@oh-my-pi/pi-agent-core 17.1.3 → 17.1.5

package/CHANGELOG.md

@@ -2,6 +2,26 @@
 
 ## [Unreleased]
 
+## [17.1.5] - 2026-07-27
+
+### Fixed
+
+- Fixed proxy-stream clients dropping finalized provider-only content blocks, including Anthropic native web-search history, by allowing `done` and `error` events to carry terminal assistant content while retaining delta-reconstructed content from older proxy servers that omit it ([#6703](https://github.com/can1357/oh-my-pi/issues/6703)).
+
+## [17.1.4] - 2026-07-26
+
+### Changed
+
+- Steering is now woken by an event instead of polled on a fixed interval while a tool batch runs. `AgentLoopConfig.waitForSteeringMessages` resolves when a steer is enqueued, so an interruption is observed as soon as it arrives rather than at the next tick, and idle batches stop burning wakeups. The interval timer remains for the IRC interrupt queue, which has no wake callback, and checks only IRC while the event watcher owns steering. Waits are raced against local abort, so a callback that does not observe its signal cannot hang batch teardown.
+
+### Fixed
+
+- Fixed a Cursor tool result being lost when a custom `cursorOnToolResult` transformer was still pending as the turn closed. The provider dispatches decoded messages without awaiting them, so a `message_end` from the same chunk could drain the buffer before the transformer resolved, dropping the result and leaving its `toolCall` block to be stripped as dangling on replay. The entry is now reserved synchronously and patched in place once the transformer resolves, preserving buffer order.
+- Fixed an async `cursorOnToolResult` transformer's rewrite being silently discarded when it resolved after the buffer drain. The reservation kept the call from dangling but the late patch mutated a detached entry, so the already-persisted message kept the pre-transform payload. The drain now awaits any transformer still in flight before persisting, matching the awaited exec-channel paths. A rejecting transformer is swallowed and the reserved payload stands in, so a failing hook cannot take the turn down or cost the result.
+- Fixed Cursor tool results being dropped for hosts that pass neither `cursorExecHandlers` nor `cursorOnToolResult`. Both are optional, but the Cursor provider resolves native todo calls server-side and synthesizes exec blocks regardless, marking both as resolved so no placeholder result is emitted for them. The result buffer callback was only installed when one of the options was present, so a bare SDK host discarded the provider's paired result and every rebuilt transcript stripped the interaction. It is now installed unconditionally.
+- Fixed an async `cursorOnToolResult` rewrite being lost when the provider errored mid-transform. The normal drain waits for a pending transformer, but the error path snapshotted the buffer without that await, so a transform still in flight patched an entry the catch path had already detached and the pre-transform payload was persisted. A provider error is exactly when a transform is most likely to be mid-flight.
+- Reduced oversized OpenAI native compaction requests by replacing only trailing tool-output bodies that exceed the model context window, while preserving calls, assistant history, and reasoning.
+
 ## [17.1.2] - 2026-07-24
 
 ### Added

package/dist/types/agent.d.ts

@@ -166,7 +166,16 @@ export interface AgentOptions {
     /** Additional tools Cursor executes through its MCP request-context bridge, resolved before each provider call. */
     getCursorTools?: () => AgentTool[];
     /**
-     * Cursor tool result callback for exec tool responses.
+     * Optional rewrite of Cursor exec-channel tool results. May return a Promise.
+     *
+     * The Agent reserves the original result in its Cursor result buffer first,
+     * then awaits this hook and patches the reserved entry in place. That keeps
+     * the call paired even if `message_end` arrives while the Promise is still
+     * pending, and `#emitCursorSplitAssistantMessage` waits for any transformer
+     * still in flight before persisting, so a late rewrite is not lost. A
+     * rejecting transformer is swallowed and the reserved payload stands in.
+     * Hosts that only pass `cursorExecHandlers` (the coding-agent path) never
+     * hit this hook.
      */
     cursorOnToolResult?: CursorToolResultHandler;
     /** Current working directory used by local tool execution. */

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

@@ -26,6 +26,20 @@ export declare const OPENAI_REMOTE_COMPACTION_PRESERVE_KEY = "openaiRemoteCompac
  * behind it). On timeout the caller falls back to local summarization.
  */
 export declare const REMOTE_COMPACTION_TIMEOUT_MS = 180000;
+export declare const CONTEXT_WINDOW_TRUNCATED_OUTPUT_MESSAGE: string;
+export interface TrimRemoteCompactionInputResult {
+    input: Array<Record<string, unknown>>;
+    rewrittenOutputs: number;
+    estimatedTokensBefore: number;
+    estimatedTokensAfter: number;
+}
+/**
+ * Preserve the full native transcript unless trailing tool outputs alone push a
+ * remote compaction request beyond the model window. Replacing only those
+ * outputs keeps call/result pairing and all earlier assistant/reasoning history,
+ * matching Codex's recovery path for oversized tool turns.
+ */
+export declare function trimRemoteCompactionInputToContextWindow(input: Array<Record<string, unknown>>, contextWindow: number | null | undefined, instructions: string, tools?: unknown[]): TrimRemoteCompactionInputResult;
 export type OpenAiRemoteCompactionItem = {
     type: "compaction" | "compaction_summary";
     encrypted_content?: string;

