@oh-my-pi/pi-ai 14.7.3 → 14.7.5

package/CHANGELOG.md

@@ -2,6 +2,16 @@
 
 ## [Unreleased]
 
+## [14.7.5] - 2026-05-07
+
+### Added
+
+- Added `OpenAICompat.supportsMultipleSystemMessages` so chat-completions hosts can opt out of separate leading system blocks. Auto-detected as `true` for OpenAI, Azure, OpenRouter, Cerebras, Together, Fireworks, Groq, DeepSeek, Mistral, xAI, Z.ai, GitHub Copilot, and Zenmux; `false` for MiniMax, Alibaba Dashscope, and Qwen Portal whose chat templates reject follow-up system messages. Unknown OpenAI-compatible hosts (custom vLLM/local) default to `false`; users can opt back in via `compat.supportsMultipleSystemMessages: true`.
+
+### Fixed
+
+- Fixed strict-template OpenAI-compatible hosts (e.g. Qwen 3.5+ via vLLM, MiniMax) rejecting follow-up `system`/`developer` messages by coalescing ordered system prompts into a single block joined by `\n\n` when `compat.supportsMultipleSystemMessages` is false. Canonical hosts continue to receive separate blocks so KV-cache reuse stays effective when only the trailing prompt changes ([#958](https://github.com/can1357/oh-my-pi/issues/958)).
+
 ## [14.7.2] - 2026-05-06
 
 ### Fixed

package/package.json

@@ -1,7 +1,7 @@
 {
 	"type": "module",
 	"name": "@oh-my-pi/pi-ai",
-	"version": "14.7.3",
+	"version": "14.7.5",
 	"description": "Unified LLM API with automatic model discovery and provider configuration",
 	"homepage": "https://github.com/can1357/oh-my-pi",
 	"author": "Can Boluk",
@@ -46,8 +46,8 @@
 		"@aws-sdk/credential-provider-node": "^3.972.39",
 		"@bufbuild/protobuf": "^2.12.0",
 		"@google/genai": "^1.52.0",
-		"@oh-my-pi/pi-natives": "14.7.3",
-		"@oh-my-pi/pi-utils": "14.7.3",
+		"@oh-my-pi/pi-natives": "14.7.5",
+		"@oh-my-pi/pi-utils": "14.7.5",
 		"@sinclair/typebox": "^0.34.49",
 		"@smithy/node-http-handler": "^4.6.1",
 		"ajv": "^8.20.0",

package/src/providers/openai-completions-compat.ts

@@ -99,6 +99,52 @@ export function detectOpenAICompat(model: Model<"openai-completions">, resolvedB
 	const isGrok = provider === "xai" || baseUrl.includes("api.x.ai");
 	const isMistral = provider === "mistral" || baseUrl.includes("mistral.ai");
 
