@oh-my-pi/pi-coding-agent 16.0.3 → 16.0.5

package/src/session/agent-session.ts

@@ -18,6 +18,7 @@ import * as os from "node:os";
 import * as path from "node:path";
 import { scheduler } from "node:timers/promises";
 import { isPromise } from "node:util/types";
+
 import type { InMemorySnapshotStore } from "@oh-my-pi/hashline";
 import {
 	type AfterToolCallContext,
@@ -126,6 +127,8 @@ import {
 	AdvisorRuntime,
 	type AdvisorSeverity,
 	formatAdvisorBatchContent,
+	isAdvisorInterruptImmuneTurnActive,
+	isInterruptingSeverity,
 	resolveAdvisorDeliveryChannel,
 } from "../advisor";
 import { type AsyncJob, type AsyncJobDeliveryState, AsyncJobManager } from "../async";
@@ -178,6 +181,7 @@ import type {
 	SessionBeforeCompactResult,
 	SessionBeforeSwitchResult,
 	SessionBeforeTreeResult,
+	SessionStopEventResult,
 	ToolExecutionEndEvent,
 	ToolExecutionStartEvent,
 	ToolExecutionUpdateEvent,
@@ -202,7 +206,7 @@ import { containsOrchestrate, ORCHESTRATE_NOTICE } from "../modes/orchestrate";
 import { getCurrentThemeName, theme } from "../modes/theme/theme";
 import { parseTurnBudget } from "../modes/turn-budget";
 import { containsUltrathink, ULTRATHINK_NOTICE } from "../modes/ultrathink";
-import { computeNonMessageTokens } from "../modes/utils/context-usage";
+import { computeNonMessageBreakdown, computeNonMessageTokens } from "../modes/utils/context-usage";
 import { containsWorkflow, WORKFLOW_NOTICE } from "../modes/workflow";
 import { createPlanReadMatcher } from "../plan-mode/plan-protection";
 import type { PlanModeState } from "../plan-mode/state";
@@ -261,6 +265,7 @@ import { type EditMode, resolveEditMode } from "../utils/edit-mode";
 import { resolveFileDisplayMode } from "../utils/file-display-mode";
 import { extractFileMentions, generateFileMentionMessages } from "../utils/file-mentions";
 import { normalizeModelContextImages } from "../utils/image-loading";
+import { describeAttachedImagesForTextModel } from "../utils/image-vision-fallback";
 import { buildNamedToolChoice, isToolChoiceActive } from "../utils/tool-choice";
 import type { AuthStorage } from "./auth-storage";
 import type { ClientBridge, ClientBridgePermissionOption, ClientBridgePermissionOutcome } from "./client-bridge";
@@ -295,6 +300,8 @@ import { ToolChoiceQueue } from "./tool-choice-queue";
 import { classifyUnexpectedStop, isUnexpectedStopCandidate } from "./unexpected-stop-classifier";
 import { YieldQueue } from "./yield-queue";
 
+const SESSION_STOP_CONTINUATION_CAP = 8;
+
 /** Session-specific events that extend the core AgentEvent */
 export type AgentSessionEvent =
 	| AgentEvent
@@ -338,6 +345,24 @@ const UNEXPECTED_STOP_MAX_RETRIES = 3;
 const UNEXPECTED_STOP_TIMEOUT_MS = 4000;
 const EMPTY_STOP_MAX_RETRIES = 3;
 const RETRY_BACKOFF_MAX_DELAY_MS = 8_000;
+
+type CompactionCheckResult = Readonly<{
+	deferredHandoff: boolean;
+	continuationScheduled: boolean;
+}>;
+
+const COMPACTION_CHECK_NONE: CompactionCheckResult = {
+	deferredHandoff: false,
+	continuationScheduled: false,
+};
+const COMPACTION_CHECK_DEFERRED_HANDOFF: CompactionCheckResult = {
+	deferredHandoff: true,
+	continuationScheduled: true,
+};
+const COMPACTION_CHECK_CONTINUATION: CompactionCheckResult = {
+	deferredHandoff: false,
+	continuationScheduled: true,
+};
 export type CommandMetadataChangedListener = () => void | Promise<void>;
 export type AsyncJobSnapshotItem = Pick<AsyncJob, "id" | "type" | "status" | "label" | "startTime">;
 
@@ -555,6 +580,17 @@ export interface RoleModelCycle {
 	currentIndex: number;
 }
 
+export interface ContextUsageBreakdown {
+	contextWindow: number;
+	anchored: boolean;
+	usedTokens: number;
+	systemPromptTokens: number;
+	systemToolsTokens: number;
+	systemContextTokens: number;
+	skillsTokens: number;
+	messagesTokens: number;
+}
+
 /** Session statistics for /session command */
 export interface SessionStats {
 	sessionFile: string | undefined;
@@ -976,6 +1012,10 @@ const MAGIC_KEYWORD_NOTICE_TYPES: ReadonlySet<string> = new Set([
 	"workflow-notice",
 ]);
 
+/** Custom-message type of the hidden companion carrying vision descriptions of image
+ *  attachments sent to a text-only model (see `#buildImageDescriptionNotice`). */
+const IMAGE_ATTACHMENT_DESCRIPTION_TYPE = "image-attachment-description";
+
 /**
  * A hidden, user-attributed companion of a queued user prompt: the magic-keyword
  * notices (`ultrathink`/`orchestrate`/`workflow`) enqueued alongside the user
@@ -989,7 +1029,7 @@ function isHiddenUserCompanion(message: AgentMessage): boolean {
 		message.role === "custom" &&
 		message.attribution === "user" &&
 		message.display === false &&
-		MAGIC_KEYWORD_NOTICE_TYPES.has(message.customType)
+		(MAGIC_KEYWORD_NOTICE_TYPES.has(message.customType) || message.customType === IMAGE_ATTACHMENT_DESCRIPTION_TYPE)
 	);
 }
 