package/dist/types/proxy.d.ts

@@ -53,11 +53,13 @@ export type ProxyAssistantMessageEvent = {
     type: "done";
     reason: Extract<StopReason, "stop" | "length" | "toolUse">;
     usage: AssistantMessage["usage"];
+    content?: AssistantMessage["content"];
 } | {
     type: "error";
     reason: Extract<StopReason, "aborted" | "error">;
     errorMessage?: string;
     usage: AssistantMessage["usage"];
+    content?: AssistantMessage["content"];
 };
 export interface ProxyStreamOptions extends SimpleStreamOptions {
     /** Auth token for the proxy server */

package/dist/types/tokenizer.d.ts

@@ -1 +1,2 @@
 export declare function countTokens(text: string | string[]): number;
+export declare function countTokensConservatively(text: string | string[]): number;

package/dist/types/types.d.ts

@@ -181,6 +181,13 @@ export interface AgentLoopConfig extends SimpleStreamOptions {
      * messages are still delivered at the next injection boundary.
      */
     hasSteeringMessages?: () => boolean | SteeringQueueState | Promise<boolean | SteeringQueueState>;
+    /**
+     * Wakes the in-flight tool interrupt watcher when a steering message is queued.
+     * The callback must not consume the queue; the loop still calls
+     * {@link hasSteeringMessages} before aborting and injects through
+     * {@link getSteeringMessages}.
+     */
+    waitForSteeringMessages?: (signal?: AbortSignal) => Promise<void>;
     /**
      * Peeks whether IRC messages should interrupt an interruptible waiting tool.
      *

package/package.json

@@ -1,7 +1,7 @@
 {
 	"type": "module",
 	"name": "@oh-my-pi/pi-agent-core",
-	"version": "17.1.3",
+	"version": "17.1.5",
 	"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": "17.1.3",
-		"@oh-my-pi/pi-catalog": "17.1.3",
-		"@oh-my-pi/pi-natives": "17.1.3",
-		"@oh-my-pi/pi-utils": "17.1.3",
-		"@oh-my-pi/pi-wire": "17.1.3",
-		"@oh-my-pi/snapcompact": "17.1.3",
+		"@oh-my-pi/pi-ai": "17.1.5",
+		"@oh-my-pi/pi-catalog": "17.1.5",
+		"@oh-my-pi/pi-natives": "17.1.5",
+		"@oh-my-pi/pi-utils": "17.1.5",
+		"@oh-my-pi/pi-wire": "17.1.5",
+		"@oh-my-pi/snapcompact": "17.1.5",
 		"@opentelemetry/api": "^1.9.1"
 	},
 	"devDependencies": {

package/src/agent-loop.ts

@@ -2004,7 +2004,7 @@ async function executeToolCalls(
 	// external + IRC aborts. Every other tool sees ONLY the external signal:
 	// neither queued steering nor a peer IRC ever hard-kills a partially
 	// side-effecting foreground tool (e.g. `bash`) — those get the cooperative
-	// steeringSignal above, and the message injects at the next boundary.
+	// `steeringSignal` above, and the message injects at the next boundary.
 	const nonInterruptibleSignal: AbortSignal = signal ?? new AbortController().signal;
 	const interruptibleSignal: AbortSignal = signal
 		? AbortSignal.any([signal, steeringAbortController.signal, ircAbortController.signal])
@@ -2050,6 +2050,21 @@ async function executeToolCalls(
 		};
 	});
 
+	const checkIrcInterrupts = async (): Promise<void> => {
+		// IRC only fires once: a peer interrupt already recorded on interruptState
+		// must not re-abort, and (unlike steering) never re-consumes a queue.
+		if (!shouldInterruptImmediately || signal?.aborted || interruptState.triggered) return;
+		if (hasIrcInterrupts && (await hasIrcInterrupts())) {
+			// Peer IRC hard-aborts interruptible waits only; foreground tools keep
+			// running (no partial side effects) but get the cooperative soft
+			// signal so backgroundable work can step aside for the peer message.
+			interruptState.triggered = true;
+			interruptState.source = "irc";
+			ircAbortController.abort();
+			steeringSoftController.abort();
+		}
+	};
+
 	const checkSteering = async (): Promise<void> => {
 		// `signal` (external/user abort) is checked separately from the internal
 		// abort controllers: once the run is externally aborted it is unwinding
@@ -2087,18 +2102,7 @@ async function executeToolCalls(
 			}
 			return;
 		}
-		// IRC only fires once: a peer interrupt already recorded on interruptState
-		// must not re-abort, and (unlike steering above) never re-consume a queue.
-		if (interruptState.triggered) return;
-		if (hasIrcInterrupts && (await hasIrcInterrupts())) {
-			// Peer IRC hard-aborts interruptible waits only; foreground tools keep
-			// running (no partial side effects) but get the cooperative soft
-			// signal so backgroundable work can step aside for the peer message.
-			interruptState.triggered = true;
-			interruptState.source = "irc";
-			ircAbortController.abort();
-			steeringSoftController.abort();
-		}
+		await checkIrcInterrupts();
 	};
 
