diff --git a/packages/opencode/src/session/llm-native-events.ts b/packages/opencode/src/session/llm-native-events.ts
index 0003c2a2a7..34dc5f02c8 100644
--- a/packages/opencode/src/session/llm-native-events.ts
+++ b/packages/opencode/src/session/llm-native-events.ts
@@ -35,6 +35,37 @@ const stringifyResult = (result: ToolResultValue) => {
   return JSON.stringify(result.value)
 }
 
+// Recognize the opencode `Tool.ExecuteResult` shape inside a `tool-result`
+// event's `result.value`. Native-path tool dispatchers wrap their handler
+// output in this shape so the AI-SDK-shaped session event carries the
+// real `title`, `metadata`, and `output` fields rather than the JSON
+// encoding of the whole record. Provider-executed tools (Anthropic
+// `web_search` etc.) and synthetic results that don't follow the shape
+// still go through `stringifyResult` below.
+type ExecuteShape = {
+  readonly title?: unknown
+  readonly metadata?: unknown
+  readonly output?: unknown
+}
+
+const isExecuteResult = (value: unknown): value is ExecuteShape => {
+  if (typeof value !== "object" || value === null || Array.isArray(value)) return false
+  const v = value as ExecuteShape
+  return typeof v.output === "string"
+}
+
+const toolResultOutput = (result: ToolResultValue) => {
+  if (result.type !== "json" || !isExecuteResult(result.value)) {
+    return { title: "", metadata: {}, output: stringifyResult(result) }
+  }
+  const value = result.value
+  return {
+    title: typeof value.title === "string" ? value.title : "",
+    metadata: typeof value.metadata === "object" && value.metadata !== null ? (value.metadata as Record<string, unknown>) : {},
+    output: typeof value.output === "string" ? value.output : "",
+  }
+}
+
 const response = () => ({ id: "", timestamp: new Date(0), modelId: "" })
 
 const finishReason = (reason: Extract<LLMEvent, { type: "request-finish" | "step-finish" }>["reason"]) =>
@@ -147,7 +178,7 @@ export const mapper = () => {
             toolCallId: event.id,
             toolName: event.name,
             input: state.toolInputs.get(event.id) ?? {},
-            output: { title: "", metadata: {}, output: stringifyResult(event.result) },
+            output: toolResultOutput(event.result),
           },
         ]
       case "tool-error":
diff --git a/packages/opencode/src/session/llm-native-tools.ts b/packages/opencode/src/session/llm-native-tools.ts
new file mode 100644
index 0000000000..2e58197dd3
--- /dev/null
+++ b/packages/opencode/src/session/llm-native-tools.ts
@@ -0,0 +1,248 @@
+import {
+  LLM,
+  type LLMClient,
+  type LLMError,
+  type LLMEvent,
+  type LLMRequest,
+  type FinishReason,
+  type ContentPart,
+  type RequestExecutor,
+} from "@opencode-ai/llm"
+import { Cause, Deferred, Effect, FiberSet, Queue, Stream, type Scope } from "effect"
+import type { Tool, ToolExecutionOptions } from "ai"
+
+// Maximum number of model rounds before the streaming-dispatch loop stops.
+// Mirrors `ToolRuntime.run`'s default; tweak via `maxSteps` if a caller needs
+// a different ceiling.
+export const DEFAULT_MAX_STEPS = 10
+
+// What we care about from the round's events to (a) decide whether to start
+// another round and (b) build the continuation request's message history.
+interface RoundState {
+  finishReason: FinishReason | undefined
+  // Echoed back as the next round's assistant message — text deltas merged
+  // into a single text part, reasoning deltas into a single reasoning part,
+  // tool calls appended in order. Provider-executed tool results are also
+  // appended here so the provider sees the full hosted-tool round-trip.
+  assistantContent: ContentPart[]
+  // Client-side tool dispatches. One entry per `tool-call` event we forked
+  // a handler for, populated when the handler completes.
