diff --git a/packages/junior/src/chat/pi/traced-stream.ts b/packages/junior/src/chat/pi/traced-stream.ts
new file mode 100644
index 00000000..d1f6aecf
--- /dev/null
+++ b/packages/junior/src/chat/pi/traced-stream.ts
@@ -0,0 +1,119 @@
+import type { StreamFn } from "@mariozechner/pi-agent-core";
+import {
+  type Api,
+  type AssistantMessage,
+  type Context,
+  type Model,
+  streamSimple,
+} from "@mariozechner/pi-ai";
+import * as Sentry from "@/chat/sentry";
+import {
+  extractGenAiUsageAttributes,
+  getLogContextAttributes,
+  serializeGenAiAttribute,
+} from "@/chat/logging";
+
+const PROVIDER_NAME = "vercel-ai-gateway";
+
+// Compose only the OTel GenAI attributes that are knowable at span start
+// (request-shape + system instructions). End-of-call attributes such as
+// usage and finish reasons are set after the stream resolves.
+function buildChatStartAttributes(
+  model: Model<Api>,
+  context: Context,
+): Record<string, string> {
+  const attributes: Record<string, string> = {
+    "gen_ai.operation.name": "chat",
+    "gen_ai.provider.name": PROVIDER_NAME,
+    "gen_ai.request.model": model.id,
+  };
+
+  const inputMessages = serializeGenAiAttribute(context.messages);
+  if (inputMessages) {
+    attributes["gen_ai.input.messages"] = inputMessages;
+  }
+
+  if (context.systemPrompt) {
+    const systemInstructions = serializeGenAiAttribute([
+      { type: "text", content: context.systemPrompt },
+    ]);
+    if (systemInstructions) {
+      attributes["gen_ai.system_instructions"] = systemInstructions;
+    }
+  }
+
+  return attributes;
+}
+
+// Composes post-stream attributes for the chat span. Two deferrals worth
+// flagging: `extractGenAiUsageAttributes` currently omits cache token fields,
+// and `gen_ai.response.finish_reasons` is emitted with pi-ai's raw StopReason
+// (Sentry ingest expects the OTel set). Both are tracked separately and out
+// of scope here.
+function buildChatEndAttributes(
+  message: AssistantMessage,
+): Record<string, string | string[] | number> {
+  const attributes: Record<string, string | string[] | number> = {};
+
+  const outputMessages = serializeGenAiAttribute([message]);
+  if (outputMessages) {
+    attributes["gen_ai.output.messages"] = outputMessages;
+  }
+
+  Object.assign(attributes, extractGenAiUsageAttributes(message));
+
+  if (message.stopReason) {
+    attributes["gen_ai.response.finish_reasons"] = [message.stopReason];
+  }
+
+  if (message.model) {
+    attributes["gen_ai.response.model"] = message.model;
+  }
+
+  return attributes;
+}
+
+/**
+ * Wraps pi-ai's `streamSimple` so each LLM call inside a pi-agent-core agent
+ * loop produces its own `gen_ai.chat` Sentry span. The returned function is
+ * passed to `new Agent({ streamFn: ... })` and runs once per loop iteration.
+ *
+ * The base argument exists so tests can inject a stub stream function.
+ */
+export function createTracedStreamFn(base: StreamFn = streamSimple): StreamFn {
+  return async (model, context, options) => {
+    const span = Sentry.startInactiveSpan({
+      name: `chat ${model.id}`,
+      op: "gen_ai.chat",
+      attributes: {
+        ...getLogContextAttributes(),
+        ...buildChatStartAttributes(model, context),
+      },
+    });
+
+    try {
+      const stream = await Sentry.withActiveSpan(span, () =>
+        Promise.resolve(base(model, context, options)),
+      );
+
+      stream.result().then(
+        (finalMessage) => {
+          for (const [key, value] of Object.entries(
+            buildChatEndAttributes(finalMessage),
+          )) {
+            span.setAttribute(key, value);
+          }
+          span.end();
+        },
+        () => {
+          span.end();
+        },
+      );
+
+      return stream;
+    } catch (error) {
+      span.end();
+      throw error;
+    }
+  };
+}
diff --git a/packages/junior/src/chat/respond.ts b/packages/junior/src/chat/respond.ts
index e56da16c..bc9a6895 100644
--- a/packages/junior/src/chat/respond.ts
+++ b/packages/junior/src/chat/respond.ts
@@ -50,6 +50,7 @@ import {
   resolveGatewayModel,
 } from "@/chat/pi/client";
 import type { PiMessage } from "@/chat/pi/messages";
+import { createTracedStreamFn } from "@/chat/pi/traced-stream";
 import {
   createSandboxExecutor,
