fix(ai-isolate-cloudflare): accumulate toolResults across rounds

.changeset/accumulate-tool-results.md

@@ -0,0 +1,11 @@
+---
+'@tanstack/ai-isolate-cloudflare': patch
+---
+
+fix(ai-isolate-cloudflare): accumulate `toolResults` across rounds in the driver round-trip
+
+The Cloudflare isolate driver was wiping `toolResults` between rounds. `wrap-code` uses sequential `tc_<idx>` ids that are re-derived every round when the Worker re-executes user code, so prior-round results must remain in the cache. With the wipe, multi-tool programs (e.g. `await A(); await B();`) would ping-pong between `{tc_0}` and `{tc_1}` and exhaust `maxToolRounds`, surfacing as `MaxRoundsExceeded`.
+
+Single-tool code worked because only one cache entry was ever needed in a given round. Existing tests covered single-round flows only and did not exercise real `wrap-code` ids end-to-end, so the regression slipped through.
+
+Added a `tc_<idx>`-shaped regression test that fails on the prior implementation and passes with the merge.

packages/typescript/ai-isolate-cloudflare/src/isolate-driver.ts

@@ -174,8 +174,12 @@ class CloudflareIsolateContext implements IsolateContext {
         // Collect logs from this round
         allLogs = [...allLogs, ...result.logs]
 
-        // Execute tool calls locally
-        toolResults = {}
+        // Execute tool calls locally. Accumulate across rounds so prior-round
+        // results stay cached when the Worker re-executes user code.
+        // wrap-code uses sequential `tc_<idx>` ids re-derived every round; if
+        // we wipe the cache, multi-tool programs ping-pong between missing
+        // ids and exhaust `maxToolRounds` (MaxRoundsExceeded).
+        toolResults = { ...(toolResults ?? {}) }
 
         for (const toolCall of result.toolCalls) {
           const binding = this.bindings[toolCall.name] as

packages/typescript/ai-isolate-cloudflare/tests/isolate-driver.test.ts

@@ -233,6 +233,70 @@ describe('createCloudflareIsolateDriver', () => {
       expect(body2.toolResults!['add_1']).toEqual({ success: true, value: 5 })
     })
 
+    it('accumulates toolResults across rounds for sequential tool calls', async () => {
+      // Reproduces a real bug: when user code calls two tools sequentially
+      // (e.g. `await A(); await B();`), wrap-code re-runs from the start each
+      // round and re-derives sequential `tc_<idx>` ids. If the driver wipes
+      // toolResults between rounds, round 3 misses tc_0, the wrapper re-throws,
+      // and the loop ping-pongs between {tc_0} and {tc_1} until MaxRoundsExceeded.
+      const a = makeBinding('A', async () => 'a')
+      const b = makeBinding('B', async () => 'b')
+
+      // Round 1: code re-runs, throws on tc_0
+      fetchMock.mockResolvedValueOnce({
+        ok: true,
+        json: async () =>
+          ({
+            status: 'need_tools',
+            toolCalls: [{ id: 'tc_0', name: 'A', args: {} }],
+            logs: [],
+            continuationId: 'c1',
+          }) as ExecuteResponse,
+      })
+
+      // Round 2: tc_0 cached, code re-runs, throws on tc_1
+      fetchMock.mockResolvedValueOnce({
+        ok: true,
+        json: async () =>
+          ({
+            status: 'need_tools',
+            toolCalls: [{ id: 'tc_1', name: 'B', args: {} }],
+            logs: [],
+            continuationId: 'c2',
+          }) as ExecuteResponse,
+      })
+
+      // Round 3: tc_0 + tc_1 BOTH must be present, code completes
+      fetchMock.mockResolvedValueOnce({
+        ok: true,
+        json: async () =>
+          ({
+            status: 'done',
+            success: true,
+            value: 'ab',
+            logs: [],
+          }) as ExecuteResponse,
+      })
+
+      const driver = createCloudflareIsolateDriver({ workerUrl: WORKER_URL })
+      const context = await driver.createContext({
+        bindings: { A: a, B: b },
+      })
+
+      const result = await context.execute(
+        'const x = await A({}); const y = await B({}); return x + y',
+      )
+
+      expect(result.success).toBe(true)
+      expect(result.value).toBe('ab')
+      expect(fetchMock).toHaveBeenCalledTimes(3)
+
+      // Round 3 body MUST include both tc_0 and tc_1 (regression guard)
+      const body3: ExecuteRequest = JSON.parse(fetchMock.mock.calls[2][1].body)
+      expect(body3.toolResults!['tc_0']).toEqual({ success: true, value: 'a' })
+      expect(body3.toolResults!['tc_1']).toEqual({ success: true, value: 'b' })
+    })
+
     it('handles multiple tool calls in one round', async () => {
       const getA = makeBinding('getA', async () => 'A')
       const getB = makeBinding('getB', async () => 'B')