ */
+declare const BILLING_BLOCK_REASONS: readonly ["insufficient_balance", "payment_failed", "hard_spend_limit_reached"];
+declare const payloadSchema: z.ZodObject<{
+  blockReason: z.ZodEnum<["insufficient_balance", "payment_failed", "hard_spend_limit_reached"]>;
+  message: z.ZodString;
+}, "strict", z.ZodTypeAny, {
+  blockReason: "hard_spend_limit_reached" | "insufficient_balance" | "payment_failed";
+  message: string;
+}, {
+  blockReason: "hard_spend_limit_reached" | "insufficient_balance" | "payment_failed";
+  message: string;
+}>;
+type BillingBlockedProviderSignal = z.infer<typeof payloadSchema>;
+//#endregion
+export { type A2AOptions, type AnthropicMessagesOptions, BILLING_BLOCK_REASONS, type BillingBlockedProviderSignal, type ChatCompletionsOptions, type CreateSessionArgs, DEFAULT_MODEL, DEFAULT_PROVIDER, DEFAULT_THINKING_LEVEL, KNOWN_PI_PROVIDERS, type Middleware, type ModelSpec, PI_AGENT_DIR, type PostPromptCallback, type ProtocolHandler, type ProtocolsOptions, type ResponsesOptions, type SessionFactory, type SessionSetupCallback, type TraceContext, VALID_THINKING_LEVELS, type WebHandler, buildAgentCard, buildModelFromSpec, chainMiddleware, composeHandlers, createAgentExecutor, createMapSessionRegistry, createPrewarm, createProtocolHandlers, createProtocols, deriveTraceContext, getCurrentTraceparent, logger, modelSpecSchema, mountAt, newTraceContext, parseModelSpec, parseTraceId, requestHeaders, requestLogger, requestUrl, resolveRequestModel, runInTraceContext, setCurrentTraceparent, webHandlerToMiddleware, withSpecApiKey };

package/dist/index.mjs

@@ -259,7 +259,7 @@ function parseShellEnv(request) {
 		const env = {};
 		for (const [k, v] of Object.entries(parsed)) if (typeof v === "string") env[k] = v;
 		return Object.keys(env).length > 0 ? env : null;
-	} catch {
+	} catch (_err) {
 		return null;
 	}
 }
@@ -283,12 +283,12 @@ function createSSEResponse(handler) {
 		write(chunk) {
 			try {
 				controller.enqueue(encoder.encode(chunk));
-			} catch {}
+			} catch (_err) {}
 		},
 		close() {
 			try {
 				controller.close();
-			} catch {}
+			} catch (_err) {}
 		}
 	};
 	const keepalive = setInterval(() => {
@@ -298,13 +298,13 @@ function createSSEResponse(handler) {
 		clearInterval(keepalive);
 		try {
 			controller.close();
-		} catch {}
+		} catch (_err) {}
 	}, (err) => {
 		logger.error({ err }, "SSE handler crashed — stream closed without response");
 		clearInterval(keepalive);
 		try {
 			controller.close();
-		} catch {}
+		} catch (_err) {}
 	});
 	return new Response(stream, {
 		status: 200,
@@ -373,6 +373,57 @@ async function runConversation({ session, prompt, images, log, postPrompt }) {
 	});
 }
 /**
+* Track the agent session's built-in model-call auto-retry so it leaves a
+* trace.
+*
+* `AgentSession` already restarts a failed assistant turn in place (via
+* `agent.continue()`, so no prompt is replayed and no tool re-executes) for the
+* transient provider/transport failures pi classifies as retryable -- dropped
+* streams, `terminated`, 5xx, overloaded, rate limits. It is on by default,
+* with its own budget and backoff, and it emits `auto_retry_start` /
+* `auto_retry_end` around each attempt.
+*
+* Nothing consumed those events, so a retry left no trace anywhere: a call that
+* succeeded first try and one that burned the whole budget before failing
+* produced the same terminal error, and the fleet-wide retry rate was
+* unmeasurable. That gap is why a 2026-08-16 investigation into three runs lost
+* to `provider_error: terminated` could not tell whether the budget had run out
+* (ANY-7101).
+*
+* Exposed as a handler rather than its own `session.subscribe` call so each
+* protocol feeds it from the single subscription it already owns -- one
+* subscriber, explicit ordering.
+*
+* `attempts()` reports what has been spent so far, so a terminal error can
+* carry the count to the worker, where it lands in a log group we can query
+* fleet-wide (the sandbox's own logs are not).