+	// Hosts whose chat-completions endpoints are known to accept multiple
+	// leading `system`/`developer` messages (preferred for KV-cache reuse).
+	// Anything outside this allowlist defaults to coalescing because
+	// strict chat templates (Qwen 3.5+ via vLLM, MiniMax, etc.) reject
+	// follow-up system messages with a 400.
+	const isOpenAIHost = provider === "openai" || baseUrl.includes("api.openai.com");
+	const isAzureHost =
+		provider === "azure" ||
+		baseUrl.includes(".openai.azure.com") ||
+		baseUrl.includes("models.inference.ai.azure.com") ||
+		baseUrl.includes("azure.com/openai");
+	const isOpenRouter = provider === "openrouter" || baseUrl.includes("openrouter.ai");
+	const isTogether = provider === "together" || baseUrl.includes("api.together.xyz");
+	const isFireworks = baseUrl.includes("fireworks.ai");
+	const isGroqHost = provider === "groq" || baseUrl.includes("api.groq.com");
+	const isCopilotHost = provider === "github-copilot";
+	const isZenmuxHost = provider === "zenmux";
+	// Endpoints that MUST receive a single system block. MiniMax's OpenAI
+	// endpoint returns error 2013 on multiple system messages; Alibaba's
+	// Dashscope and Qwen Portal serve Qwen models whose chat template
+	// raises "System message must be at the beginning" if any system
+	// message appears past index 0.
+	const isMiniMaxHost =
+		provider === "minimax-code" ||
+		provider === "minimax-code-cn" ||
+		baseUrl.includes("api.minimax.io") ||
+		baseUrl.includes("api.minimaxi.com");
+	const isQwenPortal = provider === "qwen-portal" || baseUrl.includes("portal.qwen.ai");
+	const supportsMultipleSystemMessagesDefault =
+		!isMiniMaxHost &&
+		!isAlibaba &&
+		!isQwenPortal &&
+		(isOpenAIHost ||
+			isAzureHost ||
+			isOpenRouter ||
+			isCerebras ||
+			isTogether ||
+			isFireworks ||
+			isGroqHost ||
+			isDeepseekFamily ||
+			isMistral ||
+			isGrok ||
+			isZai ||
+			isCopilotHost ||
+			isZenmuxHost);
+
 	const reasoningEffortMap: NonNullable<OpenAICompat["reasoningEffortMap"]> =
 		provider === "groq" && model.id === "qwen/qwen3-32b"
 			? ({
@@ -115,6 +161,7 @@ export function detectOpenAICompat(model: Model<"openai-completions">, resolvedB
 	return {
 		supportsStore: !isNonStandard,
 		supportsDeveloperRole: !isNonStandard,
+		supportsMultipleSystemMessages: supportsMultipleSystemMessagesDefault,
 		supportsReasoningEffort: !isGrok && !isZai,
 		reasoningEffortMap,
 		supportsUsageInStreaming: !isCerebras,
@@ -175,6 +222,8 @@ export function resolveOpenAICompat(
 	return {
 		supportsStore: model.compat.supportsStore ?? detected.supportsStore,
 		supportsDeveloperRole: model.compat.supportsDeveloperRole ?? detected.supportsDeveloperRole,
+		supportsMultipleSystemMessages:
+			model.compat.supportsMultipleSystemMessages ?? detected.supportsMultipleSystemMessages,
 		supportsReasoningEffort: model.compat.supportsReasoningEffort ?? detected.supportsReasoningEffort,
 		reasoningEffortMap: model.compat.reasoningEffortMap ?? detected.reasoningEffortMap,
 		supportsUsageInStreaming: model.compat.supportsUsageInStreaming ?? detected.supportsUsageInStreaming,

package/src/providers/openai-completions.ts

@@ -1191,8 +1191,18 @@ export function convertMessages(
 	if (systemPrompts.length > 0) {
 		const useDeveloperRole = model.reasoning && compat.supportsDeveloperRole;
 		const role = useDeveloperRole ? "developer" : "system";
-		for (const systemPrompt of systemPrompts) {
-			params.push({ role, content: systemPrompt });
+		// Default to one block per ordered system prompt so the leading prefix
+		// stays byte-identical between turns and the provider's KV cache can
+		// reuse it. Hosts whose chat templates reject follow-up system messages
+		// (Qwen via vLLM, MiniMax, Alibaba Dashscope, Qwen Portal, …) opt out
+		// via `compat.supportsMultipleSystemMessages = false`; in that mode we
+		// coalesce into a single message joined by `\n\n`.
+		if (compat.supportsMultipleSystemMessages) {
+			for (const systemPrompt of systemPrompts) {
+				params.push({ role, content: systemPrompt });
+			}
+		} else {
+			params.push({ role, content: systemPrompts.join("\n\n") });
 		}
 	}
 

package/src/types.ts

@@ -540,6 +540,19 @@ export interface OpenAICompat {
 	supportsStore?: boolean;
 	/** Whether the provider supports the `developer` role (vs `system`). Default: auto-detected from URL. */
 	supportsDeveloperRole?: boolean;
+	/**
+	 * Whether the provider's chat-completions endpoint accepts multiple
+	 * leading `system`/`developer` messages. When false, ordered system
+	 * prompts are coalesced into a single message joined by `\n\n` so
+	 * strict chat templates (e.g. Qwen-served via vLLM, MiniMax) accept
+	 * the request. Default: detected per provider/baseUrl. Canonical
+	 * OpenAI/Azure/OpenRouter/Cerebras/Together/Fireworks/Groq/DeepSeek/
+	 * Mistral/xAI/Z.ai/GitHub Copilot/Zenmux are treated as `true`;
+	 * unknown or strict-template hosts default to `false`. Setting this
+	 * to `true` preserves separate blocks, which is preferred for
+	 * KV-cache reuse when the trailing prompt changes between calls.
+	 */
+	supportsMultipleSystemMessages?: boolean;
 	/** Whether the provider supports `reasoning_effort`. Default: auto-detected from URL. */
 	supportsReasoningEffort?: boolean;
 	/** Optional mapping from pi-ai reasoning levels to provider/model-specific `reasoning_effort` values. */