@oh-my-pi/pi-agent-core 18.12.6 → 18.13.1

package/CHANGELOG.md

@@ -2,6 +2,31 @@
 
 ## [Unreleased]
 
+## [18.13.1] - 2026-10-27
+
+### Added
+
+- Hosts can distinguish safely paused agents from turns still draining with `Agent.isPaused` ([#17073](https://github.com/can1357/oh-my-pi/pull/17073) by [@anatoli-tsinovoy](https://github.com/anatoli-tsinovoy)).
+
+## [18.13.0] - 2026-10-26
+
+### Added
+
+- Added `SummaryOptions.remotePromptCacheKey` and `remoteCompactionPolicyModel()` so remote compaction can carry a cache key gated against the request it actually sends ([#16988](https://github.com/can1357/oh-my-pi/pull/16988) by [@rafalopilowski1](https://github.com/rafalopilowski1))
+
+### Changed
+
+- Remote compaction now sends cache and session headers for the endpoint and model the request actually uses, so a configured compaction override no longer inherits the live turn's routing ([#16988](https://github.com/can1357/oh-my-pi/pull/16988) by [@rafalopilowski1](https://github.com/rafalopilowski1))
+- Aside providers (`AgentLoopConfig.getAsideMessages`, `Agent.setAsideMessageProvider`) now receive which boundary is polled, `"work"` after tool calls or `"stop"` at the end of a turn, so they can keep a message out of a finished turn ([#16583](https://github.com/can1357/oh-my-pi/pull/16583) by [@will-bogusz](https://github.com/will-bogusz))
+
+### Fixed
+
+- Fixed `--max-time` values above about 596 hours (such as `600h`) aborting the first turn with "Deadline exceeded" right away ([#17054](https://github.com/can1357/oh-my-pi/pull/17054) by [@trukhinyuri](https://github.com/trukhinyuri))
+- Fixed repeated compaction on small or provider-capped context windows: compaction keeps only as much recent history as fits under the configured trigger, and no longer counts stale pre-compaction usage ([#16904](https://github.com/can1357/oh-my-pi/pull/16904) by [@eggpeat](https://github.com/eggpeat))
+- Fixed native compaction token estimates counting opaque encrypted-state bytes as conversation text ([#16904](https://github.com/can1357/oh-my-pi/pull/16904) by [@eggpeat](https://github.com/eggpeat))
+- Fixed split-turn compaction leaving `session.compacting` context out of the turn-prefix summary ([#16707](https://github.com/can1357/oh-my-pi/pull/16707) by [@integrITsolutions](https://github.com/integrITsolutions))
+- Fixed compaction failing as a whole, and discarding the completed summary, when the optional short summary fails or returns no answer; the compaction now completes without a short summary, and cancellation still cancels it ([#16942](https://github.com/can1357/oh-my-pi/pull/16942) by [@Mouriya-Emma](https://github.com/Mouriya-Emma))
+
 ## [18.12.4] - 2026-10-20
 
 ### Fixed

package/dist/types/agent.d.ts

@@ -3,7 +3,7 @@ import type { Dialect } from "@oh-my-pi/pi-ai/dialect";
 import type { HarmonyAuditEvent } from "@oh-my-pi/pi-ai/utils/harmony-leak";
 import type { AppendOnlyContextManager } from "./append-only-context.js";
 import { Tokenizer } from "./tokenizer.js";
-import type { AgentBeforeModelCall, AgentContext, AgentEvent, AgentLoopConfig, AgentMessage, AgentState, AgentTool, AgentToolContext, AgentTurnEndContext, AsideMessage, PrepareQueuedMessages, SpeculativeToolExecutionConfig, StreamFn, ToolCallContext, ToolChoiceDirective } from "./types.js";
+import type { AgentBeforeModelCall, AgentContext, AgentEvent, AgentLoopConfig, AgentMessage, AgentState, AgentTool, AgentToolContext, AgentTurnEndContext, AsideBoundary, AsideMessage, PrepareQueuedMessages, SpeculativeToolExecutionConfig, StreamFn, ToolCallContext, ToolChoiceDirective } from "./types.js";
 export declare class AgentBusyError extends Error {
     constructor(message?: string);
 }
@@ -240,6 +240,8 @@ export interface AgentPromptOptions {
 }
 export declare class Agent {
     #private;
+    /** True while this unfinished turn is safely parked by the process pause gate. */
+    get isPaused(): boolean;
     streamFn: StreamFn;
     getApiKey?: (model: Model) => Promise<ApiKey | undefined> | ApiKey | undefined;
     /** Prepare actual queue deliveries after dequeue gates; commit runs only while ownership remains valid. */
@@ -469,7 +471,7 @@ export declare class Agent {
      * aborts foreground tools; an interruptible `wait` may end early through the
      * peek hooks. See `AgentLoopConfig.getAsideMessages`.
      */
-    setAsideMessageProvider(fn: (() => AsideMessage[] | Promise<AsideMessage[]>) | undefined): void;
+    setAsideMessageProvider(fn: ((boundary: AsideBoundary) => AsideMessage[] | Promise<AsideMessage[]>) | undefined): void;
     emitExternalEvent(event: AgentEvent): void;
     setSystemPrompt(v: string[] | string): void;
     setModel(model: Model): void;

package/dist/types/compaction/compaction-v2-streaming.d.ts

@@ -8,6 +8,7 @@
  */
 import type { Api, CodexCompactionContext, FetchImpl, Model, ProviderSessionState } from "@oh-my-pi/pi-ai";
 import { type OpenAICodexCompactionBody } from "@oh-my-pi/pi-ai/providers/openai-codex-responses";
+import type { CatalogPolicyInputs } from "@oh-my-pi/pi-catalog/types";
 /** Retained-message budget Codex uses after streamed V2 compaction. */
 export declare const V2_RETAINED_MESSAGE_TOKEN_BUDGET = 64000;
 /** Max retries for V2 streaming compaction on transient stream errors. */
@@ -46,6 +47,15 @@ export declare function getCompactionV2Endpoint(model: Model): string | undefine
 export declare function shouldUseCompactionV2Streaming(model: Model): model is Model<"openai-responses" | "azure-openai-responses" | "openai-codex-responses">;
 /** APIs that speak the OpenAI Responses compaction protocol (V1 `/responses/compact` and V2 streaming). */
 export declare function isOpenAiRemoteCompactionApi(api: Api | undefined): boolean;
+/**
+ * Catalog-policy view of a remote compaction request (V1 and V2): the
+ * effective adapter (`remoteCompaction.api`) and the wire id the body sends
+ * (`remoteCompaction.model`). Under a model override the active lane's
+ * `requestModelId` and thinking classification describe a model that never
+ * reaches the wire, so they are dropped; `reasoning` stays because the
+ * compaction body derives its reasoning config from the active model.
+ */
+export declare function remoteCompactionPolicyModel(model: Model): CatalogPolicyInputs;
 /** Clamp the retained-message budget to Codex's known-safe 64K ceiling. */
 export declare function resolveCompactionV2RetainedMessageBudget(value: number | undefined): number;
 /** Build a V2 streaming compaction request from Responses-native history. */

