@linxiraos/pi-ai 1.1.13 → 1.1.15

package/src/error/flags.ts

@@ -156,6 +156,13 @@ const TIMEOUT_PATTERN = /\b(?:operation\s+)?timed?\s*out\b|\btimeout\b|\bstream
 const TRANSIENT_ENVELOPE_PATTERN = /anthropic stream envelope error:/i;
 const TRANSIENT_ENVELOPE_TRUNCATION_PATTERN = /before message_(?:start|stop)/i;
 export const STREAM_READ_ERROR_PATTERN = /stream[_ -]?read[_ -]?error/i;
+/** Python h2/httpx diagnostics forwarded through provider or proxy error events. */
+export const PYTHON_HTTP2_STREAM_RESET_PATTERN = /<StreamReset stream_id:\d+, error_code:(?:2|7), remote_reset:True>/;
+/** Python h11/httpx EOF while reading an HTTP/1.1 chunked response body. */
+export const PYTHON_HTTP_INCOMPLETE_CHUNK_PATTERN =
+	/peer closed connection without sending complete message body \(incomplete chunked read\)/;
+/** reqwest body-frame failures forwarded by the Codex HTTP proxy. */
+export const CODEX_HTTP_BODY_READ_ERROR_PATTERN = /\btransport error reading codex response body\b/i;
 export const TRANSIENT_TRANSPORT_PATTERN =
 	/\b(?:no[_ -]?capacity|(?:high|peak)[ _-]?demand|(?:at|over|insufficient)[ _-]?capacity|capacity[ _-]?(?:exceeded|exhausted)|peak[ _-]?load)\b|overloaded|provider.?returned.?error|rate.?limit|too many requests|auth-gateway\s+5\d{2}(?=[:\s]|$)|\b(?:429|500|502|503|504)\b|service.?unavailable|server.?error|internal.?error|retry your request|network.?error|connection.?error|connection.?refused|unable.?to.?connect\.\s*is the computer able to access the url\?|other side closed|fetch failed|upstream.?connect|upstream.?request.?failed|reset before headers|socket hang up|timed? out|timeout|terminated|retry delay|stream stall|no error details in response|HTTP2(?:StreamReset|RefusedStream|EnhanceYourCalm)|nghttp2_(?:internal_error|refused_stream)|stream closed with error code nghttp2_(?:internal_error|refused_stream)|malformed.?function.?call/i;
 const AUTH_FAILURE_PATTERN =
@@ -396,6 +403,8 @@ function isTransientErrorText(text: string): boolean {
 	return (
 		isUnexpectedSocketCloseMessage(text) ||
 		isStreamReadErrorText(text) ||
+		PYTHON_HTTP2_STREAM_RESET_PATTERN.test(text) ||
+		PYTHON_HTTP_INCOMPLETE_CHUNK_PATTERN.test(text) ||
 		(TRANSIENT_ENVELOPE_PATTERN.test(text) && TRANSIENT_ENVELOPE_TRUNCATION_PATTERN.test(text)) ||
 		TRANSIENT_TRANSPORT_PATTERN.test(text)
 	);
@@ -432,6 +441,17 @@ function matchesOverflowText(text: string): boolean {
 	return OVERFLOW_PATTERNS.some(p => p.test(text)) || OVERFLOW_NO_BODY_PATTERN.test(text);
 }
 
