fix: browser integration for AI research agent

- Dockerfile: Install Google Chrome stable instead of Playwright npm package
  (playwright-cli defaults to /opt/google/chrome/chrome path). Removed unused
  Playwright npm package, kept Playwright Python for page automation.

- bundle/bundle.md: Fix orchestrator config override with correct key path
  (session.orchestrator.config.extended_thinking: false) to properly override
  foundation's extended_thinking: true default, which was causing empty
  AI text responses.

- bundle/agents/researcher.md: Add strong pre-installed environment guardrail
  telling agent NOT to install packages. Add explicit bash() tool call examples
  showing correct playwright-cli syntax. Fix snapshot command (removed non-existent
  -ic flag), change 'type' to 'fill' for form field syntax.

- bundle/context/researcher-instructions.md: Enrich from thin pointer to
  concrete examples with correct bash() wrapping and actual playwright-cli
  commands. Add 'do NOT install' guardrail.

Validated: Agent now opens Chrome browser during research queries, noVNC panel
shows live browser activity, AI generates text responses properly.

Generated with Amplifier

Co-Authored-By: Amplifier <240397093+microsoft-amplifier@users.noreply.github.com>

.dockerignore

@@ -0,0 +1,14 @@
+**/__pycache__
+**/*.pyc
+**/*.egg-info
+.venv
+backend/.venv
+frontend/node_modules
+frontend/dist
+artifacts
+.git
+.gitea
+.pytest_cache
+.ruff_cache
+tests
+*.swp

.gitea/workflows/deploy.yml

@@ -0,0 +1,52 @@
+name: Deploy Research Workbench
+
+on:
+  push:
+    branches: [main]
+
+jobs:
+  deploy:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Pull latest code
+        working-directory: /home/ken/workspace/research-workbench
+        run: |
+          git pull origin main
+
+      - name: Stop old containers
+        working-directory: /home/ken/workspace/research-workbench
+        run: |
+          docker rm -f car-help 2>/dev/null || true
+          docker compose down 2>/dev/null || true
+
+      - name: Build and start container
+        working-directory: /home/ken/workspace/research-workbench
+        env:
+          ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
+          AUTH_USER: ${{ secrets.AUTH_USER }}
+          AUTH_PASS_HASH: ${{ secrets.AUTH_PASS_HASH }}
+        run: |
+          docker compose up --build -d
+
+      - name: Wait for healthy
+        run: |
+          echo "Waiting for container to start..."
+          for i in $(seq 1 90); do
+            if curl -sf http://localhost:8080/ > /dev/null 2>&1; then
+              echo "App is up at http://localhost:8080/"
+              exit 0
+            fi
+            sleep 2
+          done
+          echo "ERROR: App did not respond within 180s"
+          docker compose -f /home/ken/workspace/research-workbench/docker-compose.yml logs --tail=50
+          exit 1
+
+      - name: Verify new UI is served
+        run: |
+          BODY=$(curl -sf http://localhost:8080/ || echo "FETCH_FAILED")
+          if echo "$BODY" | grep -qi "car-help"; then
+            echo "ERROR: Still serving old car-help UI!"
+            exit 1
+          fi
+          echo "Deploy successful - new UI is live at https://research.ampbox.io"

Dockerfile

@@ -1,12 +1,15 @@
 FROM python:3.13-slim
 
-# Install system dependencies
+# ── System dependencies ────────────────────────────────────────────────────────
+# Includes Xvfb/VNC stack, git (needed for pip's git+https installs),
+# and Chromium runtime libs (for the playwright-managed Chromium binary).
 RUN apt-get update && apt-get install -y --no-install-recommends \
     xvfb \
     x11vnc \
     websockify \
     novnc \
     curl \
+    git \
     libnss3 \
     libatk1.0-0 \
     libatk-bridge2.0-0 \
@@ -25,40 +28,58 @@ RUN apt-get update && apt-get install -y --no-install-recommends \
     fonts-liberation \
     && rm -rf /var/lib/apt/lists/*
 
-# Install Node.js 22.x from nodesource.com
+# ── Node.js 22.x ──────────────────────────────────────────────────────────────
 RUN curl -fsSL https://deb.nodesource.com/setup_22.x | bash - \
     && apt-get install -y --no-install-recommends nodejs \
     && rm -rf /var/lib/apt/lists/*
 