@@ -1044,6 +1084,8 @@ export class AgentSession {
 	 *  suppresses advisor concern/blocker auto-resume until the user next resumes.
 	 *  Advisor advice is still recorded into the transcript, just not auto-run. */
 	#advisorAutoResumeSuppressed = false;
+	#advisorPrimaryTurnsCompleted = 0;
+	#advisorInterruptImmuneTurnStart: number | undefined;
 	#planModeState: PlanModeState | undefined;
 	#goalModeState: GoalModeState | undefined;
 	#goalRuntime: GoalRuntime;
@@ -1224,15 +1266,20 @@ export class AgentSession {
 	#unexpectedStopRetryCount = 0;
 	#promptGeneration = 0;
 	#pendingAgentEndEmit: AgentSessionEvent | undefined;
-	#pendingProviderRequestNonMessageTokens: number | undefined = undefined;
-	#lastProviderUsageNonMessage:
+	#pendingContextSnapshot:
 		| {
-				provider: AssistantMessage["provider"];
-				model: AssistantMessage["model"];
-				timestamp: AssistantMessage["timestamp"];
-				tokens: number;
+				promptTokens: number;
+				nonMessageTokens: number;
+				cutoffCount: number;
 		  }
-		| undefined;
+		| undefined = undefined;
+	#sessionStopContinuationCount = 0;
+	#sessionStopHookActive = false;
+	// Bumped whenever the pending in-flight snapshot is set/cleared. The
+	// status-line context memo includes this so clearing the snapshot on
+	// turn-end/abort invalidates the cache even though the message list is
+	// unchanged — otherwise a mid-turn estimate would survive into idle.
+	#contextUsageRevision = 0;
 	#obfuscator: SecretObfuscator | undefined;
 	#checkpointState: CheckpointState | undefined = undefined;
 	#pendingRewindReport: string | undefined = undefined;
@@ -1476,6 +1523,7 @@ export class AgentSession {
 		this.agent.setRawSseEventInterceptor(this.#onSseEvent);
 		this.agent.setOnTurnEnd(async (messages, signal) => {
 			if (signal?.aborted) return;
+			this.#advisorPrimaryTurnsCompleted++;
 			if (this.#advisorRuntime && !this.#advisorRuntime.disposed) {
 				this.#advisorRuntime.onTurnEnd(messages);
 				const syncBacklog = this.settings.get("advisor.syncBacklog");
@@ -1608,6 +1656,27 @@ export class AgentSession {
 	// -------------------------------------------------------------------------
 	// Advisor runtime lifecycle
 	// -------------------------------------------------------------------------
+	#advisorImmuneTurnLimit(): number {
+		const immuneTurns = this.settings.get("advisor.immuneTurns") as number;
+		if (!Number.isFinite(immuneTurns) || immuneTurns <= 0) return 0;
+		return Math.trunc(immuneTurns);
+	}
+
+	#isAdvisorInterruptImmuneTurnActive(): boolean {
+		return isAdvisorInterruptImmuneTurnActive({
+			completedTurns: this.#advisorPrimaryTurnsCompleted,
+			immuneTurnStart: this.#advisorInterruptImmuneTurnStart,
+			immuneTurns: this.#advisorImmuneTurnLimit(),
+		});
+	}
+
+	// The next primary turn number starts the immune-turn window. While the
+	// interrupting steer is still in flight, completedTurns is lower than this
+	// start, so duplicate concern/blocker advice is also downgraded.
+	#recordAdvisorInterruptDelivered(): void {
+		this.#advisorInterruptImmuneTurnStart = this.#advisorPrimaryTurnsCompleted + 1;
+	}
+
 	#buildAdvisorRuntime(seedToCurrent = false): boolean {
 		if (this.#isDisposed) return false;
 		if (this.#advisorRuntime) return true;
@@ -1637,6 +1706,7 @@ export class AgentSession {
 		// strand the advice and dump the backlog as one burst at the next prompt. A
 		// plain nit always rides the non-interrupting YieldQueue aside.
 		const enqueueAdvice = (note: string, severity?: AdvisorSeverity) => {
+			const interrupting = isInterruptingSeverity(severity);
 			const channel = resolveAdvisorDeliveryChannel({
 				severity,
 				autoResumeSuppressed: this.#advisorAutoResumeSuppressed,
@@ -1647,6 +1717,7 @@ export class AgentSession {
 				// auto-resume it despite the user's interrupt.
 				streaming: this.agent.state.isStreaming,
 				aborting: this.#abortInProgress,
+				interruptImmuneTurnActive: interrupting && this.#isAdvisorInterruptImmuneTurnActive(),
 			});
 			if (channel === "aside") {
 				this.yieldQueue.enqueue("advisor", { note, severity });
@@ -1667,6 +1738,7 @@ export class AgentSession {
 				});
 				return;
 			}
+			this.#recordAdvisorInterruptDelivered();
 			void this.sendCustomMessage(
 				{ customType: "advisor", content, display: true, attribution: "agent", details },
 				{ deliverAs: "steer", triggerTurn: true },
@@ -1682,6 +1754,7 @@ export class AgentSession {
 		if (this.#advisorWatchdogPrompt) {
 			systemPrompt.push(this.#advisorWatchdogPrompt);
 		}
+		const advisorSessionId = this.sessionId ? `${this.sessionId}-advisor` : undefined;
 		const advisorAgent = new Agent({
 			initialState: {
 				systemPrompt,
@@ -1690,15 +1763,8 @@ export class AgentSession {
 				tools: [adviseTool, ...advisorReadOnlyTools],
 			},
 			appendOnlyContext,
-			sessionId: this.sessionId ? `${this.sessionId}-advisor` : undefined,
-			getApiKey: async provider => {
-				const key = await this.#modelRegistry.getApiKeyForProvider(
-					provider,
-					this.sessionId ? `${this.sessionId}-advisor` : undefined,
-				);
-				if (!key) throw new Error(`No API key for advisor provider "${provider}"`);
-				return key;
-			},
+			sessionId: advisorSessionId,
+			getApiKey: requestModel => this.#modelRegistry.resolver(requestModel, advisorSessionId),
 			intentTracing: false,
 		});
 		advisorAgent.setDisableReasoning(shouldDisableReasoning(advisorThinkingLevel));
@@ -2354,6 +2420,15 @@ export class AgentSession {
 				event.message.role === "fileMention"
 			) {
 				// Regular LLM message - persist as SessionMessageEntry
+				if (event.message.role === "assistant") {
+					const assistantMsg = event.message as AssistantMessage;
+					if (assistantMsg.stopReason !== "aborted" && assistantMsg.stopReason !== "error" && assistantMsg.usage) {
+						assistantMsg.contextSnapshot = {
+							promptTokens: calculatePromptTokens(assistantMsg.usage),
+							nonMessageTokens: this.#pendingContextSnapshot?.nonMessageTokens ?? computeNonMessageTokens(this),
+						};
+					}
+				}
 				this.sessionManager.appendMessage(event.message);
 			}
 			// Other message types (bashExecution, compactionSummary, branchSummary) are persisted elsewhere
@@ -2362,14 +2437,6 @@ export class AgentSession {
 			if (event.message.role === "assistant") {
 				this.#lastAssistantMessage = event.message;
 				const assistantMsg = event.message as AssistantMessage;
-				if (assistantMsg.stopReason !== "aborted" && assistantMsg.stopReason !== "error" && assistantMsg.usage) {
-					this.#lastProviderUsageNonMessage = {
-						provider: assistantMsg.provider,
-						model: assistantMsg.model,
-						timestamp: assistantMsg.timestamp,
-						tokens: this.#pendingProviderRequestNonMessageTokens ?? computeNonMessageTokens(this),
-					};
-				}
 				const currentGrantsAnthropicPriority =
 					this.serviceTier === "priority" || this.serviceTier === "claude-only";
 				if (assistantMsg.disabledFeatures?.includes("priority") && currentGrantsAnthropicPriority) {
@@ -2412,7 +2479,6 @@ export class AgentSession {
 					this.#retryAttempt = 0;
 				}
 			}
-
 			if (event.message.role === "toolResult") {
 				const { toolName, details, isError, content } = event.message as {
 					toolName?: string;
@@ -2472,6 +2538,9 @@ export class AgentSession {
 
 		// Check auto-retry and auto-compaction after agent completes
 		if (event.type === "agent_end") {
+			const emitAgentEndNotification = async () => {
+				await this.#emitAgentEndNotification(event.messages);
+			};
 			const usage = this.getSessionStats().tokens;
 			await this.#goalRuntime.onAgentEnd({
 				currentUsage: {
@@ -2488,6 +2557,7 @@ export class AgentSession {
 			this.#lastAssistantMessage = undefined;
 			if (!msg) {
 				this.#lastSuccessfulYieldToolCallId = undefined;
+				await emitAgentEndNotification();
 				return;
 			}
 
@@ -2504,60 +2574,81 @@ export class AgentSession {
 			if (this.#skipPostTurnMaintenanceAssistantTimestamp === msg.timestamp) {
 				this.#skipPostTurnMaintenanceAssistantTimestamp = undefined;
 				this.#lastSuccessfulYieldToolCallId = undefined;
+				await emitAgentEndNotification();
 				return;
 			}
 
 			if (this.#assistantEndedWithSuccessfulYield(msg)) {
 				this.#lastSuccessfulYieldToolCallId = undefined;
+				await emitAgentEndNotification();
 				return;
 			}
 			this.#lastSuccessfulYieldToolCallId = undefined;
 
 			if (await this.#handleEmptyAssistantStop(msg)) {
+				await emitAgentEndNotification();
 				return;
 			}
 			if (await this.#handleUnexpectedAssistantStop(msg)) {
+				await emitAgentEndNotification();
 				return;
 			}
 
 			if (this.#isRetryableReasonlessAbort(msg)) {
 				const didRetry = await this.#handleRetryableError(msg, { allowModelFallback: false });
-				if (didRetry) return;
+				if (didRetry) {
+					await emitAgentEndNotification();
+					return;
+				}
 			}
 
