@oh-my-pi/pi-ai 17.0.4 → 17.0.6

package/src/providers/anthropic.ts

@@ -100,6 +100,8 @@ export type AnthropicHeaderOptions = {
 	isCloudflareAiGateway?: boolean;
 	claudeCodeSessionId?: string;
 	claudeCodeBetas?: readonly string[];
+	/** Allow explicit fingerprint headers to replace OAuth defaults on non-official endpoints. */
+	allowAnthropicHeaderOverrides?: boolean;
 };
 
 export function normalizeAnthropicBaseUrl(baseUrl?: string): string | undefined {
@@ -144,7 +146,8 @@ const claudeCodeAgentBetaDefaults = [
 	midConversationSystemBeta,
 	"advanced-tool-use-2025-11-20",
 ] as const;
-const claudeCodeAgentPostEffortBetas = ["extended-cache-ttl-2025-04-11"] as const;
+const extendedCacheTtlBeta = "extended-cache-ttl-2025-04-11";
+const claudeCodeAgentPostEffortBetas = [extendedCacheTtlBeta] as const;
 const fineGrainedToolStreamingBeta = "fine-grained-tool-streaming-2025-05-14";
 const interleavedThinkingBeta = "interleaved-thinking-2025-05-14";
 // Asks the API to redact thinking blocks from responses. Only sent when the
@@ -225,22 +228,36 @@ export function buildAnthropicHeaders(options: AnthropicHeaderOptions): Record<s
 	const acceptHeader = oauthToken ? "application/json" : stream ? "text/event-stream" : "application/json";
 	const isCloudflare = options.isCloudflareAiGateway ?? false;
 	const honorAuthorization = !oauthToken && !isCloudflare;
+	const allowAnthropicHeaderOverrides =
+		oauthToken &&
+		options.allowAnthropicHeaderOverrides === true &&
+		!isCloudflare &&
+		!isOfficialAnthropicApiUrl(options.baseUrl);
 	const honorApiKey = !isCloudflare;
 	const modelHeaders: Record<string, string> = {};
+	const anthropicHeaderOverrides: Record<string, string> = {};
 	const filteredEnforcedKeys: string[] = [];
-	for (const [key, value] of Object.entries(options.modelHeaders ?? {})) {
-		const lowerKey = key.toLowerCase();
-		if (enforcedHeaderKeys.has(lowerKey)) {
-			// user-agent is always re-applied explicitly. authorization / x-api-key
-			// are silently re-applied in honoring branches and dropped + logged
-			// where the branch enforces its own credential.
-			if (lowerKey === "user-agent") continue;
-			if (lowerKey === "authorization" && honorAuthorization) continue;
-			if (lowerKey === "x-api-key" && honorApiKey) continue;
-			filteredEnforcedKeys.push(key);
-			continue;
+	const headerSource = options.modelHeaders;
+	if (headerSource) {
+		for (const key in headerSource) {
+			const value = headerSource[key];
+			const lowerKey = key.toLowerCase();
+			if (enforcedHeaderKeys.has(lowerKey)) {
+				if (allowAnthropicHeaderOverrides && overridableAnthropicHeaderKeys.has(lowerKey)) {
+					anthropicHeaderOverrides[key] = value;
+					continue;
+				}
+				// user-agent is always re-applied explicitly. authorization / x-api-key
+				// are silently re-applied in honoring branches and dropped + logged
+				// where the branch enforces its own credential.
+				if (lowerKey === "user-agent") continue;
+				if (lowerKey === "authorization" && honorAuthorization) continue;
+				if (lowerKey === "x-api-key" && honorApiKey) continue;
+				filteredEnforcedKeys.push(key);
+				continue;
+			}
+			modelHeaders[key] = value;
 		}
-		modelHeaders[key] = value;
 	}
 	if (filteredEnforcedKeys.length > 0) {
 		// Caller/env-supplied values (options.headers, ANTHROPIC_CUSTOM_HEADERS)
@@ -266,7 +283,7 @@ export function buildAnthropicHeaders(options: AnthropicHeaderOptions): Record<s
 		const userAgent = isClaudeCodeClientUserAgent(incomingUserAgent)
 			? incomingUserAgent
 			: `claude-cli/${claudeCodeVersion} (external, local-agent, agent-sdk/${claudeAgentSdkVersion})`;
-		return {
+		const headers = {
 			...modelHeaders,
 			...claudeCodeHeaders,
 			Accept: acceptHeader,
@@ -278,6 +295,7 @@ export function buildAnthropicHeaders(options: AnthropicHeaderOptions): Record<s
 			"User-Agent": userAgent,
 			...(incomingApiKey ? { "X-Api-Key": incomingApiKey } : {}),
 		};
+		return allowAnthropicHeaderOverrides ? mergeHeaders(headers, anthropicHeaderOverrides) : headers;
 	} else if (!isOfficialAnthropicApiUrl(options.baseUrl)) {
 		return {
 			...modelHeaders,
@@ -424,7 +442,15 @@ function getCacheControl(
 	cacheRetention: CacheRetention | undefined,
 	isOAuthToken: boolean,
 ): { retention: CacheRetention; cacheControl?: AnthropicCacheControl } {
-	const retention = cacheRetention ?? (isOAuthToken ? "long" : resolveCacheRetention(undefined));
+	// OAuth mirrors Claude Code and always defaults to 1h retention. API-key
+	// requests also default to 1h where the endpoint supports it (canonical
+	// Anthropic API, `compat.supportsLongCacheRetention`): agent sessions
+	// routinely idle past 5 minutes waiting on background jobs, and a 5m
+	// breakpoint cold-misses the entire prefix on resume. PI_CACHE_RETENTION
+	// still overrides the API-key default in either direction.
+	const retention = isOAuthToken
+		? (cacheRetention ?? "long")
+		: resolveCacheRetention(cacheRetention, model.compat.supportsLongCacheRetention ? "long" : "short");
 	if (retention === "none") {
 		return { retention };
 	}
@@ -512,6 +538,10 @@ const enforcedHeaderKeys = new Set(
 	].map(key => key.toLowerCase()),
 );
 
