@oh-my-pi/pi-agent-core 18.3.0 → 18.3.2

package/src/agent-loop.ts

@@ -22,6 +22,7 @@ import {
 	type ToolResultProviderMetadata,
 	type TSchema,
 	toolWireSchema,
+	type UserMessage,
 	validateToolArguments,
 } from "@oh-my-pi/pi-ai";
 import {
@@ -38,6 +39,7 @@ import {
 	getStreamingPartialJson,
 	kCursorExecResolved,
 } from "@oh-my-pi/pi-ai/utils/block-symbols";
+import { schemaDefinesProperty } from "@oh-my-pi/pi-ai/utils/schema/json-schema-validator";
 import { stamp } from "@oh-my-pi/pi-ai/utils/schema/stamps";
 import {
 	createHarmonyAuditEvent,
@@ -51,6 +53,7 @@ import {
 } from "@oh-my-pi/pi-ai/utils/harmony-leak";
 import { logger, sanitizeText, structuredCloneJSON } from "@oh-my-pi/pi-utils";
 import { INTENT_FIELD } from "@oh-my-pi/pi-wire";
+import { LiveSteeringChannel } from "./live-steering";
 import { agentPauseGate } from "./pause";
 import { type AgentRunCoverage, type AgentRunSummary, ToolCallBlockedError } from "./run-collector";
 import { SpeculativeOperationCoordinator } from "./speculative-execution";
@@ -70,6 +73,7 @@ import {
 	startExecuteToolSpan,
 	startInvokeAgentSpan,
 } from "./telemetry";
+import { createAdditionalContextMessage, isNonBlankContext, joinAdditionalContext } from "./tool-context";
 import type {
 	AgentContext,
 	AgentEvent,
@@ -748,6 +752,7 @@ async function emitTurnEnd(
 	await config.onTurnEnd?.(currentContext.messages, terminalYield ? undefined : signal, {
 		message,
 		toolResults,
+		additionalMessages: [],
 		willContinue: false,
 		...context,
 	});
@@ -1018,6 +1023,13 @@ function resolveIntentMode(intent: AgentTool["intent"]): "require" | "optional"
 	return "require";
 }
 
+/**
+ * Longest `i` value accepted as an intent. The injected field is described as
+ * a "concise intent" (INTENT_FIELD_DESCRIPTION); anything past this is a tool
+ * payload the model put in the wrong field, not a label.
+ */
+const MAX_INTENT_LENGTH = 200;
+
 function extractIntent(args: Record<string, unknown>): { intent?: string; strippedArgs: Record<string, unknown> } {
 	const { [INTENT_FIELD]: intent, ...strippedArgs } = args;
 	if (typeof intent !== "string") {
@@ -1096,6 +1108,26 @@ function emitInputMessages(stream: EventStream<AgentEvent, AgentMessage[]>, mess
 	}
 }
 
+/**
+ * Append passive tool-call context after its results as a developer message.
+ * Returns the injected message for turn-end bookkeeping, or undefined when
+ * there is nothing to inject. Shared by the normal tool-call path and the
+ * resume-tail replay so replayed calls deliver context identically.
+ */
+function injectExecutionAdditionalContext(
+	currentContext: AgentContext,
+	newMessages: AgentMessage[],
+	stream: EventStream<AgentEvent, AgentMessage[]>,
+	additionalContext: string | undefined,
+): AgentMessage | undefined {
+	if (additionalContext === undefined) return undefined;
+	const contextMessage = createAdditionalContextMessage(additionalContext);
+	currentContext.messages.push(contextMessage);
+	newMessages.push(contextMessage);
+	emitInputMessages(stream, [contextMessage]);
+	return contextMessage;
+}
+
 /**
  * Resolve aside entries at the moment the loop is about to inject them. Each entry
  * is either a ready {@link AgentMessage} or a sync thunk evaluated here so the
@@ -1159,6 +1191,10 @@ async function runLoopBody(
 	let preserveSoftRequirementState = false;
 
