agent-accelerator 0.2.0 → 0.4.0

package/src/agent/agent.ts

@@ -1,4 +1,4 @@
-import type { AgentConfig, AgentRunOptions } from "../types/agent.ts";
+import type { AgentConfig, AgentRunOptions, MidSessionConfig } from "../types/agent.ts";
 import type { ToolDefinition } from "../types/tool.ts";
 import type { ThinkingLevel, ThinkingConfig, CacheConfig, ServiceTier } from "../types/core.ts";
 import type { ContentPart } from "../types/message.ts";
@@ -8,10 +8,16 @@ import { AgentContext } from "./context.ts";
 import { resolveModel } from "../providers/registry.ts";
 import { buildAgentTools, createSubagentSpawnTool } from "./delegation.ts";
 import { convert_document_to_markdown } from "../utils/documents.ts";
-import { runAgentLoop, streamAgentLoop } from "./loop.ts";
+import { runAgentLoop, streamAgentLoop, type SteerEntry } from "./loop.ts";
 import { createSessionId } from "../utils/session.ts";
+import { getSubAgentTrace, subscribeToSubAgent, listSubAgentTraceIds, type SubAgentTrace } from "./delegation.ts";
+import { saveSessionDir, saveSessionFile } from "../session/store.ts";
+import { clearSessionRouting } from "../providers.ts";
+import { clearInteractionChains } from "../providers/google.ts";
 import { getModel, getSubModel } from "../utils/env.ts";
 import { validateModelThinking, ensureModelCatalogFresh } from "../models/catalog.ts";
+import { buildSessionData, type PersistedAgentSession, type SessionTotals,} from "../session/store.ts";
+import { SessionTelemetry } from "../session/store.ts";
 
 /**
  * Resolves a per-run thinking override without mutating agent config.
@@ -26,8 +32,49 @@ export function resolveEffectiveThinking(
   return { enabled: true, level: overrideLevel as ThinkingConfig["level"] };
 }
 
+import { ConfigError } from "../types/errors.ts";
+
+/**
+ * Disposers for the manual user-signal forward listeners (Q-21 fallback when
+ * `AbortSignal.any` is unavailable). Keyed by run controller so no expando
+ * properties are attached to AbortController instances.
+ */
+const mergeDisposers = new WeakMap<AbortController, () => void>();
+
+/** Max named sessions kept per Agent (Q-19a LRU). */
+const SESSION_LRU_MAX = 100;
+/** Idle TTL for named sessions (Q-19a). */
+const SESSION_TTL_MS = 60 * 60 * 1000;
+
+/** Normalized per-run limits (Q-22). Sub-agents inherit these unless overridden. */
+export interface AgentRunLimits {
+  maxTurns?: number;
+  deadlineMs?: number;
+  maxToolCallsPerRun?: number;
+  requestTimeoutMs: number;
+  firstByteTimeoutMs: number;
+  streamIdleTimeoutMs: number;
+}
+
+/** Resolves `(config as any).limits` with Q-22 defaults (timeouts tolerate reasoning). */
+function resolveAgentLimits(config: { limits?: Record<string, unknown> }): AgentRunLimits {
+  const raw = (config as { limits?: Record<string, unknown> }).limits ?? {};
+  const num = (v: unknown): number | undefined =>
+    typeof v === "number" && Number.isFinite(v) ? v : undefined;
+  return {
+    ...(num(raw.maxTurns) !== undefined ? { maxTurns: Math.max(1, Math.floor(num(raw.maxTurns)!)) } : {}),
+    ...(num(raw.deadlineMs) !== undefined ? { deadlineMs: num(raw.deadlineMs)! } : {}),
+    ...(num(raw.maxToolCallsPerRun) !== undefined
+      ? { maxToolCallsPerRun: Math.max(1, Math.floor(num(raw.maxToolCallsPerRun)!)) }
+      : {}),
+    requestTimeoutMs: num(raw.requestTimeoutMs) ?? 300_000,
+    firstByteTimeoutMs: num(raw.firstByteTimeoutMs) ?? 120_000,
+    streamIdleTimeoutMs: num(raw.streamIdleTimeoutMs) ?? 300_000,
+  };
+}
+
 /** Thrown when dynamic sub-agent spawning is enabled without an explicit model. */
-export class SubAgentModelError extends Error {
+export class SubAgentModelError extends ConfigError {
   readonly parentModel?: string;
 
   /**
@@ -39,20 +86,12 @@ export class SubAgentModelError extends Error {
   constructor(parentModel?: string) {
     const parentName = parentModel || "your main model";
     const formatted =
-      `\x1b[31m[Agent Accelerator] Missing Configuration: a sub-agent model is required when dynamic sub-agents are enabled\x1b[0m\n` +
-      `  \x1b[1mMain Agent Model:\x1b[0m ${parentName}\n` +
-      `  \x1b[1mIssue:\x1b[0m            Dynamic sub-agent delegation was enabled (dynamicSubagents.enabled), but no model was assigned for sub-agents.\n` +
-      `                    Sub-agents must never run on unverified models or default implicitly.\n\n` +
-      `  \x1b[36m💡 How to fix:\x1b[0m\n` +
-      `    1. Pass a model in your Agent configuration:\n` +
-      `       const agent = new Agent({\n` +
-      `         model: "${parentName}",\n` +
-      `         dynamicSubagents: { enabled: true, model: "provider/model-id" },\n` +
-      `       });\n\n` +
-      `    2. Or set the SUB_AGENT_MODEL environment variable in your .env or shell:\n` +
-      `       SUB_AGENT_MODEL="provider/model-id"`;
-
-    super(formatted);
+      `[Agent Accelerator] Missing Configuration: a sub-agent model is required when dynamic sub-agents are enabled. ` +
+      `Main Agent Model: ${parentName}. ` +
+      `Dynamic sub-agent delegation was enabled (dynamicSubagents.enabled), but no model was assigned for sub-agents. ` +
+      `Fix: pass dynamicSubagents.model in Agent config or set SUB_AGENT_MODEL.`;
+
+    super(formatted, { parentModel: parentName });
     this.name = "SubAgentModelError";
     this.parentModel = parentModel;
 
