@demicodes/agent 0.15.0 → 0.17.0

package/dist/client-entry.d.mts

@@ -1,4 +1,4 @@
-import { b as AgentMetadata, d as TranscriptPatch, l as ServerFrame, n as AgentServerTransport, o as ClientFrame, r as AgentTransport, s as ClientSessionEvent, t as AgentClientTransport, u as ShellCommandStatusLike } from "./transport-gkyOzXn-.mjs";
-import { a as AgentClient, i as AgentActionOptions, n as createWebSocketClientTransport, o as AgentClientListener, r as createWebSocketServerTransport, t as JsonWebSocket } from "./websocket-transport-C6bu3Ufc.mjs";
+import { b as AgentMetadata, d as TranscriptPatch, l as ServerFrame, n as AgentServerTransport, o as ClientFrame, r as AgentTransport, s as ClientSessionEvent, t as AgentClientTransport, u as ShellCommandStatusLike } from "./transport-DtBCCS_A.mjs";
+import { a as AgentClient, i as AgentActionOptions, n as createWebSocketClientTransport, o as AgentClientListener, r as createWebSocketServerTransport, t as JsonWebSocket } from "./websocket-transport-DdfGgBjO.mjs";
 import { ProviderSelection } from "@demicodes/provider";
 export { type AgentActionOptions, AgentClient, type AgentClientListener, type AgentClientTransport, type AgentMetadata, type AgentServerTransport, type AgentTransport, type ClientFrame, type ClientSessionEvent, type JsonWebSocket, type ProviderSelection, type ServerFrame, type ShellCommandStatusLike, type TranscriptPatch, createWebSocketClientTransport, createWebSocketServerTransport };

package/dist/client-entry.mjs

@@ -1,2 +1,2 @@
-import { n as createWebSocketServerTransport, r as AgentClient, t as createWebSocketClientTransport } from "./websocket-transport-BjSj3U0I.mjs";
+import { n as createWebSocketServerTransport, r as AgentClient, t as createWebSocketClientTransport } from "./websocket-transport-c6P3NBuf.mjs";
 export { AgentClient, createWebSocketClientTransport, createWebSocketServerTransport };

