@oh-my-pi/pi-ai 8.13.0 → 9.1.0

package/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@oh-my-pi/pi-ai",
-	"version": "8.13.0",
+	"version": "9.1.0",
 	"description": "Unified LLM API with automatic model discovery and provider configuration",
 	"type": "module",
 	"main": "./src/index.ts",
@@ -63,7 +63,7 @@
 		"@connectrpc/connect-node": "^2.1.1",
 		"@google/genai": "^1.38.0",
 		"@mistralai/mistralai": "^1.13.0",
-		"@oh-my-pi/pi-utils": "8.13.0",
+		"@oh-my-pi/pi-utils": "9.1.0",
 		"@sinclair/typebox": "^0.34.48",
 		"@smithy/node-http-handler": "^4.4.8",
 		"ajv": "^8.17.1",

package/src/providers/anthropic.ts

@@ -914,7 +914,7 @@ function convertTools(tools: Tool[], isOAuthToken: boolean): Anthropic.Messages.
 	});
 }
 
-function mapStopReason(reason: Anthropic.Messages.StopReason): StopReason {
+function mapStopReason(reason: Anthropic.Messages.StopReason | string): StopReason {
 	switch (reason) {
 		case "end_turn":
 			return "stop";
@@ -928,9 +928,10 @@ function mapStopReason(reason: Anthropic.Messages.StopReason): StopReason {
 			return "stop";
 		case "stop_sequence":
 			return "stop"; // We don't supply stop sequences, so this should never happen
-		default: {
-			const _exhaustive: never = reason;
-			throw new Error(`Unhandled stop reason: ${_exhaustive}`);
-		}
+		case "sensitive": // Content flagged by safety filters (not yet in SDK types)
+			return "error";
+		default:
+			// Handle unknown stop reasons gracefully (API may add new values)
+			throw new Error(`Unhandled stop reason: ${reason}`);
 	}
 }

package/src/providers/google-gemini-cli.ts

@@ -70,7 +70,7 @@ const GEMINI_CLI_HEADERS = {
 
 // Headers for Antigravity (sandbox endpoint) - requires specific User-Agent
 const ANTIGRAVITY_HEADERS = {
-	"User-Agent": "antigravity/1.11.5 darwin/arm64",
+	"User-Agent": "antigravity/1.15.8 darwin/arm64",
 	"X-Goog-Api-Client": "google-cloud-sdk vscode_cloudshelleditor/0.1",
 	"Client-Metadata": JSON.stringify({
 		ideType: "IDE_UNSPECIFIED",

package/src/providers/openai-completions.ts

@@ -789,6 +789,7 @@ function detectCompat(model: Model<"openai-completions">): ResolvedOpenAICompat
 		provider === "mistral" ||
 		baseUrl.includes("mistral.ai") ||
 		baseUrl.includes("chutes.ai") ||
+		baseUrl.includes("deepseek.com") ||
 		isZai ||
 		provider === "opencode" ||
 		baseUrl.includes("opencode.ai");

package/src/providers/openai-responses.ts

@@ -33,6 +33,22 @@ import { sanitizeSurrogates } from "../utils/sanitize-unicode";
 import { mapToOpenAIResponsesToolChoice } from "../utils/tool-choice";
 import { transformMessages } from "./transform-messages";
 
+/**
+ * Get prompt cache retention based on PI_CACHE_RETENTION env var.
+ * Only applies to direct OpenAI API calls (api.openai.com).
+ * Returns '24h' for long retention, undefined for default (in-memory).
+ */
+function getPromptCacheRetention(baseUrl: string): "24h" | undefined {
+	if (
+		typeof process !== "undefined" &&
+		process.env.PI_CACHE_RETENTION === "long" &&
+		baseUrl.includes("api.openai.com")
+	) {
+		return "24h";
+	}
+	return undefined;
+}
+
 // OpenAI Responses-specific options
 export interface OpenAIResponsesOptions extends StreamOptions {
 	reasoningEffort?: "minimal" | "low" | "medium" | "high" | "xhigh";
@@ -395,6 +411,7 @@ function buildParams(model: Model<"openai-responses">, context: Context, options
 		input: messages,
 		stream: true,
 		prompt_cache_key: options?.sessionId,
+		prompt_cache_retention: getPromptCacheRetention(model.baseUrl),
 	};
 
 	if (options?.maxTokens) {

package/src/providers/transform-messages.ts

@@ -6,13 +6,33 @@ import type { Api, AssistantMessage, Message, Model, ToolCall, ToolResultMessage
  * Anthropic APIs require IDs matching ^[a-zA-Z0-9_-]+$ (max 64 chars).
  */
 function normalizeToolCallId(id: string): string {
+	// Handle pipe-separated IDs from OpenAI Responses API
+	// Format: {call_id}|{item_id} where {item_id} can be 400+ chars with special chars (+, /, =)
+	// Extract just the call_id part and normalize it
+	if (id.includes("|")) {
+		const [callId] = id.split("|");
+		// Sanitize to allowed chars and truncate to 40 chars (OpenAI limit)
+		return callId.replace(/[^a-zA-Z0-9_-]/g, "_").slice(0, 40);
+	}
 	return id.replace(/[^a-zA-Z0-9_-]/g, "").slice(0, 40);
 }
 