+  toolResults: Array<{ id: string; name: string; result: unknown }>
+}
+
+const appendStreamingText = (state: RoundState, type: "text" | "reasoning", text: string) => {
+  const last = state.assistantContent.at(-1)
+  if (last?.type === type) {
+    state.assistantContent[state.assistantContent.length - 1] = { ...last, text: `${last.text}${text}` }
+    return
+  }
+  state.assistantContent.push({ type, text })
+}
+
+const accumulate = (state: RoundState, event: LLMEvent) => {
+  if (event.type === "text-delta") return appendStreamingText(state, "text", event.text)
+  if (event.type === "reasoning-delta") return appendStreamingText(state, "reasoning", event.text)
+  if (event.type === "tool-call") {
+    state.assistantContent.push(
+      LLM.toolCall({
+        id: event.id,
+        name: event.name,
+        input: event.input,
+        providerExecuted: event.providerExecuted,
+      }),
+    )
+    return
+  }
+  if (event.type === "tool-result" && event.providerExecuted) {
+    state.assistantContent.push(
+      LLM.toolResult({
+        id: event.id,
+        name: event.name,
+        result: event.result,
+        providerExecuted: true,
+      }),
+    )
+    return
+  }
+  if (event.type === "request-finish") {
+    state.finishReason = event.reason
+  }
+}
+
+// Dispatch a single client-side tool call. Returns the synthetic LLMEvent
+// that should be injected back into the round's stream — either a
+// `tool-result` (success) or `tool-error` (handler threw / unknown tool).
+// Errors from the AI SDK execute handler are caught and turned into
+// `tool-error` so the round survives and the model can self-correct on
+// the next step.
+const dispatchTool = (
+  call: { readonly id: string; readonly name: string; readonly input: unknown },
+  tools: Record<string, Tool>,
+  abort: AbortSignal,
+): Effect.Effect<LLMEvent> =>
+  Effect.gen(function* () {
+    const tool = tools[call.name]
+    if (!tool || typeof tool.execute !== "function") {
+      return {
+        type: "tool-error",
+        id: call.id,
+        name: call.name,
+        message: `Unknown tool: ${call.name}`,
+      } satisfies LLMEvent
+    }
+    const options: ToolExecutionOptions = {
+      toolCallId: call.id,
+      messages: [],
+      abortSignal: abort,
+    }
+    return yield* Effect.tryPromise({
+      try: () => Promise.resolve(tool.execute!(call.input as never, options)),
+      catch: (err) => err,
+    }).pipe(
+      Effect.map(
+        (result): LLMEvent => ({
+          type: "tool-result",
+          id: call.id,
+          name: call.name,
+          result: { type: "json", value: result },
+        }),
+      ),
+      Effect.catch(
+        (err): Effect.Effect<LLMEvent> =>
+          Effect.succeed({
+            type: "tool-error",
+            id: call.id,
+            name: call.name,
+            message: err instanceof Error ? err.message : String(err),
+          }),
+      ),
+    )
+  })
+
+// Drive one model round. Streams every LLM event in real time; each
+// non-provider-executed `tool-call` event forks a dispatcher fiber that
+// pushes the resulting `tool-result` (or `tool-error`) event back into the
+// same stream as soon as the handler completes. The round ends when:
+//   1. the LLM stream completes, AND
+//   2. every forked dispatcher has finished.
+// At that point the queue is closed (consumers see end-of-stream) and
+// `done` resolves with the accumulated state so the multi-round driver can
+// decide whether to recurse.
+const runOneRound = (
+  client: LLMClient,
+  request: LLMRequest,
+  tools: Record<string, Tool>,
+  abort: AbortSignal,
+): Effect.Effect<
+  {
+    readonly events: Stream.Stream<LLMEvent, LLMError>
+    readonly done: Deferred.Deferred<RoundState>
+  },
+  never,
+  Scope.Scope | RequestExecutor.Service
+> =>
+  Effect.gen(function* () {
+    const queue = yield* Queue.unbounded<LLMEvent, LLMError | Cause.Done>()
+    const fiberSet = yield* FiberSet.make<unknown, never>()
+    const state: RoundState = { finishReason: undefined, assistantContent: [], toolResults: [] }
+    const done = yield* Deferred.make<RoundState>()
+
+    yield* Effect.forkScoped(
+      Effect.gen(function* () {
+        yield* client.stream(request).pipe(
+          Stream.runForEach((event) =>
+            Effect.gen(function* () {
+              accumulate(state, event)
+              yield* Queue.offer(queue, event)
+              if (event.type === "tool-call" && !event.providerExecuted) {
+                yield* FiberSet.run(
+                  fiberSet,
+                  dispatchTool(event, tools, abort).pipe(
+                    Effect.flatMap((resultEvent) =>
+                      Effect.gen(function* () {
+                        if (resultEvent.type === "tool-result") {
+                          state.toolResults.push({
+                            id: resultEvent.id,
+                            name: resultEvent.name,
+                            result: (resultEvent.result as { readonly value: unknown }).value,
+                          })
+                        }
+                        yield* Queue.offer(queue, resultEvent)
+                      }),
+                    ),
+                  ),
+                )
+              }
+            }),
+          ),
+          Effect.catchCause((cause) =>
+            Effect.gen(function* () {
+              yield* Queue.failCause(queue, cause)
+              yield* Deferred.succeed(done, state)
+            }),
+          ),
+        )
+        yield* FiberSet.awaitEmpty(fiberSet)
+        yield* Queue.end(queue)
+        yield* Deferred.succeed(done, state)
+      }),
+    )
+
+    return { events: Stream.fromQueue(queue), done }
+  })
+
+// Build the next round's `LLMRequest` by appending the assistant message that
+// echoes everything the round produced (text, reasoning, tool calls, hosted
+// tool results) plus a `tool` role message per dispatched result. Lowering
+// of these LLM-shaped messages back to the provider wire format is handled
+// inside the existing adapter `prepare` step.