package/dist/index.d.mts

@@ -1,5 +1,5 @@
-import { A as AgentTool, B as TranscriptOptions, C as AgentSessionCheckpoint, D as AgentSessionRestoreParams, E as AgentSessionParams, F as SessionEvent, G as resolveRetryPolicy, H as DEFAULT_TURN_RETRY_POLICY, I as SessionEventListener, K as retryDelayMs, L as CompactionWindow, M as AgentToolInvokeContext, N as AgentToolInvokeResult, O as AgentSessionStore, P as ExternalMutationReservation, R as DrainedTranscriptPatches, S as AgentReferenceResolveContext, T as AgentSessionOptions, U as TurnRetryPolicy, V as estimateTranscriptBlockTokens, W as isRetryableCode, _ as AgentHarnessRuntime, a as createInProcessTransportPair, b as AgentMetadata, c as ConversationSummary, d as TranscriptPatch, f as AbortResult, g as AgentHarnessContext, h as AgentHarness, i as InProcessTransportPair, j as AgentToolContext, k as AgentSystemPromptContext, l as ServerFrame, m as AgentDisposeContext, n as AgentServerTransport, o as ClientFrame, p as AbortTarget, r as AgentTransport, s as ClientSessionEvent, t as AgentClientTransport, u as ShellCommandStatusLike, v as AgentHostContext, w as AgentSessionCloneParams, x as AgentPromptContext, y as AgentLifecycleEvent, z as TranscriptLog } from "./transport-gkyOzXn-.mjs";
-import { a as AgentClient, i as AgentActionOptions, n as createWebSocketClientTransport, o as AgentClientListener, r as createWebSocketServerTransport, t as JsonWebSocket } from "./websocket-transport-C6bu3Ufc.mjs";
+import { A as AgentTool, B as TranscriptLog, C as AgentSessionCheckpoint, D as AgentSessionRestoreParams, E as AgentSessionParams, F as ModelSwitchApply, G as isRetryableCode, H as estimateTranscriptBlockTokens, I as SessionEvent, K as resolveRetryPolicy, L as SessionEventListener, M as AgentToolInvokeContext, N as AgentToolInvokeResult, O as AgentSessionStore, P as ExternalMutationReservation, R as CompactionWindow, S as AgentReferenceResolveContext, T as AgentSessionOptions, U as DEFAULT_TURN_RETRY_POLICY, V as TranscriptOptions, W as TurnRetryPolicy, _ as AgentHarnessRuntime, a as createInProcessTransportPair, b as AgentMetadata, c as ConversationSummary, d as TranscriptPatch, f as AbortResult, g as AgentHarnessContext, h as AgentHarness, i as InProcessTransportPair, j as AgentToolContext, k as AgentSystemPromptContext, l as ServerFrame, m as AgentDisposeContext, n as AgentServerTransport, o as ClientFrame, p as AbortTarget, q as retryDelayMs, r as AgentTransport, s as ClientSessionEvent, t as AgentClientTransport, u as ShellCommandStatusLike, v as AgentHostContext, w as AgentSessionCloneParams, x as AgentPromptContext, y as AgentLifecycleEvent, z as DrainedTranscriptPatches } from "./transport-DtBCCS_A.mjs";
+import { a as AgentClient, i as AgentActionOptions, n as createWebSocketClientTransport, o as AgentClientListener, r as createWebSocketServerTransport, t as JsonWebSocket } from "./websocket-transport-DdfGgBjO.mjs";
 import { BashAuditEvent, BashEnvironment, BashEnvironmentOptions, CommandMetadataRecord, Host, ShellCommandStatus, ShellOutputChunk } from "@demicodes/shell";
 import { AgentProvider, Provider } from "@demicodes/provider";
 import { Block, Model, ModelMediaKind, ModelSelection, ProviderErrorDiagnostics, QueuedMessage, SessionPhase, UserContentBlock } from "@demicodes/core";
@@ -48,6 +48,7 @@ declare class AgentSession<State> {
   private readonly store?;
   private readonly compactionKeepRecentTokens;
   private readonly compactionThresholdRatio;
+  private readonly compactionThresholdTokens;
   private readonly listeners;
   private readonly pendingActions;
   private readonly queued;
@@ -122,18 +123,19 @@ declare class AgentSession<State> {
   }): Promise<void>;
   cancelPendingSteer(id: string): boolean;
   /**
-   * Switches the model (and optionally the provider) for the rest of the session. Queued as an
-   * action so it lands at a turn boundary, never mid-tool-continuation; the next turn uses it.
-   * A non-null `provider` replaces the current one (its predecessor is disposed); pass null to
-   * keep the same provider instance and only change the model.
-   */
-  /**
-   * Records a model/provider switch. It is applied at the next turn's preflight (see
-   * applyPendingModelSwitch): if the new model can't hold the history, compaction runs there
-   * with the current (pre-switch) model first, then the swap happens. Recording is cheap and
-   * non-blocking so it never holds up an in-flight turn or other frames.
+   * Records a model/provider switch for the rest of the session. A non-null `provider`
+   * replaces the current one (its predecessor is disposed); pass null to keep the same
+   * provider instance and only change the model.
+   *
+   * `apply` picks the boundary the switch lands on: 'next_turn' (default) waits for the
+   * start of the next queued action, so a running turn finishes entirely on the old model;
+   * 'immediate' also applies mid-turn at the next sampling/tool continuation, so the very
+   * next request already runs on the new model (the same "as soon as possible" semantics
+   * steering has). Either way, if the new model can't hold the history, compaction runs
+   * first with the current (pre-switch) model, then the swap happens. Recording is cheap
+   * and non-blocking so it never holds up an in-flight turn or other frames.
    */
-  updateModel(provider: AgentProvider | null, model: ModelSelection): void;
+  updateModel(provider: AgentProvider | null, model: ModelSelection, apply?: ModelSwitchApply): void;
   abort(): Promise<AbortResult>;