@@ -120,6 +159,37 @@ export class Agent {
   readonly context: AgentContext;
   /** Whether history is cleared around each run. */
   readonly stateless: boolean;
+  /** Durable persistence target: session file rewritten after every step. */
+  readonly persistConfig?: { dir?: string; file?: string };
+  /** Mid-session queue policy for new queries arriving while a run is active. */
+  readonly midSessionConfig: Required<Pick<MidSessionConfig, "mode">> & MidSessionConfig & { maxSteers?: number };
+  /** Max buffered steer entries per run (Q-20, default 20). */
+  readonly maxSteers: number;
+  /**
+   * Normalized run limits (Q-22). Read from `(config as any).limits`;
+   * sub-agents inherit these unless overridden. `maxTurns`/`deadlineMs`/
+   * `maxToolCallsPerRun` default to unlimited; timeouts default to
+   * request 300s / first-byte 120s / idle 300s (tolerate reasoning).
+   */
+  readonly runLimits: AgentRunLimits;
+  private currentRun: {
+    steerInbox: SteerEntry[];
+    outerStream?: AssistantMessageEventStream;
+    runPromise?: Promise<AgentResponse>;
+    abortController: AbortController;
+  } | null = null;
+  private pendingQueue: Array<{
+    prompt: string | ContentPart[];
+    options?: AgentRunOptions;
+    isStream: boolean;
+    deferredStream?: AssistantMessageEventStream;
+    resolve: (res: AgentResponse) => void;
+    reject: (err: unknown) => void;
+  }> = [];
+  /** Named session contexts for `sessionId` runs (Q-19a, LRU 100 + TTL). */
+  private sessions = new Map<string, { context: AgentContext; lastAccess: number }>();
+  /** Coalesced persistence single-flight flag (Q-28). */
+  private persistScheduled = false;
 
   /**
    * Creates an agent and registers its model, tools, cache, and delegation settings.
@@ -197,8 +267,32 @@ export class Agent {
     this.apiKey = (rawModel as any)?.apiKey || config.apiKey;
     this.baseUrl = (rawModel as any)?.baseUrl || config.baseUrl;
     this.customHeaders = config.headers;
-    this.maxTurns = config.maxTurns ?? 10;
+    this.maxTurns =
+      config.maxTurns === undefined || config.maxTurns === 0 || config.maxTurns === Infinity
+        ? Infinity
+        : Number.isFinite(config.maxTurns as number)
+          ? Math.max(1, Math.floor(config.maxTurns as number))
+          : Infinity;
     this.sessionId = config.sessionId || config.cache?.sessionId || createSessionId();
+    (this as any).subagentTraces = {};
+    this.persistConfig = config.persist;
+    const rawMode = config.midSession?.mode ?? "auto";
+    const normalizedMode = rawMode === "steer" || rawMode === "queue" ? rawMode : "auto";
+    const rawMaxQueued = (config.midSession as any)?.maxQueued;
+    const rawMaxSteers = (config.midSession as any)?.maxSteers;
+    this.midSessionConfig = {
+      mode: normalizedMode,
+      ...(Number.isFinite(rawMaxQueued) && (rawMaxQueued as number) >= 0
+        ? { maxQueued: Math.floor(rawMaxQueued as number) }
+        : { maxQueued: 20 }),
+      ...(Number.isFinite(rawMaxSteers) && (rawMaxSteers as number) >= 0
+        ? { maxSteers: Math.floor(rawMaxSteers as number) }
+        : { maxSteers: 20 }),
+    };
+    // Q-20: bound on buffered steer entries (default 20).
+    this.maxSteers = this.midSessionConfig.maxSteers ?? 20;
+    // Q-22: run limits (sub-agents inherit unless overridden).
+    this.runLimits = resolveAgentLimits(config as unknown as { limits?: Record<string, unknown> });
 
     // DX4: single thinkingLevel flag — also inherit from ModelProviderInstance when omitted
     const mpThinking = (rawModel as any)?.thinkingLevel;
@@ -301,41 +395,751 @@ export class Agent {
 
   /** Clears conversation messages and provider thought signatures while keeping configuration. */
   reset(): void {
+    const sid = this.sessionId;
     this.context.messages = [];
     this.context.thoughtSignatures = [];
     this.context.cachedContentId = (this.cacheConfig as any)?.cachedContentId;
     this.context.systemPrompt = this.getFullInstructions();
+    // Q-25: reset also drops session-scoped provider chaining/routing.
+    try { clearInteractionChains(sid); } catch {}
+    try { clearSessionRouting(sid); } catch {}
+  }
+
+  /**
+   * Returns a named session handle (Q-19a). `session()` creates an isolated
+   * `AgentContext` bound to `id` (separate histories per id); `agent.run`
+   * with `{ sessionId }` selects/creates the same entry. Sessions are
+   * ephemeral in-memory (LRU 100 + 1h TTL). `agent.run`/`stream` without a
+   * `sessionId` remain sugar over the default session for compat.
+   */
+  session(id?: string): { id: string; context: AgentContext; runController: AbortController | null } {
+    const sid = id || this.sessionId;
+    const ctx = this.getSessionContext(sid);
+    return { id: sid, context: ctx, runController: null };
+  }
+
+  /** Drops one named session (Q-19a). No-op for unknown ids. */
+  disposeSession(id: string): void {
+    try { this.sessions.delete(id); } catch {}
+    try { clearInteractionChains(id); } catch {}
+    try { clearSessionRouting(id); } catch {}
+  }
+
+  /** Drops all named sessions except the default context (Q-19a). */
+  clearSessions(): void {
+    try {
+      for (const id of [...this.sessions.keys()]) {
+        try { clearInteractionChains(id); } catch {}
+        try { clearSessionRouting(id); } catch {}
+      }
+      this.sessions.clear();
+    } catch {}
+  }
+
+  /** Resolves the context for a `sessionId` run (default session = `this.context`). */
+  private getSessionContext(sessionId?: string): AgentContext {
+    const sid = sessionId || this.sessionId;
+    if (sid === this.sessionId) return this.context;
+    const now = Date.now();
+    // TTL eviction.
+    try {
+      for (const [key, entry] of [...this.sessions.entries()]) {
+        if (now - entry.lastAccess > SESSION_TTL_MS) {
+          this.sessions.delete(key);
+          try { clearInteractionChains(key); } catch {}
+          try { clearSessionRouting(key); } catch {}
+        }
+      }
+    } catch {}
+    const existing = this.sessions.get(sid);
+    if (existing) {
+      existing.lastAccess = now;
+      return existing.context;
+    }
+    const fresh = new AgentContext(this.getFullInstructions());
+    try {
+      const cached = (this.cacheConfig as unknown as { cachedContentId?: string })?.cachedContentId;
+      if (cached) fresh.cachedContentId = cached;
+    } catch {}
+    this.sessions.set(sid, { context: fresh, lastAccess: now });
+    // LRU eviction (oldest lastAccess first).
+    try {
+      while (this.sessions.size > SESSION_LRU_MAX) {
+        let oldest: string | undefined;
+        let oldestTs = Infinity;
+        for (const [key, entry] of this.sessions) {
+          if (entry.lastAccess < oldestTs) { oldestTs = entry.lastAccess; oldest = key; }
+        }
+        if (!oldest) break;
+        this.sessions.delete(oldest);
+        try { clearInteractionChains(oldest); } catch {}
+        try { clearSessionRouting(oldest); } catch {}
+      }
+    } catch {}
+    return fresh;
+  }
+
+  /** Rejects + drops all queued follow-ups (Q-20). Returns the cleared count. */
+  clearQueue(): number {
+    const n = this.pendingQueue.length;
+    if (n === 0) return 0;
+    const err = Object.assign(new Error("Queue cleared"), { name: "AbortError" });
+    const pending = this.pendingQueue.splice(0);
+    for (const entry of pending) {
+      try { entry.reject(err); } catch {}
+      try { entry.deferredStream?.fail(err instanceof Error ? err : new Error(String(err))); } catch {}
+    }
+    return n;
+  }
+
+  /** Throws when the steer inbox is at its Q-20 bound. */
+  private assertSteerRoom(inbox: SteerEntry[]): void {
+    const max = this.maxSteers ?? 20;
+    if (inbox.length >= max) {
+      throw new Error(
+        `[Agent Accelerator] Steer inbox is full (${inbox.length}/${max}). Wait for the active run to drain or increase midSession.maxSteers.`
+      );
+    }
+  }
+
+  /**
+   * True while a `run()`/`stream()` turn is active on this agent.
+   * Use with `steer()` (current-turn redirect) vs `queue()` (next-turn follow-up).
+   */
+  get isBusy(): boolean {
+    return this.currentRun !== null;
+  }
+
+  /** Number of queued follow-up requests waiting for the active run to finish. */
+  get pendingCount(): number {
+    return this.pendingQueue.length;
+  }
+
+  /** Number of steer messages buffered for injection into the active turn. */
+  get steerPendingCount(): number {
+    return this.currentRun?.steerInbox.length ?? 0;
+  }
+
+  /** Effective mid-session policy (`steer` | `queue` | `auto`). */
+  get midSessionMode(): "steer" | "queue" | "auto" {
+    return this.midSessionConfig.mode;
   }
 