-# Install @anthropic-ai/playwright-cli globally via npm
-RUN npm install -g @anthropic-ai/playwright-cli
+# ── npm global tools: @playwright/cli + Chrome ────────────────────────────────
+# @playwright/cli  — token-efficient browser CLI used by research agents
+#                    (playwright-cli -s=research open/snapshot/click/type …)
+# playwright-cli defaults to Chrome at /opt/google/chrome/chrome. Install it directly.
+RUN npm install -g @playwright/cli \
+    && curl -fsSL https://dl.google.com/linux/direct/google-chrome-stable_current_amd64.deb -o /tmp/chrome.deb \
+    && apt-get update && apt-get install -y /tmp/chrome.deb \
+    && rm -f /tmp/chrome.deb && rm -rf /var/lib/apt/lists/*
 
-# Install uv via pip
+# ── Python tooling ────────────────────────────────────────────────────────────
 RUN pip install uv
+# overlayfs inside Docker doesn't support reflinks → use copy mode to avoid EAGAIN
+ENV UV_LINK_MODE=copy
 
-# Copy backend dependency files and install Python dependencies
+# ── amplifierd: Amplifier session daemon ──────────────────────────────────────
+# Not on PyPI — source is github.com/microsoft/amplifierd.
+# uv pip --system installs to /usr/local/bin/amplifierd (on PATH by default).
+#
+# amplifier-foundation is also not on PyPI (it's a pure git package that
+# amplifierd depends on), so we install it explicitly in the same step.
+# amplifier-core IS on PyPI and will be resolved automatically.
+RUN uv pip install --system \
+    "amplifierd @ git+https://github.com/microsoft/amplifierd" \
+    "amplifier-module-provider-anthropic @ git+https://github.com/microsoft/amplifier-module-provider-anthropic"
+
+# ── Frontend: install deps and build ──────────────────────────────────────────
+# Copy package manifests first so npm ci is cached independently of source changes.
+COPY frontend/package.json frontend/package-lock.json /app/frontend/
+WORKDIR /app/frontend
+RUN npm ci
+COPY frontend/ /app/frontend/
+RUN npm run build
+
+# ── Backend: install Python deps ──────────────────────────────────────────────
 COPY backend/pyproject.toml backend/uv.lock /app/backend/
 WORKDIR /app/backend
 RUN uv sync --frozen
 
-# Install Playwright chromium
-RUN uv run playwright install chromium
-
-# Install amplifierd
-RUN uv pip install amplifierd
-
-# Copy application files
+# ── Application files ─────────────────────────────────────────────────────────
 COPY backend/*.py /app/backend/
 COPY bundle/ /app/bundle/
-COPY frontend/dist/ /app/frontend/dist/
 COPY entrypoint.sh /entrypoint.sh
 RUN chmod +x /entrypoint.sh
 
-# Create artifacts directory
 RUN mkdir -p /app/artifacts
 
 WORKDIR /app
-
 EXPOSE 8080 6080
 
 ENV DISPLAY=:99

backend/agui_adapter.py

@@ -40,13 +40,19 @@ class AmplifierdEventTranslator:
         """Translate a single amplifierd SSE event into zero or more AG-UI events.
 
         Args:
-            event: Dict with 'event_type' and 'data' keys.
+            event: Dict with 'event_type' and 'data' keys.  The 'data' value is
+                the raw SSE payload from amplifierd, which wraps the event-specific
+                fields under a nested ``"data"`` key
+                (e.g. ``{"event": "...", "data": {...}, "session_id": ..., ...}``).
 
         Returns:
             List of ag_ui.core event objects (may be empty for unknown events).
         """
         event_type = event.get("event_type", "")
-        data = event.get("data", {})
+        outer = event.get("data", {})
+        # amplifierd wraps event-specific fields in a nested "data" key.
+        # Fall back to the outer dict for tests that pass data directly.
+        data = outer.get("data", outer)
 
         if event_type == "content_block:start":
             return self._handle_block_start(data)
@@ -62,9 +68,13 @@ class AmplifierdEventTranslator:
             return []
 
     def _handle_block_start(self, data: dict[str, Any]) -> list[Any]:
-        """Handle content_block:start — emit TEXT_MESSAGE_START or TOOL_CALL_START."""
-        index = data.get("index", 0)
-        block_type = data.get("type", "text")
+        """Handle content_block:start — register block; defer TEXT_MESSAGE_START to first delta.
+
+        Supports both amplifierd wire format (``block_index`` / ``block_type``) and the
+        legacy test format (``index`` / ``type``) via fallbacks.
+        """
+        index = data.get("block_index", data.get("index", 0))
+        block_type = data.get("block_type", data.get("type", "text"))
 