+/**
+ * A 4xx the provider rejected as a deterministic client error — every 4xx
+ * except 408 (Request Timeout) and 429 (Too Many Requests), the retryable
+ * pair. Mirrors the 4xx policy in {@link isProviderRetryableError}: such a
+ * request replays identically, so a transient signal riding on it (e.g. a
+ * truncation phrase in the body) must not flip it to retryable.
+ */
+function isTerminalClientErrorStatus(status: number | undefined): boolean {
+	return status !== undefined && status >= 400 && status < 500 && status !== 408 && status !== 429;
+}
+
 function classifyText(
 	errorMessage: string | undefined,
 	errorStatus: number | undefined,
@@ -483,6 +503,24 @@ function classifyText(
 		}
 		if (isTimeoutText(errorMessage)) kinds |= Flag.Transient | Flag.Timeout;
 		else if (isTransientErrorText(errorMessage)) kinds |= Flag.Transient;
+		// A stream truncation or forwarded Codex HTTP body-read failure may not
+		// match TRANSIENT_TRANSPORT_PATTERN. Flag it explicitly so AIError.retriable and
+		// the turn-recovery layer treat it as retryable, matching the provider
+		// retry path (isProviderRetryableError). Separate `if` (not chained onto
+		// the else-if) so a timeout whose text also reads as a truncation keeps
+		// Flag.Timeout alongside Flag.Transient. The string arm applies the strict
+		// STREAM_PARSE_DIAGNOSTIC_PATTERN, per the rationale on isTransientStreamParseError.
+		// Skip a truncation phrase that rides on a terminal 4xx (e.g. a malformed
+		// request rejected as "400 unexpected EOF"): that is a deterministic client
+		// error that replays identically, so keep it terminal. classify() carries
+		// the outer terminal status down the cause chain so a wrapped truncation
+		// (ProviderHttpError 400 → cause "unexpected EOF") is caught here too.
+		if (
+			!isTerminalClientErrorStatus(statusClean) &&
+			(isTransientStreamParseError(errorMessage) || CODEX_HTTP_BODY_READ_ERROR_PATTERN.test(errorMessage))
+		) {
+			kinds |= Flag.Transient;
+		}
 		// A concurrency cap (e.g. Vertex "Online prediction concurrent requests
 		// quota exceeded") is transient — shed-and-backoff. The bare wording need
 		// not match TRANSIENT_TRANSPORT_PATTERN, so flag it explicitly to keep
@@ -521,6 +559,11 @@ export function classify(error: unknown, api?: Api): number {
 	const seen = new Set<object>();
 	const causeTokenEvidence = hasCauseTokenContextOverflowEvidence(error);
 	let link: unknown = error;
+	// A terminal 4xx on an outer link governs its own cause diagnostics: a
+	// wrapped truncation is describing why the deterministic request failed,
+	// not an independently retryable transport fault. Carry it down so the
+	// stream-parse guard in classifyText sees it on the status-less cause.
+	let governingTerminalStatus: number | undefined;
 	while (link !== undefined && link !== null) {
 		if (typeof link === "object") {
 			if (seen.has(link)) break;
@@ -595,8 +638,10 @@ export function classify(error: unknown, api?: Api): number {
 			linkMessage = (link as { message: string }).message;
 		}
 
-		const textId = classifyText(linkMessage, status(link), causeTokenEvidence, api);
+		const linkStatus = status(link);
+		const textId = classifyText(linkMessage, linkStatus ?? governingTerminalStatus, causeTokenEvidence, api);
 		kinds |= textId & KIND_MASK;
+		if (isTerminalClientErrorStatus(linkStatus)) governingTerminalStatus = linkStatus;
 
 		link = typeof link === "object" && "cause" in link ? (link as { cause: unknown }).cause : undefined;
 	}

package/src/error/rate-limit.ts

@@ -30,6 +30,13 @@ const INSUFFICIENT_BALANCE_PATTERN = /insufficient.?balance/i;
 // "credits exhausted". Account-local, so rotate to a sibling credential.
 const CREDITS_EXHAUSTED_PATTERN =
 	/\b(?:exceed\w*|insufficient|not enough)\b[^\n]{0,40}\bcredits?\b|\bcredits?\b[^\n]{0,40}\b(?:exhausted|depleted)\b/i;
+// Anthropic subscription entitlement wall: "Usage credits are required for this
+// model" with `error_code: credits_required`. The account cannot serve the model
+// at all, so rotate to a sibling rather than backing off on this one. Bounded to
+// the documented sentence and the exact code: bare "usage credits" also appears
+// in unrelated diagnostics ("Failed to fetch usage credits from billing
+// service"), which must not rotate a healthy credential.
+const ANTHROPIC_CREDITS_REQUIRED_PATTERN = /\busage credits are required\b|\bcredits_required\b/i;
 const SPEND_LIMIT_PATTERN = /spend.?limit/i;
 const SUBSCRIPTION_CAP_PATTERN =
 	/\b(?:subscription|plan|membership)\b[^\n]{0,80}\b(?:rate.?limits?|quota|cap)\b|\b(?:rate.?limits?|quota|cap)\b[^\n]{0,80}\b(?:subscription|plan|membership)\b/i;
@@ -234,6 +241,10 @@ export function parseRateLimitReason(errorMessage: string): RateLimitReason {
 		return "QUOTA_EXHAUSTED";
 	}
 