 	const emitToolResult = (record: (typeof records)[number], result: AgentToolResult<any>, isError: boolean): void => {
@@ -2443,13 +2447,60 @@ async function executeToolCalls(
 	// mode; checkSteering is idempotent (no-op once triggered).
 	const watchSteeringWhileRunning =
 		shouldInterruptImmediately && (hasSteeringMessages !== undefined || hasIrcInterrupts !== undefined);
-	const steeringWatchTimer = watchSteeringWhileRunning
-		? setInterval(() => void checkSteering(), STEERING_INTERRUPT_POLL_MS)
+	const eventDrivenSteeringWatch =
+		watchSteeringWhileRunning && config.waitForSteeringMessages !== undefined && hasSteeringMessages !== undefined;
+	const steeringWatchAbortController = new AbortController();
+	const steeringWatchSignal = signal
+		? AbortSignal.any([signal, steeringWatchAbortController.signal])
+		: steeringWatchAbortController.signal;
+	// Race every wait against one local abort promise. The callback contract does
+	// not require an implementation to observe the signal, and one that resolves
+	// only on the next queue event would otherwise never settle once the batch
+	// finishes, so awaiting it during teardown would hang a batch with no steer.
+	const { promise: watchAborted, resolve: resolveWatchAbort } = Promise.withResolvers<void>();
+	if (steeringWatchSignal.aborted) {
+		resolveWatchAbort();
+	} else {
+		steeringWatchSignal.addEventListener("abort", () => resolveWatchAbort(), { once: true });
+	}
+	const watchAbortedFalse = watchAborted.then(() => false);
+	const steeringWatchPromise = eventDrivenSteeringWatch
+		? (async (): Promise<void> => {
+				while (!steeringWatchSignal.aborted) {
+					// Subscribe before checking queue state. This closes the edge
+					// race where a steer arrives after a check but before listener
+					// registration: the subsequent check observes queued state,
+					// while later arrivals resolve this already-installed wait.
+					const steeringQueued = config.waitForSteeringMessages?.(steeringWatchSignal).then(
+						() => true,
+						() => false,
+					);
+					const steeringChecked = checkSteering().then(
+						() => true,
+						() => false,
+					);
+					if (!(await Promise.race([steeringChecked, watchAbortedFalse]))) return;
+					if (steeringWatchSignal.aborted || interruptState.triggered) return;
+					if (!(await Promise.race([steeringQueued, watchAbortedFalse]))) return;
+				}
+			})()
 		: undefined;
+	// IRC interrupt records have a separate session-owned queue and no wake
+	// callback. Keep its established timer fallback when that queue is present;
+	// system steering uses the event-driven path above and does not poll.
+	const steeringWatchTimer =
+		watchSteeringWhileRunning && (!eventDrivenSteeringWatch || hasIrcInterrupts !== undefined)
+			? setInterval(
+					() => void (eventDrivenSteeringWatch ? checkIrcInterrupts() : checkSteering()),
+					STEERING_INTERRUPT_POLL_MS,
+				)
+			: undefined;
 	try {
 		await Promise.allSettled(tasks);
 	} finally {
-		if (steeringWatchTimer !== undefined) clearInterval(steeringWatchTimer);
+		steeringWatchAbortController.abort();
+		await steeringWatchPromise?.catch(() => undefined);
+		clearInterval(steeringWatchTimer);
 	}
 	// Yield after batch tool execution to let GC and I/O catch up,
 	// especially when tool results are large (e.g. bash output).

package/src/agent.ts

@@ -276,7 +276,16 @@ export interface AgentOptions {
 	getCursorTools?: () => AgentTool[];
 
 	/**
-	 * Cursor tool result callback for exec tool responses.
+	 * Optional rewrite of Cursor exec-channel tool results. May return a Promise.
+	 *
+	 * The Agent reserves the original result in its Cursor result buffer first,
+	 * then awaits this hook and patches the reserved entry in place. That keeps
+	 * the call paired even if `message_end` arrives while the Promise is still
+	 * pending, and `#emitCursorSplitAssistantMessage` waits for any transformer
+	 * still in flight before persisting, so a late rewrite is not lost. A
+	 * rejecting transformer is swallowed and the reserved payload stands in.
+	 * Hosts that only pass `cursorExecHandlers` (the coding-agent path) never
+	 * hit this hook.
 	 */
 	cursorOnToolResult?: CursorToolResultHandler;
 