-  private prepareTurn(prompt: string | ContentPart[], options?: AgentRunOptions): void {
+  private resolveEnqueueMode(requested?: "steer" | "queue"): "steer" | "queue" {
+    const enforced = this.midSessionConfig.mode;
+    if (enforced === "steer" || enforced === "queue") return enforced;
+    return requested === "steer" ? "steer" : "queue";
+  }
+
+  private assertPromptValid(prompt: string | ContentPart[]): void {
     if (Array.isArray(prompt) && prompt.length === 0) {
       throw new Error(
         "[Agent Accelerator] Prompt cannot be empty: pass a non-empty string or at least one content part."
       );
     }
+  }
+
+  private previewPrompt(prompt: string | ContentPart[]): string {
+    if (typeof prompt === "string") return prompt.slice(0, 500);
+    try {
+      const texts = prompt
+        .filter((p) => (p as { type?: unknown }).type === "text")
+        .map((p) => (p as { text?: string }).text ?? "")
+        .join("\n");
+      return (texts || "[multipart]").slice(0, 500);
+    } catch {
+      return "[multipart]";
+    }
+  }
+
+  /**
+   * Interrupts the active run immediately (STOP. Do this instead).
+   * Cancels the active stream and aborts provider/tool/sub-agent work via
+   * the run's internal `AbortController`. No-op when idle. The interrupted
+   * run rejects with `AbortError`; queued follow-ups still run afterwards.
+   */
+  interrupt(reason?: unknown): void {
+    const active = this.currentRun;
+    if (!active) return;
+    try {
+      active.abortController.abort(reason instanceof Error ? reason : new Error("Agent interrupted"));
+    } catch {}
+    try {
+      active.outerStream?.cancel(reason);
+    } catch {}
+  }
+
+  /**
+   * Steers the active turn (While you're doing that, change direction).
+   * The current provider/tool work finishes first, then `prompt` is injected
+   * as a user message into the SAME run — same `AgentResponse`, same
+   * `maxTurns` budget. Resolves to the active run's final response.
+   * When idle, behaves like `run(prompt, options)`.
+   * Throws when the developer enforces `midSession: { mode: "queue" }`.
+   * Note: `options` only honors `additionalContext` mid-turn (thinking level,
+   * session, and streaming callbacks stay with the active run). For a
+   * streaming follow-up use `stream(prompt, { enqueue: "steer" })`.
+   */
+  steer(prompt: string | ContentPart[], options?: AgentRunOptions): Promise<AgentResponse> {
+    this.assertPromptValid(prompt);
+    if (options?.stream === true) {
+      throw new Error(
+        '[Agent Accelerator] steer() is non-streaming. Use stream(prompt, { enqueue: "steer" }) for a streaming steer.'
+      );
+    }
+    if (this.midSessionConfig.mode === "queue") {
+      throw new Error(
+        '[Agent Accelerator] steer() is disabled by midSession.mode "queue". Use queue() or set mode to "auto"/"steer".'
+      );
+    }
+    const active = this.currentRun;
+    if (!active) return this.run(prompt, options) as Promise<AgentResponse>;
+    this.assertSteerRoom(active.steerInbox);
+    active.steerInbox.push({ prompt, options });
+    this.persistNow();
+    if (active.runPromise) return active.runPromise;
+    if (active.outerStream) return active.outerStream.result();
+    return this.run(prompt, options) as Promise<AgentResponse>;
+  }
+
+  /**
+   * Queues a follow-up for the next turn (When you're done, do this next).
+   * The active run finishes completely first; this prompt then starts a new
+   * run with its own turn budget and `AgentResponse`. Resolves to that next
+   * turn's response. When idle, behaves like `run(prompt, options)`.
+   * Throws when the developer enforces `midSession: { mode: "steer" }` or
+   * when `maxQueued` is exceeded. For a streaming follow-up use
+   * `stream(prompt, { enqueue: "queue" })`.
+   */
+  queue(prompt: string | ContentPart[], options?: AgentRunOptions): Promise<AgentResponse> {
+    this.assertPromptValid(prompt);
+    if (options?.stream === true) {
+      throw new Error(
+        '[Agent Accelerator] queue() is non-streaming. Use stream(prompt, { enqueue: "queue" }) for a streaming follow-up.'
+      );
+    }
+    if (this.midSessionConfig.mode === "steer") {
+      throw new Error(
+        '[Agent Accelerator] queue() is disabled by midSession.mode "steer". Use steer() or set mode to "auto"/"queue".'
+      );
+    }
+    if (!this.currentRun) return this.run(prompt, options) as Promise<AgentResponse>;
+    const max = this.midSessionConfig.maxQueued ?? 20;
+    if (this.pendingQueue.length >= max) {
+      throw new Error(
+        `[Agent Accelerator] Queue is full (${this.pendingQueue.length}/${max}). Wait for the active run to finish or increase midSession.maxQueued.`
+      );
+    }
+    return new Promise<AgentResponse>((resolve, reject) => {
+      this.pendingQueue.push({ prompt, options, isStream: false, resolve, reject });
+      try {
+        this.currentRun?.outerStream?.push({
+          type: "queued",
+          queuedPrompt: this.previewPrompt(prompt),
+          queueLength: this.pendingQueue.length,
+        } as any);
+      } catch {}
+    });
+  }
+
+  private pumpQueue(): void {
+    if (this.currentRun !== null) return;
+    const next = this.pendingQueue.shift();
+    if (!next) return;
+    // Q-20: reject queued items whose signal aborted before start.
+    try {
+      if ((next.options as { signal?: AbortSignal } | undefined)?.signal?.aborted) {
+        const reason = ((next.options as { signal?: AbortSignal }).signal as unknown as { reason?: unknown })?.reason;
+        const abortErr = Object.assign(
+          reason instanceof Error ? reason : new Error("Queued run aborted before start"),
+          { name: "AbortError" }
+        );
+        try { next.reject(abortErr); } catch {}
+        try { next.deferredStream?.fail(abortErr instanceof Error ? abortErr : new Error(String(abortErr))); } catch {}
+        queueMicrotask(() => this.pumpQueue());
+        return;
+      }
+    } catch {}
+    if (next.isStream && next.deferredStream) {
+      const deferred = next.deferredStream;
+      try {
+        const real = this.startStreamingRun(next.prompt, next.options);
+        try { (deferred as any).__piped = true; } catch {}
+        const forward = (e: StreamEvent) => {
+          try { deferred.push(e); } catch {}
+        };
+        real.on("*", forward as any);
+        try {
+          deferred.onCancel(() => {
+            try { real.cancel(); } catch {}
+          });
+        } catch {}
+        real.result().then(
+          (res) => {
+            try { next.resolve(res); } catch {}
+            try { deferred.end(res); } catch {}
+          },
+          (err) => {
+            try { next.reject(err); } catch {}
+            // Pipe failures into the deferred stream so for-await sees them.
+            try { deferred.fail(err instanceof Error ? err : new Error(String(err))); } catch {}
+          }
+        );
+      } catch (err) {
+        try { next.reject(err); } catch {}
+        try { deferred.fail(err instanceof Error ? err : new Error(String(err))); } catch {}
+        // Keep draining even if starting failed.
+        queueMicrotask(() => this.pumpQueue());
+      }
+      return;
+    }
+    this.startNonStreamingRun(next.prompt, next.options).then(next.resolve, next.reject);
+  }
+
+  private finishCurrentRun(): void {
+    // Q-20: leftover steer entries become queued follow-ups instead of dropping.
+    let leftover: SteerEntry[] = [];
+    try { leftover = this.currentRun?.steerInbox?.splice(0) ?? []; } catch { leftover = []; }
+    this.currentRun = null;
+    if (leftover.length > 0) {
+      try {
+        const max = this.midSessionConfig.maxQueued ?? 20;
+        for (const entry of leftover) {
+          try {
+            if ((entry.options as { signal?: AbortSignal } | undefined)?.signal?.aborted) continue;
+          } catch { continue; }
+          if (this.pendingQueue.length >= max) break;
+          this.pendingQueue.push({
+            prompt: entry.prompt,
+            options: entry.options,
+            isStream: false,
+            resolve: () => {},
+            reject: () => {},
+          });
+        }
+      } catch {}
+    }
+    // Defer pump so the completing run's .then handlers settle first.
+    queueMicrotask(() => this.pumpQueue());
+  }
+
+  private cleanupMerge(runController: AbortController): void {
+    try {
+      const dispose = mergeDisposers.get(runController);
+      if (dispose) {
+        try { dispose(); } catch {}
+        mergeDisposers.delete(runController);