+*/
+function createAutoRetryObserver(log) {
+	let attempts = 0;
+	return {
+		observe(event) {
+			if (event.type === "auto_retry_start") {
+				attempts = typeof event.attempt === "number" ? event.attempt : attempts + 1;
+				log.warn({
+					event: "model_call_auto_retry",
+					attempt: event.attempt,
+					max_attempts: event.maxAttempts,
+					delay_ms: event.delayMs,
+					error_message: event.errorMessage
+				}, "retrying failed model call in place");
+				return;
+			}
+			if (event.type === "auto_retry_end") log.warn({
+				event: "model_call_auto_retry_end",
+				attempt: event.attempt,
+				success: event.success,
+				final_error: event.finalError
+			}, event.success ? "model call recovered after retry" : "model call retries exhausted");
+		},
+		attempts: () => attempts
+	};
+}
+/**
 * Hard-stop the in-flight turn for a session. Shared by every protocol's
 * `/:id/abort` route: a cancel signals the stop explicitly instead of relying
 * on a dropped connection. `session.abort()` interrupts the turn and resolves
@@ -419,7 +470,10 @@ async function extractParts(parts, log) {
 		}
 		if (part.url != null && part.mediaType?.startsWith("image/")) {
 			try {
-				images.push(await fetchImageAsBase64({ url: part.url }));
+				images.push(await fetchImageAsBase64({
+					url: part.url,
+					userAgent: null
+				}));
 			} catch (err) {
 				log.warn({
 					event: "a2a_image_fetch_failed",
@@ -877,6 +931,7 @@ function classifyCode(status, message, code) {
 	if (status === 429 || code === "rate_limit_exceeded" || /rate[_ ]limit|too many requests|resource[_ ]exhausted/i.test(message)) return "rate_limited";
 	if (status === 529 || code === "provider_overloaded" || /overloaded/i.test(message)) return "provider_overloaded";
 	if (status === 503 || status === 504 || code === "provider_unavailable" || /no healthy upstream|upstream request timeout|stream timeout|service unavailable|unavailable|gateway/i.test(message)) return "provider_unavailable";
+	if (code === "content_filter" || /finish_reason:\s*content_filter|content[_ ]filter|content[_ ]management[_ ]policy/i.test(message)) return "content_filter";
 	return "provider_error";
 }
 /**
@@ -891,7 +946,7 @@ function isModelProviderError(err) {
 	if (numericStatus(e) !== null) return true;
 	if (e.error && typeof e.error === "object") return true;
 	const msg = messageText(err);
-	return /prompt is too long|context[_ ]length[_ ]exceeded|exceeds maximum token limit|input token count|rate[_ ]limit|overloaded|no healthy upstream|upstream request timeout|insufficient credits|requires more credits|credit balance is too low|payment required|purchase more credits/i.test(msg);
+	return /prompt is too long|context[_ ]length[_ ]exceeded|exceeds maximum token limit|input token count|rate[_ ]limit|overloaded|no healthy upstream|upstream request timeout|insufficient credits|requires more credits|credit balance is too low|payment required|purchase more credits|finish_reason:\s*content_filter|content[_ ]filter|content[_ ]management[_ ]policy/i.test(msg);
 }
 /** Build the structured, forwardable error from a thrown model-provider error. */
 function toModelProviderError(err) {
@@ -947,7 +1002,10 @@ async function extractContent$1(content, log) {
 				data: src.data
 			});
 			else if (src?.type === "url" && src.url) try {
-				images.push(await fetchImageAsBase64({ url: src.url }));
+				images.push(await fetchImageAsBase64({
+					url: src.url,
+					userAgent: null
+				}));
 			} catch (err) {
 				log.warn({
 					event: "image_fetch_failed",
@@ -1036,7 +1094,7 @@ async function handleMessages(request, ctx, registry) {
 	let parsed;
 	try {
 		parsed = JSON.parse(await request.text());
-	} catch {
+	} catch (_err) {
 		return jsonError(400, "invalid json");
 	}
 	const { messages, stream = false, model: modelInput, system, x_model: modelSpecInput } = parsed ?? {};
@@ -1059,7 +1117,7 @@ async function handleMessages(request, ctx, registry) {
 	const baseSystemPrompt = extractSystemPrompt(system);
 	const { sessionId } = parseSessionId(request);
 	const shellEnv = parseShellEnv(request);