   scheduleYieldWakeup(durationMs: number, metadata?: AgentMetadata | null): AgentToolInvokeResult;
   waitUntilDone(): Promise<void>;
@@ -169,13 +171,20 @@ declare class AgentSession<State> {
   private runWorker;
   private executeAction;
   /**
-   * Applies a queued model/provider switch at a turn boundary (preflight). If the new model's
-   * context window can't hold the current history, compaction runs FIRST with the current
-   * (pre-switch) model + provider — which can still load it to summarize — and only then do we
-   * swap. Doing it the other way would ask the smaller model to summarize a history it may not
-   * be able to load.
+   * Applies a queued model/provider switch at an action boundary (turn preflight). If the
+   * new model's context window can't hold the current history, compaction runs FIRST with
+   * the current (pre-switch) model + provider — which can still load it to summarize — and
+   * only then do we swap. Doing it the other way would ask the smaller model to summarize
+   * a history it may not be able to load. Returns whether that compaction actually ran
+   * (false when no switch was pending).
    */
   private applyPendingModelSwitch;
+  /**
+   * Turn-loop variant for mid-turn continuation boundaries: only switches recorded with
+   * apply 'immediate' land here; 'next_turn' switches stay pending until the running turn
+   * finishes and the next action starts.
+   */
+  private applyImmediateModelSwitch;
   private runWithCompactingPhase;
   private executeSend;
   private resolveReferences;
@@ -231,6 +240,7 @@ interface AgentServerSessionOptions {
   compaction?: {
     keepRecentTokens?: number;
     preflightThresholdRatio?: number;
+    preflightThresholdTokens?: number;
   };
   persistIntervalMs?: number;
 }
@@ -375,4 +385,4 @@ interface ShellToolView {
 declare function finishShellToolResult<State>(environment: BashEnvironment, result: ShellCommandStatus, ctx: AgentToolInvokeContext<State>): Promise<AgentToolInvokeResult>;
 declare function shellCommandHandleRequired(result: ShellCommandStatus, budgetTokens: number): boolean;
 //#endregion
-export { AbortResult, AbortTarget, ActiveTurnPhase, AgentActionOptions, AgentClient, AgentClientListener, AgentClientTransport, AgentDisposeContext, AgentHarness, AgentHarnessContext, AgentHarnessRuntime, AgentHostContext, AgentLifecycleEvent, AgentMetadata, AgentPromptContext, AgentReferenceResolveContext, AgentServer, AgentServerOptions, AgentServerSessionOptions, AgentServerTransport, AgentSession, AgentSessionCheckpoint, AgentSessionCloneParams, AgentSessionOptions, AgentSessionParams, AgentSessionRestoreParams, AgentSessionStore, AgentSystemPromptContext, AgentTool, AgentToolContext, AgentToolInvokeContext, AgentToolInvokeResult, AgentTransport, AgentTransportBinding, ClientFrame, ClientSessionEvent, CompactionWindow, ConversationSummary, DEFAULT_MAX_MEDIA_BYTES, DEFAULT_TURN_RETRY_POLICY, DrainedTranscriptPatches, ExternalMutationReservation, InProcessTransportPair, JsonWebSocket, PrepareShell, PrepareShellContext, ProviderStreamError, ResumePoint, RunCommandLineCommandNotRegisteredError, RunCommandLineOptions, RunCommandLineResult, RunCommandLineShellNotFoundError, RunCommandLineTimeoutError, SHELL_VIEW_MAX_CHARS, ServerFrame, SessionEvent, SessionEventListener, ShellCommandStatusLike, ShellToolResultOptions, ShellToolView, StandardAgentToolOptions, TranscriptLog, TranscriptOptions, TranscriptPatch, TurnRetryPolicy, applyTranscriptPatches, cloneBlocks, createInProcessTransportPair, createStandardAgentTools, createWebSocketClientTransport, createWebSocketServerTransport, estimateTranscriptBlockTokens, findResumePoint, finishShellToolResult, isContextLengthExceeded, isRetryableCode, resolveRetryPolicy, retryDelayMs, shellCommandHandleRequired, shellPreviewBudgetTokens, toShellToolResult };
+export { AbortResult, AbortTarget, ActiveTurnPhase, AgentActionOptions, AgentClient, AgentClientListener, AgentClientTransport, AgentDisposeContext, AgentHarness, AgentHarnessContext, AgentHarnessRuntime, AgentHostContext, AgentLifecycleEvent, AgentMetadata, AgentPromptContext, AgentReferenceResolveContext, AgentServer, AgentServerOptions, AgentServerSessionOptions, AgentServerTransport, AgentSession, AgentSessionCheckpoint, AgentSessionCloneParams, AgentSessionOptions, AgentSessionParams, AgentSessionRestoreParams, AgentSessionStore, AgentSystemPromptContext, AgentTool, AgentToolContext, AgentToolInvokeContext, AgentToolInvokeResult, AgentTransport, AgentTransportBinding, ClientFrame, ClientSessionEvent, CompactionWindow, ConversationSummary, DEFAULT_MAX_MEDIA_BYTES, DEFAULT_TURN_RETRY_POLICY, DrainedTranscriptPatches, ExternalMutationReservation, InProcessTransportPair, JsonWebSocket, ModelSwitchApply, PrepareShell, PrepareShellContext, ProviderStreamError, ResumePoint, RunCommandLineCommandNotRegisteredError, RunCommandLineOptions, RunCommandLineResult, RunCommandLineShellNotFoundError, RunCommandLineTimeoutError, SHELL_VIEW_MAX_CHARS, ServerFrame, SessionEvent, SessionEventListener, ShellCommandStatusLike, ShellToolResultOptions, ShellToolView, StandardAgentToolOptions, TranscriptLog, TranscriptOptions, TranscriptPatch, TurnRetryPolicy, applyTranscriptPatches, cloneBlocks, createInProcessTransportPair, createStandardAgentTools, createWebSocketClientTransport, createWebSocketServerTransport, estimateTranscriptBlockTokens, findResumePoint, finishShellToolResult, isContextLengthExceeded, isRetryableCode, resolveRetryPolicy, retryDelayMs, shellCommandHandleRequired, shellPreviewBudgetTokens, toShellToolResult };