+	if (ANTHROPIC_CREDITS_REQUIRED_PATTERN.test(errorMessage)) {
+		return "QUOTA_EXHAUSTED";
+	}
+
 	if (
 		lower.includes("per minute") ||
 		lower.includes("rate limit") ||
@@ -253,6 +264,7 @@ export function parseRateLimitReason(errorMessage: string): RateLimitReason {
 		lower.includes("out of credits") ||
 		lower.includes("spending-limit") ||
 		lower.includes("spending limit") ||
+		lower.includes("access_terminated_error") ||
 		INSUFFICIENT_BALANCE_PATTERN.test(errorMessage) ||
 		CREDITS_EXHAUSTED_PATTERN.test(errorMessage)
 	) {
@@ -294,7 +306,7 @@ export function calculateRateLimitBackoffMs(reason: RateLimitReason): number {
 
 /** Detect usage/quota limit errors in error messages (persistent, requires credential switch). */
 const USAGE_LIMIT_PATTERN =
-	/usage.?limit|usage_limit_reached|usage_not_included|limit_reached|quota.?(?:exceeded|reached|insufficient)|额度不足|额度耗尽|resource.?exhausted|exhausted your capacity|quota will reset|insufficient.?(?:balance|quota)|balance.?exhausted|run out of credits|out of credits|spending[- _]?limit|personal-team-blocked|clinepass limit|free limit reached on model/i;
+	/usage.?limit|usage_limit_reached|usage_not_included|limit_reached|quota.?(?:exceeded|reached|insufficient)|额度不足|额度耗尽|resource.?exhausted|exhausted your capacity|quota will reset|insufficient.?(?:balance|quota)|balance.?exhausted|run out of credits|out of credits|spending[- _]?limit|personal-team-blocked|clinepass limit|free limit reached on model|access_terminated_error/i;
 
 /**
  * HTTP status codes that, absent richer body classification, represent an
@@ -397,6 +409,7 @@ export function matchesUsageLimitText(errorMessage: string): boolean {
 	if (isDashScopeTokenLimitText(errorMessage)) return false;
 	return (
 		USAGE_LIMIT_PATTERN.test(errorMessage) ||
+		ANTHROPIC_CREDITS_REQUIRED_PATTERN.test(errorMessage) ||
 		CREDITS_EXHAUSTED_PATTERN.test(errorMessage) ||
 		(CN_QUOTA_EXHAUSTED_PATTERN.test(errorMessage) && !CN_TRANSIENT_CAP_PATTERN.test(errorMessage)) ||
 		SPEND_LIMIT_PATTERN.test(errorMessage) ||

package/src/error/retryable.ts

@@ -1,5 +1,6 @@
 import { isRetryableError, isUnexpectedSocketCloseMessage } from "@linxiraos/pi-utils";
 import {
+	CODEX_HTTP_BODY_READ_ERROR_PATTERN,
 	isRetryableStreamEnvelopeError,
 	isTransientStreamParseError,
 	isUsageLimit,
@@ -51,6 +52,7 @@ export function isProviderRetryableError(error: unknown): boolean {
 		isUnexpectedSocketCloseMessage(msg) ||
 		isTransientTransportMessage(msg) ||
 		TRANSIENT_TRANSPORT_PATTERN.test(msg) ||
+		CODEX_HTTP_BODY_READ_ERROR_PATTERN.test(msg) ||
 		PROVIDER_TRANSIENT_EXTRA_PATTERN.test(msg) ||
 		isTransientStreamParseError(error) ||
 		isRetryableStreamEnvelopeError(error)