-	const { session, sessionManager: sm } = await ctx.createSession({
+	const { session } = await ctx.createSession({
 		cwd: ctx.cwd,
 		sessionId,
 		perRequestApiKeys: withSpecApiKey(resolvePerRequestApiKeys(request), modelSpec),
@@ -1082,7 +1140,6 @@ async function handleMessages(request, ctx, registry) {
 	if (stream) {
 		const response = runStream$2({
 			session,
-			sm,
 			prompt,
 			images,
 			id,
@@ -1098,7 +1155,6 @@ async function handleMessages(request, ctx, registry) {
 	}
 	const response = await runBlocking$2({
 		session,
-		sm,
 		prompt,
 		images,
 		id,
@@ -1112,7 +1168,7 @@ async function handleMessages(request, ctx, registry) {
 	if (traceId) response.headers.set("x-trace-id", traceId);
 	return response;
 }
-function runStream$2({ session, sm, prompt, images, id, sessionId, modelName, log, postPrompt, registry }) {
+function runStream$2({ session, prompt, images, id, sessionId, modelName, log, postPrompt, registry }) {
 	return createSSEResponse(async (writer) => {
 		let contentBlockIndex = 0;
 		let textBlockOpen = false;
@@ -1158,12 +1214,14 @@ function runStream$2({ session, sm, prompt, images, id, sessionId, modelName, lo
 			textBlockOpen = false;
 		};
 		let capturedModelError = null;
+		const autoRetry = createAutoRetryObserver(log);
 		const emitProviderError = (providerError) => {
 			log.warn({
 				event: "model_provider_error",
 				code: providerError.code,
 				upstream_status: providerError.upstreamStatus,
-				provider: providerError.provider
+				provider: providerError.provider,
+				retry_attempts: autoRetry.attempts()
 			}, "forwarding model-provider error to client");
 			sseEvent(writer, "error", {
 				type: "error",
@@ -1173,13 +1231,15 @@ function runStream$2({ session, sm, prompt, images, id, sessionId, modelName, lo
 					x_model_provider_error: {
 						code: providerError.code,
 						provider: providerError.provider,
-						upstream_status: providerError.upstreamStatus
+						upstream_status: providerError.upstreamStatus,
+						retry_attempts: autoRetry.attempts()
 					}
 				}
 			});
 		};
 		session.subscribe((event) => {
 			const ev = event;
+			autoRetry.observe(ev);
 			const endedMessage = ev.message ?? (Array.isArray(ev.messages) ? ev.messages[ev.messages.length - 1] : null);
 			if (endedMessage?.stopReason === "error" && typeof endedMessage.errorMessage === "string" && endedMessage.errorMessage.length > 0) capturedModelError = endedMessage.errorMessage;
 			if (ev.type !== "message_update") return;
@@ -1266,7 +1326,7 @@ function runStream$2({ session, sm, prompt, images, id, sessionId, modelName, lo
 		sseEvent(writer, "message_stop", { type: "message_stop" });
 	});
 }
-async function runBlocking$2({ session, sm, prompt, images, id, sessionId, modelName, log, postPrompt, registry }) {
+async function runBlocking$2({ session, prompt, images, id, sessionId, modelName, log, postPrompt, registry }) {
 	let text = "";
 	const toolUseBlocks = [];
 	const tcByContentIdx = /* @__PURE__ */ new Map();
@@ -1317,7 +1377,7 @@ async function runBlocking$2({ session, sm, prompt, images, id, sessionId, model
 	}
 	for (const block of toolUseBlocks) if (block.type === "tool_use" && typeof block.input === "string") try {
 		block.input = JSON.parse(block.input);
-	} catch {
+	} catch (_err) {
 		block.input = {};
 	}
 	const content = [];
@@ -1379,6 +1439,87 @@ function create$2(options) {
 	};
 }
 //#endregion
+//#region src/billing-blocked-provider-signal.ts
+const BILLING_BLOCKED_PROVIDER_SIGNAL_PREFIX = "ANYONE_BILLING_BLOCKED_V1:";
+/**
+* Duplicated from the billing gate's `BlockReason` (apps/anyone/billing
+* gate/block-reason.ts): pi-server runs inside the sandbox and must not
+* depend on platform packages. A reason the proxy sends that predates this
+* build fails the enum and degrades to the untyped `billing_blocked`
+* handling — never a crash. Exported so agent-lifecycle's chat.test.ts can
+* pin this copy against its own (which billing's block-reason.test.ts in
+* turn pins against the gate), keeping every copy drift-checked.
