threepointone · Apr 4, 2026 · Apr 3, 2026 · Apr 3, 2026 · Apr 4, 2026 · Apr 4, 2026
diff --git a/.changeset/unified-fiber-architecture.md b/.changeset/unified-fiber-architecture.md
@@ -0,0 +1,20 @@
+---
+"agents": minor
+"@cloudflare/ai-chat": minor
+"@cloudflare/think": minor
+---
+
+Add durable fiber execution to the Agent base class.
+
+`runFiber(name, fn)` registers work in SQLite, holds a `keepAlive` ref, and enables recovery via `onFiberRecovered` after DO eviction. `ctx.stash()` and `this.stash()` checkpoint progress that survives eviction.
+
+`AIChatAgent` gains `durableStreaming` — when enabled, each chat turn is wrapped in a fiber. `onChatRecovery` provides provider-specific recovery (Workers AI continuation, OpenAI response retrieval, Anthropic synthetic message). `continueLastTurn()` appends to the interrupted assistant message seamlessly.
+
+`Think` now extends `Agent` directly (no mixin). Fiber support is inherited from the base class.
+
+**Breaking (experimental APIs only):**
+
+- Removed `withFibers` mixin (`agents/experimental/forever`)
+- Removed `withDurableChat` mixin (`@cloudflare/ai-chat/experimental/forever`)
+- Removed `./experimental/forever` export from both packages
+- Think no longer has a `fibers` flag — recovery is automatic via alarm housekeeping
diff --git a/design/think.md b/design/think.md
diff --git a/experimental/README.md b/experimental/README.md
@@ -7,6 +7,7 @@ The code here is for exploration, prototyping, and validating patterns that may
 ## Contents
 
 - **[gadgets.md](./gadgets.md)** — Exploration of facets, isolation, and structural safety for agents. Covers the Gatekeeper/ApprovalQueue pattern, Worker Loader sandboxing, sub-agent facets, multi-room chat, and other patterns worth pulling into the SDK.
-- **[forever.md](./forever.md)** — Design doc for durable long-running execution. Covers keepAlive, fibers (spawnFiber/stashFiber), eviction recovery, and AIChatAgent integration. Implemented as mixins in `agents/experimental/forever` and `@cloudflare/ai-chat/experimental/forever`.
+- **[forever.md](./forever.md)** — Design doc for durable long-running execution. Covers `keepAlive`, `runFiber` (checkpointing, eviction recovery), and `AIChatAgent` chat recovery via `onChatRecovery`. Built into the `Agent` and `AIChatAgent` base classes.
+- **[forever-fibers/](./forever-fibers/)** — Example of `Agent.runFiber()` for durable background work with real-time progress tracking.
+- **[forever-chat/](./forever-chat/)** — Example of `AIChatAgent` with `durableStreaming` for multi-provider chat recovery (Workers AI, OpenAI, Anthropic).
 - **[session-memory](./session-memory/)** — Example of the Session API for conversation history with automatic compaction. Demonstrates `agents/experimental/memory/session` with micro-compaction and LLM-based summarization.
-- **[workspace-chat](../examples/workspace-chat/)** — _(moved to examples/)_ AIChatAgent with a persistent virtual filesystem.
diff --git a/experimental/forever-chat/README.md b/experimental/forever-chat/README.md
@@ -1,24 +1,22 @@
 # Forever Chat — Durable AI Streaming
 
-> **WARNING: EXPERIMENTAL.** This example uses APIs from `agents/experimental/forever` and `@cloudflare/ai-chat/experimental/forever` that are under active development and **will break** between releases. Do not use in production. Pin your package versions and expect to rewrite your code when upgrading.
-
-AI chat using `withDurableChat(AIChatAgent)` — the mixin adds keepAlive during streaming and automatic recovery after DO eviction.
+AI chat with `durableStreaming` enabled — wraps each chat turn in `runFiber` for automatic keepAlive during streaming and recovery after DO eviction.
 
 See [forever.md](../forever.md) for the full design doc.
 