package/src/providers/amazon-bedrock.ts

@@ -836,7 +836,7 @@ function handleMetadata(event: MetadataEvent, model: Model<"bedrock-converse-str
 		output.usage.cacheRead = event.usage.cacheReadInputTokens || 0;
 		output.usage.cacheWrite = event.usage.cacheWriteInputTokens || 0;
 		output.usage.totalTokens = event.usage.totalTokens || output.usage.input + output.usage.output;
-		calculateCost(model, output.usage);
+		calculateCost(model, output.usage, output.timestamp);
 	}
 }
 
@@ -924,6 +924,34 @@ function buildSystemPrompt(
 	return blocks;
 }
 
+function buildToolResultBlock(
+	message: ToolResultMessage,
+	model: Model<"bedrock-converse-stream">,
+	hoistedImages: ImageBlockWire[],
+): ToolResultBlockWire {
+	const content: Array<TextBlockWire | ImageBlockWire> = [];
+	for (const block of message.content) {
+		if (block.type === "image") {
+			const image: ImageBlockWire = { image: createImageBlock(block.mimeType, block.data) };
+			if (model.requiresToolResultImageHoisting) {
+				content.push({ text: "(see attached image)" });
+				hoistedImages.push(image);
+			} else {
+				content.push(image);
+			}
+		} else {
+			content.push({ text: block.text.toWellFormed() });
+		}
+	}
+	return {
+		toolResult: {
+			toolUseId: normalizeToolCallId(message.toolCallId),
+			content,
+			status: message.isError ? "error" : "success",
+		},
+	};
+}
+
 function convertMessages(
 	context: Context,
 	model: Model<"bedrock-converse-stream">,
@@ -1022,38 +1050,20 @@ function convertMessages(
 			case "toolResult": {
 				// Collect all consecutive toolResult messages into a single user message —
 				// Bedrock requires all tool results to be in one message.
-				const toolResults: ToolResultBlockWire[] = [];
-				toolResults.push({
-					toolResult: {
-						toolUseId: normalizeToolCallId(m.toolCallId),
-						content: m.content.map(c =>
-							c.type === "image"
-								? { image: createImageBlock(c.mimeType, c.data) }
-								: { text: c.text.toWellFormed() },
-						),
-						status: m.isError ? "error" : "success",
-					},
-				});
+				const contentBlocks: UserContent[] = [];
+				const hoistedImages: ImageBlockWire[] = [];
+				contentBlocks.push(buildToolResultBlock(m, model, hoistedImages));
 
 				let j = i + 1;
 				while (j < transformedMessages.length && transformedMessages[j].role === "toolResult") {
 					const nextMsg = transformedMessages[j] as ToolResultMessage;
-					toolResults.push({
-						toolResult: {
-							toolUseId: normalizeToolCallId(nextMsg.toolCallId),
-							content: nextMsg.content.map(c =>
-								c.type === "image"
-									? { image: createImageBlock(c.mimeType, c.data) }
-									: { text: c.text.toWellFormed() },
-							),
-							status: nextMsg.isError ? "error" : "success",
-						},
-					});
+					contentBlocks.push(buildToolResultBlock(nextMsg, model, hoistedImages));
 					j++;
 				}
 				i = j - 1;
 
-				result.push({ role: "user", content: toolResults });
+				contentBlocks.push(...hoistedImages);
+				result.push({ role: "user", content: contentBlocks });
 				break;
 			}
 			default:

package/src/providers/anthropic-wire.ts

@@ -168,6 +168,22 @@ export type FallbackBlockParam = {
 	to: { model: string };
 };
 