+*/
+const BILLING_BLOCK_REASONS = [
+	"insufficient_balance",
+	"payment_failed",
+	"hard_spend_limit_reached"
+];
+const payloadSchema = z.object({
+	blockReason: z.enum(BILLING_BLOCK_REASONS),
+	message: z.string()
+}).strict();
+const nestedProxyEnvelopeSchema = z.object({
+	error: z.object({
+		type: z.literal("billing_blocked"),
+		code: z.literal("billing_blocked"),
+		message: z.string()
+	}).strict(),
+	blockReason: z.enum(BILLING_BLOCK_REASONS),
+	message: z.string()
+}).strict();
+/**
+* pi-ai's openai-compatible providers (openrouter, xai, groq, deepseek, …) can
+* NOT fold a proxy's non-2xx body into `error.message`, so `formatProviderError`
+* (@earendil-works/pi-ai utils/error-body) composes the display string as
+* `"<status>: <body>"` or, with a provider label, `"<prefix> (<status>): <body>"`.
+* The billing gate's 402 body therefore reaches us wrapped, e.g.
+* `"402: {\"error\":{...},\"blockReason\":\"insufficient_balance\",\"message\":...}"`
+* or `"OpenRouter (402): {...}"`. Neither the bare `JSON.parse` nor the `"402 "`
+* (space) strip below recognizes that, so a real billing block from an
+* openrouter-routed model degrades to the generic model-provider error. Peel a
+* single leading `"<status>: "` / `"<prefix> (<status>): "` wrapper off the
+* front so the recovered body flows through the existing shape checks. Returns
+* the message unchanged when no wrapper is present.
+*/
+function unwrapOpenAICompatStatusPrefix(message) {
+	const withPrefix = message.match(/^.+ \(\d{3}\): ([\s\S]+)$/);
+	if (withPrefix?.[1] !== void 0) return withPrefix[1];
+	const bare = message.match(/^\d{3}: ([\s\S]+)$/);
+	if (bare?.[1] !== void 0) return bare[1];
+	return message;
+}
+function parsePrefixedSignal(message) {
+	const normalized = message.startsWith("402 ") ? message.slice(4) : message;
+	if (!normalized.startsWith(BILLING_BLOCKED_PROVIDER_SIGNAL_PREFIX)) return null;
+	try {
+		const parsed = payloadSchema.safeParse(JSON.parse(normalized.slice(26)));
+		return parsed.success ? parsed.data : null;
+	} catch (_err) {
+		return null;
+	}
+}
+/**
+* Decode only the versioned platform envelope. Anthropic and OpenAI prefix the
+* nested error message with HTTP 402. Google's SDK instead preserves the full
+* proxy response as JSON, so that outer shape is validated separately.
+*/
+function parseBillingBlockedProviderSignal(message) {
+	const unwrapped = unwrapOpenAICompatStatusPrefix(message);
+	const direct = parsePrefixedSignal(unwrapped);
+	if (direct) return direct;
+	try {
+		const envelope = nestedProxyEnvelopeSchema.safeParse(JSON.parse(unwrapped));
+		if (!envelope.success) return null;
+		const nested = parsePrefixedSignal(envelope.data.error.message);
+		const topLevel = parsePrefixedSignal(envelope.data.message);
+		if (!nested || nested.blockReason !== envelope.data.blockReason || envelope.data.message !== nested.message && (!topLevel || topLevel.blockReason !== nested.blockReason || topLevel.message !== nested.message)) return null;
+		return nested;
+	} catch (_err) {
+		return null;
+	}
+}
+//#endregion
 //#region src/protocols/chat-completions.ts
 /**
 * OpenAI Chat Completions–compatible protocol handler.
@@ -1563,7 +1704,7 @@ async function loadHistoryIntoSession({ sm, messages, upToIdxExclusive, modelNam
 				let args = {};
 				try {
 					args = tc.function.arguments ? JSON.parse(tc.function.arguments) : {};
-				} catch {
+				} catch (_err) {
 					args = { _raw: tc.function.arguments };
 				}
 				contentArr.push({
@@ -1605,6 +1746,19 @@ async function loadHistoryIntoSession({ sm, messages, upToIdxExclusive, modelNam
 		}
 	}
 }
+/**
+* Each pi/session message carries the real wall-clock `timestamp` the
+* provider stamped when it produced that message (see @earendil-works/pi-ai
+* Message). Forward it verbatim as `x_created_at` (epoch ms) on the OAI
+* trailer message so the consumer (anyone-messaging completeRun) can stamp
+* each persisted segment at its true emit time and interleave a mid-run steer
+* by createdAt — instead of positionally reconstructing per-turn times, which
+* drifts on parallel tool calls and provider-split replies.