   type SandboxAcquiredState,
@@ -918,6 +919,7 @@ export async function generateAssistantReply(
     // ── Agent execution ──────────────────────────────────────────────
     agent = new Agent({
       getApiKey: () => getPiGatewayApiKeyOverride(),
+      streamFn: createTracedStreamFn(),
       initialState: {
         systemPrompt: baseInstructions,
         model: resolveGatewayModel(botConfig.modelId),
diff --git a/packages/junior/tests/unit/chat/pi/traced-stream.test.ts b/packages/junior/tests/unit/chat/pi/traced-stream.test.ts
new file mode 100644
index 00000000..2bbbd852
--- /dev/null
+++ b/packages/junior/tests/unit/chat/pi/traced-stream.test.ts
@@ -0,0 +1,207 @@
+import type { StreamFn } from "@mariozechner/pi-agent-core";
+import { afterEach, describe, expect, it, vi } from "vitest";
+import {
+  createAssistantMessageEventStream,
+  type AssistantMessage,
+  type Model,
+} from "@mariozechner/pi-ai";
+
+const { startInactiveSpan, withActiveSpan } = vi.hoisted(() => {
+  const span = {
+    setAttribute: vi.fn(),
+    setAttributes: vi.fn(),
+    setStatus: vi.fn(),
+    end: vi.fn(),
+  };
+  return {
+    startInactiveSpan: vi.fn((_options: unknown) => span),
+    withActiveSpan: vi.fn(<T>(_s: unknown, cb: () => T) => cb()),
+  };
+});
+
+vi.mock("@/chat/sentry", () => ({
+  startInactiveSpan,
+  withActiveSpan,
+}));
+
+function fakeModel(id: string): Model<"anthropic-messages"> {
+  return { id } as unknown as Model<"anthropic-messages">;
+}
+
+function fakeMessage(): AssistantMessage {
+  return {
+    role: "assistant",
+    content: [{ type: "text", text: "hi" }],
+    api: "anthropic-messages",
+    provider: "vercel-ai-gateway",
+    model: "openai/gpt-5.4",
+    usage: {
+      input: 100,
+      output: 5,
+      cacheRead: 0,
+      cacheWrite: 0,
+      totalTokens: 105,
+      cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },
+    },
+    stopReason: "stop",
+    timestamp: Date.now(),
+  };
+}
+
+type SpanMock = {
+  setAttribute: ReturnType<typeof vi.fn>;
+  setAttributes: ReturnType<typeof vi.fn>;
+  setStatus: ReturnType<typeof vi.fn>;
+  end: ReturnType<typeof vi.fn>;
+};
+
+function getSpan(): SpanMock {
+  return startInactiveSpan.mock.results[0]!.value as SpanMock;
+}
+
+describe("createTracedStreamFn", () => {
+  afterEach(() => {
+    vi.clearAllMocks();
+    vi.resetModules();
+  });
+
+  it("opens a gen_ai.chat span when invoked", async () => {
+    const { createTracedStreamFn } = await import("@/chat/pi/traced-stream");
+    const stream = createAssistantMessageEventStream();
+    const base = vi.fn(() => stream);
+
+    const traced = createTracedStreamFn(base as unknown as StreamFn);
+    const returned = await traced(
+      fakeModel("openai/gpt-5.4"),
+      { messages: [{ role: "user", content: "hi", timestamp: 0 }] },
+      undefined,
+    );
+
+    expect(returned).toBe(stream);
+    expect(startInactiveSpan).toHaveBeenCalledTimes(1);
+    const opts = startInactiveSpan.mock.calls[0]?.[0] as unknown as {
+      name: string;
+      op: string;
+    };
+    expect(opts.op).toBe("gen_ai.chat");
+    expect(opts.name).toBe("chat openai/gpt-5.4");
+  });
+
+  it("sets gen_ai.input.messages and gen_ai.system_instructions on the chat span", async () => {
+    const { createTracedStreamFn } = await import("@/chat/pi/traced-stream");
+    const stream = createAssistantMessageEventStream();
+    const base = vi.fn(() => stream);
+
+    const traced = createTracedStreamFn(base as unknown as StreamFn);
+    await traced(
+      fakeModel("openai/gpt-5.4"),
+      {
+        systemPrompt: "you are junior",
+        messages: [{ role: "user", content: "hello", timestamp: 0 }],
+      },
+      undefined,
+    );
+
+    const opts = startInactiveSpan.mock.calls[0]?.[0] as unknown as {
+      attributes: Record<string, unknown>;
+    };
+    expect(opts.attributes["gen_ai.provider.name"]).toBe("vercel-ai-gateway");
+    expect(typeof opts.attributes["gen_ai.input.messages"]).toBe("string");
+    expect(opts.attributes["gen_ai.input.messages"]).toContain("hello");
+    expect(typeof opts.attributes["gen_ai.system_instructions"]).toBe("string");
+    expect(opts.attributes["gen_ai.system_instructions"]).toContain(
