@oh-my-pi/pi-ai 14.7.4 → 14.7.6

package/CHANGELOG.md

@@ -2,6 +2,25 @@
 
 ## [Unreleased]
 
+## [14.7.6] - 2026-05-07
+
+### Added
+
+- Added `hideThinkingSummary` option to `SimpleStreamOptions`. When true, `streamSimple` requests that the underlying provider omit reasoning/thinking summaries: Anthropic receives `thinking.display = "omitted"` (where supported), and OpenAI Responses / Azure / Codex providers leave `reasoning.summary` unset so the server skips emitting the human-readable summary stream entirely.
+
+### Changed
+
+- Changed OpenAI Responses, Azure OpenAI Responses, and OpenAI Codex providers to omit `reasoning.summary` from requests when `reasoningSummary` is explicitly `null` (previously fell back to `"auto"`).
+## [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": "14.7.4",
+	"version": "14.7.6",
 	"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": "14.7.4",
-		"@oh-my-pi/pi-utils": "14.7.4",
+		"@oh-my-pi/pi-natives": "14.7.6",
+		"@oh-my-pi/pi-utils": "14.7.6",
 		"@sinclair/typebox": "^0.34.49",
 		"@smithy/node-http-handler": "^4.6.1",
 		"ajv": "^8.20.0",

package/src/providers/azure-openai-responses.ts

@@ -317,11 +317,14 @@ function buildParams(
 		// See: https://github.com/can1357/oh-my-pi/issues/41
 		params.include = ["reasoning.encrypted_content"];
 
-		if (options?.reasoning || options?.reasoningSummary) {
-			params.reasoning = {
+		if (options?.reasoning || options?.reasoningSummary !== undefined) {
+			const reasoningParams: NonNullable<typeof params.reasoning> = {
 				effort: options?.reasoning || "medium",
-				summary: options?.reasoningSummary || "auto",
 			};
+			if (options?.reasoningSummary !== null) {
+				reasoningParams.summary = options?.reasoningSummary || "auto";
+			}
+			params.reasoning = reasoningParams;
 		} else {
 			if (model.name.toLowerCase().startsWith("gpt-5")) {
 				// Jesus Christ, see https://community.openai.com/t/need-reasoning-false-option-for-gpt-5/1351588/7

package/src/providers/openai-codex/request-transformer.ts

@@ -4,7 +4,7 @@ import type { Api, Model } from "../../types";
 
 export interface ReasoningConfig {
 	effort: "none" | "minimal" | "low" | "medium" | "high" | "xhigh";
-	summary: "auto" | "concise" | "detailed" | null;
+	summary?: "auto" | "concise" | "detailed";
 }
 
 export interface CodexRequestOptions {
@@ -52,11 +52,14 @@ export interface RequestBody {
 }
 
 function getReasoningConfig(model: Model<Api>, options: CodexRequestOptions): ReasoningConfig {
-	return {
+	const config: ReasoningConfig = {
 		effort:
 			options.reasoningEffort === "none" ? "none" : requireSupportedEffort(model, options.reasoningEffort as Effort),
-		summary: options.reasoningSummary ?? "detailed",
 	};
+	if (options.reasoningSummary !== null) {
+		config.summary = options.reasoningSummary ?? "detailed";
+	}
+	return config;
 }
 
 function filterInput(input: InputItem[] | undefined): InputItem[] | undefined {

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/providers/openai-responses.ts

@@ -430,13 +430,16 @@ function buildParams(
 		// See: https://github.com/can1357/oh-my-pi/issues/41
 		params.include = ["reasoning.encrypted_content"];
 
-		if (options?.reasoning || options?.reasoningSummary) {
-			params.reasoning = {
+		if (options?.reasoning || options?.reasoningSummary !== undefined) {
+			const reasoningParams: NonNullable<typeof params.reasoning> = {
 				effort: mapReasoningEffort(options?.reasoning || "medium", model.compat?.reasoningEffortMap) as NonNullable<
 					OpenAIResponsesSamplingParams["reasoning"]
 				>["effort"],
-				summary: options?.reasoningSummary || "auto",
 			};
+			if (options?.reasoningSummary !== null) {
+				reasoningParams.summary = options?.reasoningSummary || "auto";
+			}
+			params.reasoning = reasoningParams;
 		} else if (model.name.startsWith("gpt-5")) {
 			// Jesus Christ, see https://community.openai.com/t/need-reasoning-false-option-for-gpt-5/1351588/7
 			messages.push({

package/src/stream.ts

@@ -462,6 +462,7 @@ function mapOptionsForApi<TApi extends Api>(
 					...base,
 					thinkingEnabled: false,
 					toolChoice: mapAnthropicToolChoice(options?.toolChoice),
+					thinkingDisplay: options?.hideThinkingSummary ? "omitted" : undefined,
 				});
 			}
 