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

@@ -67,6 +67,18 @@ export interface CompactionSettings {
 }
 /** Reserve applied when {@link CompactionSettings.reserveTokens} is unset. */
 export declare const DEFAULT_RESERVE_TOKENS = 16384;
+/**
+ * Hysteresis band under the compaction trigger. Maintenance counts as having
+ * resolved threshold pressure only when residual context lands at or below
+ * `COMPACTION_RECOVERY_BAND × threshold`. Re-checking against the raw threshold
+ * lets a pass keep reclaiming a trickle of the previous turn's output and land
+ * just under the line every turn, sustaining the auto-continue dead loop
+ * reported in #2275; the same band stops the context-full / snapcompact tail
+ * from re-firing on a history whose single most-recent kept turn already
+ * exceeds the threshold. {@link prepareCompaction} sizes the kept tail to land
+ * inside it.
+ */
+export declare const COMPACTION_RECOVERY_BAND = 0.8;
 /**
  * Hard ceiling on a generated compaction summary.
  *
@@ -231,8 +243,15 @@ export interface SummaryOptions {
     thinkingLevel?: ThinkingLevel;
     /** Session routing key for remote compaction transports with sticky provider sessions. */
     sessionId?: string;
-    /** Prompt-cache key for remote compaction transports that support provider prefix caching. */
+    /** Prompt-cache key for conventional summary requests, gated against the summarizing model. */
     promptCacheKey?: string;
+    /**
+     * Prompt-cache key for remote compaction requests, gated against the request
+     * they send (`remoteCompaction.api`/`model` override, see
+     * `remoteCompactionPolicyModel`). Defaults to {@link promptCacheKey} only when
+     * omitted; an explicit `undefined` means the remote target gets no key.
+     */
+    remotePromptCacheKey?: string;
     /** Mutable provider state used to keep Codex compaction on the live session identity. */
     providerSessionState?: Map<string, ProviderSessionState>;
     /** Whether Codex remote compaction should prefer the provider WebSocket transport. */

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

@@ -30,8 +30,8 @@ export declare const REMOTE_COMPACTION_TIMEOUT_MS = 300000;
 export declare const CONTEXT_WINDOW_TRUNCATED_OUTPUT_MESSAGE: string;
 /**
  * Tokens of stored Responses replacement history. Retained `input_image` parts
- * count at their image estimate rather than as base64 text; everything else,
- * opaque compaction state included, counts as decoded text.
+ * count at their image estimate, not their base64 size. Opaque encrypted state
+ * stays excluded, matching ordinary message estimates and remote fit checks.
  */
 export declare function countResponsesHistoryTokens(items: unknown[], tokenizer: Tokenizer): number;
 export interface TrimRemoteCompactionInputResult {
@@ -87,6 +87,13 @@ export interface OpenAiRemoteCompactionRequest {
     model: string;
     input: Array<Record<string, unknown>>;
     instructions: string;
+    /**
+     * Responses body cache identity. Mirrors the V2 body: the explicit
+     * prompt-cache key when supplied (normalized to the 64-char wire limit),
+     * else the session-derived cache id — retention-gated through
+     * `getOpenAIPromptCacheKey`; absent when it resolves to nothing.
+     */
+    prompt_cache_key?: string;
     reasoning?: {
         context?: string;
         [key: string]: unknown;
@@ -124,6 +131,13 @@ export declare function requestOpenAiRemoteCompaction(model: Model, apiKey: stri
     fetch?: FetchImpl;
     timeoutMs?: number;
     sessionId?: string;
+    /**
+     * Explicit prompt-cache identity. Feeds the Go family affinity header
+     * (`x-opencode-session`) and the Responses body `prompt_cache_key`
+     * under the same KDL gate as the V2 request; an absent/empty key keeps
+     * the child conversation id.
+     */
+    promptCacheKey?: string;
     providerSessionState?: Map<string, ProviderSessionState>;
     codexCompaction?: CodexCompactionContext;
     /** Forwarded to {@link trimRemoteCompactionInputToContextWindow}. */

package/dist/types/speculative-execution.d.ts

@@ -23,6 +23,9 @@ export declare class SpeculativeOperationCoordinator {
     #private;
     readonly config: SpeculativeToolExecutionConfig;
     readonly environment?: CoordinatorEnvironment | undefined;
+    /** Observer of existing speculative execution draining; no separate work ledger. */
+    onIdle?: () => void;
+    get isRunning(): boolean;
     constructor(config: SpeculativeToolExecutionConfig, environment?: CoordinatorEnvironment | undefined);
     get maxInFlight(): number;
     get size(): number;

package/dist/types/types.d.ts

@@ -33,6 +33,8 @@ export type CommittableAsideMessage = AgentMessage & {
  * (e.g. dropping late diagnostics a newer edit superseded).
  */
 export type AsideMessage = CommittableAsideMessage | (() => CommittableAsideMessage | null);
+/** Where the loop polls asides: after a tool batch (`"work"`) or after a turn without tool calls (`"stop"`). */
+export type AsideBoundary = "work" | "stop";
 export interface AgentTurnEndContext {
     /** Assistant/user message that just completed this turn boundary. */
     message: AgentMessage;
@@ -124,6 +126,8 @@ export interface SteeringQueueState {
  */
 export interface AgentLoopConfig extends SimpleStreamOptions {
     model: Model;
+    /** Reports whether this loop has reached a safe process-pause boundary. */
+    onPauseChange?: (paused: boolean) => void;
     /**
      * When to interrupt tool execution for steering messages.
      * - "immediate" = cut interruptible waits short and raise the cooperative
@@ -308,8 +312,12 @@ export interface AgentLoopConfig extends SimpleStreamOptions {
      * should reach the model between requests without waiting for the agent to
      * fully stop. Returned messages are appended to the context with normal
      * message events and keep the loop running so the model can react.
+     *
+     * `boundary` is `"work"` after a tool batch and `"stop"` when the model ended
+     * its turn without tool calls; a provider may hold back messages that should
+     * not extend a finished turn.
      */
-    getAsideMessages?: () => Promise<AsideMessage[]>;
+    getAsideMessages?: (boundary: AsideBoundary) => Promise<AsideMessage[]>;
     /**
      * Hook fired right before the loop would exit.
      *