+const overridableAnthropicHeaderKeys = new Set(
+	[...Object.keys(claudeCodeHeaders), "anthropic-beta", "User-Agent", "x-app"].map(key => key.toLowerCase()),
+);
+
 const CLAUDE_BILLING_HEADER_PREFIX = "x-anthropic-billing-header:";
 
 function createClaudeBillingHeader(firstUserMessageText: string): string {
@@ -1853,6 +1883,17 @@ const streamAnthropicOnce = (
 				) {
 					extraBetas.push(contextManagementBeta);
 				}
+				// `ttl: "1h"` requires the extended-cache-ttl beta on API-key
+				// requests. OAuth requests never add it here: agent requests
+				// already carry it in the Claude Code beta list, and utility
+				// requests must not deviate from CC's header fingerprint.
+				if (
+					!(options?.isOAuth ?? isAnthropicOAuthToken(apiKey)) &&
+					getCacheControl(model, options?.cacheRetention, false).cacheControl?.ttl === "1h" &&
+					!extraBetas.includes(extendedCacheTtlBeta)
+				) {
+					extraBetas.push(extendedCacheTtlBeta);
+				}
 				// Server-side fallback beta chain: opt-in via `options.fallbacks`.
 				// Nested overrides (`speed`, `output_config.effort`,
 				// `output_config.task_budget`) reuse the same top-level betas
@@ -2791,6 +2832,7 @@ export function buildAnthropicClientOptions(args: AnthropicClientOptionsArgs): A
 			dynamicHeaders,
 		),
 		isCloudflareAiGateway: model.provider === "cloudflare-ai-gateway",
+		allowAnthropicHeaderOverrides: model.compat.allowAnthropicHeaderOverrides,
 		claudeCodeSessionId,
 		claudeCodeBetas: oauthToken
 			? buildClaudeCodeBetas(

package/src/providers/cursor.ts

@@ -2868,7 +2868,7 @@ function buildGrpcRequest(
 		conversationId: state.conversationId,
 	});
 