+*/
+function messageTimestampMs(m) {
+	const ts = m?.timestamp;
+	return typeof ts === "number" && Number.isFinite(ts) ? ts : null;
+}
 function piMessagesToOpenAI(messages) {
 	const out = [];
 	for (const m of messages) {
@@ -1633,6 +1787,8 @@ function piMessagesToOpenAI(messages) {
 			};
 			if (toolCalls.length) msg.tool_calls = toolCalls;
 			if (thinking.length) msg.x_thinking = thinking;
+			const createdAtMs = messageTimestampMs(m);
+			if (createdAtMs !== null) msg.x_created_at = createdAtMs;
 			out.push(msg);
 			continue;
 		}
@@ -1644,6 +1800,8 @@ function piMessagesToOpenAI(messages) {
 				content: text
 			};
 			if (m.isError) tm.x_is_error = true;
+			const createdAtMs = messageTimestampMs(m);
+			if (createdAtMs !== null) tm.x_created_at = createdAtMs;
 			out.push(tm);
 			continue;
 		}
@@ -1658,7 +1816,7 @@ async function handleChatCompletions(request, ctx, registry, pendingSteerIds) {
 	let parsed;
 	try {
 		parsed = JSON.parse(await request.text());
-	} catch {
+	} catch (_err) {
 		return jsonError(400, "invalid json");
 	}
 	const { messages, stream = false, model: modelInput, thinkingLevel: thinkingInput, context_window: contextWindowInput, x_model: modelSpecInput } = parsed ?? {};
@@ -1670,7 +1828,7 @@ async function handleChatCompletions(request, ctx, registry, pendingSteerIds) {
 	}
 	if (lastUserIdx < 0) return jsonError(400, "no user message in history");
 	const lastUser = messages[lastUserIdx];
-	const reqUA = request.headers.get("user-agent") ?? void 0;
+	const reqUA = request.headers.get("user-agent") ?? null;
 	const { text: prompt, images } = await extractContent(lastUser.content, {
 		userAgent: reqUA,
 		log
@@ -1701,7 +1859,7 @@ async function handleChatCompletions(request, ctx, registry, pendingSteerIds) {
 	const sessionId = parsedSession.source === "header" ? parsedSession.sessionId : `chatcmpl-${parsedSession.sessionId}`;
 	const shellEnv = parseShellEnv(request);
 	const sessionSetupStart = performance.now();
-	const [{ session, sessionManager: sm }, customTools] = await Promise.all([ctx.createSession({
+	const [{ session, sessionManager: sm }] = await Promise.all([ctx.createSession({
 		cwd: ctx.cwd,
 		sessionId,
 		perRequestApiKeys: withSpecApiKey(resolvePerRequestApiKeys(request), modelSpec),
@@ -1788,6 +1946,7 @@ function runStream$1({ session, sm, prompt, images, sessionId, created, reqStart
 		let usage = null;
 		let lastFollowUpCount = 0;
 		let capturedModelError = null;
+		const autoRetry = createAutoRetryObserver(log);
 		const sendChunk = (delta, finishReason) => {
 			if (firstChunkAt === null) {
 				firstChunkAt = performance.now();
@@ -1812,12 +1971,13 @@ function runStream$1({ session, sm, prompt, images, sessionId, created, reqStart
 		};
 		const ensureRole = () => {
 			if (roleEmitted) return;
-			sendChunk({ role: "assistant" });
+			sendChunk({ role: "assistant" }, null);
 			roleEmitted = true;
 		};
 		let _evCount = 0;
 		session.subscribe((event) => {
 			const ev = event;
+			autoRetry.observe(ev);
 			const endedMessage = ev.message ?? (Array.isArray(ev.messages) ? ev.messages[ev.messages.length - 1] : null);
 			if (endedMessage?.stopReason === "error" && typeof endedMessage.errorMessage === "string" && endedMessage.errorMessage.length > 0) capturedModelError = endedMessage.errorMessage;
 			if (ev.type === "queue_update") {
@@ -1835,7 +1995,7 @@ function runStream$1({ session, sm, prompt, images, sessionId, created, reqStart
 					steering,
 					follow_up: followUp,
 					...consumedIds.length > 0 ? { consumed_steer_ids: consumedIds } : {}
-				} });
+				} }, null);
 				return;
 			}
 			_evCount++;
@@ -1858,7 +2018,7 @@ function runStream$1({ session, sm, prompt, images, sessionId, created, reqStart
 				...ev.type === "message_start" || ev.type === "message_end" ? { messageJson: JSON.stringify(ev.message).slice(0, 500) } : {}
 			}, "harness session event");
 			if (ev.type === "tool_execution_start") {
-				sendChunk({ x_tool_execution_start: { tool_call_id: String(ev.toolCallId ?? "") } });
+				sendChunk({ x_tool_execution_start: { tool_call_id: String(ev.toolCallId ?? "") } }, null);