 			// A deliberate abort should settle the current turn, not trigger queued continuations.
 			if (msg.stopReason === "aborted") {
 				this.#resolveRetry();
+				this.#resetSessionStopContinuationState();
+				await emitAgentEndNotification();
 				return;
 			}
 			// Check for retryable errors first (overloaded, rate limit, server errors)
 			if (this.#isRetryableError(msg)) {
 				const didRetry = await this.#handleRetryableError(msg);
-				if (didRetry) return; // Retry was initiated, don't proceed to compaction
+				if (didRetry) {
+					await emitAgentEndNotification();
+					return;
+				}
 			}
 			this.#resolveRetry();
 
 			const compactionTask = this.#checkCompaction(msg);
 			this.#trackPostPromptTask(compactionTask);
-			const compactionDeferredHandoff = await compactionTask;
+			const compactionResult = await compactionTask;
 			// Check for incomplete todos only after a final assistant stop, not intermediate tool-use turns.
 			const hasToolCalls = msg.content.some(content => content.type === "toolCall");
 			if (hasToolCalls) {
+				await emitAgentEndNotification();
 				return;
 			}
-			// When checkCompaction scheduled a deferred handoff, skip the rewind/todo passes:
-			// any reminder we append here would race the handoff's session reset, and
-			// #scheduleAgentContinue would start a fresh streaming turn alongside the handoff
-			// LLM call (visible as "Auto-handoff" loader + an assistant message still streaming).
-			if (compactionDeferredHandoff) {
+			// When compaction queued recovery, skip the rewind/todo/session_stop passes:
+			// any reminder or hook continuation we append here would race the handoff,
+			// retry, auto-continue prompt, or queued-message drain that already owns the
+			// next turn.
+			if (compactionResult.deferredHandoff || compactionResult.continuationScheduled) {
+				await emitAgentEndNotification();
 				return;
 			}
 			if (msg.stopReason !== "error") {
 				if (this.#enforceRewindBeforeYield()) {
+					await emitAgentEndNotification();
+					return;
+				}
+				const todoContinuationScheduled = await this.#checkTodoCompletion();
+				if (todoContinuationScheduled) {
+					await emitAgentEndNotification();
 					return;
 				}
-				await this.#checkTodoCompletion();
 			}
+			await this.#emitSessionStopEvent(event.messages);
+			await emitAgentEndNotification();
 		}
 	};
 
@@ -3513,6 +3604,83 @@ export class AgentSession {
 		}
 	}
 
+	#resetSessionStopContinuationState(): void {
+		this.#sessionStopContinuationCount = 0;
+		this.#sessionStopHookActive = false;
+	}
+
+	#clearPendingSessionStopContinuations(): void {
+		if (!this.#pendingNextTurnMessages.some(message => message.customType === "session-stop-continuation")) {
+			return;
+		}
+		this.#pendingNextTurnMessages = this.#pendingNextTurnMessages.filter(
+			message => message.customType !== "session-stop-continuation",
+		);
+	}
+
+	#sessionStopContinuationContext(result: SessionStopEventResult | undefined): string | undefined {
+		if (!result) return undefined;
+		const additionalContext =
+			typeof result.additionalContext === "string" && result.additionalContext.length > 0
+				? result.additionalContext
+				: undefined;
+		const reason = typeof result.reason === "string" && result.reason.length > 0 ? result.reason : undefined;
+		if (result.continue === true) {
+			return additionalContext ?? reason;
+		}
+		if (result.decision === "block") {
+			return reason ?? additionalContext;
+		}
+		return undefined;
+	}
+
+	async #emitAgentEndNotification(messages: AgentMessage[]): Promise<void> {
+		await this.#extensionRunner?.emit({ type: "agent_end", messages });
+	}
+
+	async #emitSessionStopEvent(messages: AgentMessage[]): Promise<void> {
+		if (this.#agentKind === "sub" || !this.#extensionRunner?.hasHandlers("session_stop")) return;
+		const generation = this.#promptGeneration;
+		const result = await this.#extensionRunner.emitSessionStop({
+			messages,
+			turn_id: Math.max(0, this.#turnIndex - 1),
+			last_assistant_message: this.getLastAssistantMessage(),
+			session_id: this.sessionId,
+			session_file: this.sessionFile,
+			stop_hook_active: this.#sessionStopHookActive,
+		});
+		if (this.#promptGeneration !== generation || this.#abortInProgress || this.#isDisposed) {
+			this.#resetSessionStopContinuationState();
+			return;
+		}
+		const additionalContext = this.#sessionStopContinuationContext(result);
+		if (!additionalContext) {
+			this.#resetSessionStopContinuationState();
+			return;
+		}
+		if (this.#sessionStopContinuationCount >= SESSION_STOP_CONTINUATION_CAP) {
+			logger.warn("session_stop continuation cap reached", {
+				sessionId: this.sessionId,
+				cap: SESSION_STOP_CONTINUATION_CAP,
+			});
+			this.#resetSessionStopContinuationState();
+			return;
+		}
+		this.#sessionStopContinuationCount++;
+		this.#sessionStopHookActive = true;
+		this.#queueHiddenNextTurnMessage(
+			{
+				role: "custom",
+				customType: "session-stop-continuation",
+				content: additionalContext,
+				display: false,
+				attribution: "agent",
+				timestamp: Date.now(),
+			},
+			true,
+		);
+	}
+
 	/** Emit extension events based on session events */
 	async #emitExtensionEvent(event: AgentSessionEvent): Promise<void> {
 		if (!this.#extensionRunner) return;
@@ -3520,7 +3688,9 @@ export class AgentSession {
 			this.#turnIndex = 0;
 			await this.#extensionRunner.emit({ type: "agent_start" });
 		} else if (event.type === "agent_end") {
-			await this.#extensionRunner.emit({ type: "agent_end", messages: event.messages });
+			// `agent_end` extension notification is emitted from the settled
+			// agent_end maintenance path so `session_stop` control hooks are not
+			// blocked by unrelated notification-only work.
 		} else if (event.type === "turn_start") {
 			const hookEvent: TurnStartEvent = {
 				type: "turn_start",
@@ -4791,11 +4961,24 @@ export class AgentSession {
 			openrouterRoutingPreset !== "default" && options.openrouterVariant === undefined
 				? openrouterRoutingPreset
 				: undefined;
-		if (!sessionOnPayload && !sessionOnResponse && !sessionMetadata && !sessionOnSseEvent && !openrouterVariant)
+		const antigravityEndpointMode =
+			provider === "google-antigravity" ? this.settings.get("providers.antigravityEndpoint") : undefined;
+
+		if (
+			!sessionOnPayload &&
+			!sessionOnResponse &&
+			!sessionMetadata &&
+			!sessionOnSseEvent &&
+			!openrouterVariant &&
+			!antigravityEndpointMode
+		)
 			return options;
 
-		const preparedOptions: SimpleStreamOptions =
-			openrouterVariant === undefined ? { ...options } : { ...options, openrouterVariant };
+		const preparedOptions: SimpleStreamOptions = {
+			...options,
+			...(openrouterVariant !== undefined && { openrouterVariant }),
+			...(antigravityEndpointMode !== undefined && { antigravityEndpointMode }),
+		};
 
 		// Stamp session metadata (e.g. user_id={session_id}) onto direct-call requests so
 		// they share the same session bucket as Agent.prompt-routed requests on Anthropic
@@ -5114,6 +5297,62 @@ export class AgentSession {
 		return normalizeModelContextImages(images, { model: this.model });
 	}
 