+/** Beta enabling server-side compaction (`compact_20260112` edit, `compaction` blocks). */
+export const COMPACTION_BETA = "compact-2026-01-12";
+
+/**
+ * Server-side compaction summary (compact-2026-01-12). Returned at the start
+ * of the assistant response that crossed the trigger; on replay the API drops
+ * every block that precedes it, so it may open the messages array. The
+ * `encrypted_content` is opaque provider state, round-tripped verbatim.
+ */
+export type CompactionBlockParam = {
+	type: "compaction";
+	content: string;
+	encrypted_content?: string | null;
+	cache_control?: CacheControlEphemeral | null;
+};
+
 export type ContentBlockParam =
 	| TextBlockParam
 	| ImageBlockParam
@@ -180,7 +196,8 @@ export type ContentBlockParam =
 	| ToolRemovalBlockParam
 	| ThinkingBlockParam
 	| RedactedThinkingBlockParam
-	| FallbackBlockParam;
+	| FallbackBlockParam
+	| CompactionBlockParam;
 
 /**
  * A single conversation turn.
@@ -276,9 +293,19 @@ export type FallbackParam = {
 	speed?: "fast";
 };
 
+/** Server-side compaction edit (compact-2026-01-12). */
+export type CompactionEdit = {
+	type: "compact_20260112";
+	/** `input_tokens` is the only trigger; `value` must be at least 50,000. */
+	trigger?: { type: "input_tokens"; value: number };
+	pause_after_compaction?: boolean;
+	/** Replaces the API's default summarization prompt entirely. */
+	instructions?: string;
+};
+
 /** Claude Code context-management beta payload. */
 export type ContextManagement = {
-	edits: Array<{ type: "clear_thinking_20251015"; keep: "all" }>;
+	edits: Array<{ type: "clear_thinking_20251015"; keep: "all" } | CompactionEdit>;
 };
 
 export type MessageCreateParams = {
@@ -322,7 +349,8 @@ export type StopReason =
 	| "pause_turn"
 	| "refusal"
 	| "sensitive"
-	| "model_context_window_exceeded";
+	| "model_context_window_exceeded"
+	| "compaction";
 
 export type CacheCreation = {
 	ephemeral_5m_input_tokens?: number | null;
@@ -336,12 +364,15 @@ export type ServerToolUsage = {
 
 /**
  * Per-attempt token accounting inside a multi-run turn
- * (server-side-fallback-2026-06-01). Populated whenever a fallback chain
- * ran, including sticky-served turns with no `fallback` content block.
- * A `fallback_message` entry is the definitive "served by fallback" signal.
+ * (server-side-fallback-2026-06-01, compact-2026-01-12). Populated whenever
+ * a fallback chain ran, including sticky-served turns with no `fallback`
+ * content block, and whenever the compaction beta is active. A
+ * `fallback_message` entry is the definitive "served by fallback" signal; a
+ * `compaction` entry is the summarization sampling the top-level usage
+ * excludes.
  */
 export type UsageIteration = {
-	type?: "message" | "fallback_message" | string;
+	type?: "message" | "fallback_message" | "compaction" | string;
 	model?: string | null;
 	input_tokens?: number | null;
 	output_tokens?: number | null;
@@ -401,13 +432,15 @@ export type ResponseContentBlock =
 	| ServerToolUseBlockParam
 	| WebSearchToolResultBlockParam
 	| ToolSearchToolResultBlockParam
-	| { type: "fallback"; from: { model: string }; to: { model: string } };
+	| { type: "fallback"; from: { model: string }; to: { model: string } }
+	| { type: "compaction"; content?: string | null; encrypted_content?: string | null };
 
 export type ContentBlockDelta =
 	| { type: "text_delta"; text: string }
 	| { type: "input_json_delta"; partial_json: string }
 	| { type: "thinking_delta"; thinking: string }
-	| { type: "signature_delta"; signature: string };
+	| { type: "signature_delta"; signature: string }
+	| { type: "compaction_delta"; content?: string | null; encrypted_content?: string | null };
 
 export type StopDetails = {
 	type: string;