-	options?.onPayload?.(runRequest);
+	options?.onPayload?.(runRequest, model);
 
 	// Tools are sent later via requestContext (exec handshake)
 

package/src/providers/devin.ts

@@ -43,9 +43,11 @@ import type {
 	Tool,
 	ToolCall,
 } from "../types";
+import { isDemotedThinking } from "../utils/block-symbols";
 import { deterministicUuid } from "../utils/deterministic-id";
 import { AssistantMessageEventStream } from "../utils/event-stream";
 import { toolWireSchema } from "../utils/schema/wire";
+import { transformMessages } from "./transform-messages";
 
 /** Base host for Codeium/Windsurf's Cascade chat API (Connect protocol over HTTP/1.1). */
 export const DEVIN_API_URL = "https://server.codeium.com";
@@ -78,6 +80,13 @@ const CONNECT_END_STREAM_FLAG = 0x02;
  * fails fast instead of consuming memory.
  */
 const MAX_CONNECT_FRAME_PAYLOAD = 16 * 1024 * 1024;
+/**
+ * Recovery heuristic for opaque Devin `invalid_argument` trailers. This is not
+ * asserted to be the backend's hard limit: small requests can hit the same
+ * intermittent error, while compactable message history this large is likely
+ * to benefit from the existing context-overflow maintenance path.
+ */
+const LARGE_HISTORY_RECOVERY_BYTES = 512 * 1024;
 
 export const streamDevin: StreamFunction<"devin-agent"> = (
 	model: Model<"devin-agent">,
@@ -157,10 +166,14 @@ export const streamDevin: StreamFunction<"devin-agent"> = (
 			const auth = await fetchDevinAuthMetadata(apiKey, baseUrl, fetchImpl, options?.signal);
 			const chatBaseUrl = auth.baseUrl ?? baseUrl;
 			const request = buildDevinChatRequest(model, context, options, apiKey, auth.userJwt);
-			logger.debug("devin: sending chat request", { model: model.id, tools: context.tools?.length ?? 0 });
-
 			const reqBytes = toBinary(GetChatMessageRequestSchema, request);
 			const gz = gzipSync(reqBytes);
+			logger.debug("devin: sending chat request", {
+				model: model.id,
+				tools: context.tools?.length ?? 0,
+				requestBytes: reqBytes.byteLength,
+				compressedBytes: gz.byteLength,
+			});
 			const frame = Buffer.alloc(5 + gz.length);
 			frame[0] = CONNECT_COMPRESSED_FLAG;
 			frame.writeUInt32BE(gz.length, 1);
@@ -223,7 +236,53 @@ export const streamDevin: StreamFunction<"devin-agent"> = (
 					if (flag & CONNECT_END_STREAM_FLAG) {
 						const trailerBytes = flag & CONNECT_COMPRESSED_FLAG ? gunzipSync(payload) : payload;
 						const trailerError = readConnectTrailerError(trailerBytes.toString("utf8").trim());
-						if (trailerError) throw new AIError.ValidationError(trailerError);
+						if (trailerError) {
+							const error = new AIError.ValidationError(trailerError.formatted);
+							if (
+								firstTokenTime === undefined &&
+								trailerError.code.toLowerCase() === "invalid_argument" &&
+								/\binternal error\b/i.test(trailerError.message)
+							) {
+								// The full protobuf also contains the system prompt and tool
+								// schemas, which history maintenance cannot shrink. Re-encode
+								// only the repeated history field before choosing recovery.
+								let activeTailCount = 0;
+								const lastRole = context.messages.at(-1)?.role;
+								if (lastRole === "user" || lastRole === "developer") {
+									activeTailCount = 1;
+									// A trailing developer message can accompany the current user
+									// prompt. Earlier user-role records may instead be flushed
+									// execution history and must remain eligible for compaction.
+									if (lastRole === "developer") {
+										for (let i = context.messages.length - 2; i >= 0; i--) {
+											const role = context.messages[i].role;
+											if (role !== "user" && role !== "developer") break;
+											activeTailCount++;
+										}
+									}
+								}
+								const shrinkablePrompts =
+									activeTailCount > 0
+										? request.chatMessagePrompts.slice(0, -activeTailCount)
+										: request.chatMessagePrompts;
+								const historyBytes = toBinary(
+									GetChatMessageRequestSchema,
+									create(GetChatMessageRequestSchema, {
+										chatMessagePrompts: shrinkablePrompts,
+									}),
+								).byteLength;
+								if (historyBytes >= LARGE_HISTORY_RECOVERY_BYTES) {
+									AIError.attach(error, AIError.create(AIError.Flag.ContextOverflow));
+									logger.warn("devin: treating large-history invalid_argument as context overflow", {
+										model: model.id,
+										historyBytes,
+										requestBytes: reqBytes.byteLength,
+										compressedBytes: gz.byteLength,
+									});
+								}
+							}
+							throw error;
+						}
 						continue;
 					}
 
