chore: release api v1.0.0

- Event-based streaming for task agent
- Retry logic when LLM responds without calling tools
- Hardened prompts to enforce tool use
- Working directory context in all agent prompts
- Project paused: local LLMs not capable enough for agentic use

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

webber-api/CHANGELOG.md

@@ -7,6 +7,29 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+## [1.0.0] - 2026-01-15
+
+### Added
+- Event-based streaming for task agent (`StreamEvent` objects instead of raw text)
+  - New `tools_streaming.py` with all tools emitting structured events
+  - Event types: `tool_start`, `tool_done`, `thinking`, `response`, `error`, `done`
+- Retry logic when LLM responds without calling tools (max 2 retries)
+  - Tracks `tools_called` counter on TaskContext
+  - Stronger retry prompt forces tool use
+- Working directory context injected into all agent prompts
+
+### Changed
+- Hardened system prompts to enforce tool use before responding
+  - Added "CRITICAL RULE" section requiring tool calls first
+  - Made "MANDATORY WORKFLOW" more emphatic
+- Updated explore and plan agents with `_build_prompt_with_context()` method
+
+### Fixed
+- Agent path hallucination - now explicitly communicates working directory to LLM
+
+### Note
+- Project paused: Local LLMs (Mistral Nemo 12B on available hardware) are not capable enough for reliable agentic tool use. Models frequently hallucinate responses instead of calling tools, even with prompt hardening and retry logic. Would require larger models (70B+) or cloud API integration to continue.
+
 ## [0.4.2] - 2026-01-11
 
 ### Added

webber-api/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "webber-api"
-version = "0.4.2"
+version = "1.0.0"
 description = "Webber API - Multi-Agent AI Development Server"
 authors = [
     {name = "jpmschweitzer"}

webber-api/src/domains/agents/explore/agent.py

@@ -79,6 +79,14 @@ class ExploreAgentImpl(BaseAgent):
         from src.domains.agents.explore.tools import register_explore_tools
         register_explore_tools(agent)
 
+    def _build_prompt_with_context(self, prompt: str, working_dir: str) -> str:
+        """Build the prompt with working directory context."""
+        return f"""Working directory: {working_dir}
+
+Use paths within this working directory for file operations.
+
+User request: {prompt}"""
+
     @logged()
     async def run(
         self,
@@ -98,16 +106,20 @@ class ExploreAgentImpl(BaseAgent):
         Returns:
             Agent response with findings
         """
+        effective_working_dir = working_dir or os.getcwd()
+
         ctx = ExploreContext(
-            working_dir=working_dir or os.getcwd(),
+            working_dir=effective_working_dir,
             allowed_paths=allowed_paths or self._settings.effective_allowed_paths,
             timeout_seconds=self._settings.tool_timeout_seconds,
         )
 
+        full_prompt = self._build_prompt_with_context(prompt, effective_working_dir)
+
         async with trace_span("explore_agent_run"):
             try:
                 # Use run() not run_stream() - Ollama has bugs with streaming + tools
-                result = await self.agent.run(prompt, deps=ctx)
+                result = await self.agent.run(full_prompt, deps=ctx)
                 return result.output
             except Exception as e:
                 logger.exception(f"Explore agent error: {e}")
@@ -126,15 +138,19 @@ class ExploreAgentImpl(BaseAgent):
 
         Yields text chunks as they become available.
         """
+        effective_working_dir = working_dir or os.getcwd()
+
         ctx = ExploreContext(
-            working_dir=working_dir or os.getcwd(),
+            working_dir=effective_working_dir,
             allowed_paths=allowed_paths or self._settings.effective_allowed_paths,
             timeout_seconds=self._settings.tool_timeout_seconds,
         )
 
+        full_prompt = self._build_prompt_with_context(prompt, effective_working_dir)
+
         async with trace_span("explore_agent_stream"):
             try:
-                async with self.agent.run_stream(prompt, deps=ctx) as result:
+                async with self.agent.run_stream(full_prompt, deps=ctx) as result:
                     async for chunk in result.stream_text():
                         yield chunk
             except Exception as e:

webber-api/src/domains/agents/plan/agent.py

@@ -82,6 +82,14 @@ class PlanAgentImpl(BaseAgent):
         from src.domains.agents.plan.tools import register_plan_tools
         register_plan_tools(agent)
 
+    def _build_prompt_with_context(self, prompt: str, working_dir: str) -> str:
+        """Build the prompt with working directory context."""
+        return f"""Working directory: {working_dir}
+
+Use paths within this working directory for file operations.
+
+User request: {prompt}"""
+
     @logged()
     async def run(
         self,
@@ -101,16 +109,20 @@ class PlanAgentImpl(BaseAgent):
         Returns:
             Implementation plan with steps and critical files
         """
+        effective_working_dir = working_dir or os.getcwd()
+
         ctx = PlanContext(
-            working_dir=working_dir or os.getcwd(),
+            working_dir=effective_working_dir,
             allowed_paths=allowed_paths or self._settings.effective_allowed_paths,
             timeout_seconds=self._settings.tool_timeout_seconds,
         )
 
+        full_prompt = self._build_prompt_with_context(prompt, effective_working_dir)
+
         async with trace_span("plan_agent_run"):
             try:
                 # Use run() not run_stream() - Ollama has bugs with streaming + tools
-                result = await self.agent.run(prompt, deps=ctx)
+                result = await self.agent.run(full_prompt, deps=ctx)
                 return result.output
             except Exception as e:
                 logger.exception(f"Plan agent error: {e}")
@@ -128,15 +140,19 @@ class PlanAgentImpl(BaseAgent):
 
         Yields text chunks as they become available.
         """
+        effective_working_dir = working_dir or os.getcwd()
+
         ctx = PlanContext(
-            working_dir=working_dir or os.getcwd(),
+            working_dir=effective_working_dir,
             allowed_paths=allowed_paths or self._settings.effective_allowed_paths,
             timeout_seconds=self._settings.tool_timeout_seconds,
         )
 