package/dist/index.mjs

@@ -1,4 +1,4 @@
-import { a as cloneBlocks, i as applyTranscriptPatches, n as createWebSocketServerTransport, o as ProviderStreamError, r as AgentClient, s as isContextLengthExceeded, t as createWebSocketClientTransport } from "./websocket-transport-BjSj3U0I.mjs";
+import { a as cloneBlocks, i as applyTranscriptPatches, n as createWebSocketServerTransport, o as ProviderStreamError, r as AgentClient, s as isContextLengthExceeded, t as createWebSocketClientTransport } from "./websocket-transport-c6P3NBuf.mjs";
 import { AbortError, abortable, asError, asRecord, asString, bytesToBase64, createId, delay, isAbortError, isRecord, noop, parseJsonOrString, safeJsonStringify, sliceHead, sliceTail, throwIfAborted, toWellFormedText, truncate } from "@demicodes/utils";
 import { BashEnvironment, CommandRegistry, MAX_TIMEOUT_MS, heredocDelimiter, shellQuote } from "@demicodes/shell";
 import { providerRuntime } from "@demicodes/provider";
@@ -793,6 +793,20 @@ var PendingSteerQueue = class {
 function estimateTokens(text) {
 	return Math.ceil(text.length / 4);
 }
+/**
+* Resolves the token count at which auto-compaction fires.
+*
+* - A non-finite `thresholdRatio` disables auto-compaction (used by summary clones).
+* - When `thresholdTokens` is a finite number, it is used as an absolute threshold,
+*   clamped to `[0, contextWindow]`.
+* - Otherwise the threshold is `floor(contextWindow * thresholdRatio)`.
+*/
+function resolveCompactionThreshold(contextWindow, thresholdRatio, thresholdTokens) {
+	if (contextWindow <= 0) return 0;
+	if (!Number.isFinite(thresholdRatio)) return Number.POSITIVE_INFINITY;
+	if (thresholdTokens != null && Number.isFinite(thresholdTokens)) return Math.min(Math.max(0, Math.floor(thresholdTokens)), contextWindow);
+	return Math.floor(contextWindow * thresholdRatio);
+}
 /** The next (smaller) cut point to retry compaction with, or null when nothing more can be compacted. */
 function nextSmallerCompactionCutPoint(startIndex, cutPoint) {
 	const compactedBlockCount = cutPoint - startIndex;
@@ -821,21 +835,29 @@ var CompactionController = class {
 	constructor(host) {
 		this.host = host;
 	}
-	/** Compacts (up to 8 passes) until the history fits `targetModel`'s context, if over threshold. */
+	/**
+	* Compacts (up to 8 passes) until the history fits `targetModel`'s context, if over
+	* threshold. Returns whether anything was compacted.
+	*/
 	async compactToFit(targetModel) {
 		const contextWindow = targetModel.model.contextWindow;
-		if (contextWindow <= 0) return;
-		const threshold = Math.floor(contextWindow * this.host.thresholdRatio);
-		if (this.host.transcript.estimateContextTokens(contextWindow) < threshold) return;
-		await this.host.runWithCompactingPhase(async () => {
-			for (let attempt = 0; attempt < 8 && this.host.transcript.estimateContextTokens(contextWindow) >= threshold; attempt += 1) if (!await this.run()) break;
+		if (contextWindow <= 0) return false;
+		const threshold = resolveCompactionThreshold(contextWindow, this.host.thresholdRatio, this.host.thresholdTokens);
+		if (this.host.transcript.estimateContextTokens(contextWindow) < threshold) return false;
+		return this.host.runWithCompactingPhase(async () => {
+			let compacted = false;
+			for (let attempt = 0; attempt < 8 && this.host.transcript.estimateContextTokens(contextWindow) >= threshold; attempt += 1) {
+				if (!await this.run()) break;
+				compacted = true;
+			}
+			return compacted;
 		});
 	}
 	/** Runs one compaction pass before a turn when the current model is over threshold. */
 	async preflight() {
 		const contextWindow = this.host.model.model.contextWindow;
 		if (contextWindow <= 0) return;
-		const threshold = Math.floor(contextWindow * this.host.thresholdRatio);
+		const threshold = resolveCompactionThreshold(contextWindow, this.host.thresholdRatio, this.host.thresholdTokens);
 		if (this.host.transcript.estimateContextTokens(contextWindow) < threshold) return;