 	let pendingMessages: AgentMessage[] = [];
+	// Steering the provider took from the queue during the last response:
+	// `liveAccepted` reached the model inside it, `liveDeferred` did not.
+	let liveAccepted: AgentMessage[] = [];
+	let liveDeferred: AgentMessage[] = [];
 	try {
 		let messagesToEmit = [...initialMessages];
 		if (isDeadlineExceeded(config.deadline)) {
@@ -1216,8 +1252,15 @@ async function runLoopBody(
 				currentContext.messages.push(result);
 				newMessages.push(result);
 			}
+			const resumeContextMessage = injectExecutionAdditionalContext(
+				currentContext,
+				newMessages,
+				stream,
+				executionResult.additionalContext,
+			);
 			await emitTurnEnd(stream, currentContext, resumeTail, executionResult.toolResults, config, signal, {
 				willContinue: !isDeadlineExceeded(config.deadline),
+				...(resumeContextMessage ? { additionalMessages: [resumeContextMessage] } : {}),
 			});
 			turnOpen = false;
 			// A tool hook may mark its completed result as terminal (e.g. subagent
@@ -1296,6 +1339,7 @@ async function runLoopBody(
 					}
 
 					preparedProviderCall = await prepareProviderCall(currentContext, config, signal);
+					preparedProviderCall.liveSteering = openLiveSteering(config, signal, preparedProviderCall);
 					gateResult = (await config.beforeModelCall?.(preparedProviderCall.context, signal)) || undefined;
 				} catch (error) {
 					if (!turnOpen) {
@@ -1422,6 +1466,12 @@ async function runLoopBody(
 						harmonyRetryAttempt++;
 						continue;
 					}
+				} finally {
+					const channel = preparedProviderCall.liveSteering;
+					if (channel) {
+						liveAccepted.push(...channel.accepted);
+						liveDeferred.push(...channel.deferred);
+					}
 				}
 				if (recovered) {
 					message = snapshotAssistantMessage(message);
@@ -1527,6 +1577,7 @@ async function runLoopBody(
 				const softNonCompliant = softGateActive && !calledOnlyRequiredTool;
 
 				const toolResults: ToolResultMessage[] = [];
+				const additionalMessages: AgentMessage[] = [];
 				if (softNonCompliant && softRequiredTool !== undefined) {
 					SpeculativeOperationCoordinator.discardForMessage(message, "soft tool requirement deferred execution");
 					if (softRequirementState.escalations >= MAX_SOFT_TOOL_ESCALATIONS) {
@@ -1569,13 +1620,19 @@ async function runLoopBody(
 						telemetry,
 						invokeAgentSpan,
 					);
-
 					toolResults.push(...executionResult.toolResults);
 
 					for (const result of toolResults) {
 						currentContext.messages.push(result);
 						newMessages.push(result);
 					}
+					const injectedContext = injectExecutionAdditionalContext(
+						currentContext,
+						newMessages,
+						stream,
+						executionResult.additionalContext,
+					);
+					if (injectedContext) additionalMessages.push(injectedContext);
 				} else if (toolCalls.length > 0) {
 					SpeculativeOperationCoordinator.discardForMessage(
 						message,
@@ -1626,6 +1683,7 @@ async function runLoopBody(
 				}
 
 				await emitTurnEnd(stream, currentContext, message, toolResults, config, signal, {
+					additionalMessages,
 					willContinue: hasMoreToolCalls && !isDeadlineExceeded(config.deadline),
 				});
 				turnOpen = false;
@@ -1640,16 +1698,36 @@ async function runLoopBody(
 				// instantly aborts — message lands in history, agent never responds. The
 				// mid-batch interrupt poll only peeks (hasSteeringMessages), so the queue
 				// still owns every message until this dequeue.
-				const steering = signal?.aborted ? [] : (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?.());
-					pendingMessages = asides.length > 0 ? [...steering, ...asides] : steering;
+				// Aborted: live-taken steering stays unrecorded, so the agent returns
+				// it to the queue for the continuation run.
+				const live = signal?.aborted ? [] : [...liveAccepted, ...liveDeferred];
+				const liveReachedModel = !signal?.aborted && liveAccepted.length > 0;
+				if (liveReachedModel) {
+					for (const message of liveAccepted) {
+						if (message.role === "user") message.liveSteered = true;
+					}
+				}
+				liveAccepted = [];
+				liveDeferred = [];
+				if (liveReachedModel) {
+					// The server continues from exactly this steering; anything else
+					// queued now would not line up with its continuation, so it waits
+					// for the next boundary (or is steered into that response).
+					pendingMessages = live;
 				} else {
-					// Stop boundary: only steering (live user input) forces another turn here. Leave
-					// asides for the outer drain below so a passive aside can't trigger an extra model
-					// turn ahead of a queued follow-up — the outer drain batches asides + follow-ups together.
-					pendingMessages = steering;
+					const steering = signal?.aborted
+						? []
+						: [...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?.());
+						pendingMessages = asides.length > 0 ? [...steering, ...asides] : steering;
+					} else {
+						// Stop boundary: only steering (live user input) forces another turn here. Leave
+						// asides for the outer drain below so a passive aside can't trigger an extra model
+						// turn ahead of a queued follow-up — the outer drain batches asides + follow-ups together.
+						pendingMessages = steering;
+					}
 				}
 			}
 