+        full_prompt = self._build_prompt_with_context(prompt, effective_working_dir)
+
         async with trace_span("plan_agent_stream"):
             try:
-                async with self.agent.run_stream(prompt, deps=ctx) as result:
+                async with self.agent.run_stream(full_prompt, deps=ctx) as result:
                     async for chunk in result.stream_text():
                         yield chunk
             except Exception as e:

webber-api/src/domains/agents/router.py

@@ -5,6 +5,9 @@ Supports permission modes for controlling agent tool access:
 - default: All tools available (approval may be required)
 - plan: Read-only tools only
 - auto_accept: All tools, no approval prompts
+
+Streaming uses structured events instead of raw text to avoid
+garbled output during tool execution.
 """
 import json
 from fastapi import APIRouter, HTTPException
@@ -22,6 +25,7 @@ from src.domains.agents.schemas import (
     AgentInfo,
     AgentListResponse,
     PermissionMode,
+    StreamEvent,
 )
 
 
@@ -98,13 +102,16 @@ async def stream_agent(request: AgentRunRequest) -> StreamingResponse:
     """
     Run an agent with streaming response.
 
-    Returns Server-Sent Events (SSE) with text chunks.
+    Returns Server-Sent Events (SSE) with structured events.
     Permission mode controls which tools are available.
 
-    Event types:
-    - "chunk": Text chunk from the agent
-    - "done": Stream complete
-    - "error": Error occurred
+    Event types (from StreamEvent):
+    - tool_start: Tool execution beginning
+    - tool_done: Tool execution complete
+    - thinking: Agent status update
+    - response: Final response text chunk
+    - error: Error occurred
+    - done: Stream complete
     """
     agent = get_agent(request.agent_type)
     if not agent:
@@ -118,21 +125,23 @@ async def stream_agent(request: AgentRunRequest) -> StreamingResponse:
 
     async def generate():
         try:
-            async for chunk in agent.run_stream(
+            async for event in agent.run_stream(
                 request.prompt,
                 working_dir=request.working_dir,
                 mode=mode,
             ):
-                # SSE format: data: {json}\n\n
-                event = {"event": "chunk", "data": chunk}
-                yield f"data: {json.dumps(event)}\n\n"
-
-            # Signal completion
-            yield f"data: {json.dumps({'event': 'done', 'mode': mode.value})}\n\n"
+                # Handle both StreamEvent objects and legacy string chunks
+                if isinstance(event, StreamEvent):
+                    # New structured event format
+                    event_data = event.model_dump(exclude_none=True)
+                    yield f"data: {json.dumps(event_data)}\n\n"
+                else:
+                    # Legacy string chunk (for explore/plan agents)
+                    yield f"data: {json.dumps({'event': 'chunk', 'data': event})}\n\n"
 
         except Exception as e:
             logger.exception(f"Stream error: {e}")
-            error_event = {"event": "error", "data": str(e)}
+            error_event = {"event": "error", "error_message": str(e)}
             yield f"data: {json.dumps(error_event)}\n\n"
 
     return StreamingResponse(

webber-api/src/domains/agents/schemas.py

@@ -127,3 +127,36 @@ class AgentInfo(BaseSchema):
 class AgentListResponse(BaseSchema):
     """List of available agents."""
     agents: list[AgentInfo]
+
+
+# Streaming event types for event-based streaming
+class StreamEventType(str, Enum):
+    """
+    Event types for structured agent streaming.
+
+    Instead of streaming raw text (which gets garbled during tool calls),
+    we emit structured events that the CLI can render appropriately.
+    """
+    tool_start = "tool_start"      # Tool execution starting
+    tool_done = "tool_done"        # Tool execution complete
+    thinking = "thinking"          # Agent reasoning status
+    response = "response"          # Final response text chunk
+    error = "error"                # Error occurred
+    done = "done"                  # Stream complete
+
+
+class StreamEvent(BaseSchema):
+    """
+    Structured streaming event from agent execution.
+
+    Events are emitted instead of raw text to provide clean
+    progress feedback during multi-tool agent loops.
+    """
+    event: StreamEventType
+    tool: str | None = None           # Tool name (for tool_start/tool_done)
+    args: dict | None = None          # Tool arguments (for tool_start)
+    result_summary: str | None = None # Brief result (for tool_done)
+    message: str | None = None        # Status message (for thinking)
+    text: str | None = None           # Response text (for response)
+    error_message: str | None = None  # Error details (for error)
+    mode: str | None = None           # Permission mode (for done)

webber-api/src/domains/agents/task/agent.py

@@ -5,7 +5,9 @@ Full orchestrator agent that can:
 - Execute multi-step tasks autonomously
 - Use all tools (read + write) based on permission mode
 - Spawn sub-agents (Explore, Plan) for focused work
+- Stream structured events instead of raw text
 """
+import asyncio
 import os
 from collections.abc import AsyncIterator
 from dataclasses import dataclass, field
@@ -15,7 +17,7 @@ from pydantic_ai import Agent
 from pydantic_ai.models.openai import OpenAIModel
 
 from src.domains.agents.base import BaseAgent, AgentContext, register_agent
-from src.domains.agents.schemas import PermissionMode
+from src.domains.agents.schemas import PermissionMode, StreamEvent, StreamEventType
 from src.domains.agents.task.prompts import TASK_SYSTEM_PROMPT, TASK_PLAN_MODE_PROMPT
 from src.ollama.provider import get_ollama_provider
 from src.shared.config import get_settings
@@ -30,11 +32,33 @@ class TaskContext(AgentContext):
     Context for task agent tools.
 