@@ -330,6 +339,13 @@ export interface AgentPromptOptions {
 /** Buffered Cursor exec-channel tool result waiting to be emitted after the assistant message. */
 interface CursorToolResultEntry {
 	toolResult: ToolResultMessage;
+	/**
+	 * Set while an async `cursorOnToolResult` transformer is still running for
+	 * this entry, and cleared once it settles. The drain awaits it so a
+	 * transformer that rewrites the payload is not silently discarded when
+	 * `message_end` lands in the same chunk as the tool result.
+	 */
+	pending?: Promise<void>;
 }
 
 export class Agent {
@@ -353,6 +369,8 @@ export class Agent {
 	#transformProviderContext?: (context: Context, model: Model) => Context | Promise<Context>;
 	#steeringQueue: AgentMessage[] = [];
 	#followUpQueue: AgentMessage[] = [];
+	#steeringWaiters = new Set<() => void>();
+
 	#steeringMode: "all" | "one-at-a-time";
 	#followUpMode: "all" | "one-at-a-time";
 	#interruptMode: "immediate" | "wait";
@@ -869,6 +887,7 @@ export class Agent {
 	replaceQueues(steering: AgentMessage[], followUp: AgentMessage[]) {
 		this.#steeringQueue = steering.slice();
 		this.#followUpQueue = followUp.slice();
+		this.#notifySteeringWaiters();
 	}
 
 	appendMessage(m: AgentMessage) {
@@ -889,6 +908,7 @@ export class Agent {
 	 */
 	steer(m: AgentMessage) {
 		this.#steeringQueue.push(m);
+		this.#notifySteeringWaiters();
 	}
 
 	/**
@@ -901,6 +921,7 @@ export class Agent {
 
 	clearSteeringQueue() {
 		this.#steeringQueue = [];
+		this.#notifySteeringWaiters();
 	}
 
 	clearFollowUpQueue() {
@@ -910,6 +931,7 @@ export class Agent {
 	clearAllQueues() {
 		this.#steeringQueue = [];
 		this.#followUpQueue = [];
+		this.#notifySteeringWaiters();
 	}
 
 	hasQueuedMessages(): boolean {
@@ -990,6 +1012,29 @@ export class Agent {
 		return this.#runningPrompt ?? Promise.resolve();
 	}
 
+	/**
+	 * Wait for a steering message without consuming the steering queue.
+	 *
+	 * The signal releases the waiter when the prompt ends, so an in-flight
+	 * tool watcher never survives the tool batch that owns it.
+	 */
+	#waitForSteeringMessages(signal?: AbortSignal): Promise<void> {
+		if (this.#steeringQueue.length > 0 || signal?.aborted) return Promise.resolve();
+		const { promise, resolve } = Promise.withResolvers<void>();
+		const onAbort = (): void => resolve();
+		this.#steeringWaiters.add(resolve);
+		signal?.addEventListener("abort", onAbort, { once: true });
+		return promise.finally(() => {
+			this.#steeringWaiters.delete(resolve);
+			signal?.removeEventListener("abort", onAbort);
+		});
+	}
+
+	#notifySteeringWaiters(): void {
+		const waiters = [...this.#steeringWaiters];
+		for (const resolve of waiters) resolve();
+	}
+
 	reset() {
 		this.#state.messages.length = 0;
 		this.#state.isStreaming = false;
@@ -998,6 +1043,7 @@ export class Agent {
 		this.#state.error = undefined;
 		this.#steeringQueue = [];
 		this.#followUpQueue = [];
+		this.#notifySteeringWaiters();
 	}
 
 	/** Send a prompt with an AgentMessage */
@@ -1127,26 +1173,51 @@ export class Agent {
 			tools: this.#state.tools,
 		};
 
-		const cursorOnToolResult =
-			this.#cursorExecHandlers || this.#cursorOnToolResult
-				? async (message: ToolResultMessage) => {
-						let finalMessage = message;
-						if (this.#cursorOnToolResult) {
-							try {
-								const updated = await this.#cursorOnToolResult(message);
-								if (updated) {
-									finalMessage = updated;
-								}
-							} catch {}
-						}
-						// Cursor executes tools server-side during streaming. We buffer
-						// each toolResult and emit them right after the assistant message
-						// closes (see `#emitCursorSplitAssistantMessage`), so replay
-						// receives (assistant with interleaved toolCall blocks) → results.
-						this.#cursorToolResultBuffer.push({ toolResult: finalMessage });
-						return finalMessage;
-					}
-				: undefined;
+		// Installed unconditionally. Both `cursorExecHandlers` and
+		// `cursorOnToolResult` are optional, but the Cursor provider resolves
+		// native todo calls server-side and synthesizes exec blocks regardless,
+		// marking both `kCursorExecResolved` — `agent-loop.ts` then emits no
+		// placeholder result for them. The provider always offers a paired result
+		// for those blocks (its todo fallback, and every `resolveExecHandler`
+		// exit including the no-handler one), but only through this sink: without
+		// it the result is dropped on the floor, the assistant block is left
+		// unpaired, and `buildSessionContext` strips the whole interaction on
+		// replay. A non-Cursor provider never calls this, so the closure costs
+		// nothing.
+		const cursorOnToolResult = async (message: ToolResultMessage) => {
+			// Cursor executes tools server-side during streaming. We buffer each
+			// toolResult and emit them right after the assistant message closes
+			// (see `#emitCursorSplitAssistantMessage`), so replay receives
+			// (assistant with interleaved toolCall blocks) → results.
+			//
+			// The entry is reserved SYNCHRONOUSLY, before awaiting the optional
+			// transformer. The provider's data loop dispatches messages with
+			// `void handleServerMessage(...)`, so a `message_end` decoded from the
+			// same chunk can drain the buffer while a transformer is still pending
+			// — pushing afterwards would drop the result and strip its toolCall
+			// block as dangling on replay.
+			//
+			// The transformer's in-flight promise is recorded on the entry so the
+			// drain can await it (`#emitCursorSplitAssistantMessage`). Without
+			// that, a transformer resolving after the swap would patch a detached
+			// object while the persisted result kept the original payload — the
+			// rewrite silently lost.
+			const entry: CursorToolResultEntry = { toolResult: message };
+			this.#cursorToolResultBuffer.push(entry);
+			const transform = this.#cursorOnToolResult;
+			if (transform) {
+				const pending = (async () => {
+					try {
+						const updated = await transform(message);
+						if (updated) entry.toolResult = updated;
+					} catch {}
+				})();
+				entry.pending = pending;
+				await pending;
+				entry.pending = undefined;
+			}
+			return entry.toolResult;
+		};
 