+      }
+    } catch {}
+  }
+
+  private mergeUserSignal(userSignal?: AbortSignal, runController?: AbortController): AbortSignal | undefined {
+    if (!runController) return userSignal;
+    if (!userSignal) return runController.signal;
+    if (userSignal.aborted) {
+      try {
+        runController.abort((userSignal as any).reason);
+      } catch {
+        try { runController.abort(); } catch {}
+      }
+      return runController.signal;
+    }
+    // Q-21: prefer AbortSignal.any (no expando, no listener leak).
+    try {
+      const anyFn = (AbortSignal as unknown as { any?: (signals: AbortSignal[]) => AbortSignal }).any;
+      if (typeof anyFn === "function") {
+        return anyFn.call(AbortSignal, [userSignal, runController.signal]);
+      }
+    } catch {}
+    // Fallback: manual forward without expando (disposer in WeakMap).
+    const forward = () => {
+      try {
+        runController.abort((userSignal as any).reason);
+      } catch {
+        try { runController.abort(); } catch {}
+      }
+    };
+    try { userSignal.addEventListener("abort", forward, { once: true }); } catch {}
+    try {
+      mergeDisposers.set(runController, () => {
+        try { userSignal.removeEventListener("abort", forward); } catch {}
+      });
+    } catch {}
+    return runController.signal;
+  }
+
+  private startNonStreamingRun(
+    prompt: string | ContentPart[],
+    options?: AgentRunOptions
+  ): Promise<AgentResponse> {
+    const runController = new AbortController();
+    const steerInbox: SteerEntry[] = [];
+    const mergedSignal = this.mergeUserSignal(options?.signal, runController);
+    const runOptions = { ...options, signal: mergedSignal };
+    // Q-19a: route to the named session context (default = this.context).
+    const targetContext = this.getSessionContext(runOptions?.sessionId || this.sessionId);
+    const task = (async (): Promise<AgentResponse> => {
+      await ensureModelCatalogFresh();
+      const resolved = resolveModel(this.modelStringOrSpec);
+      const effectiveLevel = runOptions?.thinkingLevel || this.thinkingConfig?.level;
+      if (effectiveLevel) {
+        validateModelThinking(resolved.provider.id, resolved.modelId, effectiveLevel);
+      }
+      this.prepareTurn(prompt, runOptions, targetContext);
+      this.persistNow();
+
+      const providerOptions = {
+        apiKey: this.apiKey,
+        baseUrl: this.baseUrl,
+        headers: { ...(this.customHeaders ?? {}), ...(runOptions?.headers ?? {}) },
+        thinking: resolveEffectiveThinking(this.thinkingConfig, runOptions?.thinkingLevel),
+        cache: this.stateless
+          ? { sessionId: runOptions?.sessionId || this.sessionId }
+          : { ...this.cacheConfig, sessionId: runOptions?.sessionId || this.sessionId },
+        serviceTier: this.serviceTier,
+        sessionId: runOptions?.sessionId || this.sessionId,
+      };
+
+      const loopConfig = {
+        agentName: this.name,
+        provider: resolved.provider,
+        modelId: resolved.modelId,
+        context: targetContext,
+        tools: this.tools,
+        options: providerOptions,
+        runOptions,
+        maxTurns: this.maxTurns,
+        bypassInputFileModality: this.bypassInputFileModality,
+        onProgress: () => this.persistNow(),
+        steerInbox,
+        onSteerInjected: () => this.persistNow(),
+        // Q-22: loop enforces these (plus local timeout races).
+        limits: { ...this.runLimits },
+      };
+
+      const res = await runAgentLoop(loopConfig);
+      this.mergeSubagentTraces(res.subagents as any);
+      this.persistNow();
+      if (this.stateless) {
+        targetContext.messages = [];
+        targetContext.thoughtSignatures = [];
+      }
+      return res;
+    })();
+
+    this.currentRun = { steerInbox, abortController: runController, runPromise: task };
+    const cleanup = () => {
+      try { this.cleanupMerge(runController); } catch {}
+      if (this.currentRun?.runPromise === task) this.finishCurrentRun();
+    };
+    task.then(cleanup, cleanup);
+    return task;
+  }
+
+  private startStreamingRun(
+    prompt: string | ContentPart[],
+    options?: AgentRunOptions
+  ): AssistantMessageEventStream {
+    const resolved = resolveModel(this.modelStringOrSpec);
+    const effectiveLevel = options?.thinkingLevel || this.thinkingConfig?.level;
+    if (effectiveLevel) {
+      validateModelThinking(resolved.provider.id, resolved.modelId, effectiveLevel);
+    }
+    // Q-19a: named session routing (default session for compat).
+    const targetContext = this.getSessionContext(options?.sessionId || this.sessionId);
+    this.prepareTurn(prompt, options, targetContext);
+    this.persistNow();
+
+    const runController = new AbortController();
+    const mergedSignal = this.mergeUserSignal(options?.signal, runController);
+    const runOptions = { ...options, signal: mergedSignal };
+    const steerInbox: SteerEntry[] = [];
+
+    const providerOptions = {
+      apiKey: this.apiKey,
+      baseUrl: this.baseUrl,
+      headers: { ...(this.customHeaders ?? {}), ...(runOptions?.headers ?? {}) },
+      thinking: resolveEffectiveThinking(this.thinkingConfig, runOptions?.thinkingLevel),
+      cache: this.stateless
+        ? { sessionId: runOptions?.sessionId || this.sessionId }
+        : { ...this.cacheConfig, sessionId: runOptions?.sessionId || this.sessionId },
+      serviceTier: this.serviceTier,
+      sessionId: runOptions?.sessionId || this.sessionId,
+    };
+
+    const loopConfig = {
+      agentName: this.name,
+      provider: resolved.provider,
+      modelId: resolved.modelId,
+      context: targetContext,
+      tools: this.tools,
+      options: providerOptions,
+      runOptions,
+      maxTurns: this.maxTurns,
+      bypassInputFileModality: this.bypassInputFileModality,
+      onProgress: () => this.persistNow(),
+      steerInbox,
+      onSteerInjected: () => this.persistNow(),
+      // Q-22: loop enforces these (plus local timeout races).
+      limits: { ...this.runLimits },
+    };
+
+    const s = streamAgentLoop(loopConfig);
+    this.currentRun = { steerInbox, outerStream: s, abortController: runController };
+    s.result().then(
+      (res) => {
+        try {
+          this.mergeSubagentTraces(res.subagents as any);
+          this.persistNow();
+        } catch {}
+        try { this.cleanupMerge(runController); } catch {}
+        if (this.stateless) {
+          targetContext.messages = [];
+          targetContext.thoughtSignatures = [];
+        }
+        if (this.currentRun?.outerStream === s) this.finishCurrentRun();
+      },
+      () => {
+        try { this.cleanupMerge(runController); } catch {}
+        if (this.currentRun?.outerStream === s) this.finishCurrentRun();
+      }
+    );
+    return s;
+  }
+
+  /**
+   * Returns a snapshot of one sub-agent worker's trace by tracking id
+   * (`SUBAGENT-NAME-{32hex}` or the raw 32-hex suffix). Sees live running
+   * state: steps appear as the worker acts, not just after it finishes.
+   *
+   * @example `const trace = agent.track("RAM-RESEARCH-AGENT-9f2c…");`
+   */
+  track(trackingId: string): SubAgentTrace | undefined {
+    if (!trackingId) return undefined;
+    try {
+      const own = (this as any).subagentTraces as Record<string, SubAgentTrace> | undefined;
+      const hit = own?.[trackingId];
+      if (hit) return { ...hit, steps: hit.steps.map((s) => ({ ...s })) };
+    } catch {}
+    return getSubAgentTrace(trackingId);
+  }
+
+  /** Lists TrackingIDs of worker traces visible to this agent (deduped). */
+  listTrackedSubAgents(): string[] {
+    const rawOf = (id: string): string => {
+      const tail = id.split("-").pop() ?? "";
+      return /^[0-9a-f]{32}$/.test(tail) ? tail : id;
+    };
+    const seen = new Set<string>();
+    const out: string[] = [];
+    for (const id of [...Object.keys((this as any).subagentTraces ?? {}), ...listSubAgentTraceIds()]) {
+      const raw = rawOf(id);
+      if (!seen.has(raw)) {
+        seen.add(raw);
+        out.push(raw);
+      }
+    }
+    return out;
+  }
+
+  /**
+   * Subscribes to live updates of one worker trace. The callback fires on
+   * every logged step; returns an unsubscribe fn.
+   *
+   * @example `const off = agent.subscribeToSubAgent(id, (t) => render(t));`
+   */
+  subscribeToSubAgent(trackingId: string, fn: (trace: SubAgentTrace) => void): () => void {
+    return subscribeToSubAgent(trackingId, fn);
+  }
+
+  /**
+   * Rewrites the configured session file now (when `persist` is set).
+   * Called automatically after every step during runs; safe to call