@@ -1718,6 +1796,45 @@ interface PreparedProviderCall {
 	context: Context;
 	promptToolWireTools: Context["tools"];
 	ownedDialect: Dialect | undefined;
+	/** Steering source offered to the provider for this call. */
+	liveSteering?: LiveSteeringChannel;
+}
+
+/**
+ * Offer queued steering to a provider that can deliver it into the response it
+ * is streaming. Latency decides whether steering lands before the model commits
+ * to its next output, so claims convert only the steering batch — message-level
+ * transforms (steering envelope, redaction) — never the whole transcript.
+ * Provider-context transforms rewrite images, so image-bearing steering waits
+ * for the boundary rather than risk bytes the next request would not replay.
+ */
+function openLiveSteering(
+	config: AgentLoopConfig,
+	loopSignal: AbortSignal | undefined,
+	prepared: PreparedProviderCall,
+): LiveSteeringChannel | undefined {
+	const { getSteeringMessages, waitForSteeringMessages } = config;
+	if (!getSteeringMessages || !waitForSteeringMessages || prepared.ownedDialect) return undefined;
+	const bound = (signal: AbortSignal): AbortSignal => (loopSignal ? AbortSignal.any([signal, loopSignal]) : signal);
+	return new LiveSteeringChannel({
+		wait: signal => waitForSteeringMessages(bound(signal)),
+		take: signal => getSteeringMessages(bound(signal)),
+		toProvider: async (messages, signal) => {
+			const transformed = config.transformContext
+				? await config.transformContext(messages, bound(signal))
+				: messages;
+			const converted = normalizeMessagesForProvider(await config.convertToLlm(transformed), prepared.model);
+			const userMessages: UserMessage[] = [];
+			for (const message of converted) {
+				if (message.role !== "user") return undefined;
+				if (typeof message.content !== "string" && message.content.some(part => part.type === "image")) {
+					return undefined;
+				}
+				userMessages.push(message);
+			}
+			return userMessages.length > 0 ? userMessages : undefined;
+		},
+	});
 }
 
 async function prepareProviderCall(
@@ -1733,7 +1850,8 @@ async function prepareProviderCall(
 