     Passed to all tool functions via RunContext.
-    Extends base AgentContext with permission mode.
+    Extends base AgentContext with permission mode and event queue.
     """
     mode: PermissionMode = PermissionMode.default
     # Prep for approval flow - tools can check this
     pending_approvals: list[str] = field(default_factory=list)
+    # Event queue for streaming events from tools
+    event_queue: asyncio.Queue | None = field(default=None, repr=False)
+    # Track tool calls for retry logic
+    tools_called: int = 0
+
+
+def _emit_event(ctx: AgentContext, event: StreamEvent) -> None:
+    """Emit an event to the queue if available."""
+    if hasattr(ctx, 'event_queue') and ctx.event_queue is not None:
+        ctx.event_queue.put_nowait(event)
+
+
+def _summarize_result(result: str, max_len: int = 80) -> str:
+    """Create a brief summary of a tool result."""
+    # Count lines if multiline
+    lines = result.strip().split('\n')
+    if len(lines) > 1:
+        return f"{len(lines)} lines"
+    # Single line - truncate if needed
+    if len(result) > max_len:
+        return result[:max_len] + "..."
+    return result
 
 
 class TaskAgentImpl(BaseAgent):
@@ -102,14 +126,41 @@ class TaskAgentImpl(BaseAgent):
 
     def _register_tools(self, agent: Agent[TaskContext, str], mode: PermissionMode) -> None:
         """Register tools with the agent based on permission mode."""
-        from src.domains.agents.task.tools import register_task_tools, register_readonly_tools
+        from src.domains.agents.task.tools_streaming import (
+            register_task_tools_streaming,
+            register_readonly_tools_streaming,
+        )
 
         if mode == PermissionMode.plan:
             # Plan mode: read-only tools only
-            register_readonly_tools(agent)
+            register_readonly_tools_streaming(agent)
         else:
             # Default and auto_accept: all tools
-            register_task_tools(agent)
+            register_task_tools_streaming(agent)
+
+    # Maximum retries when no tools are called
+    MAX_NO_TOOL_RETRIES = 2
+
+    def _build_prompt_with_context(self, prompt: str, working_dir: str) -> str:
+        """Build the prompt with working directory context."""
+        return f"""Working directory: {working_dir}
+
+When using file tools, use paths relative to or within this working directory.
+For example, to read a file at {working_dir}/README.md, use file_path="{working_dir}/README.md".
+
+User request: {prompt}"""
+
+    def _build_retry_prompt(self, prompt: str, working_dir: str) -> str:
+        """Build a stronger prompt for retry after no tool calls."""
+        return f"""Working directory: {working_dir}
+
+IMPORTANT: Your previous response was REJECTED because you did not call any tools.
+You MUST call a tool (like glob_files, bash_readonly, or read_file) BEFORE responding.
+DO NOT answer from memory. DO NOT fabricate information.
+
+Call a tool NOW to gather real information, then respond based on the results.
+
+User request: {prompt}"""
 
     @logged()
     async def run(
@@ -132,24 +183,48 @@ class TaskAgentImpl(BaseAgent):
         Returns:
             Consolidated task summary with results
         """
-        ctx = TaskContext(
-            working_dir=working_dir or os.getcwd(),
-            allowed_paths=allowed_paths or self._settings.effective_allowed_paths,
-            timeout_seconds=self._settings.tool_timeout_seconds,
-            mode=mode,
-        )
+        effective_working_dir = working_dir or os.getcwd()
 
         # Get agent configured for this mode
         agent = self._get_agent_for_mode(mode)
 
         async with trace_span("task_agent_run"):
+            retries = 0
+            while retries <= self.MAX_NO_TOOL_RETRIES:
+                # Create fresh context for each attempt
+                ctx = TaskContext(
+                    working_dir=effective_working_dir,
+                    allowed_paths=allowed_paths or self._settings.effective_allowed_paths,
+                    timeout_seconds=self._settings.tool_timeout_seconds,
+                    mode=mode,
+                )
+
+                # Build prompt - use retry prompt if this is a retry
+                if retries == 0:
+                    full_prompt = self._build_prompt_with_context(prompt, effective_working_dir)
+                else:
+                    full_prompt = self._build_retry_prompt(prompt, effective_working_dir)
+                    logger.warning(f"Retry {retries}/{self.MAX_NO_TOOL_RETRIES}: No tools called, retrying with stronger prompt")
+
                 try:
-                result = await agent.run(prompt, deps=ctx)
+                    result = await agent.run(full_prompt, deps=ctx)
+
+                    # Check if tools were called
+                    if ctx.tools_called == 0 and retries < self.MAX_NO_TOOL_RETRIES:
+                        retries += 1
+                        continue
+
+                    if ctx.tools_called == 0:
+                        logger.warning("Agent responded without calling tools after all retries")
+
                     return result.output
                 except Exception as e:
                     logger.exception(f"Task agent error: {e}")
                     raise
 
+            # Should not reach here, but just in case
+            return result.output
+
     async def run_stream(
         self,
         prompt: str,
@@ -157,9 +232,12 @@ class TaskAgentImpl(BaseAgent):
         allowed_paths: list[str] | None = None,
         mode: PermissionMode = PermissionMode.default,
         **kwargs: Any
-    ) -> AsyncIterator[str]:
+    ) -> AsyncIterator[StreamEvent]:
         """
-        Run the task agent with streaming output.
+        Run the task agent with structured event streaming.
+
+        Instead of streaming raw text (which gets garbled during tool calls),
+        yields structured events that clients can render appropriately.
 