@@ -322,7 +381,8 @@ export const streamDevin: StreamFunction<"devin-agent"> = (
 						output.usage.output = Number(msg.usage.outputTokens);
 						output.usage.cacheRead = Number(msg.usage.cacheReadTokens);
 						output.usage.cacheWrite = Number(msg.usage.cacheWriteTokens);
-						output.usage.totalTokens = output.usage.input + output.usage.output;
+						output.usage.totalTokens =
+							output.usage.input + output.usage.output + output.usage.cacheRead + output.usage.cacheWrite;
 					}
 				}
 
@@ -442,6 +502,7 @@ function buildDevinChatRequest(
 		options?.stopSequences && options.stopSequences.length > 0
 			? [...DEVIN_DEFAULT_STOP_PATTERNS, ...options.stopSequences]
 			: DEVIN_DEFAULT_STOP_PATTERNS;
+	const messages = transformMessages(context.messages, model);
 	return create(GetChatMessageRequestSchema, {
 		metadata: create(MetadataSchema, {
 			apiKey,
@@ -453,7 +514,7 @@ function buildDevinChatRequest(
 			locale: "en",
 		}),
 		prompt: (context.systemPrompt ?? []).join("\n\n"),
-		chatMessagePrompts: buildChatMessagePrompts(context.messages, cascadeId),
+		chatMessagePrompts: buildChatMessagePrompts(messages, cascadeId, model),
 		chatModelUid: options?.chatModelUid ?? model.requestModelId ?? model.id,
 		requestType: ChatMessageRequestType.CASCADE,
 		plannerMode: ConversationalPlannerMode.DEFAULT,
@@ -485,7 +546,11 @@ function buildDevinChatRequest(
 }
 