 ## What it shows
 
-- `withDurableChat` mixin from `@cloudflare/ai-chat/experimental/forever`
+- `durableStreaming = true` on `AIChatAgent` — wraps chat turns in fibers
 - keepAlive during streaming — DO stays alive for long LLM responses
 - `onChatRecovery` — provider-specific recovery after eviction
 - `continueLastTurn()` — seamlessly continues the interrupted assistant message inline
 - Multi-provider support with a dropdown selector:
 
-| Provider   | Model             | Recovery strategy                                                                       |
-| ---------- | ----------------- | --------------------------------------------------------------------------------------- |
-| Workers AI | kimi-k2.5         | Persist partial + inline continuation via `continueLastTurn()`                          |
-| OpenAI     | gpt-5.4           | Retrieve completed response via Responses API (`store: true`) — zero wasted tokens      |
-| Anthropic  | claude-sonnet-4.6 | Persist partial + continue via synthetic user message (reasoning disabled for recovery) |
+| Provider   | Model             | Recovery strategy                                                                                 |
+| ---------- | ----------------- | ------------------------------------------------------------------------------------------------- |
+| Workers AI | kimi-k2.5         | Persist partial + continue via `continueLastTurn()` (text + reasoning merge into existing blocks) |
+| OpenAI     | gpt-5.4           | Retrieve completed response via Responses API (`store: true`) — zero wasted tokens                |
+| Anthropic  | claude-sonnet-4.6 | Persist partial + continue via synthetic user message (reasoning disabled for recovery)           |
 
 ## Run it
 

diff --git a/experimental/forever-chat/src/server.ts b/experimental/forever-chat/src/server.ts
@@ -2,25 +2,25 @@
  * Forever Chat — Durable AI streaming with multi-provider recovery.
  *
  * Demonstrates three recovery strategies after DO eviction:
- * - Workers AI: persist partial + inline continuation via continueLastTurn()
+ * - Workers AI: persist partial + continue via continueLastTurn()
+ *              (text + reasoning merge into existing blocks)
  * - OpenAI: retrieve completed response via Responses API (store: true)
  * - Anthropic: persist partial + continue via synthetic user message
  *              (no prefill support, reasoning disabled for recovery)
  *
- * Uses the withDurableChat mixin for automatic keepAlive during
+ * Uses durableStreaming for automatic keepAlive during
  * streaming and onChatRecovery for provider-specific recovery.
  */
 import { createWorkersAI } from "workers-ai-provider";
 import { createOpenAI } from "@ai-sdk/openai";
 import { createAnthropic } from "@ai-sdk/anthropic";
 import { routeAgentRequest } from "agents";
-import { AIChatAgent } from "@cloudflare/ai-chat";
-import type { OnChatMessageOptions } from "@cloudflare/ai-chat";
 import {
-  withDurableChat,
+  AIChatAgent,
   type ChatRecoveryContext,
-  type ChatRecoveryOptions
-} from "@cloudflare/ai-chat/experimental/forever";
+  type ChatRecoveryOptions,
+  type OnChatMessageOptions
+} from "@cloudflare/ai-chat";
 import {
   streamText,
   convertToModelMessages,
@@ -100,11 +100,13 @@ const SYSTEM_PROMPT =
   "You can check the weather and perform calculations. " +
   "For calculations with large numbers (over 1000), you need user approval first.";
 
-// ── Agent ─────────────────────────────────────────────────────────────
+const RECOVERY_SUFFIX =
+  " Do not think or reason — continue the text output directly.";
 
-const DurableChatAgent = withDurableChat(AIChatAgent);
+// ── Agent ─────────────────────────────────────────────────────────────
 
-export class ForeverChatAgent extends DurableChatAgent<Env, AgentState> {
+export class ForeverChatAgent extends AIChatAgent<Env, AgentState> {
+  override durableStreaming = true;
   maxPersistedMessages = 200;
 
   // ── Recovery ────────────────────────────────────────────────────
@@ -124,11 +126,24 @@ export class ForeverChatAgent extends DurableChatAgent<Env, AgentState> {
       return { continue: false };
     }
 
+    // Workers AI: use continueLastTurn() (prefill works). Set the
+    // recovering flag so onChatMessage can hint the model to skip
+    // reasoning. Even if the model still reasons, the framework merges
+    // reasoning into the existing block automatically.
+    if (provider === "workersai" || !provider) {
+      await this.schedule(0, "_continueWorkersAI", undefined, {
+        idempotent: true
+      });
+      return { continue: false };
+    }
+
     if (provider !== "openai") return {};
 