         Args:
             prompt: Task description
@@ -168,27 +246,129 @@ class TaskAgentImpl(BaseAgent):
             mode: Permission mode controlling tool access
 
         Yields:
-            Text chunks as they become available.
+            StreamEvent objects for tool progress and final response.
+
+        Event types:
+        - tool_start: Tool execution beginning
+        - tool_done: Tool execution complete with summary
+        - thinking: Agent status update
+        - response: Final response text
+        - error: Error occurred
+        - done: Stream complete
         """
-        ctx = TaskContext(
-            working_dir=working_dir or os.getcwd(),
-            allowed_paths=allowed_paths or self._settings.effective_allowed_paths,
-            timeout_seconds=self._settings.tool_timeout_seconds,
-            mode=mode,
-        )
+        effective_working_dir = working_dir or os.getcwd()
 
         # Get agent configured for this mode
         agent = self._get_agent_for_mode(mode)
 
         async with trace_span("task_agent_stream"):
+            # Emit initial thinking event
+            yield StreamEvent(
+                event=StreamEventType.thinking,
+                message="Starting task execution..."
+            )
+
+            retries = 0
+            response = ""
+
+            while retries <= self.MAX_NO_TOOL_RETRIES:
+                # Create fresh event queue and context for each attempt
+                event_queue: asyncio.Queue[StreamEvent] = asyncio.Queue()
+
+                ctx = TaskContext(
+                    working_dir=effective_working_dir,
+                    allowed_paths=allowed_paths or self._settings.effective_allowed_paths,
+                    timeout_seconds=self._settings.tool_timeout_seconds,
+                    mode=mode,
+                    event_queue=event_queue,
+                )
+
+                # Build prompt - use retry prompt if this is a retry
+                if retries == 0:
+                    full_prompt = self._build_prompt_with_context(prompt, effective_working_dir)
+                else:
+                    full_prompt = self._build_retry_prompt(prompt, effective_working_dir)
+                    yield StreamEvent(
+                        event=StreamEventType.thinking,
+                        message=f"Retrying (attempt {retries + 1})..."
+                    )
+
+                # Run agent in background task so we can yield events
+                async def run_agent() -> str:
                     try:
-                async with agent.run_stream(prompt, deps=ctx) as result:
-                    async for chunk in result.stream_text():
-                        yield chunk
+                        result = await agent.run(full_prompt, deps=ctx)
+                        return result.output
                     except Exception as e:
                         logger.exception(f"Task agent stream error: {e}")
                         raise
 
+                agent_task = asyncio.create_task(run_agent())
+
+                # Yield events from queue while agent runs
+                try:
+                    while not agent_task.done():
+                        try:
+                            # Check for events with timeout
+                            event = await asyncio.wait_for(
+                                event_queue.get(),
+                                timeout=0.1
+                            )
+                            yield event
+                        except asyncio.TimeoutError:
+                            # No events, check if agent is done
+                            continue
+
+                    # Drain remaining events
+                    while not event_queue.empty():
+                        yield event_queue.get_nowait()
+
+                    # Get final result
+                    response = await agent_task
+
+                    # Check if tools were called - if not, retry
+                    if ctx.tools_called == 0 and retries < self.MAX_NO_TOOL_RETRIES:
+                        logger.warning(f"No tools called, retrying ({retries + 1}/{self.MAX_NO_TOOL_RETRIES})")
+                        retries += 1
+                        continue
+
+                    if ctx.tools_called == 0:
+                        logger.warning("Agent responded without calling tools after all retries")
+
+                    # Success - break out of retry loop
+                    break
+
+                except Exception as e:
+                    logger.exception(f"Stream error: {e}")
+                    yield StreamEvent(
+                        event=StreamEventType.error,
+                        error_message=str(e)
+                    )
+                    # Cancel agent if still running
+                    if not agent_task.done():
+                        agent_task.cancel()
+                        try:
+                            await agent_task
+                        except asyncio.CancelledError:
+                            pass
+                    return
+
+            # Yield response in chunks for streaming feel
+            chunk_size = 100
+            for i in range(0, len(response), chunk_size):
+                chunk = response[i:i + chunk_size]
+                yield StreamEvent(
+                    event=StreamEventType.response,
+                    text=chunk
+                )
+                # Small delay for streaming effect
+                await asyncio.sleep(0.01)
+
+            # Signal completion
+            yield StreamEvent(
+                event=StreamEventType.done,
+                mode=mode.value
+            )
+
 
 # Create and register the singleton instance
 task_agent = TaskAgentImpl()
@@ -208,7 +388,7 @@ async def task_stream(
     prompt: str,
     working_dir: str | None = None,
     **kwargs: Any
-) -> AsyncIterator[str]:
-    """Run task execution with streaming."""
-    async for chunk in task_agent.run_stream(prompt, working_dir=working_dir, **kwargs):
-        yield chunk
+) -> AsyncIterator[StreamEvent]:
+    """Run task execution with event streaming."""
+    async for event in task_agent.run_stream(prompt, working_dir=working_dir, **kwargs):
+        yield event

webber-api/src/domains/agents/task/prompts.py

@@ -9,6 +9,11 @@ The Task agent is a full orchestrator that can:
 
 TASK_PLAN_MODE_PROMPT = """You are a codebase analysis and planning agent in READ-ONLY mode.
 
+CRITICAL RULE: You MUST call a tool BEFORE responding to ANY request.
+- NEVER answer from memory or assumptions
+- NEVER fabricate file structures, code, or content
+- If you respond without calling a tool first, YOUR ANSWER IS WRONG
+
 You can explore and analyze code but CANNOT modify files or execute write operations.
 
 AVAILABLE TOOLS (read-only):
@@ -24,12 +29,10 @@ Shell:
 Orchestration:
 - spawn_agent: Launch sub-agents for focused tasks (explore, plan only)
 
-WORKFLOW:
-1. Understand the request
-2. Explore the codebase to gather context
-3. Analyze code structure and patterns
-4. Create detailed implementation plans
-5. Return findings with actionable recommendations
+MANDATORY WORKFLOW:
+1. FIRST: Call a tool to gather real information
+2. THEN: Analyze the actual tool results
+3. FINALLY: Respond based only on what tools returned
 
 TOOL CALL EXAMPLES:
 
@@ -42,22 +45,25 @@ To search for a function:
 To check git status:
   Call bash_readonly with command="git status"
 
+To list directory contents:
+  Call bash_readonly with command="ls -la"
+
 To get deeper analysis:
   Call spawn_agent with agent_type="explore" and prompt="find authentication code"
 
 RULES:
-- ALWAYS use tools first, then analyze results
-- Never guess file contents - read them first
+- ALWAYS call a tool FIRST - no exceptions
+- Never guess or fabricate - only report what tools return
 - Be thorough in exploration
-- Provide specific file paths and line numbers in findings
+- Provide specific file paths and line numbers from tool results
 
 OUTPUT FORMAT:
 Structure your response with:
 
 ### Analysis
-- What was found
+- What was found (from tool results)
 - Key patterns identified
-- Relevant files
+- Relevant files (actual paths from tools)
 
 ### Recommendations
 - Suggested approach
@@ -67,6 +73,11 @@ Structure your response with:
 
 TASK_SYSTEM_PROMPT = """You are an autonomous task execution agent.
 