         if block_type == "tool_use":
             tool_id = data.get("id", "")
@@ -81,21 +91,27 @@ class AmplifierdEventTranslator:
                     parent_message_id=self._current_message_id,
                 )
             ]
-        else:
-            # Text block
+        elif block_type == "text":
+            # Lazy start: don't emit TEXT_MESSAGE_START until we know there's actual content.
             message_id = str(uuid.uuid4())
-            self._blocks[index] = {"type": "text", "message_id": message_id}
+            self._blocks[index] = {"type": "text", "message_id": message_id, "started": False}
             self._current_message_id = message_id
-            return [
-                TextMessageStartEvent(
-                    message_id=message_id,
-                    role="assistant",
-                )
-            ]
+            return []
+        else:
+            # thinking, image, or any other block type — skip entirely
+            self._blocks[index] = {"type": "skip"}
+            return []
 
     def _handle_block_delta(self, data: dict[str, Any]) -> list[Any]:
-        """Handle content_block:delta — emit TEXT_MESSAGE_CONTENT or TOOL_CALL_ARGS."""
-        index = data.get("index", 0)
+        """Handle content_block:delta — emit TEXT_MESSAGE_CONTENT or TOOL_CALL_ARGS.
+
+        For text blocks, TEXT_MESSAGE_START is emitted lazily here on the first
+        non-empty token so that thinking/empty text blocks produce zero AG-UI events.
+
+        Supports both amplifierd wire format (``block_index``) and legacy test format
+        (``index``) via fallback.
+        """
+        index = data.get("block_index", data.get("index", 0))
         delta = data.get("delta", {})
         delta_type = delta.get("type", "")
 
@@ -112,6 +128,9 @@ class AmplifierdEventTranslator:
                 ]
             return []
 
+        if block_info["type"] == "skip":
+            return []
+
         if block_info["type"] == "tool_use":
             return [
                 ToolCallArgsEvent(
@@ -120,10 +139,22 @@ class AmplifierdEventTranslator:
                 )
             ]
         else:
-            # Text block
+            # Text block — lazy start: emit TEXT_MESSAGE_START on the first real token
             token = delta.get("text", "")
             if not token:
                 return []
+            if not block_info.get("started"):
+                block_info["started"] = True
+                return [
+                    TextMessageStartEvent(
+                        message_id=block_info["message_id"],
+                        role="assistant",
+                    ),
+                    TextMessageContentEvent(
+                        message_id=block_info["message_id"],
+                        delta=token,
+                    ),
+                ]
             return [
                 TextMessageContentEvent(
                     message_id=block_info["message_id"],
@@ -132,17 +163,46 @@ class AmplifierdEventTranslator:
             ]
 
     def _handle_block_end(self, data: dict[str, Any]) -> list[Any]:
-        """Handle content_block:end — emit TEXT_MESSAGE_END or TOOL_CALL_END."""
-        index = data.get("index", 0)
+        """Handle content_block:end — emit TEXT_MESSAGE_END or TOOL_CALL_END.
+
+        Supports two modes:
+        - **Streaming mode** (delta events were received): the block was already opened
+          via ``_handle_block_delta``; just close it with TEXT_MESSAGE_END.
+        - **Batch mode** (no delta events, full text in ``block.text``): emit
+          TEXT_MESSAGE_START + TEXT_MESSAGE_CONTENT + TEXT_MESSAGE_END in one shot.
+          This is the format used by amplifierd, which sends complete blocks at end time.
+
+        Text blocks with no content in either mode are silently dropped.
+
+        Supports both amplifierd wire format (``block_index``) and legacy test format
+        (``index``) via fallback.
+        """
+        index = data.get("block_index", data.get("index", 0))
         block_info = self._blocks.pop(index, None)
 
         if block_info is None:
             return []
 
+        if block_info["type"] == "skip":
+            return []
+
         if block_info["type"] == "tool_use":
             return [ToolCallEndEvent(tool_call_id=block_info["tool_id"])]
         else:
-            return [TextMessageEndEvent(message_id=block_info["message_id"])]
+            message_id = block_info["message_id"]
+            if block_info.get("started"):
+                # Streaming mode: deltas already opened the message — just close it.
+                return [TextMessageEndEvent(message_id=message_id)]
+            # Batch mode: full text is in the "block" field (amplifierd wire format).
+            block = data.get("block", {})
+            text = block.get("text", "")
+            if text:
+                return [
+                    TextMessageStartEvent(message_id=message_id, role="assistant"),
+                    TextMessageContentEvent(message_id=message_id, delta=text),
+                    TextMessageEndEvent(message_id=message_id),
+                ]
+            return []
 
     def _handle_tool_result(self, data: dict[str, Any]) -> list[Any]:
         """Handle tool:result — emit TOOL_CALL_RESULT with optional MCP App metadata."""