+	/**
+	 * Build a hidden companion message describing image attachments for a text-only
+	 * model. Each image is saved under local:// and a vision-capable model describes
+	 * it; the descriptions are returned as a `display: false` custom message (so the
+	 * model reads them but the TUI does not render the blob) carrying one
+	 * `<image path="local://…">…</image>` block per image. Returns `undefined` when
+	 * the active model already accepts images, the feature is disabled, or no
+	 * description could be produced. Never throws.
+	 */
+	async #buildImageDescriptionNotice(
+		normalizedImages: ImageContent[],
+		signal?: AbortSignal,
+	): Promise<CustomMessage | undefined> {
+		const model = this.model;
+		const shouldDescribe =
+			!!model &&
+			!model.input.includes("image") &&
+			!this.settings.get("images.blockImages") &&
+			this.settings.get("images.describeForTextModels");
+		if (!shouldDescribe || !model) {
+			return undefined;
+		}
+		let blocks: TextContent[];
+		try {
+			blocks = await describeAttachedImagesForTextModel(
+				normalizedImages,
+				{
+					activeModel: model,
+					modelRegistry: this.#modelRegistry,
+					settings: this.settings,
+					localProtocolOptions: this.#localProtocolOptions(),
+					activeModelString: formatModelString(model),
+					telemetryConfig: this.agent.telemetry,
+					sessionId: this.sessionId,
+				},
+				signal,
+			);
+		} catch (err) {
+			logger.warn("image attachment vision fallback failed; image left undescribed", {
+				error: err instanceof Error ? err.message : String(err),
+			});
+			return undefined;
+		}
+		if (blocks.length === 0) {
+			return undefined;
+		}
+		return {
+			role: "custom",
+			customType: IMAGE_ATTACHMENT_DESCRIPTION_TYPE,
+			content: blocks,
+			display: false,
+			attribution: "user",
+			timestamp: Date.now(),
+		};
+	}
+
 	async #normalizeMessageContentImages(
 		content: string | (TextContent | ImageContent)[],
 	): Promise<string | (TextContent | ImageContent)[]> {
@@ -5261,9 +5500,14 @@ export class AgentSession {
 		const normalizedImages = await this.#normalizeImagesForModel(options?.images);
 
 		const userContent: (TextContent | ImageContent)[] = [{ type: "text", text: expandedText }];
-		if (normalizedImages) {
+		if (normalizedImages?.length) {
 			userContent.push(...normalizedImages);
 		}
+		// Text-only model + image attachment: describe via a vision model and inject the
+		// description as a hidden companion (the image stays in the visible user message).
+		const imageDescriptionNotice = normalizedImages?.length
+			? await this.#buildImageDescriptionNotice(normalizedImages)
+			: undefined;
 
 		const promptAttribution = options?.attribution ?? (options?.synthetic ? "agent" : "user");
 		const message = options?.synthetic
@@ -5288,8 +5532,8 @@ export class AgentSession {
 				...options,
 				images: normalizedImages,
 				prependMessages:
-					preludeMessages.length > 0 || keywordNotices.length > 0
-						? [...preludeMessages, ...keywordNotices]
+					preludeMessages.length > 0 || keywordNotices.length > 0 || imageDescriptionNotice
+						? [...preludeMessages, ...keywordNotices, ...(imageDescriptionNotice ? [imageDescriptionNotice] : [])]
 						: undefined,
 			});
 		} finally {
@@ -5510,11 +5754,23 @@ export class AgentSession {
 			}
 
 			const agentPromptOptions = options?.toolChoice ? { toolChoice: options.toolChoice } : undefined;
-			this.#pendingProviderRequestNonMessageTokens = computeNonMessageTokens(this);
+			const nonMessageTokens = computeNonMessageTokens(this);
+			const contextWindow = this.model?.contextWindow ?? 0;
+			const breakdown = this.getContextBreakdown({ contextWindow, pendingMessages: messages });
+			const promptTokens =
+				breakdown?.usedTokens ??
+				nonMessageTokens +
+					this.messages.reduce((sum, msg) => sum + estimateTokens(msg), 0) +
+					messages.reduce((sum, msg) => sum + estimateTokens(msg), 0);
+			this.#setPendingContextSnapshot({
+				promptTokens,
+				nonMessageTokens,
+				cutoffCount: this.messages.length + messages.length,
+			});
 			try {
 				await this.#promptAgentWithIdleRetry(messages, agentPromptOptions);
 			} finally {
-				this.#pendingProviderRequestNonMessageTokens = undefined;
+				this.#setPendingContextSnapshot(undefined);
 			}
 			if (!options?.skipPostPromptRecoveryWait) {
 				await this.#waitForPostPromptRecovery(generation);
@@ -5699,7 +5955,13 @@ export class AgentSession {
 		if (normalizedImages?.length) {
 			content.push(...normalizedImages);
 		}
+		// Text-only model + image attachment: describe via a vision model and enqueue the
+		// description as a hidden companion immediately before the user message.
+		const imageDescriptionNotice = normalizedImages?.length
+			? await this.#buildImageDescriptionNotice(normalizedImages)
+			: undefined;
 		if (mode === "followUp") {
+			if (imageDescriptionNotice) this.agent.followUp(imageDescriptionNotice);
 			this.agent.followUp({
 				role: "user",
 				content,
@@ -5707,6 +5969,7 @@ export class AgentSession {
 				timestamp: Date.now(),
 			});
 		} else {
+			if (imageDescriptionNotice) this.agent.steer(imageDescriptionNotice);
 			this.agent.steer({
 				role: "user",
 				content,
@@ -5857,6 +6120,16 @@ export class AgentSession {
 		}
 	}
 
+	async #promptAgentInitiatedMessage(message: CustomMessage): Promise<void> {
+		this.#beginInFlight();
+		try {
+			await this.agent.prompt(message);
+			await this.#waitForPostPromptRecovery();
+		} finally {
+			this.#endInFlight();
+		}
+	}
+
 	/**
 	 * Send a custom message to the session. Creates a CustomMessageEntry.
 	 *
@@ -5916,7 +6189,7 @@ export class AgentSession {
 					this.#queueHiddenNextTurnMessage(normalizedAppMessage, false);
 					return false;
 				}
-				await this.agent.prompt(normalizedAppMessage);
+				await this.#promptAgentInitiatedMessage(normalizedAppMessage);
 				return true;
 			}
 			this.agent.appendMessage(normalizedAppMessage);
@@ -5935,7 +6208,7 @@ export class AgentSession {
 				this.#queueHiddenNextTurnMessage(normalizedAppMessage, false);
 				return false;
 			}
-			await this.agent.prompt(normalizedAppMessage);
+			await this.#promptAgentInitiatedMessage(normalizedAppMessage);
 			return true;
 		}
 
@@ -6158,6 +6431,8 @@ export class AgentSession {
 			// block runs, but nested prompt setup/finalizers may still be unwinding. Without this,
 			// a subsequent prompt() can incorrectly observe the session as busy after an abort.
 			this.#resetInFlight();
+			this.#resetSessionStopContinuationState();
+			this.#clearPendingSessionStopContinuations();
 			// Safety net: if the agent loop aborted without producing an assistant
 			// message (e.g. failed before the first stream), the in-flight yield was
 			// never resolved or rejected by the normal message_end path. Reject it now
@@ -7458,39 +7733,12 @@ export class AgentSession {
 		}
 	}
 
-	#estimatePendingPromptTokens(messages: AgentMessage[]): number {
-		let tokens = computeNonMessageTokens(this);
-		for (const message of this.messages) {
-			tokens += estimateTokens(message);
-		}
-		for (const message of messages) {
-			tokens += estimateTokens(message);
-		}
-		return tokens;
-	}
-
 	#estimatePrePromptContextTokens(messages: AgentMessage[], contextWindow: number): number {
-		const currentUsage = this.getContextUsage({ contextWindow });
-		if (typeof currentUsage?.tokens !== "number" || !Number.isFinite(currentUsage.tokens)) {
-			return this.#estimatePendingPromptTokens(messages);
-		}
-
-		const currentEstimate = this.#estimateContextTokens();
-		if (!currentEstimate.providerAnchored) {
-			return this.#estimatePendingPromptTokens(messages);
-		}
-
-		let tokens = currentUsage.tokens;
-		const previousNonMessageTokens = currentEstimate.providerNonMessageTokens;
-		if (previousNonMessageTokens !== undefined) {
-			const currentNonMessageTokens = computeNonMessageTokens(this);
-			const nonMessageTokenGrowth = Math.max(0, currentNonMessageTokens - previousNonMessageTokens);
-			tokens += nonMessageTokenGrowth;
-		}
-		for (const message of messages) {
-			tokens += estimateTokens(message);
-		}
-		return tokens;
+		const breakdown = this.getContextBreakdown({ contextWindow, pendingMessages: messages });
+		return (
+			breakdown?.usedTokens ??
+			computeNonMessageTokens(this) + messages.reduce((sum, msg) => sum + estimateTokens(msg), 0)
+		);
 	}
 