package/package.json

@@ -1,7 +1,7 @@
 {
 	"type": "module",
 	"name": "@oh-my-pi/pi-agent-core",
-	"version": "18.12.6",
+	"version": "18.13.1",
 	"description": "General-purpose agent with transport abstraction, state management, and attachment support",
 	"homepage": "https://omp.sh",
 	"author": {
@@ -38,16 +38,16 @@
 		"fmt": "oxfmt --no-error-on-unmatched-pattern 'src/**/*.{ts,tsx}' '{test,bench,examples,scripts}/**/*.ts' '*.ts'"
 	},
 	"dependencies": {
-		"@oh-my-pi/pi-ai": "18.12.6",
-		"@oh-my-pi/pi-catalog": "18.12.6",
-		"@oh-my-pi/pi-natives": "18.12.6",
-		"@oh-my-pi/pi-utils": "18.12.6",
-		"@oh-my-pi/pi-wire": "18.12.6",
-		"@oh-my-pi/snapcompact": "18.12.6",
+		"@oh-my-pi/pi-ai": "18.13.1",
+		"@oh-my-pi/pi-catalog": "18.13.1",
+		"@oh-my-pi/pi-natives": "18.13.1",
+		"@oh-my-pi/pi-utils": "18.13.1",
+		"@oh-my-pi/pi-wire": "18.13.1",
+		"@oh-my-pi/snapcompact": "18.13.1",
 		"@opentelemetry/api": "^1.9.1"
 	},
 	"devDependencies": {
-		"@oh-my-pi/omptype": "18.12.6",
+		"@oh-my-pi/omptype": "18.13.1",
 		"@opentelemetry/context-async-hooks": "^2.9.0",
 		"@opentelemetry/sdk-trace-base": "^2.9.0",
 		"@types/bun": "^1.3.14"

package/src/agent-loop.ts

@@ -110,6 +110,17 @@ import {
 	SPECULATIVE_STREAM_SESSION,
 } from "./types";
 import { yieldIfDue } from "./utils/yield";
+
+async function waitForProcessPause(config: AgentLoopConfig, signal?: AbortSignal): Promise<void> {
+	try {
+		while (agentPauseGate.paused && !signal?.aborted) {
+			config.onPauseChange?.(true);
+			await agentPauseGate.waitUntilResumed(signal);
+		}
+	} finally {
+		config.onPauseChange?.(false);
+	}
+}
 /** Stop-details marker for a provider error after assistant content/tool args already streamed. */
 export const STREAM_INTERRUPTED_AFTER_CONTENT_STOP_DETAIL = "stream_interrupted_after_content";
 
@@ -1125,6 +1136,9 @@ interface StepCounter {
 	count: number;
 }
 
+/** Largest setTimeout delay before 32-bit overflow (Node/Bun fire in ~1 ms past this). */
+const MAX_TIMER_MS = 2_147_483_647;
+
 function isDeadlineExceeded(deadline: number | undefined): boolean {
 	return deadline !== undefined && Date.now() >= deadline;
 }
@@ -1211,16 +1225,17 @@ async function runLoopBody(
 ): Promise<void> {
 	let deadlineTimer: Timer | undefined;
 	if (config.deadline !== undefined) {
+		const deadline = config.deadline;
 		const deadlineAbortController = new AbortController();
 		const deadlineReason = new DOMException("Deadline exceeded", "TimeoutError");
-		const delay = config.deadline - Date.now();
-		if (delay <= 0) {
-			deadlineAbortController.abort(deadlineReason);
-		} else {
-			deadlineTimer = setTimeout(() => {
-				deadlineAbortController.abort(deadlineReason);
-			}, delay);
-		}
+		// setTimeout overflows past 2^31-1 ms and fires in ~1 ms, so a valid `--max-time 600h`
+		// aborted the first turn at once: wait in bounded slices and recheck the deadline.
+		const armDeadline = (): void => {
+			const delay = deadline - Date.now();
+			if (delay <= 0) deadlineAbortController.abort(deadlineReason);
+			else deadlineTimer = setTimeout(armDeadline, Math.min(delay, MAX_TIMER_MS));
+		};
+		armDeadline();
 		signal = signal ? AbortSignal.any([signal, deadlineAbortController.signal]) : deadlineAbortController.signal;
 	}
 
@@ -1327,7 +1342,7 @@ async function runLoopBody(
 				// Park at the turn boundary while the process-wide pause gate is
 				// engaged (host /pause). An external abort releases the park so a
 				// cancelled run still unwinds while everything else stays frozen.
-				if (agentPauseGate.paused) await agentPauseGate.waitUntilResumed(signal);
+				if (agentPauseGate.paused) await waitForProcessPause(config, signal);
 
 				// Build the provider-bound context before opening the turn. Queue
 				// messages are added now but their events remain deferred until
@@ -1439,6 +1454,7 @@ async function runLoopBody(
 				let recovered: HarmonyRecoveredToolCall | undefined;
 				let message: AssistantMessage;
 				try {
+					if (agentPauseGate.paused) await waitForProcessPause(config, signal);
 					message = await streamAssistantResponse(
 						currentContext,
 						config,
@@ -1789,7 +1805,7 @@ async function runLoopBody(
 						: [...live, ...((await config.getSteeringMessages?.(signal)) || [])];
 					if (hasMoreToolCalls) {
 						// Mid-work: fold any non-interrupting asides into the next turn alongside steering.
-						const asides = signal?.aborted ? [] : resolveAsides(await config.getAsideMessages?.());
+						const asides = signal?.aborted ? [] : resolveAsides(await config.getAsideMessages?.("work"));
 						pendingMessages = asides.length > 0 ? [...steering, ...asides] : steering;
 					} else {
 						// Stop boundary: only steering (live user input) forces another turn here. Leave
@@ -1817,7 +1833,7 @@ async function runLoopBody(
 			// above and this yield point (e.g. queued while onBeforeYield ran). Without
 			// this poll it would strand in the queue until the next manual prompt.
 			const lateSteering = signal?.aborted ? [] : (await config.getSteeringMessages?.(signal)) || [];
-			const asideMessages = signal?.aborted ? [] : resolveAsides(await config.getAsideMessages?.());
+			const asideMessages = signal?.aborted ? [] : resolveAsides(await config.getAsideMessages?.("stop"));
 			const followUpMessages = signal?.aborted ? [] : (await config.getFollowUpMessages?.(signal)) || [];
 			if (lateSteering.length > 0 || asideMessages.length > 0 || followUpMessages.length > 0) {
 				// Set as pending so the inner loop processes them before stopping.
@@ -3363,6 +3379,18 @@ async function executeToolCalls(
 		};
 	});
 