     // OpenAI Responses API: the generation continues server-side even
-    // after our connection drops. Retrieve the completed response by ID.
-    const responseId = this._getStoredResponseId();
+    // after our connection drops. Retrieve the completed response by ID
+    // (stashed via this.stash() during streaming).
+    const responseId = (ctx.recoveryData as { responseId?: string } | null)
+      ?.responseId;
     if (!responseId) return {};
 
     try {
@@ -173,6 +188,14 @@ export class ForeverChatAgent extends DurableChatAgent<Env, AgentState> {
     }
   }
 
+  async _continueWorkersAI() {
+    const ready = await this.waitUntilStable({ timeout: 10_000 });
+    if (!ready) return;
+
+    this._lastBody = { ...this._lastBody, recovering: true };
+    await this.continueLastTurn();
+  }
+
   async _continueWithUserMessage() {
     const ready = await this.waitUntilStable({ timeout: 10_000 });
     if (!ready) return;
@@ -206,7 +229,7 @@ export class ForeverChatAgent extends DurableChatAgent<Env, AgentState> {
 
     const result = streamText({
       model: this._getModel(provider),
-      system: SYSTEM_PROMPT,
+      system: recovering ? SYSTEM_PROMPT + RECOVERY_SUFFIX : SYSTEM_PROMPT,
       messages: pruneMessages({
         messages: await convertToModelMessages(this.messages),
         toolCalls: "before-last-2-messages",
@@ -265,7 +288,7 @@ export class ForeverChatAgent extends DurableChatAgent<Env, AgentState> {
             | { type?: string; response?: { id?: string } }
             | undefined;
           if (raw?.type === "response.created" && raw.response?.id) {
-            this._storeResponseId(raw.response.id);
+            this.stash({ responseId: raw.response.id });
           }
         }
       };
@@ -281,23 +304,6 @@ export class ForeverChatAgent extends DurableChatAgent<Env, AgentState> {
     }
     return {};
   }
-
-  // ── OpenAI response ID persistence ─────────────────────────────
-
-  private _storeResponseId(id: string) {
-    this.sql`
-      INSERT OR REPLACE INTO cf_ai_chat_request_context (key, value)
-      VALUES ('openaiResponseId', ${id})
-    `;
-  }
-
-  private _getStoredResponseId(): string | undefined {
-    const rows = this.sql<{ value: string }>`
-      SELECT value FROM cf_ai_chat_request_context
-      WHERE key = 'openaiResponseId'
-    `;
-    return rows?.[0]?.value;
-  }
 }
 
 export default {

diff --git a/experimental/forever-fibers/README.md b/experimental/forever-fibers/README.md
@@ -1,19 +1,14 @@
 # Forever Fibers — Durable Execution Demo
 
-> **WARNING: EXPERIMENTAL.** This example uses APIs from `agents/experimental/forever` that are under active development and **will break** between releases. Do not use in production. Pin your package versions and expect to rewrite your code when upgrading.
-
-Demonstrates durable long-running execution with fibers — fire-and-forget work that survives Durable Object eviction via SQLite checkpointing and alarm-based recovery.
+Demonstrates durable long-running execution with `Agent.runFiber()` — work that survives Durable Object eviction via SQLite checkpointing and alarm-based recovery.
 
 See [forever.md](../forever.md) for the full design doc.
 
 ## What it shows
 
-- `withFibers` mixin from `agents/experimental/forever`
-- `spawnFiber()` — start a multi-step research task that runs in the background
-- `stashFiber()` — checkpoint progress after each step (persisted in SQLite)
+- `runFiber()` — start a multi-step research task that runs in the background
+- `ctx.stash()` — checkpoint progress after each step (persisted in SQLite)
 - `onFiberRecovered()` — automatically resume from the last checkpoint after eviction
-- `cancelFiber()` — stop a running fiber
-- Simulated eviction — demonstrates the recovery flow via the UI button
 - Real eviction testing — kill the wrangler process externally and restart; alarms persist to disk, so recovery happens automatically (same as production)
 