 		await this.host.runWithCompactingPhase(() => this.run());
 	}
@@ -922,6 +944,10 @@ var ProviderTurnLoop = class {
 		let autoCompactions = 0;
 		while (true) {
 			throwIfAborted(this.host.currentSignal());
+			if (await this.host.applyImmediateModelSwitch()) {
+				this.host.transcript.pushResumeTurn(this.host.currentTurnId(), this.host.model);
+				await this.host.commitTranscript();
+			}
 			await this.host.materializePendingSteers();
 			const steerContinuationBeforeStream = this.host.steerContinuationCount;
 			const shouldAutoRecover = await this.streamProviderOnce();
@@ -1146,7 +1172,7 @@ var ProviderTurnLoop = class {
 	isUsageNearLimit(usage) {
 		const contextWindow = this.host.model.model.contextWindow;
 		if (contextWindow <= 0) return false;
-		return usage.inputTokens + usage.outputTokens + usage.cacheReadTokens + usage.cacheWriteTokens >= Math.floor(contextWindow * this.host.thresholdRatio);
+		return usage.inputTokens + usage.outputTokens + usage.cacheReadTokens + usage.cacheWriteTokens >= resolveCompactionThreshold(contextWindow, this.host.thresholdRatio, this.host.thresholdTokens);
 	}
 };
 function toToolDefinition(tool) {
@@ -1172,6 +1198,7 @@ var AgentSession = class AgentSession {
 	store;
 	compactionKeepRecentTokens;
 	compactionThresholdRatio;
+	compactionThresholdTokens;
 	listeners = /* @__PURE__ */ new Set();
 	pendingActions = [];
 	queued = [];
@@ -1237,6 +1264,7 @@ var AgentSession = class AgentSession {
 		this.retryPolicy = resolveRetryPolicy(options.retry);
 		this.compactionKeepRecentTokens = options.compaction?.keepRecentTokens ?? DEFAULT_KEEP_RECENT_TOKENS;
 		this.compactionThresholdRatio = options.compaction?.preflightThresholdRatio ?? DEFAULT_PREFLIGHT_THRESHOLD_RATIO;
+		this.compactionThresholdTokens = options.compaction?.preflightThresholdTokens === void 0 ? null : options.compaction.preflightThresholdTokens;
 		const transcriptOptions = {
 			idFactory: this.idFactory,
 			now: options.now
@@ -1256,6 +1284,9 @@ var AgentSession = class AgentSession {
 			get thresholdRatio() {
 				return self.compactionThresholdRatio;
 			},
+			get thresholdTokens() {
+				return self.compactionThresholdTokens;
+			},
 			currentSignal: () => self.currentSignal(),
 			clone: (transcript) => self.clone({
 				transcript,
@@ -1294,6 +1325,9 @@ var AgentSession = class AgentSession {
 			get thresholdRatio() {
 				return self.compactionThresholdRatio;
 			},
+			get thresholdTokens() {
+				return self.compactionThresholdTokens;
+			},
 			get steerContinuationCount() {
 				return self.steerQueue.continuationCount;
 			},
@@ -1316,6 +1350,7 @@ var AgentSession = class AgentSession {
 				self.activeProviderRun = run;
 			},
 			streamProvider: (request, run) => self.providerEvents(request, run),
+			applyImmediateModelSwitch: () => self.applyImmediateModelSwitch(),
 			runCompaction: () => self.compaction.run(),
 			runWithCompactingPhase: (fn) => self.runWithCompactingPhase(fn),
 			commitTranscript: () => self.commitTranscript(),
@@ -1449,21 +1484,23 @@ var AgentSession = class AgentSession {
 		return false;
 	}
 	/**
-	* Switches the model (and optionally the provider) for the rest of the session. Queued as an
-	* action so it lands at a turn boundary, never mid-tool-continuation; the next turn uses it.
-	* A non-null `provider` replaces the current one (its predecessor is disposed); pass null to
-	* keep the same provider instance and only change the model.
-	*/
-	/**
-	* Records a model/provider switch. It is applied at the next turn's preflight (see
-	* applyPendingModelSwitch): if the new model can't hold the history, compaction runs there
-	* with the current (pre-switch) model first, then the swap happens. Recording is cheap and
-	* non-blocking so it never holds up an in-flight turn or other frames.
+	* Records a model/provider switch for the rest of the session. A non-null `provider`