+	// Derive batch readiness from existing dispatch records, not extra participants.
+	const refreshPauseState = (): void => {
+		config.onPauseChange?.(
+			agentPauseGate.paused &&
+				!signal?.aborted &&
+				!speculationCoordinator?.isRunning &&
+				records.some(record => !record.started && !record.skipped && !record.toolResultMessage) &&
+				!records.some(record => record.started && !record.toolResultMessage),
+		);
+	};
+	if (speculationCoordinator) speculationCoordinator.onIdle = refreshPauseState;
+
 	const checkAsideInterrupts = async (): Promise<void> => {
 		// Asides only fire once: an interrupt already recorded on interruptState
 		// must not re-abort, and (unlike steering) never re-consumes a queue.
@@ -3513,7 +3541,11 @@ async function executeToolCalls(
 		// Park before starting this tool while the process-wide pause gate is
 		// engaged. Tools already executing are unaffected (pausing never aborts);
 		// a batch interrupted mid-pause unwinds via the signal checks below.
-		if (agentPauseGate.paused) await agentPauseGate.waitUntilResumed(record.signal);
+		while (agentPauseGate.paused && !record.signal.aborted) {
+			refreshPauseState();
+			await agentPauseGate.waitUntilResumed(record.signal);
+		}
+		config.onPauseChange?.(false);
 
 		const { toolCall, tool } = record;
 		// Validation (and the beforeToolCall hook) ran in the prepare phase; a
@@ -3847,7 +3879,13 @@ async function executeToolCalls(
 			concurrency = concurrencyMode ?? "shared";
 		}
 		const start = concurrency === "exclusive" ? Promise.all([lastExclusive, ...sharedTasks]) : lastExclusive;
-		const task = start.then(() => runTool(record, index));
+		const task = start.then(async () => {
+			try {
+				await runTool(record, index);
+			} finally {
+				refreshPauseState();
+			}
+		});
 		tasks.push(task);
 		if (concurrency === "exclusive") {
 			lastExclusive = task;
@@ -3859,6 +3897,8 @@ async function executeToolCalls(
 	try {
 		await Promise.allSettled(tasks);
 	} finally {
+		if (speculationCoordinator) speculationCoordinator.onIdle = undefined;
+		config.onPauseChange?.(false);
 		steeringWatchAbortController.abort();
 		await steeringWatchPromise?.catch(() => undefined);
 		clearInterval(steeringWatchTimer);

package/src/agent.ts

@@ -39,6 +39,7 @@ import {
 	unpairedToolCallTail,
 } from "./agent-loop";
 import type { AppendOnlyContextManager } from "./append-only-context";
+import { agentPauseGate } from "./pause";
 import { isProviderRefusalMessage } from "./replay-policy";
 import { SentToolDefinitions } from "./sent-tool-definitions";
 import { Tokenizer } from "./tokenizer";
@@ -58,6 +59,7 @@ import type {
 	AgentTool,
 	AgentToolContext,
 	AgentTurnEndContext,
+	AsideBoundary,
 	AsideMessage,
 	PrepareQueuedMessages,
 	QueuedMessageQueue,
@@ -473,6 +475,12 @@ export class Agent {
 	#cwdResolver?: () => string | undefined;
 
 	#runningPrompt?: Promise<void>;
+	#isPaused = false;
+
+	/** True while this unfinished turn is safely parked by the process pause gate. */
+	get isPaused(): boolean {
+		return this.#isPaused && agentPauseGate.paused && !this.#abortController?.signal.aborted;
+	}
 	#resolveRunningPrompt?: () => void;
 	#kimiApiFormat?: "openai" | "anthropic";
 	#preferWebsockets?: boolean;
@@ -499,7 +507,7 @@ export class Agent {
 	#onModelCallSystemPrompt?: (systemPrompt: string[]) => void;
 	#beforeModelCall?: AgentBeforeModelCall;
 	#additionalBeforeModelCalls = new Set<AgentBeforeModelCall>();
-	#asideMessageProvider?: () => AsideMessage[] | Promise<AsideMessage[]>;
+	#asideMessageProvider?: (boundary: AsideBoundary) => AsideMessage[] | Promise<AsideMessage[]>;
 	#telemetry?: AgentLoopConfig["telemetry"];
 	#appendOnlyContext?: AppendOnlyContextManager;
 	#beforeQueuedMessageDequeueHooks = new Set<(signal?: AbortSignal) => Promise<void> | void>();
@@ -1204,7 +1212,9 @@ export class Agent {
 	 * aborts foreground tools; an interruptible `wait` may end early through the
 	 * peek hooks. See `AgentLoopConfig.getAsideMessages`.
 	 */
-	setAsideMessageProvider(fn: (() => AsideMessage[] | Promise<AsideMessage[]>) | undefined): void {
+	setAsideMessageProvider(
+		fn: ((boundary: AsideBoundary) => AsideMessage[] | Promise<AsideMessage[]>) | undefined,
+	): void {
 		this.#asideMessageProvider = fn;
 	}
 
@@ -1780,6 +1790,7 @@ export class Agent {
 		this.#state.isStreaming = true;
 		this.#state.streamMessage = null;
 		this.#state.error = undefined;
+		this.#isPaused = false;
 
 		// Clear Cursor tool result buffer at start of each run
 		this.#cursorToolResultBuffer = [];
@@ -1952,6 +1963,9 @@ export class Agent {
 			onAssistantMessageEvent: this.#onAssistantMessageEvent,
 			onHarmonyLeak: this.#onHarmonyLeak,
 			onTurnEnd: (messages, signal, context) => this.#onTurnEnd?.(messages, signal, context),
+			onPauseChange: paused => {
+				this.#isPaused = paused;
+			},
 			getToolChoice,
 			softToolRequirementState: this.#softToolRequirementState,
 			onToolChoiceRejected: () => {
@@ -1981,7 +1995,7 @@ export class Agent {
 			hasBackgroundCompletions: this.hasBackgroundCompletions,
 			hasQueuedAsides: this.hasQueuedAsides,
 			getFollowUpMessages: signal => this.#dequeueFollowUpMessagesAfterHooks(signal ?? loopSignal),
-			getAsideMessages: async () => (await this.#asideMessageProvider?.()) ?? [],
+			getAsideMessages: async boundary => (await this.#asideMessageProvider?.(boundary)) ?? [],
 			onBeforeYield: () => this.#onBeforeYield?.(),
 			onDetachedToolUpdate: event => this.emitExternalEvent(event),
 			telemetry: this.#telemetry,
@@ -2202,6 +2216,7 @@ export class Agent {
 				this.#emit({ type: "agent_end", messages: [errorMsg] });
 			}
 		} finally {
+			this.#isPaused = false;
 			this.#restoreUndeliveredQueuedMessages(loopAbortController);
 			resolveRun?.();
 			if (this.#abortController === loopAbortController) {

