@oh-my-pi/pi-agent-core 17.1.7 → 17.2.0

package/CHANGELOG.md

@@ -2,6 +2,12 @@
 
 ## [Unreleased]
 
+## [17.2.0] - 2026-07-30
+
+### Fixed
+
+- Provider-native compaction failures now surface their transport error instead of silently switching to generic summarization; streaming V2 still falls back to native V1 when available.
+
 ## [17.1.7] - 2026-07-27
 
 ### Changed

package/dist/types/compaction/compaction.d.ts

@@ -53,6 +53,8 @@ export interface CompactionSettings {
 /** Reserve applied when {@link CompactionSettings.reserveTokens} is unset. */
 export declare const DEFAULT_RESERVE_TOKENS = 16384;
 export declare const DEFAULT_COMPACTION_SETTINGS: CompactionSettings;
+/** Whether a compaction candidate preserves provider-native transport under the effective settings. */
+export declare function shouldUseProviderNativeCompaction(model: Model, settings: Pick<CompactionSettings, "remoteEnabled" | "remoteStreamingV2Enabled">): boolean;
 /**
  * Calculate total context tokens from usage.
  * Uses the native totalTokens field when available, falls back to computing from components.

package/dist/types/compaction/errors.d.ts

@@ -13,6 +13,18 @@ export declare class CompactionCancelledError extends Error {
     readonly name: "CompactionCancelledError";
     constructor(message?: string);
 }
+/**
+ * A provider-native compaction request failed after every native protocol
+ * available for the selected model was exhausted.
+ *
+ * The cause stays attached so AI error classification can still recognize
+ * authentication failures. Non-auth failures remain distinguishable from
+ * ordinary summarization errors and must not fall through to another provider.
+ */
+export declare class NativeCompactionError extends Error {
+    readonly name: "NativeCompactionError";
+    constructor(cause: unknown);
+}
 /**
  * Outcome of a compaction attempt, surfaced by `CommandController.executeCompaction`
  * so callers (e.g. the plan-mode approval flow) can distinguish a deliberate abort

package/package.json

@@ -1,7 +1,7 @@
 {
 	"type": "module",
 	"name": "@oh-my-pi/pi-agent-core",
-	"version": "17.1.7",
+	"version": "17.2.0",
 	"description": "General-purpose agent with transport abstraction, state management, and attachment support",
 	"homepage": "https://omp.sh",
 	"author": "Can Boluk",
@@ -35,12 +35,12 @@
 		"fmt": "biome format --write ."
 	},
 	"dependencies": {
-		"@oh-my-pi/pi-ai": "17.1.7",
-		"@oh-my-pi/pi-catalog": "17.1.7",
-		"@oh-my-pi/pi-natives": "17.1.7",
-		"@oh-my-pi/pi-utils": "17.1.7",
-		"@oh-my-pi/pi-wire": "17.1.7",
-		"@oh-my-pi/snapcompact": "17.1.7",
+		"@oh-my-pi/pi-ai": "17.2.0",
+		"@oh-my-pi/pi-catalog": "17.2.0",
+		"@oh-my-pi/pi-natives": "17.2.0",
+		"@oh-my-pi/pi-utils": "17.2.0",
+		"@oh-my-pi/pi-wire": "17.2.0",
+		"@oh-my-pi/snapcompact": "17.2.0",
 		"@opentelemetry/api": "^1.9.1"
 	},
 	"devDependencies": {

package/src/compaction/compaction-v2-streaming.ts

@@ -8,7 +8,7 @@
  */
 
 import type { Api, CodexCompactionContext, FetchImpl, Model, ProviderSessionState } from "@oh-my-pi/pi-ai";
-import { isTransientStatus, ProviderHttpError } from "@oh-my-pi/pi-ai/error";
+import * as AIError from "@oh-my-pi/pi-ai/error";
 import { applyCodexResponsesLiteShape } from "@oh-my-pi/pi-ai/providers/openai-codex/request-transformer";
 import {
 	createOpenAICodexCompactionRequestContext,
@@ -21,6 +21,7 @@ import {
 	parseAzureDeploymentNameMap,
 	resolveOpenAIRequestSetup,
 } from "@oh-my-pi/pi-ai/providers/openai-shared";
+import { captureOpenAIHttpError } from "@oh-my-pi/pi-ai/utils/openai-http";
 import {
 	CODEX_BASE_URL,
 	getCodexAccountId,
@@ -334,18 +335,19 @@ async function attemptCompactionV2Streaming(
 	});
 
 	if (!response.ok) {
-		const errorText = await response.text().catch(() => "");
+		const cause = await captureOpenAIHttpError(response);
 		logger.warn("V2 remote compaction failed", {
 			endpoint,
 			status: response.status,
 			statusText: response.statusText,
-			errorText,
+			errorText: cause.captured.bodyText ?? "",
 		});
-		throw new ProviderHttpError(
+		throw new AIError.ProviderHttpError(
 			`V2 remote compaction failed (${response.status} ${response.statusText})`,
 			response.status,
 			{
 				headers: response.headers,
+				cause,
 			},
 		);
 	}
@@ -560,6 +562,10 @@ function formatCompactionV2Failure(event: Record<string, unknown>, type: string)
 }
 