 ## Run it

diff --git a/experimental/forever-fibers/src/app.tsx b/experimental/forever-fibers/src/app.tsx
@@ -77,9 +77,8 @@ function ModeToggle() {
 export default function App() {
   const [topic, setTopic] = useState("quantum computing");
   const [steps, setSteps] = useState<StepInfo[]>([]);
-  const [fiberId, setFiberId] = useState<string | null>(null);
   const [status, setStatus] = useState<
-    "idle" | "running" | "complete" | "cancelled" | "recovered"
+    "idle" | "running" | "complete" | "recovered"
   >("idle");
   const [results, setResults] = useState<ResearchStep[]>([]);
   const [connectionStatus, setConnectionStatus] =
@@ -94,7 +93,6 @@ export default function App() {
 
         switch (msg.type) {
           case "research:started":
-            setFiberId(msg.fiberId);
             setStatus("running");
             setResults([]);
             setSteps(
@@ -137,10 +135,6 @@ export default function App() {
             setResults(msg.results);
             break;
 
-          case "research:cancelled":
-            setStatus("cancelled");
-            break;
-
           case "research:failed":
             setStatus("idle");
             break;
@@ -158,22 +152,6 @@ export default function App() {
     await agent.call("startResearch", [topic.trim()]);
   }, [topic, agent]);
 
-  const handleCancel = useCallback(async () => {
-    if (!agent) return;
-    await agent.call("cancelResearch", []);
-  }, [agent]);
-
-  const handleKillAndRecover = useCallback(async () => {
-    if (!agent) return;
-    setStatus("recovered");
-    await agent.call("simulateKillAndRecover", []);
-    // Recovery is async — the fiber restarts in the background
-    // and broadcasts progress updates as it resumes
-    setTimeout(() => {
-      if (status === "recovered") setStatus("running");
-    }, 2000);
-  }, [agent, status]);
-
   const isRunning = status === "running" || status === "recovered";
 
   return (
@@ -206,33 +184,20 @@ export default function App() {
                 if (e.key === "Enter") handleStart();
               }}
             />
-            {isRunning ? (
-              <Button variant="destructive" onClick={handleCancel}>
-                Cancel
-              </Button>
-            ) : (
-              <Button
-                variant="primary"
-                onClick={handleStart}
-                disabled={!topic.trim()}
-              >
-                Start Research
-              </Button>
-            )}
+            <Button
+              variant="primary"
+              onClick={handleStart}
+              disabled={!topic.trim() || isRunning}
+            >
+              {isRunning ? "Running..." : "Start Research"}
+            </Button>
           </div>
         </div>
 
         {/* Steps */}
         {steps.length > 0 && (
           <div className="rounded-lg border border-kumo-line p-4">
-            <div className="flex items-center justify-between">
-              <Text variant="heading3">Progress</Text>
-              {fiberId && (
-                <span className="font-mono text-xs text-kumo-inactive">
-                  {fiberId.slice(0, 8)}...
-                </span>
-              )}
-            </div>
+            <Text variant="heading3">Progress</Text>
 
             <div className="mt-3 flex flex-col gap-2">
               {steps.map((step, i) => (
@@ -261,23 +226,6 @@ export default function App() {
           </div>
         )}
 
-        {/* Eviction Demo */}
-        {isRunning && (
-          <div className="rounded-lg border border-kumo-line p-4">
-            <Text variant="heading3">Simulate Eviction</Text>
-            <p className="mt-1 text-sm text-kumo-inactive">
-              In production, Durable Objects can be evicted by code updates or
-              inactivity. This simulates that — the fiber state persists in
-              SQLite, and recovery picks up from the last checkpoint.
-            </p>
-            <div className="mt-3">
-              <Button variant="secondary" onClick={handleKillAndRecover}>
-                Simulate Kill &amp; Recover
-              </Button>
-            </div>
-          </div>
-        )}
-
         {/* Recovery banner */}
         {status === "recovered" && (
           <div className="rounded-lg border border-green-500/30 bg-green-500/5 p-4">
@@ -307,26 +255,6 @@ export default function App() {
                   setSteps([]);
                   setResults([]);
                   setStatus("idle");
-                  setFiberId(null);
-                }}
-              >
-                Start New Research
-              </Button>
-            </div>
-          </div>
-        )}
-
-        {/* Cancelled */}
-        {status === "cancelled" && (
-          <div className="rounded-lg border border-kumo-line p-4">
-            <Text variant="body">Research cancelled.</Text>
-            <div className="mt-3">
-              <Button
-                variant="secondary"
-                onClick={() => {
-                  setSteps([]);
-                  setStatus("idle");
-                  setFiberId(null);
                 }}
               >
                 Start New Research