 		const getToolChoice = (): ToolChoiceDirective | undefined => {
 			const queued = this.#getToolChoice?.();
@@ -1241,6 +1312,7 @@ export class Agent {
 				}
 				return { queued: true, source: "system" };
 			},
+			waitForSteeringMessages: signal => this.#waitForSteeringMessages(signal),
 			hasIrcInterrupts: this.hasIrcInterrupts,
 			getFollowUpMessages: async () => this.#dequeueFollowUpMessages(),
 			getAsideMessages: async () => (await this.#asideMessageProvider?.()) ?? [],
@@ -1277,7 +1349,7 @@ export class Agent {
 						// Check if this is an assistant message with buffered Cursor tool results.
 						// If so, split the message to emit tool results at the correct position.
 						if (event.message.role === "assistant" && this.#cursorToolResultBuffer.length > 0) {
-							this.#emitCursorSplitAssistantMessage(event.message as AssistantMessage);
+							await this.#emitCursorSplitAssistantMessage(event.message as AssistantMessage);
 							continue; // Skip default emit - split method handles everything
 						}
 						this.#state.streamMessage = null;
@@ -1334,6 +1406,15 @@ export class Agent {
 			const shouldEmitVisibleError = !stoppedForAbort;
 			const assistantPartial = partial?.role === "assistant" ? partial : undefined;
 			const hadAssistantStart = assistantPartial !== undefined;
+			// Same contract as the normal drain in `#emitCursorSplitAssistantMessage`:
+			// a transformer still in flight must be awaited before the payload is
+			// snapshotted, or its rewrite patches an entry this catch path already
+			// detached and the original is persisted instead. A provider error is
+			// exactly when a transform is most likely to be mid-flight.
+			const pendingTransforms = this.#cursorToolResultBuffer
+				.filter(entry => entry.pending !== undefined)
+				.map(entry => entry.pending);
+			if (pendingTransforms.length > 0) await Promise.all(pendingTransforms);
 			const bufferedCursorResults = this.#cursorToolResultBuffer.map(({ toolResult }) => toolResult);
 			const retainedToolCallIds = new Set(completedToolCallIds);
 			for (const { toolCallId } of bufferedCursorResults) retainedToolCallIds.add(toolCallId);
@@ -1459,10 +1540,22 @@ export class Agent {
 	 * toolCall blocks; it also copied `preambleText` into every text block on
 	 * multi-text turns, producing duplicated text on replay.
 	 */
-	#emitCursorSplitAssistantMessage(assistantMessage: AssistantMessage): void {
+	async #emitCursorSplitAssistantMessage(assistantMessage: AssistantMessage): Promise<void> {
+		// Snapshot and detach immediately so a still-pending `cursorOnToolResult`
+		// cannot push into a drained buffer. Entries already reserved stay paired
+		// with their toolCall.
 		const buffer = this.#cursorToolResultBuffer;
 		this.#cursorToolResultBuffer = [];
 
+		// Await any transformer still running for a reserved entry before reading
+		// its payload. The provider dispatches with `void handleServerMessage(…)`,
+		// so a `message_end` from the same chunk can reach this point while a
+		// transformer is mid-flight; without the await its rewrite would land on
+		// the detached entry after the original was already appended and emitted.
+		// Each `pending` swallows its own rejection, so this cannot throw.
+		const pending = buffer.filter(entry => entry.pending !== undefined).map(entry => entry.pending);
+		if (pending.length > 0) await Promise.all(pending);
+
 		this.#state.streamMessage = null;
 		this.appendMessage(assistantMessage);
 		this.#emit({ type: "message_end", message: assistantMessage });