 /** Map omp `Message` history onto Cascade `ChatMessagePrompt`s (USER / SYSTEM / TOOL channels). */
-function buildChatMessagePrompts(messages: Message[], cascadeId: string): ChatMessagePrompt[] {
+function buildChatMessagePrompts(
+	messages: Message[],
+	cascadeId: string,
+	model: Model<"devin-agent">,
+): ChatMessagePrompt[] {
 	const prompts: ChatMessagePrompt[] = [];
 	// messageId seeds are `cascadeId\0index\0role[...]` — prompt text is excluded
 	// so ids stay stable across content edits / history rebuilds.
@@ -513,16 +578,18 @@ function buildChatMessagePrompts(messages: Message[], cascadeId: string): ChatMe
 				}),
 			);
 		} else if (msg.role === "assistant") {
+			const isNativeDevinMessage =
+				msg.api === model.api && msg.provider === model.provider && msg.model === model.id;
 			let promptText = "";
 			let thinkingText = "";
 			let signature = "";
 			const toolCalls: ChatToolCall[] = [];
 			for (const part of msg.content) {
 				if (part.type === "text") {
-					promptText += part.text;
+					promptText += `${part.text}${isDemotedThinking(part) ? "\n" : ""}`;
 				} else if (part.type === "thinking") {
 					thinkingText += part.thinking;
-					if (!signature && part.thinkingSignature) signature = part.thinkingSignature;
+					if (isNativeDevinMessage && !signature && part.thinkingSignature) signature = part.thinkingSignature;
 				} else if (part.type === "toolCall") {
 					toolCalls.push(
 						create(ChatToolCallSchema, {
@@ -533,9 +600,13 @@ function buildChatMessagePrompts(messages: Message[], cascadeId: string): ChatMe
 					);
 				}
 			}
+			if (!promptText && !thinkingText && !signature && toolCalls.length === 0) continue;
 			prompts.push(
 				create(ChatMessagePromptSchema, {
-					messageId: msg.responseId ?? `bot-${deterministicUuid(`${cascadeId}\0${index}\0assistant`)}`,
+					messageId:
+						isNativeDevinMessage && msg.responseId
+							? msg.responseId
+							: `bot-${deterministicUuid(`${cascadeId}\0${index}\0assistant`)}`,
 					source: ChatMessageSource.SYSTEM,
 					prompt: promptText,
 					thinking: thinkingText,
@@ -569,12 +640,18 @@ function buildChatMessagePrompts(messages: Message[], cascadeId: string): ChatMe
 	return prompts;
 }
 
+interface ConnectTrailerError {
+	code: string;
+	message: string;
+	formatted: string;
+}
+
 /**
- * Parse a Connect end-of-stream JSON trailer and return a human-readable error
- * string when it carries `{ error: { code, message } }`, else `null`. The trailer
- * is untrusted server output, so the shape is checked with guards rather than asserted.
+ * Parse a Connect end-of-stream JSON trailer and return its structured error
+ * when it carries `{ error: { code, message } }`, else `null`. The trailer is
+ * untrusted server output, so the shape is checked with guards rather than asserted.
  */
-function readConnectTrailerError(text: string): string | null {
+function readConnectTrailerError(text: string): ConnectTrailerError | null {
 	if (text.length === 0) return null;
 	let parsed: unknown;
 	try {
@@ -588,5 +665,9 @@ function readConnectTrailerError(text: string): string | null {
 	const code = "code" in err && typeof err.code === "string" ? err.code : "";
 	const message = "message" in err && typeof err.message === "string" ? err.message : "";
 	if (!code && !message) return null;
-	return `Devin stream error${code ? ` ${code}` : ""}: ${message}`;
+	return {
+		code,
+		message,
+		formatted: `Devin stream error${code ? ` ${code}` : ""}: ${message}`,
+	};
 }

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

@@ -3913,7 +3913,8 @@ function redactHeaders(headers: Headers): Record<string, string> {
 	return redacted;
 }
 
-function resolveCodexResponsesUrl(baseUrl: string | undefined): string {
+/** Resolve a Codex Responses endpoint exactly as the chat and compaction transports do. */
+export function resolveCodexResponsesUrl(baseUrl: string | undefined): string {
 	const raw = baseUrl && baseUrl.trim().length > 0 ? baseUrl : CODEX_BASE_URL;
 	const normalized = raw.replace(/\/+$/, "");
 	if (normalized.endsWith("/codex/responses")) return normalized;