 	async #runPrePromptCompactionIfNeeded(messages: AgentMessage[]): Promise<void> {
@@ -7544,19 +7792,19 @@ export class AgentSession {
 	 *   on the pre-prompt path (where the next agent turn is about to start) set it to false
 	 *   to avoid racing the deferred handoff against the new turn.
 	 * @param autoContinue Whether maintenance may schedule the agent-authored continuation prompt.
-	 * @returns true when a deferred handoff was scheduled. Callers MUST then skip any
-	 *   subsequent `#scheduleAgentContinue` / reminder appends for this turn — the
-	 *   handoff will replace session state and a concurrent `agent.continue()` would
-	 *   stream into the soon-to-be-discarded session.
+	 * @returns whether compaction/recovery scheduled a handoff, retry, auto-continue, or
+	 *   queued-message drain that already owns the next turn. Callers MUST skip
+	 *   `session_stop` and other agent continuations when `continuationScheduled`
+	 *   is true.
 	 */
 	async #checkCompaction(
 		assistantMessage: AssistantMessage,
 		skipAbortedCheck = true,
 		allowDefer = true,
 		autoContinue = true,
-	): Promise<boolean> {
+	): Promise<CompactionCheckResult> {
 		// Skip if message was aborted (user cancelled) - unless skipAbortedCheck is false
-		if (skipAbortedCheck && assistantMessage.stopReason === "aborted") return false;
+		if (skipAbortedCheck && assistantMessage.stopReason === "aborted") return COMPACTION_CHECK_NONE;
 		const contextWindow = this.model?.contextWindow ?? 0;
 		const generation = this.#promptGeneration;
 		// Skip overflow check if the message came from a different model.
@@ -7585,15 +7833,15 @@ export class AgentSession {
 			if (promoted) {
 				// Retry on the promoted (larger) model without compacting
 				this.#scheduleAgentContinue({ delayMs: 100, generation });
-				return false;
+				return COMPACTION_CHECK_CONTINUATION;
 			}
 
 			// No promotion target available fall through to compaction
 			const compactionSettings = this.settings.getGroup("compaction");
 			if (compactionSettings.enabled && compactionSettings.strategy !== "off") {
-				await this.#runAutoCompaction("overflow", true, false, allowDefer, { autoContinue });
+				return await this.#runAutoCompaction("overflow", true, false, allowDefer, { autoContinue });
 			}
-			return false;
+			return COMPACTION_CHECK_NONE;
 		}
 
 		// Case 3: Output-side incomplete — `response.incomplete` from OpenAI Responses
@@ -7614,7 +7862,7 @@ export class AgentSession {
 					from: `${assistantMessage.provider}/${assistantMessage.model}`,
 				});
 				this.#scheduleAgentContinue({ delayMs: 100, generation });
-				return false;
+				return COMPACTION_CHECK_CONTINUATION;
 			}
 
 			const incompleteCompactionSettings = this.settings.getGroup("compaction");
@@ -7623,18 +7871,17 @@ export class AgentSession {
 					model: `${assistantMessage.provider}/${assistantMessage.model}`,
 					strategy: incompleteCompactionSettings.strategy,
 				});
-				await this.#runAutoCompaction("incomplete", true, false, allowDefer, {
+				return await this.#runAutoCompaction("incomplete", true, false, allowDefer, {
 					autoContinue,
 					triggerContextTokens: calculateContextTokens(assistantMessage.usage),
 				});
-			} else {
-				// Neither promotion nor compaction is available — surface the dead-end so
-				// the user understands why the turn yielded with nothing.
-				logger.warn("response.incomplete with no recovery path (promotion + compaction both unavailable)", {
-					model: `${assistantMessage.provider}/${assistantMessage.model}`,
-				});
 			}
-			return false;
+			// Neither promotion nor compaction is available — surface the dead-end so
+			// the user understands why the turn yielded with nothing.
+			logger.warn("response.incomplete with no recovery path (promotion + compaction both unavailable)", {
+				model: `${assistantMessage.provider}/${assistantMessage.model}`,
+			});
+			return COMPACTION_CHECK_NONE;
 		}
 
 		// Stale-result pass runs every turn, before any threshold gating: it is
@@ -7643,11 +7890,11 @@ export class AgentSession {
 		const supersedeResult = await this.#pruneStaleToolResults();
 
 		const compactionSettings = this.settings.getGroup("compaction");
-		if (!compactionSettings.enabled || compactionSettings.strategy === "off") return false;
+		if (!compactionSettings.enabled || compactionSettings.strategy === "off") return COMPACTION_CHECK_NONE;
 
 		// Case 4: Threshold - turn succeeded but context is getting large
 		// Skip if this was an error (non-overflow errors don't have usage data)
-		if (assistantMessage.stopReason === "error") return false;
+		if (assistantMessage.stopReason === "error") return COMPACTION_CHECK_NONE;
 		const pruneResult = await this.#pruneToolOutputs();
 		let contextTokens = calculateContextTokens(assistantMessage.usage);
 		if (supersedeResult) {
@@ -7666,7 +7913,7 @@ export class AgentSession {
 				});
 			}
 		}
-		return false;
+		return COMPACTION_CHECK_NONE;
 	}
 	#assistantEndedWithSuccessfulYield(assistantMessage: AssistantMessage): boolean {
 		const toolCallId = this.#lastSuccessfulYieldToolCallId;
@@ -7706,7 +7953,7 @@ export class AgentSession {
 			if (assistantMessage.stopReason === "toolUse") {
 				this.#removeEmptyStopFromActiveContext(assistantMessage);
 			}
-			return true;
+			return false;
 		}
 		this.#removeEmptyStopFromActiveContext(assistantMessage);
 		this.agent.appendMessage({
@@ -8081,12 +8328,12 @@ export class AgentSession {
 	/**
 	 * Check if agent stopped with incomplete todos and prompt to continue.
 	 */
-	async #checkTodoCompletion(): Promise<void> {
+	async #checkTodoCompletion(): Promise<boolean> {
 		// Skip todo reminders when the most recent turn was driven by an explicit user force —
 		// the user wanted exactly that tool, not a follow-up nag about incomplete todos.
 		const lastServedLabel = this.#toolChoiceQueue.consumeLastServedLabel();
 		if (lastServedLabel === "user-force") {
-			return;
+			return false;
 		}
 
 		// Suppress within a self-continuation chain: if the agent's last turn was driven by a
@@ -8097,7 +8344,7 @@ export class AgentSession {
 			logger.debug("Todo completion: prior reminder still awaiting agent action; staying silent", {
 				attempt: this.#todoReminderCount,
 			});
-			return;
+			return false;
 		}
 
 		const remindersEnabled = this.settings.get("todo.reminders");
@@ -8105,20 +8352,20 @@ export class AgentSession {
 		if (!remindersEnabled || !todosEnabled) {
 			this.#todoReminderCount = 0;
 			this.#todoReminderAwaitingProgress = false;
-			return;
+			return false;
 		}
 
 		const remindersMax = this.settings.get("todo.reminders.max");
 		if (this.#todoReminderCount >= remindersMax) {
 			logger.debug("Todo completion: max reminders reached", { count: this.#todoReminderCount });
-			return;
+			return false;
 		}
 
 		const phases = this.getTodoPhases();
 		if (phases.length === 0) {
 			this.#todoReminderCount = 0;
 			this.#todoReminderAwaitingProgress = false;
-			return;
+			return false;
 		}
 
 		const incompleteByPhase = phases
@@ -8136,7 +8383,7 @@ export class AgentSession {
 		if (incomplete.length === 0) {
 			this.#todoReminderCount = 0;
 			this.#todoReminderAwaitingProgress = false;
-			return;
+			return false;
 		}
 
 		// Build reminder message
@@ -8164,15 +8411,19 @@ export class AgentSession {
 			maxAttempts: remindersMax,
 		});
 
-		this.#todoReminderAwaitingProgress = true;
-		// Inject reminder and continue the conversation
-		this.agent.appendMessage({
+		const reminderMessage: Message = {
 			role: "developer",
 			content: [{ type: "text", text: reminder }],
 			attribution: "agent",
 			timestamp: Date.now(),
-		});
+		};
+
+		this.#todoReminderAwaitingProgress = true;
+		// Inject reminder and persist it so the JSONL transcript matches model context.
+		this.agent.appendMessage(reminderMessage);
+		this.sessionManager.appendMessage(reminderMessage);
 		this.#scheduleAgentContinue({ generation: this.#promptGeneration });
+		return true;
 	}
 