+   * manually. Never throws (failures route to `onPersistError`).
+   *
+   * Q-28: stateless agents skip persistence unless
+   * `(persist as any).allowStateless` is set. Calls in the same tick are
+   * coalesced via a `queueMicrotask` single flight.
+   */
+  persistNow(): void {
+    const cfg = this.persistConfig;
+    if (!cfg || (!cfg.dir && !cfg.file)) return;
+    // Q-28: no stateless persist by default.
+    try {
+      if (this.stateless && !(cfg as unknown as { allowStateless?: unknown }).allowStateless) return;
+    } catch {}
+    if (this.persistScheduled) return;
+    this.persistScheduled = true;
+    queueMicrotask(() => {
+      this.persistScheduled = false;
+      this.doPersist();
+    });
+  }
+
+  /** Synchronous persistence write behind the coalesced `persistNow()` gate. */
+  private doPersist(): void {
+    const cfg = this.persistConfig;
+    if (!cfg || (!cfg.dir && !cfg.file)) return;
+    try {
+      if (this.stateless && !(cfg as unknown as { allowStateless?: unknown }).allowStateless) return;
+    } catch {}
+    try {
+      if (cfg.dir) saveSessionDir(cfg.dir, this as any, null);
+      else if (cfg.file) saveSessionFile(cfg.file, this as any, null);
+    } catch (err) {
+      // Q-28: surface via hook instead of swallowing silently.
+      try { (cfg as unknown as { onPersistError?: (e: unknown) => void }).onPersistError?.(err); } catch {}
+    }
+  }
+
+  /** Merges completed worker metadata (incl. steps) into this agent's trace registry. */
+  private mergeSubagentTraces(subagents: Array<{ trackingId?: string; name: string; sessionId?: string; parentSessionId?: string; task: string; steps?: SubAgentTrace["steps"]; [k: string]: unknown }>): void {
+    if (!subagents || subagents.length === 0) return;
+    const rawOf = (id?: string): string | undefined => {
+      if (!id) return undefined;
+      const tail = id.split("-").pop() ?? "";
+      return /^[0-9a-f]{32}$/.test(tail) ? tail : id;
+    };
+    try {
+      const registry = ((this as any).subagentTraces ??= {}) as Record<string, SubAgentTrace>;
+      for (const s of subagents) {
+        const raw = rawOf(s.trackingId);
+        const key = raw || s.name;
+        const existing = registry[key] ?? (raw ? registry[s.trackingId ?? ""] : undefined);
+        if (existing) {
+          existing.status = "done";
+          if (s.steps && s.steps.length > 0 && existing.steps.length === 0) {
+            existing.steps = s.steps.map((x) => ({ ...x }));
+          }
+          if (raw && !registry[raw]) registry[raw] = existing;
+        } else if (raw) {
+          registry[raw] = {
+            trackingId: raw,
+            name: s.name,
+            sessionId: s.sessionId || "",
+            parentSessionId: s.parentSessionId || this.sessionId,
+            status: "done",
+            task: s.task,
+            turns: 0,
+            steps: (s.steps ?? []).map((x) => ({ ...x })),
+          };
+        }
+      }
+    } catch {}
+  }
+
+  /**
+   * Exports this agent's conversation + config as a storable session snapshot.
+   * Pair with `importSession` / `loadSessionFile` / `saveSessionFile`.
+   *
+   * @example `saveSessionFile(".session.json", agent, telemetry)`
+   */
+  exportSession(telemetry?: SessionTelemetry | SessionTotals | null): PersistedAgentSession {
+    return buildSessionData(this as any, telemetry);
+  }
+
+  /**
+   * Restores conversation + config from `exportSession` / `loadSessionFile`.
+   * Restores messages, system prompt, cached content id, sub-agent traces,
+   * and (when present) session id, model, thinking level, instructions,
+   * cache, and worker model. No-ops on nullish input.
+   *
+   * @example `const saved = loadSessionFile(".session.json"); if (saved) agent.importSession(saved);`
+   */
+  importSession(data?: PersistedAgentSession | null): void {
+    if (!data) return;
+    const prevSid = this.sessionId;
+    if (data.sessionId) (this as any).sessionId = data.sessionId;
+    if ((data as any).parentSessionId) (this as any).parentSessionId = (data as any).parentSessionId;
+    if ((data as any).subagents && typeof (data as any).subagents === "object") {
+      try {
+        (this as any).subagentTraces = JSON.parse(JSON.stringify((data as any).subagents));
+      } catch {
+        (this as any).subagentTraces = { ...((data as any).subagents as object) };
+      }
+    }
+    if (data.model) (this as any).modelStringOrSpec = data.model;
+    if (data.thinkingLevel) {
+      const lvl = String(data.thinkingLevel);
+      (this as any).thinkingConfig =
+        lvl === "none"
+          ? { enabled: false, level: "none", budgetTokens: 0 }
+          : lvl === "dynamic"
+            ? { enabled: true, level: "dynamic", budgetTokens: -1 }
+            : { enabled: true, level: lvl };
+    }
+    if (typeof data.instructions === "string") this.instructions = data.instructions;
+    if (data.cache) {
+      (this as any).cacheConfig = {
+        ...this.cacheConfig,
+        ...data.cache,
+        sessionId: data.sessionId ?? this.sessionId,
+      };
+    }
+    if (data.subagentModel) {
+      (this as any).subagentModel = data.subagentModel;
+      const dyn = (this as any).dynamicSubagents;
+      if (dyn) dyn.model = data.subagentModel;
+    }
+    try {
+      const g: any = globalThis as any;
+      this.context.messages = Array.isArray(data.messages)
+        ? (typeof g.structuredClone === "function"
+            ? g.structuredClone(data.messages)
+            : JSON.parse(JSON.stringify(data.messages)))
+        : [];
+    } catch {
+      this.context.messages = Array.isArray(data.messages) ? [...data.messages] : [];
+    }
+    this.context.thoughtSignatures = [];
+    if (data.systemPrompt) {
+      this.context.systemPrompt = data.systemPrompt;
+    } else {
+      this.context.systemPrompt = this.getFullInstructions();
+    }
+    this.context.cachedContentId =
+      data.cachedContentId ?? (this.cacheConfig as any)?.cachedContentId;
+    // Q-25: importing a snapshot invalidates prior provider chaining/routing
+    // for both the previous and the restored session.
+    try { clearInteractionChains(prevSid); } catch {}
+    try { clearSessionRouting(prevSid); } catch {}
+    try { clearInteractionChains(this.sessionId); } catch {}
+    try { clearSessionRouting(this.sessionId); } catch {}
+  }
+
+  private prepareTurn(prompt: string | ContentPart[], options?: AgentRunOptions, target?: AgentContext): void {
+    // Q-20: prompt validation throws synchronously in both run() and
+    // stream() (same timing; documented here).
+    if (Array.isArray(prompt) && prompt.length === 0) {
+      throw new Error(
+        "[Agent Accelerator] Prompt cannot be empty: pass a non-empty string or at least one content part."
+      );
+    }
+    const ctx = target ?? this.getSessionContext(options?.sessionId || this.sessionId);
     if (this.stateless) {
-      this.context.messages = [];
-      this.context.thoughtSignatures = [];
+      // Q-19a: stateless runs use a fresh ephemeral context (never shared).
+      ctx.messages = [];
+      ctx.thoughtSignatures = [];
     }
     const fullInstructions = this.getFullInstructions();
     // C13: keep systemPrompt stable for Google implicit cache; additionalContext goes as user prefix, not system mutation
     if (options?.additionalContext) {
       // Preserve stable instructions as systemPrompt
-      this.context.systemPrompt = fullInstructions;
+      ctx.systemPrompt = fullInstructions;
       const prefix = `[Additional Context]\n${options.additionalContext}\n\n`;
       if (typeof prompt === "string") {
         prompt = prefix + prompt;
       } else if (Array.isArray(prompt)) {
         prompt = [{ type: "text", text: prefix } as ContentPart, ...prompt];
       }
-    } else if (this.context.systemPrompt !== fullInstructions) {
-      this.context.systemPrompt = fullInstructions;
+    } else if (ctx.systemPrompt !== fullInstructions) {
+      ctx.systemPrompt = fullInstructions;
     }
     // C11: keep context cachedContentId in sync with cacheConfig if updated via options
-    if (options?.headers && (this.cacheConfig as any)?.cachedContentId && !this.context.cachedContentId) {
-      this.context.cachedContentId = (this.cacheConfig as any).cachedContentId;
+    if (options?.headers && (this.cacheConfig as any)?.cachedContentId && !ctx.cachedContentId) {
+      ctx.cachedContentId = (this.cacheConfig as any).cachedContentId;
     }
-    this.context.addUserMessage(prompt);
+    ctx.addUserMessage(prompt);
   }
 