package/src/providers/openai-completions.ts

@@ -42,7 +42,13 @@ import {
 import { OpenAIHttpError, postOpenAIStream } from "../utils/openai-http";
 import { notifyProviderResponse } from "../utils/provider-response";
 import { callWithCopilotModelRetry } from "../utils/retry";
-import { adaptSchemaForStrict, NO_STRICT, normalizeSchemaForMoonshot, toolWireSchema } from "../utils/schema";
+import {
+	adaptSchemaForStrict,
+	NO_STRICT,
+	normalizeSchemaForMoonshot,
+	sanitizeSchemaForGrammar,
+	toolWireSchema,
+} from "../utils/schema";
 import {
 	type HealedToolCall,
 	StreamMarkupHealing,
@@ -671,7 +677,7 @@ const streamOpenAICompletionsOnce = (
 				}
 				activeReasoningEffortFallbackKey = reasoningEffortFallbackKey;
 				activeRequestParams = params;
-				options?.onPayload?.(params);
+				options?.onPayload?.(params, model);
 				rawRequestDump = {
 					provider: model.provider,
 					api: output.api,
@@ -1431,6 +1437,20 @@ function dropOpenRouterKimiForcedToolReasoning(
 	}
 }
 
+function hasActiveNativeKimiK3Reasoning(
+	model: Model<"openai-completions">,
+	options: OpenAICompletionsOptions | undefined,
+): boolean {
+	if (model.provider !== "kimi-code" || model.id.toLowerCase() !== "k3" || !model.reasoning) return false;
+	if (options?.reasoning === undefined || options.disableReasoning) return false;
+	try {
+		const url = new URL(model.baseUrl);
+		return url.hostname === "api.kimi.com" && (url.pathname === "/coding" || url.pathname.startsWith("/coding/"));
+	} catch {
+		return false;
+	}
+}
+
 function buildParams(
 	model: Model<"openai-completions">,
 	context: Context,
@@ -1518,6 +1538,20 @@ function buildParams(
 	) {
 		params.tool_choice = "required";
 	}
+	const forcedToolName =
+		typeof params.tool_choice === "object" && params.tool_choice !== null && "function" in params.tool_choice
+			? params.tool_choice.function.name
+			: undefined;
+	if (
+		forcedToolName !== undefined &&
+		Array.isArray(params.tools) &&
+		params.tools.some(tool => tool.type === "function" && tool.function.name === forcedToolName) &&
+		hasActiveNativeKimiK3Reasoning(model, options)
+	) {
+		// Native K3 reasoning is incompatible with selecting a specific function.
+		// Preserve the hard tool-use contract while letting K3 choose among tools.
+		params.tool_choice = "required";
+	}
 	if (isForcedToolChoice(params.tool_choice) && !initialCompat.supportsForcedToolChoice) {
 		// Some thinking-required OpenAI-compatible models reject forced
 		// `tool_choice` while still accepting tools with the default auto
@@ -1537,10 +1571,6 @@ function buildParams(
 		delete params.tool_choice;
 	}
 
-	const forcedToolName =
-		typeof params.tool_choice === "object" && params.tool_choice !== null && "function" in params.tool_choice
-			? params.tool_choice.function.name
-			: undefined;
 	if (
 		forcedToolName !== undefined &&
 		(!Array.isArray(params.tools) ||
@@ -2205,10 +2235,16 @@ function convertTools(
 					description: tool.description || "",
 					// Moonshot/Kimi native hosts validate against the stricter MFJS subset
 					// (const→enum, typed enums, no validators) and 400 otherwise.
+					// Grammar-constrained local backends (llama.cpp, LM Studio, vLLM)
+					// build a GBNF grammar from the schema and 400 with
+					// `Unrecognized schema: true` on the bare boolean subschema
+					// `toolWireSchema` emits for open fields (issue #5914).
 					parameters:
 						compat.toolSchemaFlavor === "moonshot-mfjs"
 							? (normalizeSchemaForMoonshot(wireParameters) as Record<string, unknown>)
-							: wireParameters,
+							: compat.toolSchemaFlavor === "grammar"
+								? sanitizeSchemaForGrammar(wireParameters)
+								: wireParameters,
 					// Only include strict if provider supports it. Some reject unknown fields.
 					...(includeStrict ? { strict: true } : includeExplicitFalse ? { strict: false } : {}),
 				},