 	/**
@@ -8458,9 +8709,13 @@ export class AgentSession {
 	}
 
 	#didSessionMessagesChange(previousMessages: AgentMessage[], nextMessages: AgentMessage[]): boolean {
-		return (
-			JSON.stringify(previousMessages.map(message => this.#normalizeSessionMessageForProviderReplay(message))) !==
-			JSON.stringify(nextMessages.map(message => this.#normalizeSessionMessageForProviderReplay(message)))
+		if (previousMessages.length !== nextMessages.length) return true;
+		return previousMessages.some(
+			(message, i) =>
+				!Bun.deepEquals(
+					this.#normalizeSessionMessageForProviderReplay(message),
+					this.#normalizeSessionMessageForProviderReplay(nextMessages[i]),
+				),
 		);
 	}
 
@@ -8706,14 +8961,14 @@ export class AgentSession {
 	 * Internal: Run auto-compaction with events.
 	 *
 	 * @param allowDefer If true (default), threshold-driven handoff strategy is allowed to
-	 *   schedule itself as a deferred post-prompt task and return `true` immediately. The
-	 *   caller MUST treat that as "compaction will happen async — do not also schedule
-	 *   `agent.continue()` for this turn", otherwise the deferred handoff races a fresh
-	 *   streaming turn (the symptom: "Auto-handoff" loader + assistant message still
-	 *   streaming). Callers on a path that is about to start a new agent turn (e.g.
-	 *   the pre-prompt check in `#promptWithMessage`) pass `false` to force inline
-	 *   execution so the handoff completes before the new turn begins.
-	 * @returns true when a deferred handoff was scheduled. Inline runs always return false.
+	 *   schedule itself as a deferred post-prompt task and return a deferred-handoff result
+	 *   immediately. The caller MUST treat that as "compaction will happen async — do not
+	 *   also schedule `agent.continue()` for this turn", otherwise the deferred handoff
+	 *   races a fresh streaming turn (the symptom: "Auto-handoff" loader + assistant
+	 *   message still streaming). Callers on a path that is about to start a new agent
+	 *   turn (e.g. the pre-prompt check in `#promptWithMessage`) pass `false` to force
+	 *   inline execution so the handoff completes before the new turn begins.
+	 * @returns whether auto-compaction scheduled a follow-up turn.
 	 */
 	async #runAutoCompaction(
 		reason: "overflow" | "threshold" | "idle" | "incomplete",
@@ -8721,10 +8976,10 @@ export class AgentSession {
 		deferred = false,
 		allowDefer = true,
 		options: { autoContinue?: boolean; triggerContextTokens?: number } = {},
-	): Promise<boolean> {
+	): Promise<CompactionCheckResult> {
 		const compactionSettings = this.settings.getGroup("compaction");
-		if (compactionSettings.strategy === "off") return false;
-		if (reason !== "idle" && !compactionSettings.enabled) return false;
+		if (compactionSettings.strategy === "off") return COMPACTION_CHECK_NONE;
+		if (reason !== "idle" && !compactionSettings.enabled) return COMPACTION_CHECK_NONE;
 		const generation = this.#promptGeneration;
 		const shouldAutoContinue = options.autoContinue !== false && compactionSettings.autoContinue !== false;
 		// Shake runs inline (cheap, no remote LLM). On overflow recovery, if shake
@@ -8738,7 +8993,7 @@ export class AgentSession {
 				shouldAutoContinue,
 				options.triggerContextTokens,
 			);
-			if (outcome !== "fallback") return false;
+			if (outcome !== "fallback") return outcome;
 		}
 		// "overflow" and "incomplete" force inline execution because they are recovery
 		// paths the caller wants resolved before scheduling the next turn. "idle" is
@@ -8759,7 +9014,7 @@ export class AgentSession {
 				},
 				{ generation },
 			);
-			return true;
+			return COMPACTION_CHECK_DEFERRED_HANDOFF;
 		}
 
 		// "overflow" forces context-full because the input itself is broken — a handoff
@@ -8807,7 +9062,7 @@ export class AgentSession {
 							aborted: true,
 							willRetry: false,
 						});
-						return false;
+						return COMPACTION_CHECK_NONE;
 					}
 					logger.warn("Auto-handoff returned no document; falling back to context-full maintenance", {
 						reason,
@@ -8822,10 +9077,11 @@ export class AgentSession {
 						aborted: false,
 						willRetry: false,
 					});
-					if (!autoCompactionSignal.aborted && reason !== "idle" && shouldAutoContinue) {
+					const continuationScheduled = !autoCompactionSignal.aborted && reason !== "idle" && shouldAutoContinue;
+					if (continuationScheduled) {
 						this.#scheduleAutoContinuePrompt(generation);
 					}
-					return false;
+					return continuationScheduled ? COMPACTION_CHECK_CONTINUATION : COMPACTION_CHECK_NONE;
 				}
 			}
 
@@ -8838,7 +9094,7 @@ export class AgentSession {
 					willRetry: false,
 					skipped: true,
 				});
-				return false;
+				return COMPACTION_CHECK_NONE;
 			}
 
 			const availableModels = this.#modelRegistry.getAvailable();
@@ -8851,7 +9107,7 @@ export class AgentSession {
 					willRetry: false,
 					skipped: true,
 				});
-				return false;
+				return COMPACTION_CHECK_NONE;
 			}
 
 			const pathEntries = this.sessionManager.getBranch();
@@ -8872,8 +9128,9 @@ export class AgentSession {
 						generation,
 						shouldContinue: () => this.agent.hasQueuedMessages(),
 					});
+					return COMPACTION_CHECK_CONTINUATION;
 				}
-				return false;
+				return COMPACTION_CHECK_NONE;
 			}
 
 			let hookCompaction: CompactionResult | undefined;
@@ -8897,7 +9154,7 @@ export class AgentSession {
 						aborted: true,
 						willRetry: false,
 					});
-					return false;
+					return COMPACTION_CHECK_NONE;
 				}
 
 				if (hookResult?.compaction) {
@@ -9080,7 +9337,7 @@ export class AgentSession {
 					aborted: true,
 					willRetry: false,
 				});
-				return false;
+				return COMPACTION_CHECK_NONE;
 			}
 
 			this.sessionManager.appendCompaction(
@@ -9122,8 +9379,10 @@ export class AgentSession {
 			};
 			await this.#emitSessionEvent({ type: "auto_compaction_end", action, result, aborted: false, willRetry });
 
+			let continuationScheduled = false;
 			if (!willRetry && reason !== "idle" && shouldAutoContinue) {
 				this.#scheduleAutoContinuePrompt(generation);
+				continuationScheduled = true;
 			}
 
 			if (willRetry) {
@@ -9144,6 +9403,7 @@ export class AgentSession {
 				}
 
 				this.#scheduleAgentContinue({ delayMs: 100, generation });
+				continuationScheduled = true;
 			} else if (this.agent.hasQueuedMessages()) {
 				// Auto-compaction can complete while follow-up/steering/custom messages are waiting.
 				// Kick the loop so queued messages are actually delivered.
@@ -9152,7 +9412,9 @@ export class AgentSession {
 					generation,
 					shouldContinue: () => this.agent.hasQueuedMessages(),
 				});
+				continuationScheduled = true;
 			}
+			return continuationScheduled ? COMPACTION_CHECK_CONTINUATION : COMPACTION_CHECK_NONE;
 		} catch (error) {
 			if (autoCompactionSignal.aborted) {
 				await this.#emitSessionEvent({
@@ -9162,7 +9424,7 @@ export class AgentSession {
 					aborted: true,
 					willRetry: false,
 				});
-				return false;
+				return COMPACTION_CHECK_NONE;
 			}
 			const errorMessage = error instanceof Error ? error.message : "compaction failed";
 			await this.#emitSessionEvent({
@@ -9183,7 +9445,7 @@ export class AgentSession {
 				this.#autoCompactionAbortController = undefined;
 			}
 		}
-		return false;
+		return COMPACTION_CHECK_NONE;
 	}
 