package/src/compaction/compaction.ts

@@ -28,7 +28,7 @@ import { convertTools } from "@oh-my-pi/pi-ai/providers/openai-responses";
 import { buildResponsesInput, resolveOpenAICompatPolicy } from "@oh-my-pi/pi-ai/providers/openai-shared";
 import { preferredDialect } from "@oh-my-pi/pi-catalog/identity";
 import { clampThinkingLevelForModel } from "@oh-my-pi/pi-catalog/model-thinking";
-import { logger, prompt, stringifyJson } from "@oh-my-pi/pi-utils";
+import { isRecord, logger, prompt, stringifyJson } from "@oh-my-pi/pi-utils";
 import * as snapcompact from "@oh-my-pi/snapcompact";
 import { type AgentTelemetry, instrumentedCompleteSimple } from "../telemetry";
 import { ThinkingLevel } from "../thinking";
@@ -51,6 +51,7 @@ import {
 	requestOpenAiRemoteCompaction,
 	requestRemoteCompaction,
 	shouldUseOpenAiRemoteCompaction,
+	trimRemoteCompactionInputToContextWindow,
 	withOpenAiRemoteCompactionPreserveData,
 } from "./openai";
 import autoHandoffThresholdFocusPrompt from "./prompts/auto-handoff-threshold-focus.md" with { type: "text" };
@@ -1290,7 +1291,7 @@ function buildOpenAiResponsesCompactionInput(
 	messages: Message[],
 	model: Model<"openai-responses" | "azure-openai-responses" | "openai-codex-responses">,
 	previousReplacementHistory: Array<Record<string, unknown>> | undefined,
-): unknown[] {
+): Array<Record<string, unknown>> {
 	const input = buildResponsesInput({
 		model,
 		context: { messages },
@@ -1300,7 +1301,14 @@ function buildOpenAiResponsesCompactionInput(
 		includeThinkingSignatures: true,
 		repairOrphanOutputs: true,
 	});
-	return previousReplacementHistory ? [...previousReplacementHistory, ...input] : input;
+	const nativeInput: Array<Record<string, unknown>> = [];
+	for (const item of input) {
+		if (!isRecord(item)) {
+			throw new Error("OpenAI Responses compaction input contains a non-object item");
+		}
+		nativeInput.push(item);
+	}
+	return previousReplacementHistory ? [...previousReplacementHistory, ...nativeInput] : nativeInput;
 }
 
 /**
@@ -1415,20 +1423,33 @@ export async function compact(
 		);
 		if (remoteHistory.length > 0) {
 			try {
-				const request = buildCompactionV2Request(
-					model,
+				const instructions = summaryOptions.remoteInstructions ?? SUMMARIZATION_SYSTEM_PROMPT;
+				const tools = summaryOptions.tools
+					? convertTools(summaryOptions.tools, model.compat.supportsStrictMode, model)
+					: undefined;
+				const trimmed = trimRemoteCompactionInputToContextWindow(
 					remoteHistory,
-					summaryOptions.remoteInstructions ?? SUMMARIZATION_SYSTEM_PROMPT,
-					{
-						tools: summaryOptions.tools
-							? convertTools(summaryOptions.tools, model.compat.supportsStrictMode, model)
-							: undefined,
-						reasoning: buildCompactionV2Reasoning(model, summaryOptions.thinkingLevel),
-						sessionId: summaryOptions.sessionId,
-						promptCacheKey: summaryOptions.promptCacheKey,
-						retainedMessageBudget: settings.v2RetainedMessageBudget,
-					},
+					model.contextWindow,
+					instructions,
+					tools,
 				);
+				if (trimmed.rewrittenOutputs > 0) {
+					logger.info("Rewrote trailing tool outputs before OpenAI V2 remote compaction", {
+						model: model.id,
+						provider: model.provider,
+						rewrittenOutputs: trimmed.rewrittenOutputs,
+						estimatedTokensBefore: trimmed.estimatedTokensBefore,
+						estimatedTokensAfter: trimmed.estimatedTokensAfter,
+						contextWindow: model.contextWindow,
+					});
+				}
+				const request = buildCompactionV2Request(model, trimmed.input, instructions, {
+					tools,
+					reasoning: buildCompactionV2Reasoning(model, summaryOptions.thinkingLevel),
+					sessionId: summaryOptions.sessionId,
+					promptCacheKey: summaryOptions.promptCacheKey,
+					retainedMessageBudget: settings.v2RetainedMessageBudget,
+				});
 				const remote = await withAuth(
 					apiKey,
 					key =>

package/src/compaction/openai.ts

@@ -44,7 +44,9 @@ import {
 	OPENAI_HEADER_VALUES,
 	OPENAI_HEADERS,
 } from "@oh-my-pi/pi-catalog/wire/codex";
-import { $env, logger, stringifyJson, structuredCloneJSON } from "@oh-my-pi/pi-utils";
+import { $env, isRecord, logger, prompt, stringifyJson, structuredCloneJSON } from "@oh-my-pi/pi-utils";
+import { countTokensConservatively } from "../tokenizer";
+import contextWindowTruncatedOutputPrompt from "./prompts/context-window-truncated-output.md" with { type: "text" };
 
 export * from "./compaction-v2-streaming";
 
@@ -66,6 +68,141 @@ export const REMOTE_COMPACTION_TIMEOUT_MS = 180_000;
 
 const DEFAULT_AZURE_API_VERSION = "v1";
 