 	const llmMessages = await config.convertToLlm(messages);
 	const normalizedMessages = normalizeMessagesForProvider(llmMessages, model);
-	const ownedDialect: Dialect | undefined = config.dialect ?? resolveOwnedDialectFromEnv(Bun.env.PI_DIALECT);
+	const ownedDialect: Dialect | undefined =
+		(config.getDialect ? config.getDialect(model) : config.dialect) ?? resolveOwnedDialectFromEnv(Bun.env.PI_DIALECT);
 	const pruneToolDescriptions = !!config.pruneToolDescriptions && !ownedDialect;
 	let llmContext: Context;
 	if (config.appendOnlyContext) {
@@ -1902,6 +2020,7 @@ async function streamAssistantResponse(
 				cwd: effectiveCwd,
 				signal: finalRequestSignal,
 				onResponse: captureOnResponse,
+				liveSteering: providerCall.liveSteering,
 			});
 			if (promptToolWireTools && ownedDialect) {
 				// Re-materialize in-band tool-call text as native toolCall content blocks
@@ -2572,6 +2691,12 @@ interface PreparedToolCall {
 	tool: AgentTool<any> | undefined;
 	/** Validated (possibly hook-revised) execution args; raw args when validation failed. */
 	args: Record<string, unknown>;
+	/**
+	 * Passive context returned by `beforeToolCall`. Committed after the batch
+	 * settles only when the call's final result is not an error, so a call the
+	 * tool's own approval gate denies (or that otherwise fails) injects nothing.
+	 */
+	additionalContext?: string;
 	/** Transformed args shared by final reconciliation and eventual dispatch. */
 	executionArgs?: Record<string, unknown>;
 	/** Transform failure retained for execution's scheduled error result. */
@@ -2713,6 +2838,19 @@ async function prepareToolCallDispatch(
 		if (intentTracing) {
 			const { intent, strippedArgs } = extractIntent(toolCall.arguments);
 			argsForExecution = strippedArgs;
+			// A payload in `i` would be stripped and the tool run with the leftover
+			// args. Unknown tools fall through to the not-found error; a tool that
+			// owns `i` as a real parameter has nowhere else to put the value.
+			if (
+				intent !== undefined &&
+				intent.length > MAX_INTENT_LENGTH &&
+				tool &&
+				!schemaDefinesProperty(toolWireSchema(tool), INTENT_FIELD)
+			) {
+				entry.args = strippedArgs;
+				entry.validationErrorMessage = `\`${INTENT_FIELD}\` is a short intent label (at most ${MAX_INTENT_LENGTH} chars); the value you sent is ${intent.length} chars. The tool was not run. Put that content in the tool's own parameters and retry with a brief \`${INTENT_FIELD}\`.`;
+				continue;
+			}
 			if (intent) {
 				toolCall.intent = intent;
 			} else if (typeof tool?.intent === "function") {
@@ -2766,6 +2904,9 @@ async function prepareToolCallDispatch(
 			entry.blockReason = beforeResult.reason;
 			continue;
 		}
+		if (isNonBlankContext(beforeResult?.additionalContext)) {
+			entry.additionalContext = beforeResult.additionalContext;
+		}
 		if (beforeResult?.args !== undefined) {
 			// Revalidate: a hook revision is untrusted input to the tool schema.
 			const revised = validate(beforeResult.args);
@@ -2850,7 +2991,8 @@ async function speculativeFinalCalls(
 }
 
 /**
- * Execute tool calls from an assistant message.
+ * Execute tool calls from an assistant message. Returns model-visible context
+ * only after every result has settled, preserving assistant call order.
  */
 async function executeToolCalls(
 	currentContext: AgentContext,
@@ -2860,7 +3002,7 @@ async function executeToolCalls(
 	config: AgentLoopConfig,
 	telemetry: AgentTelemetry | undefined,
 	invokeAgentSpan: Span | undefined,
-): Promise<{ toolResults: ToolResultMessage[] }> {
+): Promise<{ toolResults: ToolResultMessage[]; additionalContext?: string }> {
 	const tools = currentContext.tools;
 	const {
 		hasSteeringMessages,
@@ -2952,6 +3094,8 @@ async function executeToolCalls(
 			blocked: prepared.blocked === true,
 			blockReason: prepared.blockReason,
 			prepareError: prepared.prepareError,
+			preparedContext: prepared.additionalContext,
+			reportedContext: [] as string[],
 			executionArgs: prepared.executionArgs,
 			transformError: prepared.transformError,
 		};
@@ -3184,11 +3328,13 @@ async function executeToolCalls(
 				}
 
 				if (!completedToolExecution) {
-					// The cooperative steering signal rides the loop-owned
-					// ToolCallContext (surfacing as `ctx.toolCall.steeringSignal`):
-					// AgentToolContext itself is app-built via declaration merging, so
-					// the loop cannot construct or extend one structurally.
-					const streamSession = speculationCoordinator?.takeStreamSession(toolCall.id);
+					// The cooperative steering signal and the passive-context sink
+					// ride the loop-owned ToolCallContext (surfacing as
+					// `ctx.toolCall.*`); the host surfaces the sink on the context it
+					// builds, and the loop hands that object to the tool untouched.
+					// Wrapper-dispatched nested calls (for example `write xd://…`)
+					// inherit the context, so their passive hook context joins this
+					// root call at the batch boundary.
 					const toolContext = getToolContext?.({
 						batchId,
 						index,
@@ -3196,7 +3342,11 @@ async function executeToolCalls(
 						toolCalls: toolCallInfos,
 						steeringSignal: steeringSoftController.signal,
 						providerMetadata: toolCall.providerMetadata,
+						addAdditionalContext: context => {
+							if (isNonBlankContext(context)) record.reportedContext.push(context);
+						},
 					});
+					const streamSession = speculationCoordinator?.takeStreamSession(toolCall.id);
 					if (streamSession && toolContext) {
 						toolContext[SPECULATIVE_STREAM_SESSION] = streamSession;
 					} else if (streamSession && !streamSession.contextIndependent) {
@@ -3449,7 +3599,23 @@ async function executeToolCalls(
 	}
 	await speculationCoordinator?.discardAll("candidate was not dispatched");
 