 function isRetryableCompactionError(error: Error): boolean {
+	// The gateway's synthetic auth_unavailable is an HTTP 503, but the
+	// captured response cause classifies it as auth. Let provider fallback run
+	// immediately instead of spending the transient retry budget.
+	if (AIError.is(AIError.classify(error), AIError.Flag.AuthFailed)) return false;
 	if (
 		error.name === "AbortError" ||
 		error.name === "TimeoutError" ||
@@ -567,8 +573,8 @@ function isRetryableCompactionError(error: Error): boolean {
 	) {
 		return true;
 	}
-	if (error instanceof ProviderHttpError) {
-		return isTransientStatus(error.status);
+	if (error instanceof AIError.ProviderHttpError) {
+		return AIError.isTransientStatus(error.status);
 	}
 	const message = error.message.toLowerCase();
 	return (

package/src/compaction/compaction.ts

@@ -22,7 +22,7 @@ import {
 	type Usage,
 	withAuth,
 } from "@oh-my-pi/pi-ai";
-import { ProviderHttpError } from "@oh-my-pi/pi-ai/error";
+import * as AIError from "@oh-my-pi/pi-ai/error";
 import { createOpenAICodexCompactionRequestContext } from "@oh-my-pi/pi-ai/providers/openai-codex-responses";
 import { convertTools } from "@oh-my-pi/pi-ai/providers/openai-responses";
 import { buildResponsesInput, resolveOpenAICompatPolicy } from "@oh-my-pi/pi-ai/providers/openai-shared";
@@ -43,6 +43,7 @@ import {
 	V2_RETAINED_MESSAGE_TOKEN_BUDGET,
 } from "./compaction-v2-streaming";
 import type { CompactionEntry, SessionEntry } from "./entries";
+import { NativeCompactionError } from "./errors";
 import { isEstimateCacheable, readEstimateCache, writeEstimateCache } from "./message-cache";
 import { type ConvertToLlm, createBranchSummaryMessage, createCustomMessage, defaultConvertToLlm } from "./messages";
 import {
@@ -202,6 +203,18 @@ export const DEFAULT_COMPACTION_SETTINGS: CompactionSettings = {
 	v2RetainedMessageBudget: V2_RETAINED_MESSAGE_TOKEN_BUDGET,
 };
 
+/** Whether a compaction candidate preserves provider-native transport under the effective settings. */
+export function shouldUseProviderNativeCompaction(
+	model: Model,
+	settings: Pick<CompactionSettings, "remoteEnabled" | "remoteStreamingV2Enabled">,
+): boolean {
+	if (settings.remoteEnabled === false) return false;
+	return (
+		shouldUseOpenAiRemoteCompaction(model) ||
+		(settings.remoteStreamingV2Enabled !== false && shouldUseCompactionV2Streaming(model))
+	);
+}
+
 // ============================================================================
 // Token calculation
 // ============================================================================
@@ -725,7 +738,9 @@ function resolveCompactionEffort(model: Model, level: ThinkingLevel | undefined)
  */
 function createSummarizationError(prefix: string, response: AssistantMessage): Error {
 	const text = `${prefix}: ${response.errorMessage || "Unknown error"}`;
-	return response.errorStatus === undefined ? new Error(text) : new ProviderHttpError(text, response.errorStatus);
+	return response.errorStatus === undefined
+		? new Error(text)
+		: new AIError.ProviderHttpError(text, response.errorStatus);
 }
 
 function shouldRetryHandoffWithAutoToolChoice(response: AssistantMessage): boolean {
@@ -1332,6 +1347,17 @@ function buildCompactionV2Reasoning(
 	return { effort: reasoning.wireEffort ?? reasoning.requestedEffort, summary: "auto" };
 }
 