+CRITICAL RULE: You MUST call a tool BEFORE responding to ANY request.
+- NEVER answer from memory or assumptions
+- NEVER fabricate file structures, code, or content
+- If you respond without calling a tool first, YOUR ANSWER IS WRONG
+
 You have access to ALL tools including file editing, writing, and bash execution.
 You can also spawn sub-agents to help with complex tasks.
 
@@ -89,16 +100,21 @@ External:
 Orchestration:
 - spawn_agent: Launch sub-agents for focused tasks
 
-WORKFLOW:
-1. Understand the task requirements
-2. Break down into sub-tasks if complex
-3. Use spawn_agent for research (explore) or planning (plan)
-4. Execute implementation steps using write tools
-5. Validate changes (run tests if applicable)
-6. Return consolidated summary
+MANDATORY WORKFLOW:
+1. FIRST: Call a tool to gather real information
+2. THEN: Analyze the actual tool results
+3. Execute implementation using write tools if needed
+4. Validate changes (run tests if applicable)
+5. FINALLY: Return summary based only on what tools returned
 
 TOOL CALL EXAMPLES:
 
+To list directory contents:
+  Call bash_readonly with command="ls -la"
+
+To find all Python files:
+  Call glob_files with pattern="**/*.py"
+
 To spawn an Explore agent for research:
   Call spawn_agent with agent_type="explore" and prompt="find all config files"
 
@@ -124,8 +140,8 @@ GIT DISCIPLINE:
 - Run tests before committing
 
 RULES:
-- ALWAYS use tools first, then analyze results
-- Never guess file contents - read them first
+- ALWAYS call a tool FIRST - no exceptions
+- Never guess or fabricate - only report what tools return
 - Prefer edit_file over write_file for existing files
 - Use spawn_agent to keep context focused
 - Validate changes by running tests when applicable