 				return;
 			}
 			if (ev.type === "tool_execution_end") {
@@ -1867,7 +2027,7 @@ function runStream$1({ session, sm, prompt, images, sessionId, created, reqStart
 					tool_call_id: String(ev.toolCallId ?? ""),
 					content: text,
 					...ev.isError ? { is_error: true } : {}
-				} });
+				} }, null);
 				return;
 			}
 			if (ev.type === "agent_end") {
@@ -1879,11 +2039,11 @@ function runStream$1({ session, sm, prompt, images, sessionId, created, reqStart
 			const t = inner?.type;
 			if (t === "text_delta") {
 				ensureRole();
-				sendChunk({ content: inner.delta ?? "" });
+				sendChunk({ content: inner.delta ?? "" }, null);
 				return;
 			}
 			if (t === "thinking_delta") {
-				sendChunk({ x_thinking_delta: inner.delta ?? "" });
+				sendChunk({ x_thinking_delta: inner.delta ?? "" }, null);
 				return;
 			}
 			if (t === "toolcall_start") {
@@ -1901,7 +2061,7 @@ function runStream$1({ session, sm, prompt, images, sessionId, created, reqStart
 						name: tc?.name ?? "",
 						arguments: ""
 					}
-				}] });
+				}] }, null);
 				return;
 			}
 			if (t === "toolcall_delta") {
@@ -1910,7 +2070,7 @@ function runStream$1({ session, sm, prompt, images, sessionId, created, reqStart
 				sendChunk({ tool_calls: [{
 					index: openaiIdx,
 					function: { arguments: inner.delta ?? "" }
-				}] });
+				}] }, null);
 			}
 		});
 		try {
@@ -1922,6 +2082,22 @@ function runStream$1({ session, sm, prompt, images, sessionId, created, reqStart
 				postPrompt
 			});
 			if (capturedModelError !== null) {
+				const billingBlocked = parseBillingBlockedProviderSignal(capturedModelError);
+				if (billingBlocked) {
+					log.warn({
+						event: "billing_blocked",
+						chatcmpl_id: sessionId,
+						source: "stop_reason_error"
+					}, "forwarding platform billing block to client");
+					emitBillingBlocked({
+						writer,
+						sessionId,
+						created,
+						billingBlocked
+					});
+					writer.write("data: [DONE]\n\n");
+					return;
+				}
 				const providerError = toModelProviderError(new Error(capturedModelError));
 				log.warn({
 					event: "model_provider_error",
@@ -1929,13 +2105,15 @@ function runStream$1({ session, sm, prompt, images, sessionId, created, reqStart
 					code: providerError.code,
 					upstream_status: providerError.upstreamStatus,
 					provider: providerError.provider,
+					retry_attempts: autoRetry.attempts(),
 					source: "stop_reason_error"
 				}, "forwarding model-provider error to client");
 				emitModelProviderError({
 					writer,
 					sessionId,
 					created,
-					providerError
+					providerError,
+					retryAttempts: autoRetry.attempts()
 				});
 				writer.write("data: [DONE]\n\n");
 			} else {
@@ -1957,19 +2135,32 @@ function runStream$1({ session, sm, prompt, images, sessionId, created, reqStart
 				err
 			}, "chat error");
 			if (isModelProviderError(err)) {
+				const billingBlocked = parseBillingBlockedProviderSignal(err instanceof Error ? err.message : String(err));
+				if (billingBlocked) {
+					emitBillingBlocked({
+						writer,
+						sessionId,
+						created,
+						billingBlocked
+					});
+					writer.write("data: [DONE]\n\n");
+					return;
+				}
 				const providerError = toModelProviderError(err);
 				log.warn({
 					event: "model_provider_error",
 					chatcmpl_id: sessionId,
 					code: providerError.code,
 					upstream_status: providerError.upstreamStatus,
-					provider: providerError.provider
+					provider: providerError.provider,
+					retry_attempts: autoRetry.attempts()
 				}, "forwarding model-provider error to client");
 				emitModelProviderError({
 					writer,
 					sessionId,
 					created,
-					providerError
+					providerError,
+					retryAttempts: autoRetry.attempts()
 				});
 			} else sendChunk({ content: `\n[error: ${err?.message ?? err}]` }, "stop");
 			writer.write("data: [DONE]\n\n");
@@ -2010,7 +2201,7 @@ function emitSessionMessagesTrailer({ writer, sm, baselineMessageCount, sessionI
 * cleanly closes the stream for any OpenAI-shaped reader that ignores the
 * extension field.