bundle/agents/researcher.md

@@ -97,11 +97,24 @@ extract current data, and synthesize findings into clear, actionable results. Th
 browser runs persistently — tabs stay open between commands, and the user can watch
 and take over at any time via the noVNC view.
 
+## IMPORTANT: Pre-installed Environment
+
+`playwright-cli` is **already installed** in this container. Do NOT install anything.
+Do NOT run `npm install`, `pip install`, or any package manager. Just use the commands
+below directly via the bash tool.
+
 ## Browser Control (playwright-cli)
 
-You control the browser using `playwright-cli` with a persistent named session
-(`-s=research`). All commands operate on the same browser session, preserving tabs
-and state across the entire conversation.
+You control the browser using `playwright-cli` via the bash tool. All commands use
+a persistent named session (`-s=research`), preserving tabs and state across the
+entire conversation.
+
+**How to invoke:** Every playwright-cli command runs through the bash tool:
+```
+bash("playwright-cli -s=research open https://example.com")
+bash("playwright-cli -s=research snapshot")
+bash("playwright-cli -s=research click @e5")
+```
 
 ### Commands
 
@@ -109,9 +122,9 @@ and state across the entire conversation.
 |--------|---------|
 | Open a URL | `playwright-cli -s=research open https://example.com` |
 | Take a snapshot | `playwright-cli -s=research snapshot` |
-| Compact snapshot | `playwright-cli -s=research snapshot -ic` |
 | Click an element | `playwright-cli -s=research click @e5` |
-| Type text | `playwright-cli -s=research type @e3 "search query"` |
+| Type text | `playwright-cli -s=research type "search query"` |
+| Fill a field | `playwright-cli -s=research fill @e3 "search query"` |
 | Press a key | `playwright-cli -s=research press Enter` |
 | Scroll down | `playwright-cli -s=research scroll down` |
 | Go back | `playwright-cli -s=research back` |

bundle/bundle.md

@@ -4,6 +4,17 @@ bundle:
   version: 0.1.0
   description: General-purpose AI research tool with browser control and artifact generation
 
+providers:
+  - module: provider-anthropic
+    source: git+https://github.com/microsoft/amplifier-module-provider-anthropic@main
+    config:
+      default_model: claude-sonnet-4-20250514
+
+session:
+  orchestrator:
+    config:
+      extended_thinking: false
+
 includes:
   - bundle: git+https://github.com/microsoft/amplifier-foundation@main
   - bundle: research-workbench:behaviors/research-workbench

bundle/context/researcher-instructions.md

@@ -3,22 +3,56 @@
 This session runs inside the **Research Workbench**. You have access to the
 following capabilities:
 
-## Available Capabilities
+## CRITICAL: Pre-installed Tools
 
-**Browser** — Live Chromium browser controlled via `playwright-cli -s=research`.
-The browser is persistent across commands and visible to the user via VNC (noVNC
-panel). The user can take over at any time.
+All tools are **already installed**. Do NOT run npm install, pip install, or any
+package manager. Just use the commands directly via the bash tool.
 
-**Artifacts** — Persistent markdown documents written to `artifacts/<session_id>/`.
+## Browser Control
+
+Live Chromium browser controlled via `playwright-cli -s=research` through the
+bash tool. The browser is persistent across commands and visible to the user via
+the noVNC panel. The user can take over at any time.
+
+**Quick start — open a site and see what's on it:**
+```
+bash("playwright-cli -s=research open https://google.com")
+bash("playwright-cli -s=research snapshot")
+```
+
+**Click an element from the snapshot:**
+```
+bash("playwright-cli -s=research click @e5")
+bash("playwright-cli -s=research snapshot")
+```
+
+**Type into a field and submit:**
+```
+bash("playwright-cli -s=research fill @e3 \"search query\"")
+bash("playwright-cli -s=research press Enter")
+bash("playwright-cli -s=research snapshot")
+```
+
+Always snapshot after navigation to get fresh element refs.
+
+## Artifacts
+
+Persistent markdown documents written to `artifacts/<session_id>/`.
 They appear in the right panel and survive across sessions.
 