webber-api/src/domains/agents/task/tools_streaming.py

@@ -0,0 +1,556 @@
+"""
+Tool registrations for the Task agent with event streaming.
+
+Same tools as tools.py but emit StreamEvent events for progress tracking.
+Tools push events to the context's event_queue when available.
+"""
+from pydantic_ai import Agent, RunContext
+
+from src.domains.agents.base import AgentContext
+from src.domains.agents.schemas import StreamEvent, StreamEventType
+from src.domains.agents.task.agent import TaskContext
+from src.domains.tools.file.read import ReadFileTool
+from src.domains.tools.file.glob import GlobFilesTool
+from src.domains.tools.file.edit import EditFileTool
+from src.domains.tools.file.write import WriteFileTool
+from src.domains.tools.search.grep import GrepContentTool
+from src.domains.tools.search.web import WebSearchTool
+from src.domains.tools.shell.bash import BashReadOnlyTool
+from src.domains.tools.shell.bash_full import BashTool
+
+
+def _emit_event(ctx: AgentContext, event: StreamEvent) -> None:
+    """Emit an event to the queue if available."""
+    if hasattr(ctx, 'event_queue') and ctx.event_queue is not None:
+        ctx.event_queue.put_nowait(event)
+
+
+def _track_tool_call(ctx: AgentContext) -> None:
+    """Increment tool call counter for retry logic."""
+    if hasattr(ctx, 'tools_called'):
+        ctx.tools_called += 1
+
+
+def _summarize_result(result: str, max_len: int = 80) -> str:
+    """Create a brief summary of a tool result."""
+    lines = result.strip().split('\n')
+    if len(lines) > 3:
+        return f"{len(lines)} lines"
+    if len(result) > max_len:
+        return result[:max_len] + "..."
+    return result.replace('\n', ' ')
+
+
+def _register_read_file(agent: Agent[TaskContext, str]) -> None:
+    """Register read_file tool with event streaming."""
+    @agent.tool
+    async def read_file(
+        ctx: RunContext[TaskContext],
+        file_path: str,
+        offset: int = 0,
+        limit: int = 2000
+    ) -> str:
+        """Read contents of a file with line numbers.
+
+        Args:
+            file_path: Absolute path to the file to read
+            offset: Line number to start from (0-based, default: 0)
+            limit: Maximum number of lines to read (default: 2000)
+
+        Returns:
+            File contents with line numbers, or error message.
+
+        IMPORTANT: Always use absolute paths. Read files before editing them.
+        """
+        _track_tool_call(ctx.deps)
+        _emit_event(ctx.deps, StreamEvent(
+            event=StreamEventType.tool_start,
+            tool="read_file",
+            args={"file_path": file_path, "offset": offset, "limit": limit}
+        ))
+
+        tool = ReadFileTool(allowed_paths=ctx.deps.allowed_paths)
+        result = await tool.execute(
+            file_path=file_path,
+            offset=offset,
+            limit=limit
+        )
+        result_str = result.to_string()
+
+        _emit_event(ctx.deps, StreamEvent(
+            event=StreamEventType.tool_done,
+            tool="read_file",
+            result_summary=_summarize_result(result_str)
+        ))
+
+        return result_str
+
+
+def _register_glob_files(agent: Agent[TaskContext, str]) -> None:
+    """Register glob_files tool with event streaming."""
+    @agent.tool
+    async def glob_files(
+        ctx: RunContext[TaskContext],
+        pattern: str,
+        path: str | None = None,
+        limit: int = 100
+    ) -> str:
+        """Find files matching a glob pattern.
+
+        Args:
+            pattern: Glob pattern (e.g., "**/*.py", "src/**/*.ts", "*.md")
+            path: Directory to search in (default: working directory)
+            limit: Maximum number of files to return (default: 100)
+
+        Returns:
+            List of absolute file paths, sorted by modification time (newest first).
+
+        Examples:
+            - "**/*.py" finds all Python files
+            - "src/**/*.ts" finds TypeScript files in src/
+            - "**/test_*.py" finds all test files
+        """
+        _track_tool_call(ctx.deps)
+        _emit_event(ctx.deps, StreamEvent(
+            event=StreamEventType.tool_start,
+            tool="glob_files",
+            args={"pattern": pattern, "path": path}
+        ))
+
+        tool = GlobFilesTool(allowed_paths=ctx.deps.allowed_paths)
+        search_path = path or ctx.deps.working_dir
+        result = await tool.execute(
+            pattern=pattern,
+            path=search_path,
+            limit=limit
+        )
+        result_str = result.to_string()
+
+        _emit_event(ctx.deps, StreamEvent(
+            event=StreamEventType.tool_done,
+            tool="glob_files",
+            result_summary=_summarize_result(result_str)
+        ))
+
+        return result_str
+
+
+def _register_grep_content(agent: Agent[TaskContext, str]) -> None:
+    """Register grep_content tool with event streaming."""
+    @agent.tool
+    async def grep_content(
+        ctx: RunContext[TaskContext],
+        pattern: str,
+        path: str | None = None,
+        file_glob: str | None = None,
+        context_lines: int = 0,
+        case_sensitive: bool = True
+    ) -> str:
+        """Search file contents using regex pattern.
+
+        Args:
+            pattern: Regex pattern to search for (Python re syntax)
+            path: Directory or file to search (default: working directory)
+            file_glob: Filter files by glob (e.g., "*.py", "*.ts")
+            context_lines: Lines of context before/after matches (default: 0)
+            case_sensitive: Case-sensitive search (default: True)
+
+        Returns:
+            Matching lines with file paths and line numbers.
+            Format: "filepath:line_num: content"
+        """
+        _track_tool_call(ctx.deps)
+        _emit_event(ctx.deps, StreamEvent(
+            event=StreamEventType.tool_start,
+            tool="grep_content",
+            args={"pattern": pattern, "path": path, "file_glob": file_glob}
+        ))
+
+        tool = GrepContentTool(allowed_paths=ctx.deps.allowed_paths)
+        search_path = path or ctx.deps.working_dir
+        result = await tool.execute(
+            pattern=pattern,
+            path=search_path,
+            file_glob=file_glob,
+            context_lines=context_lines,
+            case_sensitive=case_sensitive
+        )
+        result_str = result.to_string()
+
+        _emit_event(ctx.deps, StreamEvent(
+            event=StreamEventType.tool_done,
+            tool="grep_content",
+            result_summary=_summarize_result(result_str)
+        ))
+
+        return result_str
+
+
+def _register_bash_readonly(agent: Agent[TaskContext, str]) -> None:
+    """Register bash_readonly tool with event streaming."""
+    @agent.tool
+    async def bash_readonly(
+        ctx: RunContext[TaskContext],
+        command: str,
+        cwd: str | None = None,
+        timeout: int = 30
+    ) -> str:
+        """Execute a read-only bash command.
+
+        ALLOWED commands:
+        - File inspection: ls, find, cat, head, tail, wc, file, stat, tree, du
+        - Git (read-only): git status, git log, git diff, git show, git branch
+        - Text processing: grep, awk, sed (read-only), sort, uniq
+        - System info: pwd, whoami, hostname, which
+
+        FORBIDDEN:
+        - File modification (rm, mv, cp, mkdir, touch)
+        - Redirects (>, >>)
+        - Command chaining (&&, ||, ;)
+        - Network (curl, wget)
+
+        Args:
+            command: The bash command to execute
+            cwd: Working directory (default: agent working directory)
+            timeout: Timeout in seconds (default: 30)
+        """
+        _track_tool_call(ctx.deps)
+        _emit_event(ctx.deps, StreamEvent(
+            event=StreamEventType.tool_start,
+            tool="bash_readonly",
+            args={"command": command}
+        ))
+
+        tool = BashReadOnlyTool(allowed_paths=ctx.deps.allowed_paths)
+        working_dir = cwd or ctx.deps.working_dir
+        result = await tool.execute(
+            command=command,
+            cwd=working_dir,
+            timeout=min(timeout, ctx.deps.timeout_seconds)
+        )
+        result_str = result.to_string()
+
+        _emit_event(ctx.deps, StreamEvent(
+            event=StreamEventType.tool_done,
+            tool="bash_readonly",
+            result_summary=_summarize_result(result_str)
+        ))
+
+        return result_str
+
+
+def _register_spawn_agent(agent: Agent[TaskContext, str], readonly_only: bool = False) -> None:
+    """Register spawn_agent tool with event streaming."""
+    @agent.tool
+    async def spawn_agent(
+        ctx: RunContext[TaskContext],
+        agent_type: str,
+        prompt: str,
+        working_dir: str | None = None
+    ) -> str:
+        """Spawn a sub-agent to handle a focused task.
+
+        Use this to offload work to specialized agents:
+        - "explore": Fast codebase searches and analysis (read-only)
+        - "plan": Design implementation strategies (read-only)
+
+        Args:
+            agent_type: Type of agent to spawn ("explore" or "plan")
+            prompt: Task description for the sub-agent
+            working_dir: Working directory for the sub-agent (default: current)
+
+        Returns:
+            Sub-agent's consolidated response.
+
+        Examples:
+            - spawn_agent(agent_type="explore", prompt="find all test files")
+            - spawn_agent(agent_type="plan", prompt="design user auth feature")
+
+        IMPORTANT:
+        - Use sub-agents to keep context focused and efficient