+const continuationRequest = (request: LLMRequest, state: RoundState): LLMRequest => {
+  const assistant = LLM.message({ role: "assistant", content: state.assistantContent })
+  const toolMessages = state.toolResults.map((entry) =>
+    LLM.toolMessage({ id: entry.id, name: entry.name, result: entry.result }),
+  )
+  return LLM.updateRequest(request, {
+    messages: [...request.messages, assistant, ...toolMessages],
+  })
+}
+
+/**
+ * Run a multi-round model+tool stream with streaming dispatch within each
+ * round. As each `tool-call` event arrives, the matching AI SDK tool's
+ * `execute` runs in a forked fiber and its result is injected back into the
+ * stream as a synthetic `tool-result` event. This matches the AI SDK's
+ * `streamText` UX: long-running tools don't block subsequent tool-call
+ * streaming, and consumers see results land as they complete.
+ *
+ * Stops when the model finishes a round with anything other than
+ * `tool-calls`, when `maxSteps` is reached, or when the underlying scope is
+ * interrupted (e.g. via the abort signal).
+ */
+export const runWithTools = (input: {
+  readonly client: LLMClient
+  readonly request: LLMRequest
+  readonly tools: Record<string, Tool>
+  readonly abort: AbortSignal
+  readonly maxSteps?: number
+}): Stream.Stream<LLMEvent, LLMError, RequestExecutor.Service> => {
+  const maxSteps = input.maxSteps ?? DEFAULT_MAX_STEPS
+  const round = (request: LLMRequest, step: number): Stream.Stream<LLMEvent, LLMError, RequestExecutor.Service> =>
+    Stream.unwrap(
+      Effect.gen(function* () {
+        const { events, done } = yield* runOneRound(input.client, request, input.tools, input.abort)
+        const continuation = Stream.unwrap(
+          Effect.gen(function* () {
+            const state = yield* Deferred.await(done)
+            if (state.finishReason !== "tool-calls") return Stream.empty
+            if (state.toolResults.length === 0) return Stream.empty
+            if (step + 1 >= maxSteps) return Stream.empty
+            return round(continuationRequest(request, state), step + 1)
+          }),
+        )
+        return events.pipe(Stream.concat(continuation))
+      }),
+    )
+  return round(input.request, 0)
+}
+
+export * as LLMNativeTools from "./llm-native-tools"
diff --git a/packages/opencode/src/session/llm.ts b/packages/opencode/src/session/llm.ts
index 49bb014327..988b5bbac7 100644
--- a/packages/opencode/src/session/llm.ts
+++ b/packages/opencode/src/session/llm.ts
@@ -39,6 +39,7 @@ import * as Option from "effect/Option"
 import * as OtelTracer from "@effect/opentelemetry/Tracer"
 import { LLMNative } from "./llm-native"
 import { LLMNativeEvents } from "./llm-native-events"
+import { LLMNativeTools } from "./llm-native-tools"
 
 const log = Log.create({ service: "llm" })
 export const OUTPUT_TOKEN_MAX = ProviderTransform.OUTPUT_TOKEN_MAX
@@ -517,10 +518,30 @@ const live: Layer.Layer<
       // per-element, `map.flush()` emits the remaining `*-end` events for any
       // text/reasoning/tool-input parts left open at stream close. The flush
       // stream is built lazily (`Stream.unwrap(Effect.sync(...))`) so it
-      // observes the mapper's final state after `mapConcat` has consumed every
+      // observes the mapper's final state after `flatMap` has consumed every
       // upstream event.
+      //
+      // The upstream source is one of two paths:
+      //
+      //   - When `nativeTools` is unset (zero-tool sessions), call the LLM
+      //     client directly. One model round, single stream, no dispatch.