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

@@ -24,6 +24,7 @@ import {
 	resolveOpenAIRequestSetup,
 } from "@oh-my-pi/pi-ai/providers/openai-shared";
 import { captureOpenAIHttpError } from "@oh-my-pi/pi-ai/utils/openai-http";
+import type { CatalogPolicyInputs } from "@oh-my-pi/pi-catalog/types";
 import { isBedrockOpenAIUrl } from "@oh-my-pi/pi-catalog/hosts";
 import {
 	applyCodexResidencyHeader,
@@ -131,7 +132,8 @@ export function shouldUseCompactionV2Streaming(
 	return getCompactionV2Endpoint(model) !== undefined;
 }
 
-function compactionV2Api(model: Model): Api | undefined {
+/** Effective compaction request adapter: the `remoteCompaction.api` override, else the model's own API. Always defined — `Model.api` is required. */
+function compactionV2Api(model: Model): Api {
 	return model.remoteCompaction?.api ?? model.api;
 }
 
@@ -171,6 +173,23 @@ function resolveCompactionV2Model(model: Model): string {
 	return mappedDeployment ?? requestModel;
 }
 
+/**
+ * Catalog-policy view of a remote compaction request (V1 and V2): the
+ * effective adapter (`remoteCompaction.api`) and the wire id the body sends
+ * (`remoteCompaction.model`). Under a model override the active lane's
+ * `requestModelId` and thinking classification describe a model that never
+ * reaches the wire, so they are dropped; `reasoning` stays because the
+ * compaction body derives its reasoning config from the active model.
+ */
+export function remoteCompactionPolicyModel(model: Model): CatalogPolicyInputs {
+	const api = compactionV2Api(model);
+	const overrideId = model.remoteCompaction?.model;
+	const base = { provider: model.provider, providerType: model.providerType, api, reasoning: model.reasoning };
+	return overrideId == null
+		? { ...base, id: model.id, requestModelId: model.requestModelId, thinking: model.thinking }
+		: { ...base, id: overrideId };
+}
+
 // ============================================================================
 // Request Building
 // ============================================================================
@@ -413,8 +432,9 @@ function buildCompactionV2Headers(
 					"content-type": "application/json",
 					...resolveOpenAIRequestSetup(
 						{
-							provider: model.provider,
-							id: model.id,
+							// `on-api` scoped cache policy must see the request this
+							// compaction actually sends, not the active lane.
+							...remoteCompactionPolicyModel(model),
 							baseUrl: model.baseUrl,
 							headers: model.headers,
 							compat: {
@@ -422,7 +442,16 @@ function buildCompactionV2Headers(
 									model.compat && "sessionHeader" in model.compat ? model.compat.sessionHeader : undefined,
 							},
 						},
-						{ apiKey, messages: [], sessionId: request.sessionId ?? routingSessionId, promptCacheSessionId },
+						{
+							apiKey,
+							messages: [],
+							sessionId: request.sessionId ?? routingSessionId,
+							promptCacheSessionId,
+							// The family affinity key rides only an explicit
+							// prompt-cache identity; the sessionId-derived cache-id
+							// fallback never does.
+							promptCacheAffinityKey: request.promptCacheKey ? promptCacheSessionId : undefined,
+						},
 					).headers,
 				};
 	if (api === "openai-codex-responses" || model.provider === "openai-codex") {

package/src/compaction/compaction.ts

@@ -200,6 +200,19 @@ export interface CompactionSettings {
 /** Reserve applied when {@link CompactionSettings.reserveTokens} is unset. */
 export const DEFAULT_RESERVE_TOKENS = 16384;
 
+/**
+ * Hysteresis band under the compaction trigger. Maintenance counts as having
+ * resolved threshold pressure only when residual context lands at or below
+ * `COMPACTION_RECOVERY_BAND × threshold`. Re-checking against the raw threshold
+ * lets a pass keep reclaiming a trickle of the previous turn's output and land
+ * just under the line every turn, sustaining the auto-continue dead loop
+ * reported in #2275; the same band stops the context-full / snapcompact tail
+ * from re-firing on a history whose single most-recent kept turn already
+ * exceeds the threshold. {@link prepareCompaction} sizes the kept tail to land
+ * inside it.
+ */
+export const COMPACTION_RECOVERY_BAND = 0.8;
+
 /**
  * Hard ceiling on a generated compaction summary.
  *
@@ -728,8 +741,15 @@ export interface SummaryOptions {
 	thinkingLevel?: ThinkingLevel;
 	/** Session routing key for remote compaction transports with sticky provider sessions. */
 	sessionId?: string;
-	/** Prompt-cache key for remote compaction transports that support provider prefix caching. */
+	/** Prompt-cache key for conventional summary requests, gated against the summarizing model. */
 	promptCacheKey?: string;
+	/**
+	 * Prompt-cache key for remote compaction requests, gated against the request
+	 * they send (`remoteCompaction.api`/`model` override, see
+	 * `remoteCompactionPolicyModel`). Defaults to {@link promptCacheKey} only when
+	 * omitted; an explicit `undefined` means the remote target gets no key.
+	 */
+	remotePromptCacheKey?: string;
 	/** Mutable provider state used to keep Codex compaction on the live session identity. */
 	providerSessionState?: Map<string, ProviderSessionState>;
 	/** Whether Codex remote compaction should prefer the provider WebSocket transport. */
@@ -1205,6 +1225,14 @@ export async function generateHandoff(
 	);
 }
 
+/** Result of the short-summary request before it is mapped to the optional field. */
+type ShortSummaryOutcome = { kind: "text"; text: string } | { kind: "failed"; error: string } | { kind: "aborted" };
+
+/**
+ * Generate the optional display summary. Cancellation propagates as an
+ * `AbortError`; every other failure or an empty answer yields `undefined`, so
+ * a completed full summary is never discarded for this display-only field.
+ */
 async function generateShortSummary(
 	recentMessages: AgentMessage[],
 	historySummary: string | undefined,
@@ -1213,7 +1241,7 @@ async function generateShortSummary(
 	apiKey: ApiKey,
 	signal?: AbortSignal,
 	options?: SummaryOptions,
-): Promise<string> {
+): Promise<string | undefined> {
 	const maxTokens = Math.min(512, Math.floor(0.2 * reserveTokens));
 	const llmMessages = (options?.convertToLlm ?? defaultConvertToLlm)(recentMessages);
 	const conversationText = serializeConversationForSummary(
@@ -1229,57 +1257,97 @@ async function generateShortSummary(
 	promptText += formatAdditionalContext(options?.extraContext);
 	promptText += SHORT_SUMMARY_PROMPT;
 