+        - Explore agent for research, Plan agent for design
+        - Cannot spawn nested Task agents (recursion risk)
+        """
+        from src.domains.agents.base import get_agent
+
+        _track_tool_call(ctx.deps)
+        _emit_event(ctx.deps, StreamEvent(
+            event=StreamEventType.tool_start,
+            tool="spawn_agent",
+            args={"agent_type": agent_type, "prompt": prompt[:50] + "..." if len(prompt) > 50 else prompt}
+        ))
+
+        # Validate agent type
+        allowed_types = ["explore", "plan"]
+        if agent_type not in allowed_types:
+            if agent_type == "task":
+                result = "Error: Cannot spawn nested Task agents (recursion risk)"
+            else:
+                result = f"Error: Unknown agent type '{agent_type}'. Allowed: {allowed_types}"
+            _emit_event(ctx.deps, StreamEvent(
+                event=StreamEventType.tool_done,
+                tool="spawn_agent",
+                result_summary=result
+            ))
+            return result
+
+        sub_agent = get_agent(agent_type)
+        if not sub_agent:
+            result = f"Error: Agent '{agent_type}' not found in registry"
+            _emit_event(ctx.deps, StreamEvent(
+                event=StreamEventType.tool_done,
+                tool="spawn_agent",
+                result_summary=result
+            ))
+            return result
+
+        try:
+            result = await sub_agent.run(
+                prompt=prompt,
+                working_dir=working_dir or ctx.deps.working_dir,
+                allowed_paths=ctx.deps.allowed_paths,
+            )
+            _emit_event(ctx.deps, StreamEvent(
+                event=StreamEventType.tool_done,
+                tool="spawn_agent",
+                result_summary=_summarize_result(result)
+            ))
+            return result
+        except Exception as e:
+            result = f"Sub-agent error: {e}"
+            _emit_event(ctx.deps, StreamEvent(
+                event=StreamEventType.tool_done,
+                tool="spawn_agent",
+                result_summary=result
+            ))
+            return result
+
+
+def register_readonly_tools_streaming(agent: Agent[TaskContext, str]) -> None:
+    """
+    Register read-only tools with event streaming.
+
+    Used in plan mode. Includes:
+    - read_file, glob_files, grep_content, bash_readonly
+    - spawn_agent (restricted to explore/plan)
+    """
+    _register_read_file(agent)
+    _register_glob_files(agent)
+    _register_grep_content(agent)
+    _register_bash_readonly(agent)
+    _register_spawn_agent(agent, readonly_only=True)
+
+
+def register_task_tools_streaming(agent: Agent[TaskContext, str]) -> None:
+    """
+    Register all tools with event streaming.
+
+    Includes:
+    - Read-only tools: read_file, glob_files, grep_content, bash_readonly
+    - Write tools: edit_file, write_file, bash
+    - External: web_search
+    - Orchestration: spawn_agent
+    """
+    # Register read-only tools via helpers
+    _register_read_file(agent)
+    _register_glob_files(agent)
+    _register_grep_content(agent)
+    _register_bash_readonly(agent)
+
+    # === Write tools ===
+
+    @agent.tool
+    async def edit_file(
+        ctx: RunContext[TaskContext],
+        file_path: str,
+        old_string: str,
+        new_string: str,
+        replace_all: bool = False
+    ) -> str:
+        """Make targeted edits to a file using find-and-replace.
+
+        Args:
+            file_path: Absolute path to the file to edit
+            old_string: The exact text to find and replace (must exist in file)
+            new_string: The replacement text
+            replace_all: If True, replace all occurrences. If False (default),
+                         old_string must be unique (appear exactly once).
+
+        Returns:
+            Success message with diff preview, or error.
+
+        IMPORTANT:
+        - old_string must exactly match file content (including whitespace)
+        - By default, old_string must appear exactly once (for safety)
+        - Always read the file first to verify exact content before editing
+        """
+        _track_tool_call(ctx.deps)
+        _emit_event(ctx.deps, StreamEvent(
+            event=StreamEventType.tool_start,
+            tool="edit_file",
+            args={"file_path": file_path, "replace_all": replace_all}
+        ))
+
+        tool = EditFileTool(allowed_paths=ctx.deps.allowed_paths)
+        result = await tool.execute(
+            file_path=file_path,
+            old_string=old_string,
+            new_string=new_string,
+            replace_all=replace_all
+        )
+        result_str = result.to_string()
+
+        _emit_event(ctx.deps, StreamEvent(
+            event=StreamEventType.tool_done,
+            tool="edit_file",
+            result_summary=_summarize_result(result_str)
+        ))
+
+        return result_str
+
+    @agent.tool
+    async def write_file(
+        ctx: RunContext[TaskContext],
+        file_path: str,
+        content: str
+    ) -> str:
+        """Create a new file or overwrite an existing file.
+
+        Args:
+            file_path: Absolute path to the file to create/write
+            content: The content to write to the file
+
+        Returns:
+            Success message with file path and size.
+
+        IMPORTANT:
+        - Parent directory must exist (use bash mkdir first if needed)
+        - For editing existing files, prefer edit_file instead
+        - Will overwrite existing files without confirmation
+        """
+        _track_tool_call(ctx.deps)
+        _emit_event(ctx.deps, StreamEvent(
+            event=StreamEventType.tool_start,
+            tool="write_file",
+            args={"file_path": file_path, "content_length": len(content)}
+        ))
+
+        tool = WriteFileTool(allowed_paths=ctx.deps.allowed_paths)
+        result = await tool.execute(
+            file_path=file_path,
+            content=content
+        )
+        result_str = result.to_string()
+
+        _emit_event(ctx.deps, StreamEvent(
+            event=StreamEventType.tool_done,
+            tool="write_file",
+            result_summary=_summarize_result(result_str)
+        ))
+
+        return result_str
+
+    @agent.tool
+    async def bash(
+        ctx: RunContext[TaskContext],
+        command: str,
+        cwd: str | None = None,
+        timeout: int = 60
+    ) -> str:
+        """Execute a bash command with write capabilities.
+
+        ALLOWED:
+        - File operations: ls, find, mkdir, touch, cp, mv, rm (single files)
+        - Git (full): git add, git commit, git checkout, git merge, git pull
+        - Python: python, pip install, pytest, mypy, ruff
+        - Text processing: grep, awk, sed, sort
+        - Command chaining: && and || are allowed
+
+        FORBIDDEN:
+        - sudo, su (privilege escalation)
+        - Network: curl, wget, ssh, scp, rsync
+        - Dangerous: rm -rf, chmod 777, dd, mkfs
+
+        Args:
+            command: The bash command to execute
+            cwd: Working directory (default: agent working directory)
+            timeout: Timeout in seconds (default: 60)
+
+        Examples:
+            - "mkdir -p src/utils" creates directory
+            - "git add . && git commit -m 'fix: bug'" commits changes
+            - "pytest tests/ -v" runs tests
+        """
+        _track_tool_call(ctx.deps)
+        _emit_event(ctx.deps, StreamEvent(
+            event=StreamEventType.tool_start,
+            tool="bash",
+            args={"command": command}
+        ))
+
+        tool = BashTool(allowed_paths=ctx.deps.allowed_paths)
+        working_dir = cwd or ctx.deps.working_dir
+        result = await tool.execute(
+            command=command,
+            cwd=working_dir,
+            timeout=min(timeout, ctx.deps.timeout_seconds)
+        )
+        result_str = result.to_string()
+
+        _emit_event(ctx.deps, StreamEvent(
+            event=StreamEventType.tool_done,
+            tool="bash",
+            result_summary=_summarize_result(result_str)
+        ))
+
+        return result_str
+
+    # === External tools ===
+
+    @agent.tool
+    async def web_search(
+        ctx: RunContext[TaskContext],
+        query: str,
+        num_results: int = 5,
+        categories: str | None = None
+    ) -> str:
+        """Search the web for current information.
+
+        Args:
+            query: Search query (e.g., "Python 3.12 new features")
+            num_results: Number of results to return (1-10, default: 5)
+            categories: Optional category filter ("general", "it", "news", "science")
+
+        Returns:
+            Search results with titles, URLs, and snippets.
+
+        Use this for:
+        - Current events or recent information
+        - Documentation updates
+        - Technical references with URLs
+        """
+        _track_tool_call(ctx.deps)
+        _emit_event(ctx.deps, StreamEvent(
+            event=StreamEventType.tool_start,
+            tool="web_search",
+            args={"query": query}
+        ))
+
+        tool = WebSearchTool()
+        result = await tool.execute(
+            query=query,
+            num_results=num_results,
+            categories=categories
+        )
+        result_str = result.to_string()
+
+        _emit_event(ctx.deps, StreamEvent(
+            event=StreamEventType.tool_done,
+            tool="web_search",
+            result_summary=_summarize_result(result_str)
+        ))
+
+        return result_str
+
+    # === Orchestration tools ===
+    _register_spawn_agent(agent, readonly_only=False)