   /**
@@ -362,6 +1166,9 @@ export class Agent {
     prompt: string | ContentPart[],
     options?: AgentRunOptions
   ): Promise<AgentResponse> | AssistantMessageEventStream {
+    // Q-20: prompt validation throws synchronously in both run() and
+    // stream() (same timing; async work starts only after validation).
+    this.assertPromptValid(prompt);
     const isStream = options?.stream === true;
     if (isStream) {
       const hasCallbacks = !!(options?.onDelta || options?.onThinkingDelta || options?.onEvent);
@@ -384,48 +1191,51 @@ export class Agent {
       return s;
     }
 
-    return (async (): Promise<AgentResponse> => {
-      // Refresh the catalog BEFORE thinking validation so levels are checked
-      // against live data instead of a possibly-empty cold cache.
-      await ensureModelCatalogFresh();
-      const resolved = resolveModel(this.modelStringOrSpec);
-      const effectiveLevel = options?.thinkingLevel || this.thinkingConfig?.level;
-      if (effectiveLevel) {
-        validateModelThinking(resolved.provider.id, resolved.modelId, effectiveLevel);
+    // Mid-session: busy agent serializes via steer (same turn) or queue (next turn).
+    if (this.currentRun) {
+      const enforced = this.midSessionConfig.mode;
+      const requested = options?.enqueue;
+      if (requested && (enforced === "steer" || enforced === "queue") && requested !== enforced) {
+        throw new Error(
+          `[Agent Accelerator] enqueue "${requested}" is disabled by midSession.mode "${enforced}". Use ${enforced}() or set mode to "auto"/"${requested}".`
+        );
       }
-      this.prepareTurn(prompt, options);
-
-      const providerOptions = {
-        apiKey: this.apiKey,
-        baseUrl: this.baseUrl,
-        headers: { ...(this.customHeaders ?? {}), ...(options?.headers ?? {}) },
-        thinking: resolveEffectiveThinking(this.thinkingConfig, options?.thinkingLevel),
-        cache: this.stateless
-          ? { sessionId: options?.sessionId || this.sessionId }
-          : { ...this.cacheConfig, sessionId: options?.sessionId || this.sessionId },
-        serviceTier: this.serviceTier,
-        sessionId: options?.sessionId || this.sessionId,
-      };
-
-      const loopConfig = {
-        agentName: this.name,
-        provider: resolved.provider,
-        modelId: resolved.modelId,
-        context: this.context,
-        tools: this.tools,
-        options: providerOptions,
-        runOptions: options,
-        maxTurns: this.maxTurns,
-        bypassInputFileModality: this.bypassInputFileModality,
-      };
-
-      const res = await runAgentLoop(loopConfig);
-      if (this.stateless) {
-        this.context.messages = [];
-        this.context.thoughtSignatures = [];
+      const mode = this.resolveEnqueueMode(requested);
+      if (mode === "steer") {
+        this.assertPromptValid(prompt);
+        this.assertSteerRoom(this.currentRun.steerInbox);
+        this.currentRun.steerInbox.push({ prompt, options });
+        this.persistNow();
+        const active = this.currentRun;
+        if (active.runPromise) return active.runPromise;
+        if (active.outerStream) return active.outerStream.result();
+      } else {
+        if (this.midSessionConfig.mode === "steer") {
+          throw new Error(
+            '[Agent Accelerator] enqueue "queue" is disabled by midSession.mode "steer". Use steer() or set mode to "auto"/"queue".'
+          );
+        }
+        this.assertPromptValid(prompt);
+        const max = this.midSessionConfig.maxQueued ?? 20;
+        if (this.pendingQueue.length >= max) {
+          throw new Error(
+            `[Agent Accelerator] Queue is full (${this.pendingQueue.length}/${max}). Wait for the active run to finish or increase midSession.maxQueued.`
+          );
+        }
+        return new Promise<AgentResponse>((resolve, reject) => {
+          this.pendingQueue.push({ prompt, options, isStream: false, resolve, reject });
+          try {
+            this.currentRun?.outerStream?.push({
+              type: "queued",
+              queuedPrompt: this.previewPrompt(prompt),
+              queueLength: this.pendingQueue.length,
+            } as any);
+          } catch {}
+        });
       }
-      return res;
-    })();
+    }
+
+    return this.startNonStreamingRun(prompt, options);
   }
 