+      "you are junior",
+    );
+    expect(opts.attributes["gen_ai.operation.name"]).toBe("chat");
+    expect(opts.attributes["gen_ai.request.model"]).toBe("openai/gpt-5.4");
+  });
+
+  it("sets output.messages, usage tokens, finish_reasons, response.model after stream completion", async () => {
+    const { createTracedStreamFn } = await import("@/chat/pi/traced-stream");
+    const stream = createAssistantMessageEventStream();
+    const base = vi.fn(() => stream);
+
+    const traced = createTracedStreamFn(base as unknown as StreamFn);
+    const returned = await traced(
+      fakeModel("openai/gpt-5.4"),
+      { messages: [{ role: "user", content: "hi", timestamp: 0 }] },
+      undefined,
+    );
+
+    expect(returned).toBe(stream);
+
+    // Resolve the stream's terminal Promise to trigger end-attribute population.
+    const finalMessage = fakeMessage();
+    stream.end(finalMessage);
+    await stream.result();
+    // Allow the .then callback to flush.
+    await new Promise((r) => setImmediate(r));
+
+    const span = getSpan();
+    const endAttributes = Object.fromEntries(
+      span.setAttribute.mock.calls.map((c) => [c[0], c[1]]),
+    );
+    expect(typeof endAttributes["gen_ai.output.messages"]).toBe("string");
+    expect(endAttributes["gen_ai.usage.input_tokens"]).toBe(100);
+    expect(endAttributes["gen_ai.usage.output_tokens"]).toBe(5);
+    expect(endAttributes["gen_ai.response.finish_reasons"]).toEqual(["stop"]);
+    expect(endAttributes["gen_ai.response.model"]).toBe("openai/gpt-5.4");
+    expect(span.end).toHaveBeenCalledTimes(1);
+  });
+
+  it("inherits LogContext attributes (e.g. gen_ai.conversation.id) onto the chat span", async () => {
+    const { withLogContext } = await import("@/chat/logging");
+    const { createTracedStreamFn } = await import("@/chat/pi/traced-stream");
+    const stream = createAssistantMessageEventStream();
+    const base = vi.fn(() => stream);
+    const traced = createTracedStreamFn(base as unknown as StreamFn);
+
+    await withLogContext(
+      { conversationId: "conv_123", turnId: "turn_456" },
+      async () => {
+        await traced(
+          fakeModel("openai/gpt-5.4"),
+          { messages: [{ role: "user", content: "hi", timestamp: 0 }] },
+          undefined,
+        );
+      },
+    );
+
+    const opts = startInactiveSpan.mock.calls[0]?.[0] as {
+      attributes: Record<string, unknown>;
+    };
+    expect(opts.attributes["gen_ai.conversation.id"]).toBe("conv_123");
+    expect(opts.attributes["app.turn.id"]).toBe("turn_456");
+    // wrapper-supplied attributes still present
+    expect(opts.attributes["gen_ai.operation.name"]).toBe("chat");
+    expect(opts.attributes["gen_ai.request.model"]).toBe("openai/gpt-5.4");
+  });
+
+  it("ends the span when the stream errors", async () => {
+    const { createTracedStreamFn } = await import("@/chat/pi/traced-stream");
+    const stream = createAssistantMessageEventStream();
+    const base = vi.fn(() => stream);
+
+    const traced = createTracedStreamFn(base as unknown as StreamFn);
+    await traced(
+      fakeModel("openai/gpt-5.4"),
+      { messages: [{ role: "user", content: "hi", timestamp: 0 }] },
+      undefined,
+    );
+
+    // pi-ai's AssistantMessageEventStream resolves `result()` with the carrier
+    // AssistantMessage on `error` events instead of rejecting, so the wrapper's
+    // `.then` success arm runs on the error path. The load-bearing invariant
+    // is that the span ends exactly once.
+    const errorMessage = { ...fakeMessage(), stopReason: "error" as const };
+    stream.push({ type: "error", reason: "error", error: errorMessage });
+    await stream.result();
+    await new Promise((r) => setImmediate(r));
+
+    const span = getSpan();
+    expect(span.end).toHaveBeenCalledTimes(1);
+    // End attributes are still populated because the success arm runs.
+    const endAttributeKeys = span.setAttribute.mock.calls.map((c) => c[0]);
+    expect(endAttributeKeys).toContain("gen_ai.output.messages");
+  });
+});
diff --git a/specs/logging/tracing-spec.md b/specs/logging/tracing-spec.md
index cbc829b5..5c4cf398 100644
--- a/specs/logging/tracing-spec.md
+++ b/specs/logging/tracing-spec.md
@@ -3,7 +3,7 @@
 ## Metadata
 