+	* replaces the current one (its predecessor is disposed); pass null to keep the same
+	* provider instance and only change the model.
+	*
+	* `apply` picks the boundary the switch lands on: 'next_turn' (default) waits for the
+	* start of the next queued action, so a running turn finishes entirely on the old model;
+	* 'immediate' also applies mid-turn at the next sampling/tool continuation, so the very
+	* next request already runs on the new model (the same "as soon as possible" semantics
+	* steering has). Either way, if the new model can't hold the history, compaction runs
+	* first with the current (pre-switch) model, then the swap happens. Recording is cheap
+	* and non-blocking so it never holds up an in-flight turn or other frames.
 	*/
-	updateModel(provider, model) {
+	updateModel(provider, model, apply = "next_turn") {
 		this.pendingModelSwitch = {
 			provider,
-			model
+			model,
+			apply
 		};
 	}
 	async abort() {
@@ -1835,23 +1872,34 @@ var AgentSession = class AgentSession {
 		}
 	}
 	/**
-	* Applies a queued model/provider switch at a turn boundary (preflight). If the new model's
-	* context window can't hold the current history, compaction runs FIRST with the current
-	* (pre-switch) model + provider — which can still load it to summarize — and only then do we
-	* swap. Doing it the other way would ask the smaller model to summarize a history it may not
-	* be able to load.
+	* Applies a queued model/provider switch at an action boundary (turn preflight). If the
+	* new model's context window can't hold the current history, compaction runs FIRST with
+	* the current (pre-switch) model + provider — which can still load it to summarize — and
+	* only then do we swap. Doing it the other way would ask the smaller model to summarize
+	* a history it may not be able to load. Returns whether that compaction actually ran
+	* (false when no switch was pending).
 	*/
 	async applyPendingModelSwitch() {
 		const pending = this.pendingModelSwitch;
-		if (!pending) return;
+		if (!pending) return false;
 		this.pendingModelSwitch = null;
-		await this.compaction.compactToFit(pending.model);
+		const compacted = await this.compaction.compactToFit(pending.model);
 		if (pending.provider && pending.provider !== this.provider) {
 			const previous = this.provider;
 			this.provider = pending.provider;
 			await previous.dispose?.();
 		}
 		this.model = pending.model;
+		return compacted;
+	}
+	/**
+	* Turn-loop variant for mid-turn continuation boundaries: only switches recorded with
+	* apply 'immediate' land here; 'next_turn' switches stay pending until the running turn
+	* finishes and the next action starts.
+	*/
+	async applyImmediateModelSwitch() {
+		if (this.pendingModelSwitch?.apply !== "immediate") return false;
+		return this.applyPendingModelSwitch();
 	}
 	async runWithCompactingPhase(fn) {
 		const previousPhase = this.currentPhase;
@@ -2834,7 +2882,7 @@ var AgentTransportBindingImpl = class {
 					this.session?.cancelPendingSteer(frame.steerId);
 					return;
 				case "set_provider":
-					await this.setProvider(frame.provider);
+					await this.setProvider(frame.provider, frame.apply);
 					return;
 				case "retry": {
 					const session = this.sessionFor("retry");
@@ -3024,7 +3072,7 @@ var AgentTransportBindingImpl = class {
 				return;
 		}
 	}
-	async setProvider(provider) {
+	async setProvider(provider, apply) {
 		const session = this.session;
 		if (!session) {
 			this.send({
@@ -3035,11 +3083,11 @@ var AgentTransportBindingImpl = class {
 			return;
 		}
 		if (provider.providerId === this.currentProviderId) {
-			await session.updateModel(null, provider.model);
+			session.updateModel(null, provider.model, apply);
 			return;
 		}
 		const next = await this.createRuntime(provider);
-		await session.updateModel(next, provider.model);
+		session.updateModel(next, provider.model, apply);
 		this.currentProviderId = provider.providerId;
 	}
 	async createRuntime(selection) {