-**A2UI** — Inline interactive controls (filter chips, tables, card grids, forms)
+## A2UI
+
+Inline interactive controls (filter chips, tables, card grids, forms)
 rendered directly in chat via `a2ui` code blocks.
 
-**Inline Visuals** — Rich visual output (charts, maps, comparison tables with color
+## Inline Visuals
+
+Rich visual output (charts, maps, comparison tables with color
 coding) rendered directly in chat via `visual` code blocks.
 
-## Full Command Reference
+## Research Workflow
 
-For complete command syntax, workflow guidance, and best practices, see the
-researcher agent system prompt.
+1. Understand the request — clarify scope and constraints
+2. Plan — identify 2-4 sources to consult
+3. Browse and extract — use playwright-cli via bash to navigate sites and extract data
+4. Present — use A2UI, visuals, or artifacts to deliver results

entrypoint.sh

@@ -21,8 +21,11 @@ sleep 1
 echo "[OK] noVNC on :6080"
 
 # (4) Start amplifierd and wait for it to be healthy
+# amplifierd uses a 'serve' subcommand; --bundle requires NAME=URI format
 cd /app
-amplifierd --port 8410 --bundle /app/bundle/bundle.md &
+amplifierd serve --port 8410 \
+    --bundle "research-workbench=/app/bundle/bundle.md" \
+    --default-bundle research-workbench &
 AMPLIFIERD_PID=$!
 
 echo "Waiting for amplifierd..."

frontend/src/hooks/useChat.ts

@@ -95,7 +95,7 @@ export function useChat(sessionId: string | null) {
               break
             }
             case 'TEXT_MESSAGE_END': {
-              // no-op
+              assistantId = null
               break
             }
             case 'TOOL_CALL_START': {

tests/test_agui_adapter.py

@@ -18,21 +18,17 @@ class TestTextBlocks:
         )
 
     def test_text_block_start(self):
-        """text_block_start → TEXT_MESSAGE_START with role='assistant' and messageId."""
+        """text_block_start → no events emitted (lazy: start is deferred to first delta)."""
         event = {
             "event_type": "content_block:start",
             "data": {"index": 0, "type": "text"},
         }
         result = self.translator.translate(event)
-        assert len(result) == 1
-        ev = result[0]
-        assert ev.type.value == "TEXT_MESSAGE_START"
-        assert ev.role == "assistant"
-        assert ev.message_id  # must have a non-empty message_id
+        assert result == []
 
     def test_text_delta(self):
-        """text_delta → TEXT_MESSAGE_CONTENT with delta='Hello '."""
-        # Start the text block first
+        """First text_delta → TEXT_MESSAGE_START then TEXT_MESSAGE_CONTENT (lazy start)."""
+        # Start the text block — no event emitted yet
         start_event = {
             "event_type": "content_block:start",
             "data": {"index": 0, "type": "text"},
@@ -47,19 +43,27 @@ class TestTextBlocks:
             },
         }
         result = self.translator.translate(delta_event)
-        assert len(result) == 1
-        ev = result[0]
-        assert ev.type.value == "TEXT_MESSAGE_CONTENT"
-        assert ev.delta == "Hello "
+        # Lazy start: first delta emits [TEXT_MESSAGE_START, TEXT_MESSAGE_CONTENT]
+        assert len(result) == 2
+        start_ev, content_ev = result
+        assert start_ev.type.value == "TEXT_MESSAGE_START"
+        assert start_ev.role == "assistant"
+        assert start_ev.message_id
+        assert content_ev.type.value == "TEXT_MESSAGE_CONTENT"
+        assert content_ev.delta == "Hello "
+        assert content_ev.message_id == start_ev.message_id
 
     def test_text_block_end(self):
-        """text_block_end → TEXT_MESSAGE_END."""
-        # Start the block first
-        start_event = {
+        """text_block_end after content → TEXT_MESSAGE_END with matching messageId."""
+        # Start block, send a delta (which lazily opens it), then close it
+        self.translator.translate({
             "event_type": "content_block:start",
             "data": {"index": 0, "type": "text"},
-        }
-        self.translator.translate(start_event)
+        })
+        self.translator.translate({
+            "event_type": "content_block:delta",
+            "data": {"index": 0, "delta": {"type": "text_delta", "text": "Hi"}},
+        })
 
         end_event = {
             "event_type": "content_block:end",
@@ -195,6 +199,110 @@ class TestToolBlocks:
         assert "structuredContent" not in d
 