-	return { toolResults: emittedToolResults };
+	// Skipped calls never ran. Hook-prepared context also requires a non-error
+	// final result; context the tool itself reported during execution stands.
+	// Within a call, tool-reported context (including nested `xd://` dispatch)
+	// precedes the hook's: wrappers release hook context only after the call
+	// succeeds, so this is the one order every dispatch path can produce.
+	const additionalContext = joinAdditionalContext(
+		records
+			.filter(record => !record.skipped)
+			.flatMap(record => [
+				...record.reportedContext,
+				record.toolResultMessage?.isError ? undefined : record.preparedContext,
+			]),
+	);
+	return {
+		toolResults: emittedToolResults,
+		...(additionalContext !== undefined ? { additionalContext } : {}),
+	};
 }
 
 /**

package/src/agent.ts

@@ -40,6 +40,12 @@ import type { AppendOnlyContextManager } from "./append-only-context";
 import { isProviderRefusalMessage } from "./replay-policy";
 import { SentToolDefinitions } from "./sent-tool-definitions";
 import { Tokenizer, tokenizerEncodingForModel } from "./tokenizer";
+import {
+	createAdditionalContextMessage,
+	joinAdditionalContext,
+	TOOL_RESULT_ADDITIONAL_CONTEXT,
+	type ToolResultWithAdditionalContext,
+} from "./tool-context";
 import type {
 	AgentBeforeModelCall,
 	AgentContext,
@@ -66,7 +72,7 @@ import { EventLoopKeepalive } from "./utils/yield";
 function defaultConvertToLlm(messages: AgentMessage[]): Message[] {
 	return messages.filter((m): m is Message => {
 		if (m.role === "assistant") return !isProviderRefusalMessage(m);
-		return m.role === "user" || m.role === "toolResult";
+		return m.role === "user" || m.role === "developer" || m.role === "toolResult";
 	});
 }
 
@@ -272,6 +278,11 @@ export interface AgentOptions {
 	pruneToolDescriptions?: boolean;
 	/** Owned tool-calling dialect. Undefined keeps provider-native tool calling. */
 	dialect?: Dialect;
+	/**
+	 * Per-request owned-dialect resolver, consulted with the model being requested.
+	 * Authoritative when set (like {@link serviceTierResolver}): replaces {@link dialect}.
+	 */
+	dialectResolver?: (model: Model) => Dialect | undefined;
 	/**
 	 * When owned tool calling is active and the model fabricates a tool result
 	 * mid-turn: `true` (default) aborts the provider request immediately; `false`
@@ -362,6 +373,12 @@ interface CursorToolResultEntry {
 	 * `message_end` lands in the same chunk as the tool result.
 	 */
 	pending?: Promise<void>;
+	/**
+	 * Passive context the executor attached via
+	 * {@link TOOL_RESULT_ADDITIONAL_CONTEXT}, captured before any transformer
+	 * can replace the message. Injected after the buffered results.
+	 */
+	additionalContext?: string;
 }
 
 type QueuedMessageQueue = "steering" | "followUp";
@@ -441,6 +458,7 @@ export class Agent {
 	#intentTracing: boolean;
 	#pruneToolDescriptions: boolean;
 	#dialect?: Dialect;
+	#dialectResolver?: (model: Model) => Dialect | undefined;
 	#abortOnFabricatedToolResult?: boolean;
 	#getToolChoice?: () => ToolChoiceDirective | undefined;
 	#onToolChoiceUnavailable?: () => void;
@@ -540,6 +558,7 @@ export class Agent {
 		this.#intentTracing = opts.intentTracing === true;
 		this.#pruneToolDescriptions = opts.pruneToolDescriptions === true;
 		this.#dialect = opts.dialect;
+		this.#dialectResolver = opts.dialectResolver;
 		this.#abortOnFabricatedToolResult = opts.abortOnFabricatedToolResult;
 		this.#getToolChoice = opts.getToolChoice;
 		this.#onToolChoiceUnavailable = opts.onToolChoiceUnavailable;
@@ -749,6 +768,33 @@ export class Agent {
 		this.#hideThinkingSummary = value;
 	}
 