package/dist/stdio-transport.d.mts

@@ -1,4 +1,4 @@
-import { n as AgentServerTransport, t as AgentClientTransport } from "./transport-gkyOzXn-.mjs";
+import { n as AgentServerTransport, t as AgentClientTransport } from "./transport-DtBCCS_A.mjs";
 import { Readable, Writable } from "node:stream";
 //#region src/stdio-transport.d.ts
 declare function createStdioClientTransport(readable: Readable, writable: Writable): AgentClientTransport;

package/dist/{transport-gkyOzXn-.d.mts → transport-DtBCCS_A.d.mts}

@@ -209,6 +209,14 @@ interface AgentToolInvokeResult {
   view?: unknown | null;
   stopAfterToolResult?: boolean;
 }
+/**
+ * When a recorded model/provider switch is applied:
+ * - 'next_turn' (default): the start of the next queued action — a running turn, if any,
+ *   finishes entirely on the old model.
+ * - 'immediate': the next inference boundary — mid-turn that is the next sampling/tool
+ *   continuation, so the very next request already runs on the new model.
+ */
+type ModelSwitchApply = 'immediate' | 'next_turn';
 type AbortTarget = 'active_provider_stream' | 'active_tool' | 'active_compaction' | 'active_turn' | 'queued_action' | 'queued_message' | 'pending_yield_wakeup';
 interface AbortResult {
   aborted: boolean;
@@ -308,6 +316,12 @@ interface AgentSessionOptions<State = unknown> {
   compaction?: {
     keepRecentTokens?: number;
     preflightThresholdRatio?: number;
+    /**
+     * Absolute token threshold for auto-compaction. When set, replaces the
+     * ratio-derived threshold (still clamped to the model context window).
+     * A non-finite `preflightThresholdRatio` still disables auto-compaction.
+     */
+    preflightThresholdTokens?: number;
   };
   /**
    * Maximum interval between checkpoint writes while a turn is streaming. Writes
@@ -388,6 +402,7 @@ type ClientFrame = {
 } | {
   type: 'set_provider';
   provider: ProviderSelection;
+  apply?: ModelSwitchApply;
 } | {
   type: 'abort';
 } | {
@@ -580,4 +595,4 @@ interface InProcessTransportPair {
 }
 declare function createInProcessTransportPair(): InProcessTransportPair;
 //#endregion
-export { AgentTool as A, TranscriptOptions as B, AgentSessionCheckpoint as C, AgentSessionRestoreParams as D, AgentSessionParams as E, SessionEvent as F, resolveRetryPolicy as G, DEFAULT_TURN_RETRY_POLICY as H, SessionEventListener as I, retryDelayMs as K, CompactionWindow as L, AgentToolInvokeContext as M, AgentToolInvokeResult as N, AgentSessionStore as O, ExternalMutationReservation as P, DrainedTranscriptPatches as R, AgentReferenceResolveContext as S, AgentSessionOptions as T, TurnRetryPolicy as U, estimateTranscriptBlockTokens as V, isRetryableCode as W, AgentHarnessRuntime as _, createInProcessTransportPair as a, AgentMetadata as b, ConversationSummary as c, TranscriptPatch as d, AbortResult as f, AgentHarnessContext as g, AgentHarness as h, InProcessTransportPair as i, AgentToolContext as j, AgentSystemPromptContext as k, ServerFrame as l, AgentDisposeContext as m, AgentServerTransport as n, ClientFrame as o, AbortTarget as p, AgentTransport as r, ClientSessionEvent as s, AgentClientTransport as t, ShellCommandStatusLike as u, AgentHostContext as v, AgentSessionCloneParams as w, AgentPromptContext as x, AgentLifecycleEvent as y, TranscriptLog as z };
+export { AgentTool as A, TranscriptLog as B, AgentSessionCheckpoint as C, AgentSessionRestoreParams as D, AgentSessionParams as E, ModelSwitchApply as F, isRetryableCode as G, estimateTranscriptBlockTokens as H, SessionEvent as I, resolveRetryPolicy as K, SessionEventListener as L, AgentToolInvokeContext as M, AgentToolInvokeResult as N, AgentSessionStore as O, ExternalMutationReservation as P, CompactionWindow as R, AgentReferenceResolveContext as S, AgentSessionOptions as T, DEFAULT_TURN_RETRY_POLICY as U, TranscriptOptions as V, TurnRetryPolicy as W, AgentHarnessRuntime as _, createInProcessTransportPair as a, AgentMetadata as b, ConversationSummary as c, TranscriptPatch as d, AbortResult as f, AgentHarnessContext as g, AgentHarness as h, InProcessTransportPair as i, AgentToolContext as j, AgentSystemPromptContext as k, ServerFrame as l, AgentDisposeContext as m, AgentServerTransport as n, ClientFrame as o, AbortTarget as p, retryDelayMs as q, AgentTransport as r, ClientSessionEvent as s, AgentClientTransport as t, ShellCommandStatusLike as u, AgentHostContext as v, AgentSessionCloneParams as w, AgentPromptContext as x, AgentLifecycleEvent as y, DrainedTranscriptPatches as z };