+/**
+ * Keep any non-auth native protocol failure ahead of authentication failures.
+ * Downstream may retry compaction with another provider only when every native
+ * protocol failed authentication, so a later auth error must not hide an
+ * earlier transport or protocol failure.
+ */
+function selectNativeCompactionError(previousError: unknown, nextError: unknown): unknown {
+	if (previousError === undefined) return nextError;
+	return AIError.is(AIError.classify(previousError), AIError.Flag.AuthFailed) ? nextError : previousError;
+}
+
 /**
  * Generate summaries for compaction using prepared data.
  * Returns CompactionResult - SessionManager adds id/parentId when saving.
@@ -1406,6 +1432,7 @@ export async function compact(
 		...recentMessages,
 	];
 	let usedRemoteCompaction = false;
+	let nativeCompactionError: unknown;
 	if (
 		settings.remoteEnabled !== false &&
 		settings.remoteStreamingV2Enabled !== false &&
@@ -1467,7 +1494,8 @@ export async function compact(
 				// swallowing it here would downgrade Esc into "fall back to local
 				// summarization" and keep compaction running on an aborted signal.
 				if (signal?.aborted) throw err;
-				logger.warn("OpenAI V2 remote compaction failed, falling back to V1/local summarization", {
+				nativeCompactionError = selectNativeCompactionError(nativeCompactionError, err);
+				logger.warn("OpenAI V2 remote compaction failed, falling back to V1 remote compaction", {
 					error: err instanceof Error ? err.message : String(err),
 					model: model.id,
 					provider: model.provider,
@@ -1517,7 +1545,8 @@ export async function compact(
 				// swallowing it here would downgrade Esc into "fall back to local
 				// summarization" and keep compaction running on an aborted signal.
 				if (signal?.aborted) throw err;
-				logger.warn("OpenAI remote compaction failed, falling back to local summarization", {
+				nativeCompactionError = selectNativeCompactionError(nativeCompactionError, err);
+				logger.warn("OpenAI remote compaction failed", {
 					error: err instanceof Error ? err.message : String(err),
 					model: model.id,
 					provider: model.provider,
@@ -1526,6 +1555,10 @@ export async function compact(
 		}
 	}
 
+	if (!usedRemoteCompaction && nativeCompactionError !== undefined && !summaryOptions.remoteEndpoint) {
+		throw new NativeCompactionError(nativeCompactionError);
+	}
+
 	// Generate summaries (can be parallel if both needed) and merge into one
 	let summary: string;
 

package/src/compaction/errors.ts

@@ -18,6 +18,22 @@ export class CompactionCancelledError extends Error {
 	}
 }
 
+/**
+ * A provider-native compaction request failed after every native protocol
+ * available for the selected model was exhausted.
+ *
+ * The cause stays attached so AI error classification can still recognize
+ * authentication failures. Non-auth failures remain distinguishable from
+ * ordinary summarization errors and must not fall through to another provider.
+ */
+export class NativeCompactionError extends Error {
+	readonly name = "NativeCompactionError" as const;
+
+	constructor(cause: unknown) {
+		super(cause instanceof Error ? cause.message : String(cause), { cause });
+	}
+}
+
 /**
  * Outcome of a compaction attempt, surfaced by `CommandController.executeCompaction`
  * so callers (e.g. the plan-mode approval flow) can distinguish a deliberate abort

package/src/compaction/openai.ts

@@ -38,6 +38,7 @@ import {
 	getOpenAIResponsesHistoryPayload,
 	normalizeResponsesToolCallId,
 } from "@oh-my-pi/pi-ai/utils";
+import { captureOpenAIHttpError } from "@oh-my-pi/pi-ai/utils/openai-http";
 import {
 	CODEX_BASE_URL,
 	getCodexAccountId,
@@ -840,18 +841,19 @@ export async function requestOpenAiRemoteCompaction(
 	});
 
 	if (!response.ok) {
-		const errorText = await response.text().catch(() => "");
+		const cause = await captureOpenAIHttpError(response);
 		logger.warn("OpenAI remote compaction failed", {
 			endpoint,
 			status: response.status,
 			statusText: response.statusText,
-			errorText,
+			errorText: cause.captured.bodyText ?? "",
 		});
 		throw new ProviderHttpError(
 			`Remote compaction failed (${response.status} ${response.statusText})`,
 			response.status,
 			{
 				headers: response.headers,
+				cause,
 			},
 		);
 	}

package/src/compaction/utils.ts

@@ -4,6 +4,7 @@
 
 import type { Message, ToolCall } from "@oh-my-pi/pi-ai";
 import { type Dialect, getDialectDefinition } from "@oh-my-pi/pi-ai/dialect";
+import { escapeHarmonyControlTokens } from "@oh-my-pi/pi-ai/utils/harmony-leak";
 import { formatGroupedPaths, prompt, stringifyJson } from "@oh-my-pi/pi-utils";
 import type { AgentMessage } from "../types";
 import fileOperationsTemplate from "./prompts/file-operations.md" with { type: "text" };
@@ -207,15 +208,13 @@ export function truncateToolResultForSummary(text: string): string {
 	return `${text.slice(0, TOOL_RESULT_MAX_CHARS)}\n\n[... ${truncatedChars} more characters truncated]`;
 }
 
-const HARMONY_CONTROL_TOKEN_RE = /<\|(start|end|message|channel|constrain|return|call)\|>/g;
-
 /**
  * Serialize LLM messages as plain summary input without provider control tokens.
  */
 export function serializeConversationForSummary(messages: Message[], dialect?: Dialect): string {
 	const conversation = serializeConversation(messages, dialect);
 	if (dialect !== "harmony") return conversation;
-	return conversation.replace(HARMONY_CONTROL_TOKEN_RE, "<\\|$1\\|>");
+	return escapeHarmonyControlTokens(conversation);
 }
 
 /**