+	/** Strip tool descriptions from provider-bound specs; read per request. */
+	get pruneToolDescriptions(): boolean {
+		return this.#pruneToolDescriptions;
+	}
+
+	set pruneToolDescriptions(value: boolean) {
+		this.#pruneToolDescriptions = value;
+	}
+
+	/** Inject/strip the intent field on tool calls; applies from the next prompt run. */
+	get intentTracing(): boolean {
+		return this.#intentTracing;
+	}
+
+	set intentTracing(value: boolean) {
+		this.#intentTracing = value;
+	}
+
+	/** Abort the provider request on a fabricated tool result; applies from the next prompt run. */
+	get abortOnFabricatedToolResult(): boolean | undefined {
+		return this.#abortOnFabricatedToolResult;
+	}
+
+	set abortOnFabricatedToolResult(value: boolean | undefined) {
+		this.#abortOnFabricatedToolResult = value;
+	}
+
 	/**
 	 * Get the current max retry delay in milliseconds.
 	 */
@@ -829,7 +875,9 @@ export class Agent {
 	): Promise<Context> {
 		const model = this.#state.model;
 		if (!model) throw new Error("No active model on agent");
-		const ownedDialect = this.#dialect ?? resolveOwnedDialectFromEnv(Bun.env.PI_DIALECT);
+		const ownedDialect =
+			(this.#dialectResolver ? this.#dialectResolver(model) : this.#dialect) ??
+			resolveOwnedDialectFromEnv(Bun.env.PI_DIALECT);
 		const messages = normalizeMessagesForProvider(llmMessages, model);
 		const tools = ownedDialect
 			? []
@@ -1518,7 +1566,10 @@ export class Agent {
 			// 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 };
+			const entry: CursorToolResultEntry = {
+				toolResult: message,
+				additionalContext: (message as ToolResultWithAdditionalContext)[TOOL_RESULT_ADDITIONAL_CONTEXT],
+			};
 			this.#cursorToolResultBuffer.push(entry);
 			const transform = this.#cursorOnToolResult;
 			if (transform) {
@@ -1624,6 +1675,7 @@ export class Agent {
 			intentTracing: this.#intentTracing,
 			pruneToolDescriptions: this.#pruneToolDescriptions,
 			dialect: this.#dialect,
+			getDialect: this.#dialectResolver,
 			abortOnFabricatedToolResult: this.#abortOnFabricatedToolResult,
 			appendOnlyContext: this.#appendOnlyContext,
 			beforeToolCall: this.beforeToolCall ? (ctx, signal) => this.beforeToolCall?.(ctx, signal) : undefined,
@@ -1781,6 +1833,9 @@ export class Agent {
 				.map(entry => entry.pending);
 			if (pendingTransforms.length > 0) await Promise.all(pendingTransforms);
 			const bufferedCursorResults = this.#cursorToolResultBuffer.map(({ toolResult }) => toolResult);
+			const bufferedCursorContext = joinAdditionalContext(
+				this.#cursorToolResultBuffer.map(({ additionalContext }) => additionalContext),
+			);
 			const retainedToolCallIds = new Set(completedToolCallIds);
 			for (const { toolCallId } of bufferedCursorResults) retainedToolCallIds.add(toolCallId);
 			const errorMsg: AssistantMessage =
@@ -1857,9 +1912,13 @@ export class Agent {
 					this.#emit({ type: "message_end", message: toolResult });
 					toolResults.push(toolResult);
 				}
+				const agentEndMessages: AgentMessage[] = [errorMsg, ...toolResults];
+				if (bufferedCursorContext !== undefined) {
+					agentEndMessages.push(this.#emitCursorAdditionalContext(bufferedCursorContext));
+				}
 				this.#emit({ type: "turn_end", message: errorMsg, toolResults });
 				turnOpen = false;
-				this.#emit({ type: "agent_end", messages: [errorMsg, ...toolResults] });
+				this.#emit({ type: "agent_end", messages: agentEndMessages });
 			} else {
 				this.appendMessage(errorMsg);
 				this.#state.error = errorMessage;
@@ -1931,8 +1990,24 @@ export class Agent {
 				this.appendMessage(toolResult);
 				this.#emit({ type: "message_end", message: toolResult });
 			}
+			const additionalContext = joinAdditionalContext(buffer.map(entry => entry.additionalContext));
+			if (additionalContext !== undefined) this.#emitCursorAdditionalContext(additionalContext);
 		} finally {
 			this.#cursorToolResultDrain = undefined;
 		}
 	}
+
+	/**
+	 * Append passive context reported by Cursor exec-channel tools after their
+	 * results, mirroring the loop's post-batch developer message. Cursor runs
+	 * those tools server-side mid-stream, so the context reaches the next
+	 * provider request instead of the current one.
+	 */
+	#emitCursorAdditionalContext(text: string): AgentMessage {
+		const message = createAdditionalContextMessage(text);
+		this.#emit({ type: "message_start", message });
+		this.appendMessage(message);
+		this.#emit({ type: "message_end", message });
+		return message;
+	}
 }