-	if (options?.remoteEndpoint) {
-		const endpoint = options.remoteEndpoint;
-		const remote = await withAuth(
-			apiKey,
-			key =>
-				requestRemoteCompaction(
-					endpoint,
-					{ systemPrompt: SUMMARIZATION_SYSTEM_PROMPT, prompt: promptText, maxTokens },
+	let outcome: ShortSummaryOutcome;
+	try {
+		if (options?.remoteEndpoint) {
+			const endpoint = options.remoteEndpoint;
+			const remote = await withAuth(
+				apiKey,
+				key =>
+					requestRemoteCompaction(
+						endpoint,
+						{ systemPrompt: SUMMARIZATION_SYSTEM_PROMPT, prompt: promptText, maxTokens },
+						signal,
+						{ fetch: options?.fetch, model, apiKey: key },
+					),
+				{ signal, missingKeyMessage: "Remote compaction credentials unavailable" },
+			);
+			outcome = { kind: "text", text: remote.summary };
+		} else {
+			const response = await instrumentedCompleteSimple(
+				model,
+				{
+					systemPrompt: [SUMMARIZATION_SYSTEM_PROMPT],
+					messages: [{ role: "user", content: [{ type: "text", text: promptText }], timestamp: Date.now() }],
+				},
+				{
+					maxTokens,
 					signal,
-					{ fetch: options?.fetch, model, apiKey: key },
-				),
-			{ signal, missingKeyMessage: "Remote compaction credentials unavailable" },
-		);
-		return remote.summary;
+					apiKey,
+					reasoning: resolveCompactionEffort(model, options?.thinkingLevel),
+					initiatorOverride: options?.initiatorOverride,
+					metadata: options?.metadata,
+					fetch: options?.fetch,
+					sessionId: options?.sessionId,
+					promptCacheKey: options?.promptCacheKey,
+					providerSessionState: options?.providerSessionState,
+					codexCompaction: localCodexCompaction(options),
+				},
+				{
+					telemetry: options?.telemetry,
+					oneshotKind: "compaction_short_summary",
+					completeImpl: options?.completeImpl,
+					retry: summaryOneshotRetry(options),
+				},
+			);
+			switch (response.stopReason) {
+				case "aborted":
+					outcome = { kind: "aborted" };
+					break;
+				case "error":
+					outcome = { kind: "failed", error: response.errorMessage || "Unknown error" };
+					break;
+				case "stop":
+				case "length":
+				case "toolUse":
+					outcome = {
+						kind: "text",
+						text: response.content
+							.filter((c): c is { type: "text"; text: string } => c.type === "text")
+							.map(c => c.text)
+							.join("\n"),
+					};
+					break;
+			}
+		}
+	} catch (err) {
+		outcome = signal?.aborted
+			? { kind: "aborted" }
+			: { kind: "failed", error: err instanceof Error ? err.message : String(err) };
 	}
 
-	const response = await instrumentedCompleteSimple(
-		model,
-		{
-			systemPrompt: [SUMMARIZATION_SYSTEM_PROMPT],
-			messages: [{ role: "user", content: [{ type: "text", text: promptText }], timestamp: Date.now() }],
-		},
-		{
-			maxTokens,
-			signal,
-			apiKey,
-			reasoning: resolveCompactionEffort(model, options?.thinkingLevel),
-			initiatorOverride: options?.initiatorOverride,
-			metadata: options?.metadata,
-			fetch: options?.fetch,
-			sessionId: options?.sessionId,
-			promptCacheKey: options?.promptCacheKey,
-			providerSessionState: options?.providerSessionState,
-			codexCompaction: localCodexCompaction(options),
-		},
-		{
-			telemetry: options?.telemetry,
-			oneshotKind: "compaction_short_summary",
-			completeImpl: options?.completeImpl,
-			retry: summaryOneshotRetry(options),
-		},
-	);
-
-	if (response.stopReason === "error") {
-		throw createSummarizationError("Short summary failed", response);
+	if (signal?.aborted) outcome = { kind: "aborted" };
+	switch (outcome.kind) {
+		case "aborted":
+			throw new AIError.AbortError("Short summary aborted", { cause: signal?.reason });
+		case "failed":
+			logger.warn("Compaction short summary skipped", {
+				error: outcome.error,
+				model: model.id,
+				provider: model.provider,
+			});
+			return undefined;
+		case "text":
+			if (outcome.text.trim() === "") {
+				logger.warn("Compaction short summary skipped", {
+					error: "empty answer",
+					model: model.id,
+					provider: model.provider,
+				});
+				return undefined;
+			}
+			return outcome.text;
 	}
-
-	return response.content
-		.filter((c): c is { type: "text"; text: string } => c.type === "text")
-		.map(c => c.text)
-		.join("\n");
 }
 
 // ============================================================================
@@ -1457,9 +1525,27 @@ export function prepareCompaction(
 		compactionMessages.push(message);
 	}
 
-	const lastUsage = getLastAssistantUsage(pathEntries);
+	const postCompactionStart = Math.max(prevCompactionIndex, resetBoundaryIndex) + 1;
+	const lastUsage = getLastAssistantUsage(
+		postCompactionStart > 0 ? pathEntries.slice(postCompactionStart) : pathEntries,
+	);
 	const tokensBefore = lastUsage ? calculateContextTokens(lastUsage) : 0;
 	let keepRecentTokens = settings.keepRecentTokens;
+	if (activeModel?.contextWindow && activeModel.contextWindow > 0) {
+		// The compacted context (kept tail + summary/prompt overhead) must land in
+		// the recovery band under the configured trigger, or it fires again at once.
+		// The trigger is the user's choice, so the kept tail gives way, not the trigger:
+		// it gets what the band leaves after the reserve floor (an upper bound on the
+		// summary), or half the band when the floor alone fills it.
+		const contextWindow = activeModel.contextWindow;
+		const recoveryBand = Math.floor(resolveThresholdTokens(contextWindow, settings) * COMPACTION_RECOVERY_BAND);
+		const reserveFloor = Math.min(
+			resolveBudgetReserveTokens(contextWindow, settings),
+			settings.reserveTokens ?? DEFAULT_RESERVE_TOKENS,
+		);
+		const tailBudget = recoveryBand > reserveFloor ? recoveryBand - reserveFloor : Math.floor(recoveryBand / 2);
+		keepRecentTokens = Math.min(keepRecentTokens, Math.max(1, tailBudget));
+	}
 	if (lastUsage) {
 		const estimatedTokens = tokenizer.countMessages(compactionMessages);
 		const promptTokens = calculatePromptTokens(lastUsage);
@@ -1677,6 +1763,10 @@ export async function compact(
 		thinkingLevel: options?.thinkingLevel,
 		sessionId: options?.sessionId,
 		promptCacheKey: options?.promptCacheKey,
+		remotePromptCacheKey:
+			options && Object.hasOwn(options, "remotePromptCacheKey")
+				? options.remotePromptCacheKey
+				: options?.promptCacheKey,
 		providerSessionState: options?.providerSessionState,
 		preferWebsockets: options?.preferWebsockets,
 		codexCompaction: options?.codexCompaction,
@@ -1815,7 +1905,7 @@ export async function compact(
 						forceReasoningOff: summaryOptions.thinkingLevel === ThinkingLevel.Off,
 						responsesLite: model.useResponsesLite,
 						sessionId: summaryOptions.sessionId,
-						promptCacheKey: summaryOptions.promptCacheKey,
+						promptCacheKey: summaryOptions.remotePromptCacheKey,
 						providerSessionState: summaryOptions.providerSessionState,
 						codexCompaction: createOpenAICodexCompactionRequestContext({
 							context: summaryOptions.codexCompaction,
@@ -1888,7 +1978,7 @@ export async function compact(
 			assertRemoteCompactionInputFits(trimmed, model);
 			const requestOptions = {
 				sessionId: summaryOptions.sessionId,
-				promptCacheKey: summaryOptions.promptCacheKey,
+				promptCacheKey: summaryOptions.remotePromptCacheKey,
 				retainedMessageBudget: settings.v2RetainedMessageBudget,
 				retainedUserItems,
 			};
@@ -1970,6 +2060,7 @@ export async function compact(
 						{
 							fetch: summaryOptions.fetch,
 							sessionId: summaryOptions.sessionId,
+							promptCacheKey: summaryOptions.remotePromptCacheKey,
 							providerSessionState: summaryOptions.providerSessionState,
 							codexCompaction: summaryOptions.codexCompaction,
 							rewriteOutputs,
@@ -2272,7 +2363,7 @@ export async function compact(
 