 	/**
@@ -9202,7 +9464,7 @@ export class AgentSession {
 		generation: number,
 		autoContinue: boolean,
 		triggerContextTokens?: number,
-	): Promise<"handled" | "fallback"> {
+	): Promise<CompactionCheckResult | "fallback"> {
 		const action = "shake";
 		await this.#emitSessionEvent({ type: "auto_compaction_start", reason, action });
 		this.#autoCompactionAbortController?.abort();
@@ -9219,7 +9481,7 @@ export class AgentSession {
 					aborted: true,
 					willRetry: false,
 				});
-				return "handled";
+				return COMPACTION_CHECK_NONE;
 			}
 			const reclaimed = result.toolResultsDropped + result.blocksDropped > 0;
 			// Detect the dead-loop reported in issues #2119/#2275: the threshold check
@@ -9251,7 +9513,7 @@ export class AgentSession {
 					const recoveryBand = Math.floor(thresholdTokens * SHAKE_RECOVERY_BAND);
 					stillOverThreshold = correctedTokens > recoveryBand;
 				} else {
-					const postShakeTokens = this.#estimatePendingPromptTokens([]);
+					const postShakeTokens = this.getContextUsage({ contextWindow })?.tokens ?? 0;
 					stillOverThreshold = shouldCompact(postShakeTokens, contextWindow, compactionSettings);
 				}
 			}
@@ -9280,8 +9542,10 @@ export class AgentSession {
 				skipped: !reclaimed,
 			});
 
+			let continuationScheduled = false;
 			if (!willRetry && reason !== "idle" && autoContinue) {
 				this.#scheduleAutoContinuePrompt(generation);
+				continuationScheduled = true;
 			}
 			if (willRetry) {
 				// The shake rebuild replays every entry, so a trailing error/length
@@ -9297,14 +9561,16 @@ export class AgentSession {
 					if (shouldDrop) this.agent.replaceMessages(messages.slice(0, -1));
 				}
 				this.#scheduleAgentContinue({ delayMs: 100, generation });
+				continuationScheduled = true;
 			} else if (this.agent.hasQueuedMessages()) {
 				this.#scheduleAgentContinue({
 					delayMs: 100,
 					generation,
 					shouldContinue: () => this.agent.hasQueuedMessages(),
 				});
+				continuationScheduled = true;
 			}
-			return "handled";
+			return continuationScheduled ? COMPACTION_CHECK_CONTINUATION : COMPACTION_CHECK_NONE;
 		} catch (error) {
 			if (signal.aborted) {
 				await this.#emitSessionEvent({
@@ -9314,7 +9580,7 @@ export class AgentSession {
 					aborted: true,
 					willRetry: false,
 				});
-				return "handled";
+				return COMPACTION_CHECK_NONE;
 			}
 			const message = error instanceof Error ? error.message : "shake failed";
 			await this.#emitSessionEvent({
@@ -9326,7 +9592,7 @@ export class AgentSession {
 				errorMessage: `Auto-shake failed: ${message}`,
 			});
 			// Overflow still needs recovery even if shake threw.
-			return reason === "overflow" ? "fallback" : "handled";
+			return reason === "overflow" ? "fallback" : COMPACTION_CHECK_NONE;
 		} finally {
 			if (this.#autoCompactionAbortController === controller) {
 				this.#autoCompactionAbortController = undefined;
@@ -10443,11 +10709,7 @@ export class AgentSession {
 		if (!model) {
 			throw new Error("No active model on session");
 		}
-		const apiKey = await this.#modelRegistry.getApiKey(model, this.sessionId);
-		if (!apiKey) {
-			throw new Error(`No API key for ${model.provider}/${model.id}`);
-		}
-
+		const cacheSessionId = this.sessionId;
 		const snapshot = this.#buildEphemeralSnapshot(args.promptText);
 		const llmMessages = await this.convertMessagesToLlm(snapshot, args.signal);
 		const context: Context = {
@@ -10459,10 +10721,9 @@ export class AgentSession {
 			// removes the surface entirely.
 			tools: [],
 		};
-		const cacheSessionId = this.sessionId;
 		const options = this.prepareSimpleStreamOptions(
 			{
-				apiKey,
+				apiKey: this.#modelRegistry.resolver(model, cacheSessionId),
 				// Side-channel turns must not share OpenAI/Codex append-only
 				// conversation state with the main agent turn: IRC and /btw can run
 				// while the main turn is mid-tool-call. Keep the prompt-cache key
@@ -11185,50 +11446,173 @@ export class AgentSession {
 	 * Uses the last assistant message's usage data when available,
 	 * otherwise estimates tokens for all messages.
 	 */
-	getContextUsage(options?: { contextWindow?: number }): ContextUsage | undefined {
+	getContextBreakdown(options?: {
+		contextWindow?: number;
+		pendingMessages?: AgentMessage[];
+	}): ContextUsageBreakdown | undefined {
 		const model = this.model;
 		const contextWindow = options?.contextWindow ?? model?.contextWindow ?? 0;
 		if (!Number.isFinite(contextWindow) || contextWindow <= 0) return undefined;
 
-		// After compaction, the last assistant usage reflects pre-compaction context size.
-		// We can only trust usage from an assistant that responded after the latest compaction.
-		// If no such assistant exists, context token count is unknown until the next LLM response.
+		const { skillsTokens, toolsTokens, systemContextTokens, systemPromptTokens } = computeNonMessageBreakdown(this);
+		const categoryNonMessageTokens = skillsTokens + toolsTokens + systemContextTokens + systemPromptTokens;
+		const currentNonMessageTokens = computeNonMessageTokens(this);
+
 		const branchEntries = this.sessionManager.getBranch();
 		const latestCompaction = getLatestCompactionEntry(branchEntries);
+		const compactionIndex = latestCompaction ? branchEntries.lastIndexOf(latestCompaction) : -1;
 
-		if (latestCompaction) {
-			// Check if there's a valid assistant usage after the compaction boundary
-			const compactionIndex = branchEntries.lastIndexOf(latestCompaction);
-			let hasPostCompactionUsage = false;
-			for (let i = branchEntries.length - 1; i > compactionIndex; i--) {
-				const entry = branchEntries[i];
-				if (entry.type === "message" && entry.message.role === "assistant") {
-					const assistant = entry.message;
-					if (assistant.stopReason !== "aborted" && assistant.stopReason !== "error") {
-						const contextTokens = calculateContextTokens(assistant.usage);
-						if (contextTokens > 0) {
-							hasPostCompactionUsage = true;
-						}
-						break;
-					}
+		let usedTokens = 0;
+		let anchored = false;
+
+		const pendingMessages = options?.pendingMessages ?? [];
+
+		const pending = this.#pendingContextSnapshot;
+
+		// Always locate the latest real assistant-usage anchor after the last
+		// compaction. Its provider-reported promptTokens is ground truth for
+		// everything up to that point; only the tail after it is estimated.
+		let anchorEntry: SessionMessageEntry | undefined;
+		for (let i = branchEntries.length - 1; i > compactionIndex; i--) {
+			const entry = branchEntries[i];
+			if (entry.type === "message" && entry.message.role === "assistant") {
+				const assistant = entry.message;
+				if (assistant.stopReason !== "aborted" && assistant.stopReason !== "error" && assistant.usage) {
+					anchorEntry = entry;
+					break;
 				}
 			}
+		}
 
-			if (!hasPostCompactionUsage) {
-				return { tokens: null, contextWindow, percent: null };
+		const resolvedActiveMessages = this.messages;
+		let resolvedAnchorIndex = -1;
+		let anchorAssistant: AssistantMessage | undefined;
+		if (anchorEntry) {
+			const a = anchorEntry.message as AssistantMessage;
+			anchorAssistant = a;
+			resolvedAnchorIndex = resolvedActiveMessages.indexOf(a);
+			if (resolvedAnchorIndex === -1) {
+				resolvedAnchorIndex = resolvedActiveMessages.findIndex(
+					msg => msg.role === "assistant" && msg.timestamp === a.timestamp,
+				);
 			}
 		}
 