+      //   - When `nativeTools` is set, hand both the request and the matching
+      //     AI SDK `tools` record to `LLMNativeTools.runWithTools`, which
+      //     drives the multi-round loop with streaming dispatch: each
+      //     `tool-call` event forks a tool handler fiber, and the
+      //     handler's result is injected back into the same stream as a
+      //     synthetic `tool-result` event. Long-running tools don't block
+      //     subsequent tool-call streaming.
       const map = LLMNativeEvents.mapper()
-      return nativeClient.stream(llmRequest).pipe(
+      const upstream = input.nativeTools && input.nativeTools.length > 0
+        ? LLMNativeTools.runWithTools({
+            client: nativeClient,
+            request: llmRequest,
+            tools: input.tools,
+            abort: input.abort,
+          })
+        : nativeClient.stream(llmRequest)
+      return upstream.pipe(
         Stream.flatMap((event) => Stream.fromIterable(map.map(event))),
         Stream.concat(Stream.unwrap(Effect.sync(() => Stream.fromIterable(map.flush())))),
         Stream.provideService(RequestExecutor.Service, executor),
diff --git a/packages/opencode/test/session/llm-native-stream.test.ts b/packages/opencode/test/session/llm-native-stream.test.ts
index 5809abb503..44dce8ec70 100644
--- a/packages/opencode/test/session/llm-native-stream.test.ts
+++ b/packages/opencode/test/session/llm-native-stream.test.ts
@@ -10,12 +10,14 @@ import {
   ProviderPatch,
   RequestExecutor,
 } from "@opencode-ai/llm"
-import { Effect, Layer, Schema, Stream } from "effect"
+import { Effect, Layer, Ref, Schema, Stream } from "effect"
 import { HttpClient, HttpClientResponse } from "effect/unstable/http"
+import { tool, jsonSchema } from "ai"
 import { ModelID, ProviderID } from "../../src/provider/schema"
 import { MessageID, PartID, SessionID } from "../../src/session/schema"
 import { LLMNative } from "../../src/session/llm-native"
 import { LLMNativeEvents } from "../../src/session/llm-native-events"
+import { LLMNativeTools } from "../../src/session/llm-native-tools"
 import { ProviderTest } from "../fake/provider"
 import { testEffect } from "../lib/effect"
 import type { MessageV2 } from "../../src/session/message-v2"
@@ -37,6 +39,30 @@ const fixedResponse = (body: BodyInit, init: ResponseInit = { headers: { "conten
     ),
   )
 
+// Scripted multi-response HTTP layer. Each request consumes the next body in
+// order; the final body repeats if more requests arrive. Mirrors the
+// `scriptedResponses` helper in `packages/llm/test/lib/http.ts`.
+const scriptedResponses = (bodies: ReadonlyArray<BodyInit>, init: ResponseInit = { headers: { "content-type": "text/event-stream" } }) =>
+  RequestExecutor.layer.pipe(
+    Layer.provide(
+      Layer.unwrap(
+        Effect.gen(function* () {
+          const cursor = yield* Ref.make(0)
+          return Layer.succeed(
+            HttpClient.HttpClient,
+            HttpClient.make((request) =>
+              Effect.gen(function* () {
+                const index = yield* Ref.getAndUpdate(cursor, (n) => n + 1)
+                const body = bodies[index] ?? bodies[bodies.length - 1]
+                return HttpClientResponse.fromWeb(request, new Response(body, init))
+              }),
+            ),
+          )
+        }),
+      ),
+    ),
+  )
+
 // Encode an Anthropic SSE body. Each event becomes a `data:` line; the codec
 // also expects `event:` lines but the package's SSE framing only reads the
 // data field.
@@ -150,6 +176,122 @@ describe("LLMNative stream wire-up (audit gap #4 phase 1)", () => {
     }),
   )
 
+  // Phase 2 step 2b: drives the streaming-dispatch loop end-to-end. The
+  // scripted Anthropic backend replies in two rounds — round 1 is a tool
+  // call, round 2 is text after the tool result feeds back. Asserts that
+  // `runWithTools` (a) forks the AI SDK execute when the `tool-call` event
+  // arrives, (b) injects a synthetic `tool-result` event into the same
+  // stream, (c) issues a continuation request with the tool result in
+  // history, and (d) the stream concludes with the second-round text.
+  it.effect("dispatches a tool call mid-stream and continues the conversation", () =>
+    Effect.gen(function* () {
+      const mdl = anthropicModel()
+      const lookupParameters = Schema.Struct({
+        query: Schema.String.annotate({ description: "Search query" }),
+      })
+      const lookupTool: Tool.Def<typeof lookupParameters> = {
+        id: "lookup",
+        description: "Lookup project data",
+        parameters: lookupParameters,
+        execute: () => Effect.succeed({ title: "Weather lookup", metadata: {}, output: '{"forecast":"sunny"}' }),
+      }
+
+      // AI SDK side: the same tool wrapped so `tool.execute(args, opts)`
+      // resolves with the same opencode `ExecuteResult` shape the live
+      // `prompt.ts:resolveTools` would produce. The dispatcher inside
+      // `runWithTools` calls this; the synthetic `tool-result` LLM event
+      // carries the result back into the stream.