 	return {
 		summary,
-		shortSummary,
+		...(shortSummary !== undefined ? { shortSummary } : {}),
 		firstKeptEntryId: nativeSummary !== undefined ? nativeFirstKeptEntryId : firstKeptEntryId,
 		tokensBefore,
 		details: { readFiles, modifiedFiles } as CompactionDetails,
@@ -2300,7 +2391,7 @@ async function generateTurnPrefixSummary(
 		preferredDialect(model.id),
 		resolveCatalogPolicy(model).classifierCotExtraction === true,
 	);
-	const promptText = `<conversation>\n${conversationText}\n</conversation>\n\n${TURN_PREFIX_SUMMARIZATION_PROMPT}`;
+	const promptText = `<conversation>\n${conversationText}\n</conversation>\n\n${formatAdditionalContext(options?.extraContext)}${TURN_PREFIX_SUMMARIZATION_PROMPT}`;
 	const summarizationMessages = [
 		{
 			role: "user" as const,

package/src/compaction/openai.ts

@@ -20,9 +20,14 @@ import { getCodexAttestationHeader } from "@oh-my-pi/pi-ai/providers/openai-code
 import { createOpenAICodexCompactionRequestContext } from "@oh-my-pi/pi-ai/providers/openai-codex-compaction";
 import { applyCodexResponsesLiteShape } from "@oh-my-pi/pi-ai/providers/openai-codex/request-transformer";
 import { createOpenAICodexCompatibilityMetadata } from "@oh-my-pi/pi-ai/providers/openai-codex-responses";
-import { applySessionHeader } from "@oh-my-pi/pi-ai/providers/inference-headers";
+import {
+	applyInferenceHeaders,
+	applySessionHeader,
+	resolveOpenCodeGoCacheAffinity,
+} from "@oh-my-pi/pi-ai/providers/inference-headers";
 import {
 	encodeResponsesToolResultOutput,
+	getOpenAIPromptCacheKey,
 	hoistInterleavedResponsesToolBatchMessages,
 	parseAzureDeploymentNameMap,
 	parseTextSignature,
@@ -59,7 +64,7 @@ import { dataUrlImageSize } from "../image-tokens";
 import { Tokenizer } from "../tokenizer";
 import { appendAzureApiVersion, resolveAzureOpenAiBaseUrl } from "./azure-openai-endpoint";
 import { prepareBedrockCompactionRequest } from "./bedrock";
-import { isOpenAiRemoteCompactionApi } from "./compaction-v2-streaming";
+import { isOpenAiRemoteCompactionApi, remoteCompactionPolicyModel } from "./compaction-v2-streaming";
 import contextWindowTruncatedOutputPrompt from "./prompts/context-window-truncated-output.md" with { type: "text" };
 
 export * from "./compaction-v2-streaming";
@@ -100,18 +105,13 @@ interface RemoteEstimateParts {
 	imageTokens: number;
 }
 
-function collectRemoteCompactionEstimate(
-	value: unknown,
-	tokenizer: Tokenizer,
-	dropEncrypted: boolean,
-	parts: RemoteEstimateParts,
-): void {
+function collectRemoteCompactionEstimate(value: unknown, tokenizer: Tokenizer, parts: RemoteEstimateParts): void {
 	if (typeof value === "string") {
 		if (value) parts.fragments.push(value);
 		return;
 	}
 	if (Array.isArray(value)) {
-		for (const item of value) collectRemoteCompactionEstimate(item, tokenizer, dropEncrypted, parts);
+		for (const item of value) collectRemoteCompactionEstimate(item, tokenizer, parts);
 		return;
 	}
 	if (value === null || typeof value !== "object") {
@@ -141,9 +141,9 @@ function collectRemoteCompactionEstimate(
 		// Opaque encrypted reasoning/compaction state: its local base64 size far
 		// exceeds what the provider bills, so it stays out of the fit estimate
 		// (same policy as `MessageCountOptions.excludeEncryptedReasoning`).
-		if (item === undefined || (dropEncrypted && key === "encrypted_content" && typeof item === "string")) continue;
+		if (item === undefined || (key === "encrypted_content" && typeof item === "string")) continue;
 		parts.framing.push(key);
-		collectRemoteCompactionEstimate(item, tokenizer, dropEncrypted, parts);
+		collectRemoteCompactionEstimate(item, tokenizer, parts);
 	}
 }
 
@@ -155,21 +155,20 @@ function collectRemoteCompactionEstimate(
 function remoteCompactionEstimateParts(
 	value: unknown,
 	tokenizer: Tokenizer,
-	dropEncrypted: boolean,
 ): { fragments: string[]; imageTokens: number } {
 	const parts: RemoteEstimateParts = { fragments: [], framing: [], imageTokens: 0 };
-	collectRemoteCompactionEstimate(value, tokenizer, dropEncrypted, parts);
+	collectRemoteCompactionEstimate(value, tokenizer, parts);
 	if (parts.framing.length > 0) parts.fragments.push(parts.framing.join(" "));
 	return { fragments: parts.fragments, imageTokens: parts.imageTokens };
 }
 
 /**
  * Tokens of stored Responses replacement history. Retained `input_image` parts
- * count at their image estimate rather than as base64 text; everything else,
- * opaque compaction state included, counts as decoded text.
+ * count at their image estimate, not their base64 size. Opaque encrypted state
+ * stays excluded, matching ordinary message estimates and remote fit checks.
  */
 export function countResponsesHistoryTokens(items: unknown[], tokenizer: Tokenizer): number {
-	const { fragments, imageTokens } = remoteCompactionEstimateParts(items, tokenizer, false);
+	const { fragments, imageTokens } = remoteCompactionEstimateParts(items, tokenizer);
 	return (fragments.length === 0 ? 0 : tokenizer.countTokens(fragments)) + imageTokens;
 }
 