+export const CONTEXT_WINDOW_TRUNCATED_OUTPUT_MESSAGE = prompt.render(contextWindowTruncatedOutputPrompt);
+
+const REMOTE_COMPACTION_REQUEST_OVERHEAD_TOKENS = 256;
+const REMOTE_COMPACTION_IMAGE_TOKEN_ESTIMATE = 12_000;
+const TOOL_RESULT_IMAGE_ATTACHMENT_TEXT = "Attached image(s) from tool result:";
+
+interface NormalizedEstimateValue {
+	value: unknown;
+	imageTokens: number;
+}
+
+function normalizeRemoteCompactionEstimateValue(value: unknown): NormalizedEstimateValue {
+	if (Array.isArray(value)) {
+		const normalized: unknown[] = [];
+		let imageTokens = 0;
+		for (const item of value) {
+			const result = normalizeRemoteCompactionEstimateValue(item);
+			normalized.push(result.value);
+			imageTokens += result.imageTokens;
+		}
+		return { value: normalized, imageTokens };
+	}
+	if (!value || typeof value !== "object") return { value, imageTokens: 0 };
+
+	const record = value as Record<string, unknown>;
+	if (record.type === "input_image") {
+		return {
+			value: { ...record, image_url: "<image>" },
+			imageTokens: REMOTE_COMPACTION_IMAGE_TOKEN_ESTIMATE,
+		};
+	}
+
+	const normalized: Record<string, unknown> = {};
+	let imageTokens = 0;
+	for (const [key, item] of Object.entries(record)) {
+		const result = normalizeRemoteCompactionEstimateValue(item);
+		normalized[key] = result.value;
+		imageTokens += result.imageTokens;
+	}
+	return { value: normalized, imageTokens };
+}
+
+export interface TrimRemoteCompactionInputResult {
+	input: Array<Record<string, unknown>>;
+	rewrittenOutputs: number;
+	estimatedTokensBefore: number;
+	estimatedTokensAfter: number;
+}
+
+function estimateRemoteCompactionInputTokens(
+	input: Array<Record<string, unknown>>,
+	instructions: string,
+	tools?: unknown[],
+): number {
+	const normalized = normalizeRemoteCompactionEstimateValue({ instructions, input, ...(tools ? { tools } : {}) });
+	const serialized = stringifyJson(normalized.value) ?? "";
+	return countTokensConservatively(serialized) + normalized.imageTokens + REMOTE_COMPACTION_REQUEST_OVERHEAD_TOKENS;
+}
+
+function rewriteToolOutputForContextWindow(item: Record<string, unknown>): Record<string, unknown> | undefined {
+	if (item.type === "function_call_output" || item.type === "custom_tool_call_output") {
+		return { ...item, output: CONTEXT_WINDOW_TRUNCATED_OUTPUT_MESSAGE };
+	}
+	if (item.type === "tool_search_output") {
+		return { ...item, tools: [] };
+	}
+	return undefined;
+}
+
+function isToolResultImageAttachment(item: Record<string, unknown>): boolean {
+	if (item.type !== "message" || item.role !== "user" || !Array.isArray(item.content)) return false;
+
+	let hasLabel = false;
+	let hasImage = false;
+	for (const block of item.content) {
+		if (!isRecord(block)) continue;
+		if (block.type === "input_text" && block.text === TOOL_RESULT_IMAGE_ATTACHMENT_TEXT) hasLabel = true;
+		if (block.type === "input_image") hasImage = true;
+	}
+	return hasLabel && hasImage;
+}
+
+/**
+ * Preserve the full native transcript unless trailing tool outputs alone push a
+ * remote compaction request beyond the model window. Replacing only those
+ * outputs keeps call/result pairing and all earlier assistant/reasoning history,
+ * matching Codex's recovery path for oversized tool turns.
+ */
+export function trimRemoteCompactionInputToContextWindow(
+	input: Array<Record<string, unknown>>,
+	contextWindow: number | null | undefined,
+	instructions: string,
+	tools?: unknown[],
+): TrimRemoteCompactionInputResult {
+	const estimatedTokensBefore = estimateRemoteCompactionInputTokens(input, instructions, tools);
+	if (!contextWindow || contextWindow <= 0 || estimatedTokensBefore <= contextWindow) {
+		return {
+			input,
+			rewrittenOutputs: 0,
+			estimatedTokensBefore,
+			estimatedTokensAfter: estimatedTokensBefore,
+		};
+	}
+
+	let rewrittenInput: Array<Record<string, unknown>> | undefined;
+	let estimatedTokensAfter = estimatedTokensBefore;
+	let rewrittenOutputs = 0;
+	for (let index = input.length - 1; index >= 0 && estimatedTokensAfter > contextWindow; index--) {
+		const item = input[index];
+		if (isToolResultImageAttachment(item)) continue;
+		const rewritten = rewriteToolOutputForContextWindow(item);
+		if (!rewritten) break;
+		rewrittenInput ??= input.slice();
+		rewrittenInput[index] = rewritten;
+		rewrittenOutputs++;
+		estimatedTokensAfter = estimateRemoteCompactionInputTokens(rewrittenInput, instructions, tools);
+	}
+
+	if (!rewrittenInput || estimatedTokensAfter > contextWindow) {
+		return {
+			input,
+			rewrittenOutputs: 0,
+			estimatedTokensBefore,
+			estimatedTokensAfter: estimatedTokensBefore,
+		};
+	}
+
+	return {
+		input: rewrittenInput,
+		rewrittenOutputs,
+		estimatedTokensBefore,
+		estimatedTokensAfter,
+	};
+}
+
 /** Race the caller's signal against the request timeout; `timeoutMs <= 0` disables the watchdog. */
 function withRequestTimeout(signal: AbortSignal | undefined, timeoutMs: number): AbortSignal | undefined {
 	if (timeoutMs <= 0) return signal;
@@ -583,7 +720,7 @@ export function buildOpenAiNativeHistory(
 