+      const aiTool = tool({
+        description: "Lookup project data",
+        inputSchema: jsonSchema({
+          type: "object",
+          properties: { query: { type: "string", description: "Search query" } },
+          required: ["query"],
+        }),
+        execute: async () => ({
+          title: "Weather lookup",
+          metadata: {},
+          output: '{"forecast":"sunny"}',
+        }),
+      })
+
+      const userID = MessageID.ascending()
+      const llmRequest = yield* LLMNative.request({
+        id: "smoke-tool-loop",
+        provider: ProviderTest.info({ id: ProviderID.make("anthropic"), key: "anthropic-key" }, mdl),
+        model: mdl,
+        system: ["Be concise."],
+        messages: [userMessage(mdl, userID, [userPart(userID, "What is the weather?")])],
+        tools: [lookupTool],
+      })
+
+      // Round 1: model issues `lookup` tool call.
+      const round1 = sseBody([
+        { type: "message_start", message: { usage: { input_tokens: 5 } } },
+        { type: "content_block_start", index: 0, content_block: { type: "tool_use", id: "call_1", name: "lookup" } },
+        { type: "content_block_delta", index: 0, delta: { type: "input_json_delta", partial_json: '{"query"' } },
+        { type: "content_block_delta", index: 0, delta: { type: "input_json_delta", partial_json: ':"weather"}' } },
+        { type: "content_block_stop", index: 0 },
+        { type: "message_delta", delta: { stop_reason: "tool_use" }, usage: { output_tokens: 1 } },
+        { type: "message_stop" },
+      ])
+      // Round 2: model replies with text after seeing the tool result.
+      const round2 = sseBody([
+        { type: "message_start", message: { usage: { input_tokens: 12 } } },
+        { type: "content_block_start", index: 0, content_block: { type: "text", text: "" } },
+        { type: "content_block_delta", index: 0, delta: { type: "text_delta", text: "It is sunny." } },
+        { type: "content_block_stop", index: 0 },
+        { type: "message_delta", delta: { stop_reason: "end_turn" }, usage: { output_tokens: 4 } },
+        { type: "message_stop" },
+      ])
+
+      const client = LLMClient.make({ adapters, patches: ProviderPatch.defaults })
+      const map = LLMNativeEvents.mapper()
+
+      const events = yield* LLMNativeTools.runWithTools({
+        client,
+        request: llmRequest,
+        tools: { lookup: aiTool },
+        abort: new AbortController().signal,
+      }).pipe(
+        Stream.flatMap((event) => Stream.fromIterable(map.map(event))),
+        Stream.concat(Stream.unwrap(Effect.sync(() => Stream.fromIterable(map.flush())))),
+        Stream.runCollect,
+        Effect.provide(scriptedResponses([round1, round2])),
+      )
+
+      const collected = Array.from(events)
+
+      // Round 1: tool call streams, dispatcher fires, synthetic tool-result lands.
+      const toolCall = collected.find((event) => event.type === "tool-call")
+      expect(toolCall).toMatchObject({
+        type: "tool-call",
+        toolCallId: "call_1",
+        toolName: "lookup",
+        input: { query: "weather" },
+      })
+
+      const toolResult = collected.find((event) => event.type === "tool-result")
+      expect(toolResult).toMatchObject({
+        type: "tool-result",
+        toolCallId: "call_1",
+        toolName: "lookup",
+        output: { title: "Weather lookup", output: '{"forecast":"sunny"}' },
+      })
+
+      // Round 2: text-delta arrives after the tool result.
+      const round2Text = collected.find((event) => event.type === "text-delta")
+      expect(round2Text).toMatchObject({ type: "text-delta", text: "It is sunny." })
+
+      // Final finish should be `stop`, not `tool-calls` (tool loop terminated).
+      const finalFinish = [...collected].reverse().find((event) => event.type === "finish")
+      expect(finalFinish).toMatchObject({ finishReason: "stop" })
+
+      // No errors leaked through.
+      expect(collected.some((event) => event.type === "error")).toBe(false)
+    }),
+  )
+
   // Phase 2 step 2a: verifies a tool-bearing `nativeTools` array reaches the
   // wire as Anthropic `tools[]` blocks. The model in this fixture answers with
   // plain text instead of issuing a tool call (we don't yet have dispatch).