@@ -471,6 +472,7 @@ function mapOptionsForApi<TApi extends Api>(
 					...base,
 					thinkingEnabled: false,
 					toolChoice: mapAnthropicToolChoice(options?.toolChoice),
+					thinkingDisplay: options?.hideThinkingSummary ? "omitted" : undefined,
 				});
 			}
 
@@ -483,6 +485,7 @@ function mapOptionsForApi<TApi extends Api>(
 					thinkingEnabled: true,
 					effort,
 					toolChoice: mapAnthropicToolChoice(options?.toolChoice),
+					thinkingDisplay: options?.hideThinkingSummary ? "omitted" : undefined,
 				});
 			}
 
@@ -492,6 +495,7 @@ function mapOptionsForApi<TApi extends Api>(
 					thinkingEnabled: true,
 					thinkingBudgetTokens: thinkingBudget,
 					toolChoice: mapAnthropicToolChoice(options?.toolChoice),
+					thinkingDisplay: options?.hideThinkingSummary ? "omitted" : undefined,
 				});
 			}
 
@@ -509,6 +513,7 @@ function mapOptionsForApi<TApi extends Api>(
 					...base,
 					thinkingEnabled: false,
 					toolChoice: mapAnthropicToolChoice(options?.toolChoice),
+					thinkingDisplay: options?.hideThinkingSummary ? "omitted" : undefined,
 				});
 			} else {
 				return castApi<"anthropic-messages">({
@@ -517,6 +522,7 @@ function mapOptionsForApi<TApi extends Api>(
 					thinkingEnabled: true,
 					thinkingBudgetTokens: thinkingBudget,
 					toolChoice: mapAnthropicToolChoice(options?.toolChoice),
+					thinkingDisplay: options?.hideThinkingSummary ? "omitted" : undefined,
 				});
 			}
 		}
@@ -564,6 +570,7 @@ function mapOptionsForApi<TApi extends Api>(
 				reasoning: resolveOpenAiReasoningEffort(model, options),
 				toolChoice: mapOpenAiToolChoice(options?.toolChoice),
 				serviceTier: options?.serviceTier,
+				reasoningSummary: options?.hideThinkingSummary ? null : undefined,
 			});
 
 		case "azure-openai-responses":
@@ -572,6 +579,7 @@ function mapOptionsForApi<TApi extends Api>(
 				reasoning: resolveOpenAiReasoningEffort(model, options),
 				toolChoice: mapOpenAiToolChoice(options?.toolChoice),
 				serviceTier: options?.serviceTier,
+				reasoningSummary: options?.hideThinkingSummary ? null : undefined,
 			});
 
 		case "openai-codex-responses":
@@ -581,6 +589,7 @@ function mapOptionsForApi<TApi extends Api>(
 				toolChoice: mapOpenAiToolChoice(options?.toolChoice),
 				serviceTier: options?.serviceTier,
 				preferWebsockets: options?.preferWebsockets,
+				reasoningSummary: options?.hideThinkingSummary ? null : undefined,
 			});
 
 		case "google-generative-ai": {

package/src/types.ts

@@ -255,6 +255,14 @@ export interface SimpleStreamOptions extends StreamOptions {
 	 * this way when `reasoning` is undefined.
 	 */
 	disableReasoning?: boolean;
+	/**
+	 * If true, request that the provider omit thinking/reasoning summaries
+	 * from the response (e.g. Anthropic `thinking.display = "omitted"`,
+	 * OpenAI Responses `reasoning.summary` left unset). The model still
+	 * reasons internally; only the human-readable summary stream is dropped.
+	 * Useful when the UI hides thinking blocks anyway and the summary is wasted bandwidth.
+	 */
+	hideThinkingSummary?: boolean;
 	/** Custom token budgets for thinking levels (token-based providers only) */
 	thinkingBudgets?: ThinkingBudgets;
 	/** Cursor exec handlers for local tool execution */
@@ -540,6 +548,19 @@ 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. */