 */
-function emitModelProviderError({ writer, sessionId, created, providerError }) {
+function emitModelProviderError({ writer, sessionId, created, providerError, retryAttempts }) {
 	const chunk = {
 		id: sessionId,
 		object: "chat.completion.chunk",
@@ -2025,7 +2216,25 @@ function emitModelProviderError({ writer, sessionId, created, providerError }) {
 			message: providerError.message,
 			provider: providerError.provider,
 			type: providerError.type,
-			upstream_status: providerError.upstreamStatus
+			upstream_status: providerError.upstreamStatus,
+			retry_attempts: retryAttempts
+		}
+	};
+	writer.write(`data: ${JSON.stringify(chunk)}\n\n`);
+}
+function emitBillingBlocked({ writer, sessionId, created, billingBlocked }) {
+	const chunk = {
+		id: sessionId,
+		object: "chat.completion.chunk",
+		created,
+		choices: [{
+			index: 0,
+			delta: {},
+			finish_reason: "stop"
+		}],
+		x_billing_blocked: {
+			block_reason: billingBlocked.blockReason,
+			message: billingBlocked.message
 		}
 	};
 	writer.write(`data: ${JSON.stringify(chunk)}\n\n`);
@@ -2127,7 +2336,6 @@ const steerRequestSchema = z.object({
 	mode: z.enum(["steer", "follow-up"]).default("steer"),
 	text: z.string().optional()
 }).refine((v) => v.content !== void 0 || v.text !== void 0, { message: "content or text is required" });
-z.string().uuid();
 async function handleSteer(request, sessionId, registry, pendingSteerIds) {
 	const { log } = requestLogger(request);
 	const session = registry.get(sessionId);
@@ -2135,14 +2343,14 @@ async function handleSteer(request, sessionId, registry, pendingSteerIds) {
 	let body;
 	try {
 		body = JSON.parse(await request.text());
-	} catch {
+	} catch (_err) {
 		return jsonError(400, "invalid json");
 	}
 	const parsed = steerRequestSchema.safeParse(body);
 	if (!parsed.success) return jsonError(400, parsed.error.message);
 	const { mode } = parsed.data;
 	const { text, images } = await extractContent(parsed.data.content ?? parsed.data.text ?? "", {
-		userAgent: request.headers.get("user-agent") ?? void 0,
+		userAgent: request.headers.get("user-agent") ?? null,
 		log
 	});
 	if (!text && images.length === 0) return jsonError(400, "steer has no content");
@@ -2295,7 +2503,10 @@ async function extractImagesFromContent(content, log) {
 	for (const part of content) if (part?.type === "input_image") {
 		const url = part.image_url ?? part.url;
 		if (typeof url === "string" && /^https?:\/\//i.test(url)) try {
-			images.push(await fetchImageAsBase64({ url }));
+			images.push(await fetchImageAsBase64({
+				url,
+				userAgent: null
+			}));
 		} catch (err) {
 			log.warn({
 				event: "image_fetch_failed",
@@ -2316,7 +2527,7 @@ async function handleResponses(request, ctx, registry) {
 	let parsed;
 	try {
 		parsed = JSON.parse(await request.text());
-	} catch {
+	} catch (_err) {
 		return jsonError(400, "invalid json");
 	}
 	const { input, stream = false, model: modelInput, instructions, x_model: modelSpecInput } = parsed ?? {};
@@ -2388,6 +2599,7 @@ function runStream({ session, prompt, images, responseId, sessionId, modelName,
 		const tcByContentIdx = /* @__PURE__ */ new Map();
 		let nextOutputIndex = 0;
 		let capturedModelError = null;
+		const autoRetry = createAutoRetryObserver(log);
 		const send = (event) => {
 			writer.write(`data: ${JSON.stringify(event)}\n\n`);
 		};
@@ -2396,7 +2608,8 @@ function runStream({ session, prompt, images, responseId, sessionId, modelName,
 				event: "model_provider_error",
 				code: providerError.code,