 function normalizeResponsesToolCallId(id: string): string {
 	const [callId, itemId] = id.split("|");
 	if (callId && itemId) {
-		return id;
+		// Sanitize invalid characters and ensure proper format
+		const sanitizedCallId = callId.replace(/[^a-zA-Z0-9_-]/g, "_");
+		let sanitizedItemId = itemId.replace(/[^a-zA-Z0-9_-]/g, "_");
+		// OpenAI Responses API requires item id to start with "fc"
+		if (!sanitizedItemId.startsWith("fc")) {
+			sanitizedItemId = `fc_${sanitizedItemId}`;
+		}
+		// Truncate to 64 chars and strip trailing underscores (OpenAI Codex rejects them)
+		let normalizedCallId = sanitizedCallId.length > 64 ? sanitizedCallId.slice(0, 64) : sanitizedCallId;
+		let normalizedItemId = sanitizedItemId.length > 64 ? sanitizedItemId.slice(0, 64) : sanitizedItemId;
+		normalizedCallId = normalizedCallId.replace(/_+$/, "");
+		normalizedItemId = normalizedItemId.replace(/_+$/, "");
+		return `${normalizedCallId}|${normalizedItemId}`;
 	}
 	const hash = Bun.hash.xxHash64(id).toString(36);
 	return `call_${hash}|item_${hash}`;

package/src/utils/overflow.ts

@@ -17,13 +17,16 @@ import type { AssistantMessage } from "../types";
  * - llama.cpp: "the request exceeds the available context size, try increasing it"
  * - LM Studio: "tokens to keep from the initial prompt is greater than the context length"
  * - GitHub Copilot: "prompt token count of X exceeds the limit of Y"
- * - Cerebras: Returns "400 status code (no body)" - handled separately below
- * - Mistral: Returns "400 status code (no body)" - handled separately below
+ * - MiniMax: "invalid params, context window exceeds limit"
+ * - Kimi For Coding: "Your request exceeded model token limit: X (requested: Y)"
+ * - Cerebras: Returns "400/413 status code (no body)" - handled separately below
+ * - Mistral: Returns "400/413 status code (no body)" - handled separately below
  * - z.ai: Does NOT error, accepts overflow silently - handled via usage.input > contextWindow
  * - Ollama: Silently truncates input - not detectable via error message
  */
 const OVERFLOW_PATTERNS = [
 	/prompt is too long/i, // Anthropic
+	/input is too long for requested model/i, // Amazon Bedrock
 	/exceeds the context window/i, // OpenAI (Completions & Responses API)
 	/input token count.*exceeds the maximum/i, // Google (Gemini)
 	/maximum prompt length is \d+/i, // xAI (Grok)
@@ -32,6 +35,8 @@ const OVERFLOW_PATTERNS = [
 	/exceeds the limit of \d+/i, // GitHub Copilot
 	/exceeds the available context size/i, // llama.cpp server
 	/greater than the context length/i, // LM Studio
+	/context window exceeds limit/i, // MiniMax
+	/exceeded model token limit/i, // Kimi For Coding
 	/context[_ ]length[_ ]exceeded/i, // Generic fallback
 	/too many tokens/i, // Generic fallback
 	/token limit exceeded/i, // Generic fallback
@@ -54,11 +59,12 @@ const OVERFLOW_PATTERNS = [
  * - Google Gemini: "input token count exceeds the maximum"
  * - xAI (Grok): "maximum prompt length is X but request contains Y"
  * - Groq: "reduce the length of the messages"
- * - Cerebras: 400/413/429 status code (no body)
- * - Mistral: 400/413/429 status code (no body)
+ * - Cerebras: 400/413 status code (no body)
+ * - Mistral: 400/413 status code (no body)
  * - OpenRouter (all backends): "maximum context length is X tokens"
  * - llama.cpp: "exceeds the available context size"
  * - LM Studio: "greater than the context length"
+ * - Kimi For Coding: "exceeded model token limit: X (requested: Y)"
  *
  * **Unreliable detection:**
  * - z.ai: Sometimes accepts overflow silently (detectable via usage.input > contextWindow),
@@ -89,9 +95,9 @@ export function isContextOverflow(message: AssistantMessage, contextWindow?: num
 			return true;
 		}
 
-		// Cerebras and Mistral return 400/413/429 with no body - check for status code pattern
-		// 429 can indicate token-based rate limiting which correlates with context overflow
-		if (/^4(00|13|29)\s*(status code)?\s*\(no body\)/i.test(message.errorMessage)) {
+		// Cerebras and Mistral return 400/413 with no body for context overflow
+		// Note: 429 is rate limiting (requests/tokens per time), NOT context overflow
+		if (/^4(00|13)\s*(status code)?\s*\(no body\)/i.test(message.errorMessage)) {
 			return true;
 		}
 	}