package/src/compaction/anthropic.ts

@@ -255,7 +255,7 @@ export async function requestAnthropicNativeCompaction(
 		throw new Error(
 			response.stopDetails?.type === "compaction"
 				? "Anthropic compaction returned no signed summary"
-				: "Anthropic compaction response carried no compaction block",
+				: `Anthropic compaction response carried no compaction block (stop reason: ${response.stopDetails?.type ?? response.stopReason})`,
 		);
 	}
 	return {

package/src/compaction/messages.ts

@@ -64,6 +64,13 @@ export interface CompactionSummaryMessage {
 	images?: ImageContent[];
 	/** Post-pass dead-end warning attached to this compaction (progress guard). */
 	warning?: string;
+	/**
+	 * Thinking-binding rewrite marker when it must differ from `timestamp`: a
+	 * natively replayed summary predates it before the retained tail so that
+	 * tail's bound thinking stays valid. `timestamp` remains the commit time,
+	 * which is what invalidates the tail's pre-compaction usage reports.
+	 */
+	historyRewriteAt?: number;
 	timestamp: number;
 }
 
@@ -135,6 +142,8 @@ export interface CompactionSummaryMessageOptions {
 	method?: string;
 	/** Estimated context tokens after the rewrite, for display alongside `tokensBefore`. */
 	tokensAfter?: number;
+	/** See {@link CompactionSummaryMessage.historyRewriteAt}. */
+	historyRewriteAt?: number;
 }
 
 export function createCompactionSummaryMessage(
@@ -143,7 +152,7 @@ export function createCompactionSummaryMessage(
 	timestamp: string,
 	options: CompactionSummaryMessageOptions = {},
 ): CompactionSummaryMessage {
-	const { shortSummary, providerPayload, images, blocks, warning, method, tokensAfter } = options;
+	const { shortSummary, providerPayload, images, blocks, warning, method, tokensAfter, historyRewriteAt } = options;
 	const imageBlocks =
 		blocks?.filter((block): block is ImageContent => block.type === "image") ??
 		(images && images.length > 0 ? images : undefined);
@@ -158,6 +167,7 @@ export function createCompactionSummaryMessage(
 		blocks: blocks && blocks.length > 0 ? blocks : undefined,
 		images: imageBlocks && imageBlocks.length > 0 ? imageBlocks : undefined,
 		warning,
+		historyRewriteAt,
 		timestamp: new Date(timestamp).getTime(),
 	};
 }
@@ -246,7 +256,7 @@ export function convertMessageToLlm(message: AgentMessage): Message | undefined
 									...(message.images ?? []),
 								],
 					attribution: "agent",
-					historyRewriteAt: message.timestamp,
+					historyRewriteAt: message.historyRewriteAt ?? message.timestamp,
 					providerPayload: message.providerPayload,
 					timestamp: message.timestamp,
 				};

package/src/compaction/pruning.ts

@@ -120,7 +120,7 @@ function createPrunedNotice(tokens: number): string {
  * own rules: useless already drops no-savings candidates, superseded prunes for
  * correctness regardless of size.
  */
-const MIN_PRUNE_TOKENS = 50;
+export const MIN_PRUNE_TOKENS = 50;
 
 function getToolResultMessage(entry: SessionEntry): ToolResultMessage | undefined {
 	if (entry.type !== "message") return undefined;