   /**
@@ -495,44 +1305,141 @@ export class Agent {
     prompt: string | ContentPart[],
     options?: AgentRunOptions
   ): AssistantMessageEventStream {
-    const resolved = resolveModel(this.modelStringOrSpec);
-    const effectiveLevel = options?.thinkingLevel || this.thinkingConfig?.level;
-    if (effectiveLevel) {
-      validateModelThinking(resolved.provider.id, resolved.modelId, effectiveLevel);
+    // Q-20: same sync validation timing as run().
+    this.assertPromptValid(prompt);
+    // Mid-session: busy agent serializes via steer (same turn) or queue (next turn).
+    if (this.currentRun) {
+      const enforcedStream = this.midSessionConfig.mode;
+      const requestedStream = options?.enqueue;
+      if (requestedStream && (enforcedStream === "steer" || enforcedStream === "queue") && requestedStream !== enforcedStream) {
+        throw new Error(
+          `[Agent Accelerator] enqueue "${requestedStream}" is disabled by midSession.mode "${enforcedStream}". Use ${enforcedStream}() or set mode to "auto"/"${requestedStream}".`
+        );
+      }
+      const mode = this.resolveEnqueueMode(requestedStream);
+      if (mode === "steer") {
+        this.assertPromptValid(prompt);
+        this.assertSteerRoom(this.currentRun.steerInbox);
+        this.currentRun.steerInbox.push({ prompt, options });
+        this.persistNow();
+        const activeStream = this.currentRun.outerStream;
+        if (activeStream) {
+          this.wireStreamCallbacks(activeStream, options);
+          return activeStream;
+        }
+        // Active run is non-streaming: attach a lightweight stream that ends
+        // with the same final response so streaming callers still get events.
+        const stub = new AssistantMessageEventStream();
+        this.wireStreamCallbacks(stub, options);
+        const activePromise = this.currentRun.runPromise;
+        if (activePromise) {
+          activePromise.then(
+            (res) => {
+              try {
+                stub.push({ type: "done", delta: "", usage: res.usage, finishReason: res.finishReason, responseId: res.responseId } as any);
+              } catch {}
+              try { stub.end(res); } catch {}
+            },
+            (err) => {
+              try { stub.fail(err instanceof Error ? err : new Error(String(err))); } catch {}
+            }
+          );
+        } else {
+          try { stub.fail(new Error("[Agent Accelerator] No active run to steer.")); } catch {}
+        }
+        return stub;
+      }
+      if (this.midSessionConfig.mode === "steer") {
+        throw new Error(
+          '[Agent Accelerator] enqueue "queue" is disabled by midSession.mode "steer". Use steer() or set mode to "auto"/"queue".'
+        );
+      }
+      this.assertPromptValid(prompt);
+      const max = this.midSessionConfig.maxQueued ?? 20;
+      if (this.pendingQueue.length >= max) {
+        throw new Error(
+          `[Agent Accelerator] Queue is full (${this.pendingQueue.length}/${max}). Wait for the active run to finish or increase midSession.maxQueued.`
+        );
+      }
+      const deferred = new AssistantMessageEventStream();
+      this.wireStreamCallbacks(deferred, options);
+      // Notify the active stream (if any) that a follow-up was queued.
+      try {
+        this.currentRun.outerStream?.push({
+          type: "queued",
+          queuedPrompt: this.previewPrompt(prompt),
+          queueLength: this.pendingQueue.length + 1,
+        } as any);
+      } catch {}
+      const queuedEntry: {
+        prompt: string | ContentPart[];
+        options?: AgentRunOptions;
+        isStream: boolean;
+        deferredStream?: AssistantMessageEventStream;
+        resolve: (res: AgentResponse) => void;
+        reject: (err: unknown) => void;
+      } = {
+        prompt,
+        options,
+        isStream: true,
+        deferredStream: deferred,
+        resolve: () => {},
+        reject: () => {},
+      };
+      const done = new Promise<AgentResponse>((resolve, reject) => {
+        queuedEntry.resolve = resolve;
+        queuedEntry.reject = reject;
+      });
+      // Keep result() in sync: deferred ends when the queued turn completes.
+      done.then(
+        () => {},
+        () => {}
+      );
+      // Forward cancellation before start: drop from queue and fail.
+      deferred.onCancel(() => {
+        try {
+          const idx = this.pendingQueue.indexOf(queuedEntry as any);
+          if (idx >= 0) {
+            this.pendingQueue.splice(idx, 1);
+            queuedEntry.reject(Object.assign(new Error("Stream aborted"), { name: "AbortError" }));
+            try { deferred.fail(Object.assign(new Error("Stream aborted"), { name: "AbortError" })); } catch {}
+          }
+        } catch {}
+      });
+      this.pendingQueue.push(queuedEntry as any);
+      // Attach deferred completion to queued promise without exposing it.
+      // Real events are piped in pumpQueue() when the turn starts.
+      void done.then(
+        (res) => {
+          // If pump already piped via real stream, deferred is already ended.
+          // Otherwise (edge), end it here.
+          try {
+            if ((deferred as any).__piped !== true) {
+              deferred.push({ type: "done", delta: "", usage: res.usage, finishReason: res.finishReason, responseId: res.responseId } as any);
+              deferred.end(res);
+            }
+          } catch {}
+        },
+        (err) => {
+          try {
+            if ((deferred as any).__piped !== true) {
+              deferred.fail(err instanceof Error ? err : new Error(String(err)));
+            }
+          } catch {}
+        }
+      );
+      return deferred;
     }
-    this.prepareTurn(prompt, options);
-
-    const providerOptions = {
-      apiKey: this.apiKey,
-      baseUrl: this.baseUrl,
-      headers: { ...(this.customHeaders ?? {}), ...(options?.headers ?? {}) },
-      thinking: resolveEffectiveThinking(this.thinkingConfig, options?.thinkingLevel),
-      cache: this.stateless
-        ? { sessionId: options?.sessionId || this.sessionId }
-        : { ...this.cacheConfig, sessionId: options?.sessionId || this.sessionId },
-      serviceTier: this.serviceTier,
-      sessionId: options?.sessionId || this.sessionId,
-    };
 