-		const estimate = this.#estimateContextTokens();
-		const percent = (estimate.tokens / contextWindow) * 100;
+		// A real anchor supersedes the in-flight estimate only once a step of the
+		// CURRENT turn has produced provider usage — i.e. it resolves at or after
+		// the pending cutoff. While the turn's first response is still pending (or
+		// the newest real anchor predates this turn) the pending snapshot is the
+		// only thing accounting for the just-submitted prompt, so it wins. This
+		// keeps a long tool turn from stacking an estimate of the entire tail on
+		// top of a stale turn-start prompt.
+		const useAnchor =
+			anchorAssistant !== undefined &&
+			resolvedAnchorIndex !== -1 &&
+			(!pending || resolvedAnchorIndex >= pending.cutoffCount);
+
+		if (useAnchor && anchorAssistant) {
+			const promptTokens =
+				anchorAssistant.contextSnapshot?.promptTokens ?? calculatePromptTokens(anchorAssistant.usage);
+			const nonMessageTokens = anchorAssistant.contextSnapshot?.nonMessageTokens ?? computeNonMessageTokens(this);
+			anchored = true;
+			let tailTokens = 0;
+			for (let i = resolvedAnchorIndex + 1; i < resolvedActiveMessages.length; i++) {
+				tailTokens += estimateTokens(resolvedActiveMessages[i]);
+			}
+			usedTokens =
+				promptTokens +
+				Math.max(0, currentNonMessageTokens - nonMessageTokens) +
+				tailTokens +
+				pendingMessages.reduce((sum, msg) => sum + estimateTokens(msg), 0);
+		} else if (pending) {
+			anchored = true;
+			let tailTokens = 0;
+			if (resolvedActiveMessages.length > pending.cutoffCount) {
+				for (let i = pending.cutoffCount; i < resolvedActiveMessages.length; i++) {
+					tailTokens += estimateTokens(resolvedActiveMessages[i]);
+				}
+			}
+			usedTokens =
+				pending.promptTokens +
+				Math.max(0, currentNonMessageTokens - pending.nonMessageTokens) +
+				tailTokens +
+				pendingMessages.reduce((sum, msg) => sum + estimateTokens(msg), 0);
+		}
+
+		if (!anchored && !pending && branchEntries.length === 0) {
+			// Fallback: look for the latest assistant message with usage/snapshot in this.messages (for branchless/fake sessions in tests)
+			for (let i = resolvedActiveMessages.length - 1; i >= 0; i--) {
+				const msg = resolvedActiveMessages[i];
+				if (msg.role === "assistant" && msg.stopReason !== "aborted" && msg.stopReason !== "error" && msg.usage) {
+					const promptTokens = msg.contextSnapshot?.promptTokens ?? calculatePromptTokens(msg.usage);
+					const nonMessageTokens = msg.contextSnapshot?.nonMessageTokens ?? computeNonMessageTokens(this);
+
+					let tailTokens = 0;
+					for (let j = i + 1; j < resolvedActiveMessages.length; j++) {
+						tailTokens += estimateTokens(resolvedActiveMessages[j]);
+					}
+
+					usedTokens =
+						promptTokens +
+						Math.max(0, currentNonMessageTokens - nonMessageTokens) +
+						tailTokens +
+						pendingMessages.reduce((sum, msg) => sum + estimateTokens(msg), 0);
+					anchored = true;
+					break;
+				}
+			}
+		}
+		if (!anchored) {
+			let messagesTokens = 0;
+			for (const msg of resolvedActiveMessages) {
+				messagesTokens += estimateTokens(msg);
+			}
+			usedTokens =
+				currentNonMessageTokens +
+				messagesTokens +
+				pendingMessages.reduce((sum, msg) => sum + estimateTokens(msg), 0);
+		}
+
+		const messagesTokens = Math.max(0, usedTokens - categoryNonMessageTokens);
 
 		return {
-			tokens: estimate.tokens,
 			contextWindow,
-			percent,
+			anchored,
+			usedTokens,
+			systemPromptTokens,
+			systemToolsTokens: toolsTokens,
+			systemContextTokens,
+			skillsTokens,
+			messagesTokens,
+		};
+	}
+
+	getContextUsage(options?: { contextWindow?: number }): ContextUsage | undefined {
+		const breakdown = this.getContextBreakdown(options);
+		if (!breakdown) return undefined;
+		return {
+			tokens: breakdown.usedTokens,
+			contextWindow: breakdown.contextWindow,
+			percent: breakdown.contextWindow > 0 ? (breakdown.usedTokens / breakdown.contextWindow) * 100 : 0,
 		};
 	}
 
+	/**
+	 * Monotonic counter that changes whenever the in-flight pending context
+	 * snapshot is set or cleared. Status-line context memoization keys on this so
+	 * a value computed mid-turn cannot persist after the turn ends/aborts.
+	 */
+	get contextUsageRevision(): number {
+		return this.#contextUsageRevision;
+	}
+
+	#setPendingContextSnapshot(
+		snapshot: { promptTokens: number; nonMessageTokens: number; cutoffCount: number } | undefined,
+	): void {
+		this.#pendingContextSnapshot = snapshot;
+		this.#contextUsageRevision++;
+	}
+
 	#ingestProviderUsageHeaders(response: ProviderResponseMetadata, model?: Model): void {
 		if (model?.provider !== "anthropic") return;
 		this.#modelRegistry.authStorage.ingestUsageHeaders("anthropic", response.headers, {
@@ -11241,7 +11625,17 @@ export class AgentSession {
 		const authStorage = this.#modelRegistry.authStorage;
 		if (!authStorage.fetchUsageReports) return null;
 		return authStorage.fetchUsageReports({
-			baseUrlResolver: provider => this.#modelRegistry.getProviderBaseUrl?.(provider),
+			baseUrlResolver: provider => {
+				if (provider === "google-antigravity") {
+					const mode = this.settings.get("providers.antigravityEndpoint");
+					if (mode === "sandbox") {
+						return "https://daily-cloudcode-pa.sandbox.googleapis.com";
+					} else if (mode === "production") {
+						return "https://daily-cloudcode-pa.googleapis.com";
+					}
+				}
+				return this.#modelRegistry.getProviderBaseUrl?.(provider);
+			},
 			signal,
 		});
 	}
@@ -11409,64 +11803,6 @@ export class AgentSession {
 		return run;
 	}
 
-	/**
-	 * Estimate context tokens from messages, using the last assistant usage when available.
-	 */
-	#estimateContextTokens(): {
-		tokens: number;
-		providerAnchored: boolean;
-		providerNonMessageTokens?: number;
-	} {
-		const messages = this.messages;
-
-		// Find last assistant message with valid usage.
-		let lastUsageIndex: number | null = null;
-		let lastUsage: Usage | undefined;
-		for (let i = messages.length - 1; i >= 0; i--) {
-			const msg = messages[i];
-			if (msg.role === "assistant") {
-				const assistantMsg = msg as AssistantMessage;
-				if (assistantMsg.stopReason !== "aborted" && assistantMsg.stopReason !== "error" && assistantMsg.usage) {
-					lastUsage = assistantMsg.usage;
-					lastUsageIndex = i;
-					break;
-				}
-			}
-		}
-
-		if (!lastUsage || lastUsageIndex === null) {
-			// No usage data - estimate all messages
-			let estimated = 0;
-			for (const message of messages) {
-				estimated += estimateTokens(message);
-			}
-			return {
-				tokens: estimated,
-				providerAnchored: false,
-			};
-		}
-
-		const usageTokens = calculatePromptTokens(lastUsage);
-		const providerNonMessage =
-			this.#lastProviderUsageNonMessage &&
-			messages[lastUsageIndex]?.role === "assistant" &&
-			this.#lastProviderUsageNonMessage.provider === (messages[lastUsageIndex] as AssistantMessage).provider &&
-			this.#lastProviderUsageNonMessage.model === (messages[lastUsageIndex] as AssistantMessage).model &&
-			this.#lastProviderUsageNonMessage.timestamp === (messages[lastUsageIndex] as AssistantMessage).timestamp
-				? this.#lastProviderUsageNonMessage.tokens
-				: undefined;
-		let trailingTokens = 0;
-		for (let i = lastUsageIndex + 1; i < messages.length; i++) {
-			trailingTokens += estimateTokens(messages[i]);
-		}
-
-		return {
-			tokens: usageTokens + trailingTokens,
-			providerAnchored: true,
-			providerNonMessageTokens: providerNonMessage,
-		};
-	}
-
 	/**
 	 * Export session to HTML.
 	 * @param outputPath Optional output path (defaults to session directory)