@@ -214,7 +213,6 @@ function probeRemoteCompactionInputBudget(
 	const { fragments, imageTokens } = remoteCompactionEstimateParts(
 		{ instructions, input, ...(tools ? { tools } : {}) },
 		tokenizer,
-		true,
 	);
 	const flatTokens = imageTokens + REMOTE_COMPACTION_REQUEST_OVERHEAD_TOKENS;
 	if (!contextWindow || contextWindow <= 0) {
@@ -360,6 +358,13 @@ export interface OpenAiRemoteCompactionRequest {
 	model: string;
 	input: Array<Record<string, unknown>>;
 	instructions: string;
+	/**
+	 * Responses body cache identity. Mirrors the V2 body: the explicit
+	 * prompt-cache key when supplied (normalized to the 64-char wire limit),
+	 * else the session-derived cache id — retention-gated through
+	 * `getOpenAIPromptCacheKey`; absent when it resolves to nothing.
+	 */
+	prompt_cache_key?: string;
 	reasoning?: {
 		context?: string;
 		[key: string]: unknown;
@@ -858,6 +863,13 @@ export async function requestOpenAiRemoteCompaction(
 		fetch?: FetchImpl;
 		timeoutMs?: number;
 		sessionId?: string;
+		/**
+		 * Explicit prompt-cache identity. Feeds the Go family affinity header
+		 * (`x-opencode-session`) and the Responses body `prompt_cache_key`
+		 * under the same KDL gate as the V2 request; an absent/empty key keeps
+		 * the child conversation id.
+		 */
+		promptCacheKey?: string;
 		providerSessionState?: Map<string, ProviderSessionState>;
 		codexCompaction?: CodexCompactionContext;
 		/** Forwarded to {@link trimRemoteCompactionInputToContextWindow}. */
@@ -889,6 +901,12 @@ export async function requestOpenAiRemoteCompaction(
 		});
 	}
 	assertRemoteCompactionInputFits(trimmed, model);
+	// Same cache identity as the V2 request: the resolved id prefers the
+	// explicit prompt-cache key and falls back to the session id, all
+	// retention-gated by getOpenAIPromptCacheKey (PI_CACHE_RETENTION=none
+	// resolves to nothing here, which then also drops the affinity below).
+	const cacheOptions = { sessionId: opts?.sessionId, promptCacheKey: opts?.promptCacheKey };
+	const promptCacheSessionId = getOpenAIPromptCacheKey(cacheOptions);
 	const request: OpenAiRemoteCompactionRequest = {
 		model: requestModel,
 		// Preserve the native transcript. Only oversized trailing tool outputs are
@@ -896,6 +914,7 @@ export async function requestOpenAiRemoteCompaction(
 		// reasoning, or call/result pairing.
 		input: trimmed.input,
 		instructions,
+		...(promptCacheSessionId ? { prompt_cache_key: promptCacheSessionId } : {}),
 	};
 	const isAzureOpenAiResponses = (model.remoteCompaction?.api ?? model.api) === "azure-openai-responses";
 	const isCodexResponses =
@@ -911,6 +930,22 @@ export async function requestOpenAiRemoteCompaction(
 				Authorization: `Bearer ${apiKey}`,
 				...model.headers,
 			};
+	// Same gate as the V2 request, against the request this compaction actually
+	// sends. Azure keeps its plain header shape (V2 skips the cascade there too).
+	// Only an EXPLICIT prompt-cache key feeds the affinity; the sessionId-derived
+	// cache id never does. `promptCacheSessionId` is already retention-gated.
+	if (!isAzureOpenAiResponses) {
+		applyInferenceHeaders(headers, {
+			provider: model.provider,
+			providerType: model.providerType,
+			protocol: "openai",
+			sessionId: opts?.sessionId,
+			cacheAffinity: resolveOpenCodeGoCacheAffinity(
+				remoteCompactionPolicyModel(model),
+				opts?.promptCacheKey ? promptCacheSessionId : undefined,
+			),
+		});
+	}
 	applySessionHeader(headers, model.compat, opts?.sessionId);
 
 	// Codex endpoints require additional auth headers

package/src/speculative-execution.ts

@@ -239,6 +239,12 @@ export class SpeculativeOperationCoordinator {
 	#closed = false;
 	#admissionsFinalized = false;
 	#running = 0;
+	/** Observer of existing speculative execution draining; no separate work ledger. */
+	onIdle?: () => void;
+
+	get isRunning(): boolean {
+		return this.#running > 0;
+	}
 
 	constructor(
 		readonly config: SpeculativeToolExecutionConfig,
@@ -955,6 +961,7 @@ export class SpeculativeOperationCoordinator {
 			} finally {
 				this.#running--;
 				this.#drain();
+				if (this.#running === 0) this.onIdle?.();
 			}
 		})();
 	}

package/src/types.ts

@@ -68,6 +68,9 @@ export type CommittableAsideMessage = AgentMessage & {
  */
 export type AsideMessage = CommittableAsideMessage | (() => CommittableAsideMessage | null);
 
+/** Where the loop polls asides: after a tool batch (`"work"`) or after a turn without tool calls (`"stop"`). */
+export type AsideBoundary = "work" | "stop";
+
 export interface AgentTurnEndContext {
 	/** Assistant/user message that just completed this turn boundary. */
 	message: AgentMessage;
@@ -172,6 +175,9 @@ export interface SteeringQueueState {
 export interface AgentLoopConfig extends SimpleStreamOptions {
 	model: Model;
 
+	/** Reports whether this loop has reached a safe process-pause boundary. */
+	onPauseChange?: (paused: boolean) => void;
+
 	/**
 	 * When to interrupt tool execution for steering messages.
 	 * - "immediate" = cut interruptible waits short and raise the cooperative
@@ -368,8 +374,12 @@ export interface AgentLoopConfig extends SimpleStreamOptions {
 	 * should reach the model between requests without waiting for the agent to
 	 * fully stop. Returned messages are appended to the context with normal
 	 * message events and keep the loop running so the model can react.
+	 *
+	 * `boundary` is `"work"` after a tool batch and `"stop"` when the model ended
+	 * its turn without tool calls; a provider may hold back messages that should
+	 * not extend a finished turn.
 	 */
-	getAsideMessages?: () => Promise<AsideMessage[]>;
+	getAsideMessages?: (boundary: AsideBoundary) => Promise<AsideMessage[]>;
 	/**
 	 * Hook fired right before the loop would exit.
 	 *