-    const loopConfig = {
-      agentName: this.name,
-      provider: resolved.provider,
-      modelId: resolved.modelId,
-      context: this.context,
-      tools: this.tools,
-      options: providerOptions,
-      runOptions: options,
-      maxTurns: this.maxTurns,
-      bypassInputFileModality: this.bypassInputFileModality,
-    };
+    const s = this.startStreamingRun(prompt, options);
+    this.wireStreamCallbacks(s, options);
+    return s;
+  }
 
-    const s = streamAgentLoop(loopConfig);
-    if (this.stateless) {
-      s.result().then(() => {
-        this.context.messages = [];
-        this.context.thoughtSignatures = [];
-      }).catch(() => {});
-    }
+  private wireStreamCallbacks(
+    s: AssistantMessageEventStream,
+    options?: AgentRunOptions
+  ): void {
     // Wire one-liner callbacks so `stream:true` + onDelta is enough — no manual for-await needed
     if (options?.wrapThinking) {
       // Auto-wrap reasoning as <think>…</think> per-turn — clean boundaries across multi-turn agent runs.
@@ -698,13 +1605,12 @@ export class Agent {
       if (options?.onDelta) s.on("text_delta", (e: any) => options.onDelta!(e.delta!, e));
       if (options?.onThinkingDelta) s.on("thinking_delta", (e: any) => options.onThinkingDelta!(e.thinkingDelta!, e));
     }
-    return s;
   }
 }
 
 /** Normalizes the public `thinkingLevel` flag into provider-neutral thinking settings. */
 export function normalizeThinking(config: AgentConfig, fallbackLevel?: string): ThinkingConfig | undefined {
-  // DX4: only thinkingLevel, values: none, dynamic, minimal, low, medium, high, xhigh
+  // DX4: only thinkingLevel, values: none, dynamic, minimal, low, medium, high, xhigh, max
   const rawLevel = config.thinkingLevel ?? fallbackLevel;
   const level = rawLevel as ThinkingLevel | undefined;