package/dist/{websocket-transport-C6bu3Ufc.d.mts → websocket-transport-DdfGgBjO.d.mts}

@@ -1,4 +1,4 @@
-import { b as AgentMetadata, c as ConversationSummary, f as AbortResult, n as AgentServerTransport, s as ClientSessionEvent, t as AgentClientTransport } from "./transport-gkyOzXn-.mjs";
+import { F as ModelSwitchApply, b as AgentMetadata, c as ConversationSummary, f as AbortResult, n as AgentServerTransport, s as ClientSessionEvent, t as AgentClientTransport } from "./transport-DtBCCS_A.mjs";
 import { ProviderSelection } from "@demicodes/provider";
 import { Block, UserContentBlock } from "@demicodes/core";
 //#region src/client.d.ts
@@ -35,10 +35,15 @@ declare class AgentClient {
   }): Promise<void>;
   cancelPendingSteer(steerId: string): void;
   /**
-   * Switches the provider/model for an open session. The change takes effect on the next
-   * turn (the server applies it at a turn boundary), so this is fire-and-forget.
+   * Switches the provider/model for an open session. Fire-and-forget: the server records
+   * the switch and applies it per `options.apply` — 'next_turn' (default) waits for the
+   * next queued action, so a running turn finishes on the old model; 'immediate' applies
+   * at the next inference boundary, so mid-turn the very next sampling/tool continuation
+   * already runs on the new model.
    */
-  setProvider(provider: ProviderSelection): void;
+  setProvider(provider: ProviderSelection, options?: {
+    apply?: ModelSwitchApply;
+  }): void;
   /**
    * Discards the whole latest turn and runs it again from the user's input —
    * "regenerate". NOT a recovery path: it drops text the turn already emitted and

package/dist/{websocket-transport-BjSj3U0I.mjs → websocket-transport-c6P3NBuf.mjs}

@@ -156,13 +156,17 @@ var AgentClient = class {
 		});
 	}
 	/**
-	* Switches the provider/model for an open session. The change takes effect on the next
-	* turn (the server applies it at a turn boundary), so this is fire-and-forget.
+	* Switches the provider/model for an open session. Fire-and-forget: the server records
+	* the switch and applies it per `options.apply` — 'next_turn' (default) waits for the
+	* next queued action, so a running turn finishes on the old model; 'immediate' applies
+	* at the next inference boundary, so mid-turn the very next sampling/tool continuation
+	* already runs on the new model.
 	*/
-	setProvider(provider) {
+	setProvider(provider, options = {}) {
 		this.sendFrame({
 			type: "set_provider",
-			provider
+			provider,
+			apply: options.apply
 		});
 	}
 	/**

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@demicodes/agent",
   "description": "Session runtime and transport-neutral client and server protocol for Demi.",
-  "version": "0.15.0",
+  "version": "0.17.0",
   "private": false,
   "type": "module",
   "exports": {
@@ -19,10 +19,10 @@
     }
   },
   "dependencies": {
-    "@demicodes/core": "^0.15.0",
-    "@demicodes/provider": "^0.15.0",
-    "@demicodes/shell": "^0.15.0",
-    "@demicodes/utils": "^0.15.0"
+    "@demicodes/core": "^0.17.0",
+    "@demicodes/provider": "^0.17.0",
+    "@demicodes/shell": "^0.17.0",
+    "@demicodes/utils": "^0.17.0"
   },
   "license": "Apache-2.0",
   "main": "./dist/index.mjs",