 				upstream_status: providerError.upstreamStatus,
-				provider: providerError.provider
+				provider: providerError.provider,
+				retry_attempts: autoRetry.attempts()
 			}, "forwarding model-provider error to client");
 			send({
 				type: "response.failed",
@@ -2410,7 +2623,8 @@ function runStream({ session, prompt, images, responseId, sessionId, modelName,
 						x_model_provider_error: {
 							provider: providerError.provider,
 							type: providerError.type,
-							upstream_status: providerError.upstreamStatus
+							upstream_status: providerError.upstreamStatus,
+							retry_attempts: autoRetry.attempts()
 						}
 					}
 				}
@@ -2446,6 +2660,7 @@ function runStream({ session, prompt, images, responseId, sessionId, modelName,
 		};
 		session.subscribe((event) => {
 			const ev = event;
+			autoRetry.observe(ev);
 			const endedMessage = ev.message ?? (Array.isArray(ev.messages) ? ev.messages[ev.messages.length - 1] : null);
 			if (endedMessage?.stopReason === "error" && typeof endedMessage.errorMessage === "string" && endedMessage.errorMessage.length > 0) capturedModelError = endedMessage.errorMessage;
 			if (ev.type !== "message_update") return;
@@ -2851,4 +3066,4 @@ function chainMiddleware(middlewares) {
 	};
 }
 //#endregion
-export { DEFAULT_MODEL, DEFAULT_PROVIDER, DEFAULT_THINKING_LEVEL, KNOWN_PI_PROVIDERS, PI_AGENT_DIR, VALID_THINKING_LEVELS, buildAgentCard, buildModelFromSpec, chainMiddleware, composeHandlers, createAgentExecutor, createMapSessionRegistry, createPrewarm, createProtocolHandlers, createProtocols, deriveTraceContext, getCurrentTraceparent, logger, modelSpecSchema, mountAt, newTraceContext, parseModelSpec, parseTraceId, requestHeaders, requestLogger, requestUrl, resolveRequestModel, runInTraceContext, setCurrentTraceparent, webHandlerToMiddleware, withSpecApiKey };
+export { BILLING_BLOCK_REASONS, DEFAULT_MODEL, DEFAULT_PROVIDER, DEFAULT_THINKING_LEVEL, KNOWN_PI_PROVIDERS, PI_AGENT_DIR, VALID_THINKING_LEVELS, buildAgentCard, buildModelFromSpec, chainMiddleware, composeHandlers, createAgentExecutor, createMapSessionRegistry, createPrewarm, createProtocolHandlers, createProtocols, deriveTraceContext, getCurrentTraceparent, logger, modelSpecSchema, mountAt, newTraceContext, parseModelSpec, parseTraceId, requestHeaders, requestLogger, requestUrl, resolveRequestModel, runInTraceContext, setCurrentTraceparent, webHandlerToMiddleware, withSpecApiKey };

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@skydiveai/pi-server",
-  "version": "0.1.0-beta.165",
+  "version": "0.1.0-beta.2448",
   "homepage": "https://skydive.com",
   "license": "MIT",
   "author": "Create, Inc.",
@@ -17,20 +17,13 @@
   },
   "publishConfig": {
     "access": "public",
-    "exports": {
-      ".": {
-        "types": "./dist/index.d.mts",
-        "default": "./dist/index.mjs"
-      }
-    },
     "registry": "https://registry.npmjs.org"
   },
   "scripts": {
     "build": "tsdown",
     "typecheck": "tsgo --noEmit",
     "test:unit": "vitest run --passWithNoTests",
-    "test:ci": "vitest run --coverage --coverage.reporter=lcovonly --reporter=default --reporter=github-actions --minWorkers=1 --maxWorkers=2 --passWithNoTests",
-    "publish:system-artifacts": "doppler run --preserve-env -- node ../../scripts/anyone/publish-system-artifact.mjs"
+    "test:ci": "vitest run --coverage --coverage.reporter=lcovonly --reporter=default --reporter=github-actions --minWorkers=1 --maxWorkers=2 --passWithNoTests"
   },
   "dependencies": {
     "@a2a-js/sdk": "^0.3.13",