 
+class TestExtendedThinking:
+    """Tests for extended thinking blocks — must produce zero AG-UI events."""
+
+    def setup_method(self):
+        self.translator = AmplifierdEventTranslator(
+            run_id="run_1", thread_id="thread_1"
+        )
+
+    # ------------------------------------------------------------------
+    # helpers
+    # ------------------------------------------------------------------
+
+    def _thinking_start(self, index: int = 0) -> dict:
+        return {
+            "event_type": "content_block:start",
+            "data": {"index": index, "type": "thinking"},
+        }
+
+    def _thinking_delta(self, text: str, index: int = 0) -> dict:
+        return {
+            "event_type": "content_block:delta",
+            "data": {"index": index, "delta": {"type": "thinking_delta", "thinking": text}},
+        }
+
+    def _thinking_end(self, index: int = 0) -> dict:
+        return {"event_type": "content_block:end", "data": {"index": index}}
+
+    def _text_start(self, index: int = 1) -> dict:
+        return {
+            "event_type": "content_block:start",
+            "data": {"index": index, "type": "text"},
+        }
+
+    def _text_delta(self, text: str, index: int = 1) -> dict:
+        return {
+            "event_type": "content_block:delta",
+            "data": {"index": index, "delta": {"type": "text_delta", "text": text}},
+        }
+
+    def _text_end(self, index: int = 1) -> dict:
+        return {"event_type": "content_block:end", "data": {"index": index}}
+
+    # ------------------------------------------------------------------
+    # tests
+    # ------------------------------------------------------------------
+
+    def test_thinking_block_produces_no_events(self):
+        """thinking start/delta/end → all return [] (thinking blocks are skipped)."""
+        assert self.translator.translate(self._thinking_start()) == []
+        assert self.translator.translate(self._thinking_delta("Let me reason…")) == []
+        assert self.translator.translate(self._thinking_end()) == []
+
+    def test_empty_text_block_produces_no_events(self):
+        """text start then immediate end with no deltas → no AG-UI events emitted."""
+        assert self.translator.translate(self._text_start()) == []
+        assert self.translator.translate(self._text_end()) == []
+
+    def test_thinking_then_empty_text_produces_no_events(self):
+        """Full sequence: thinking block + empty text block → zero AG-UI events total."""
+        events = [
+            self._thinking_start(index=0),
+            self._thinking_delta("I need to think…", index=0),
+            self._thinking_end(index=0),
+            self._text_start(index=1),
+            self._text_end(index=1),
+        ]
+        all_results = []
+        for ev in events:
+            all_results.extend(self.translator.translate(ev))
+        assert all_results == []
+
+    def test_thinking_then_nonempty_text_works(self):
+        """Thinking block (skip) then text block with actual content → START + CONTENT + END."""
+        # Thinking block — all silent
+        self.translator.translate(self._thinking_start(index=0))
+        self.translator.translate(self._thinking_delta("Reasoning…", index=0))
+        self.translator.translate(self._thinking_end(index=0))
+
+        # Text block — lazy open on first delta
+        self.translator.translate(self._text_start(index=1))
+
+        first_delta_result = self.translator.translate(self._text_delta("Hello", index=1))
+        assert len(first_delta_result) == 2
+        start_ev, content_ev = first_delta_result
+        assert start_ev.type.value == "TEXT_MESSAGE_START"
+        assert start_ev.role == "assistant"
+        message_id = start_ev.message_id
+        assert content_ev.type.value == "TEXT_MESSAGE_CONTENT"
+        assert content_ev.delta == "Hello"
+        assert content_ev.message_id == message_id
+
+        # Subsequent delta — only CONTENT (no second START)
+        second_delta_result = self.translator.translate(self._text_delta(" world", index=1))
+        assert len(second_delta_result) == 1
+        assert second_delta_result[0].type.value == "TEXT_MESSAGE_CONTENT"
+        assert second_delta_result[0].delta == " world"
+
+        # End — TEXT_MESSAGE_END because content was started
+        end_result = self.translator.translate(self._text_end(index=1))
+        assert len(end_result) == 1
+        assert end_result[0].type.value == "TEXT_MESSAGE_END"
+        assert end_result[0].message_id == message_id
+
+
 class TestLifecycle:
     """Tests for lifecycle events → AG-UI run lifecycle events."""