 			if (hasImages && model.input.includes("image")) {
 				const contentBlocks: Array<Record<string, unknown>> = [
-					{ type: "input_text", text: "Attached image(s) from tool result:" },
+					{ type: "input_text", text: TOOL_RESULT_IMAGE_ATTACHMENT_TEXT },
 				];
 				for (const block of message.content) {
 					if (block.type !== "image") continue;
@@ -622,12 +759,23 @@ export async function requestOpenAiRemoteCompaction(
 ): Promise<OpenAiRemoteCompactionResponse> {
 	const endpoint = resolveOpenAiCompactEndpoint(model);
 	const requestModel = resolveOpenAiCompactModel(model);
+	const trimmed = trimRemoteCompactionInputToContextWindow(compactInput, model.contextWindow, instructions);
+	if (trimmed.rewrittenOutputs > 0) {
+		logger.info("Rewrote trailing tool outputs before OpenAI remote compaction", {
+			model: model.id,
+			provider: model.provider,
+			rewrittenOutputs: trimmed.rewrittenOutputs,
+			estimatedTokensBefore: trimmed.estimatedTokensBefore,
+			estimatedTokensAfter: trimmed.estimatedTokensAfter,
+			contextWindow: model.contextWindow,
+		});
+	}
 	const request: OpenAiRemoteCompactionRequest = {
 		model: requestModel,
-		// Send full history to the endpoint - don't trim locally.
-		// The provider handles compression via the compaction endpoint.
-		// Trimming before sending loses assistant messages and thinking blocks.
-		input: compactInput,
+		// Preserve the native transcript. Only oversized trailing tool outputs are
+		// rewritten above, reducing the request without losing assistant turns,
+		// reasoning, or call/result pairing.
+		input: trimmed.input,
 		instructions,
 	};
 	const isAzureOpenAiResponses = (model.remoteCompaction?.api ?? model.api) === "azure-openai-responses";

package/src/compaction/prompts/context-window-truncated-output.md

@@ -0,0 +1 @@
+Output exceeded the available model context and was truncated

package/src/proxy.ts

@@ -56,12 +56,14 @@ export type ProxyAssistantMessageEvent =
 			type: "done";
 			reason: Extract<StopReason, "stop" | "length" | "toolUse">;
 			usage: AssistantMessage["usage"];
+			content?: AssistantMessage["content"];
 	  }
 	| {
 			type: "error";
 			reason: Extract<StopReason, "aborted" | "error">;
 			errorMessage?: string;
 			usage: AssistantMessage["usage"];
+			content?: AssistantMessage["content"];
 	  };
 
 export interface ProxyStreamOptions extends SimpleStreamOptions {
@@ -372,6 +374,7 @@ function processProxyEvent(
 		case "done":
 			partial.stopReason = proxyEvent.reason;
 			partial.usage = proxyEvent.usage;
+			if (proxyEvent.content !== undefined) partial.content = proxyEvent.content;
 			calculateCost(model, partial.usage);
 			scrubPartialJson(partial);
 			return { type: "done", reason: proxyEvent.reason, message: partial };
@@ -380,6 +383,7 @@ function processProxyEvent(
 			partial.stopReason = proxyEvent.reason;
 			partial.errorMessage = proxyEvent.errorMessage;
 			partial.usage = proxyEvent.usage;
+			if (proxyEvent.content !== undefined) partial.content = proxyEvent.content;
 			calculateCost(model, partial.usage);
 			scrubPartialJson(partial);
 			return { type: "error", reason: proxyEvent.reason, error: partial };

package/src/tokenizer.ts

@@ -15,3 +15,13 @@ export function countTokens(text: string | string[]): number {
 		return estimateTokens(text);
 	}
 }
+
+export function countTokensConservatively(text: string | string[]): number {
+	if (accurate) {
+		return countTokensNat(text);
+	} else if (Array.isArray(text)) {
+		return text.reduce((sum, value) => sum + Buffer.byteLength(value, "utf-8"), 0);
+	} else {
+		return Buffer.byteLength(text, "utf-8");
+	}
+}

package/src/types.ts

@@ -223,6 +223,14 @@ export interface AgentLoopConfig extends SimpleStreamOptions {
 	 */
 	hasSteeringMessages?: () => boolean | SteeringQueueState | Promise<boolean | SteeringQueueState>;
 
+	/**
+	 * Wakes the in-flight tool interrupt watcher when a steering message is queued.
+	 * The callback must not consume the queue; the loop still calls
+	 * {@link hasSteeringMessages} before aborting and injects through
+	 * {@link getSteeringMessages}.
+	 */
+	waitForSteeringMessages?: (signal?: AbortSignal) => Promise<void>;
+
 	/**
 	 * Peeks whether IRC messages should interrupt an interruptible waiting tool.
 	 *