 - Created: 2026-02-25
-- Last Edited: 2026-05-01
+- Last Edited: 2026-05-06
 
 ## Changelog
 
@@ -14,6 +14,7 @@
 - 2026-04-06: Added official GenAI finish-reasons, system-instructions, and tool-description tracing attributes.
 - 2026-04-28: Added MCP tool-call span attributes from the OpenTelemetry MCP semantic conventions.
 - 2026-05-01: Added `gen_ai.conversation.id` as required correlation context for GenAI spans when available.
+- 2026-05-06: Documented the `gen_ai.invoke_agent` / `gen_ai.chat` span hierarchy rule.
 
 ## Status
 
@@ -182,6 +183,14 @@ semantic conventions:
 - Sandbox spans should be nested under `ai.generate_assistant_reply` when invoked during reply generation.
 - Sandbox execution spans should be nested under the active tool-call/request span context.
 
+### GenAI Span Hierarchy
+
+- A `gen_ai.invoke_agent` span MUST have at least one `gen_ai.chat` child covering the LLM call(s) issued during its agent loop.
+- A `gen_ai.chat` span MAY appear at the top level (as a sibling of `gen_ai.invoke_agent`, or under a non-`gen_ai.*` parent such as `chat.route_thinking`) only when it represents an LLM call that is independent of an agent loop, for example a routing or classification pre-flight.
+- Every `gen_ai.chat` span MUST carry `gen_ai.input.messages` and `gen_ai.output.messages`.
+- The parent `gen_ai.invoke_agent` MAY also carry `gen_ai.input.messages` / `gen_ai.output.messages` as a high-level rollup; this is optional.
+- The per-iteration `gen_ai.chat` child span is created in `packages/junior/src/chat/pi/traced-stream.ts` via the `streamFn` injected into `pi-agent-core`'s `Agent`.
+
 ## Rollout Guidance
 
 - Start with lifecycle + I/O spans.