webber-api/tests/test_agents_api.py

@@ -276,19 +276,17 @@ class TestPermissionModeIntegration:
     @pytest.mark.anyio
     async def test_stream_with_plan_mode(self, auth_client):
         """Test streaming agent with plan mode passes through correctly."""
-        async def mock_stream(*args, **kwargs):
-            yield "chunk1"
-            yield "chunk2"
+        from src.domains.agents.schemas import StreamEvent, StreamEventType
 
-        with patch("src.domains.agents.task.agent.TaskAgentImpl._get_agent_for_mode") as mock_get_agent:
-            mock_agent = MagicMock()
-            mock_agent.run_stream = MagicMock(return_value=MagicMock(
-                __aenter__=AsyncMock(return_value=MagicMock(
-                    stream_text=lambda: mock_stream()
-                )),
-                __aexit__=AsyncMock(return_value=None)
-            ))
-            mock_get_agent.return_value = mock_agent
+        async def mock_event_stream(*args, **kwargs):
+            """Mock event-based stream."""
+            yield StreamEvent(event=StreamEventType.thinking, message="Starting...")
+            yield StreamEvent(event=StreamEventType.response, text="chunk1")
+            yield StreamEvent(event=StreamEventType.response, text="chunk2")
+            yield StreamEvent(event=StreamEventType.done, mode="plan")
+
+        with patch("src.domains.agents.task.agent.TaskAgentImpl.run_stream") as mock_run_stream:
+            mock_run_stream.return_value = mock_event_stream()
 
             response = await auth_client.post(
                 "/agents/stream",
@@ -302,7 +300,7 @@ class TestPermissionModeIntegration:
 
             assert response.status_code == 200
             assert response.headers["content-type"] == "text/event-stream; charset=utf-8"
-            mock_get_agent.assert_called_once_with(PermissionMode.plan)
+            mock_run_stream.assert_called_once()
 
 
 class TestAgentMethodSignatures:

webber-cli/webber_cli/client.py

@@ -26,6 +26,41 @@ class PermissionMode(str, Enum):
     auto_accept = "auto_accept"
 
 
+class StreamEventType(str, Enum):
+    """Event types for structured agent streaming."""
+    tool_start = "tool_start"
+    tool_done = "tool_done"
+    thinking = "thinking"
+    response = "response"
+    error = "error"
+    done = "done"
+    chunk = "chunk"  # Legacy text chunk
+
+
+@dataclass
+class StreamEvent:
+    """
+    Structured streaming event from agent execution.
+
+    Different event types carry different data:
+    - tool_start: tool, args
+    - tool_done: tool, result_summary
+    - thinking: message
+    - response: text
+    - error: error_message
+    - done: mode
+    - chunk: text (legacy)
+    """
+    event: StreamEventType
+    tool: str | None = None
+    args: dict | None = None
+    result_summary: str | None = None
+    message: str | None = None
+    text: str | None = None
+    error_message: str | None = None
+    mode: str | None = None
+
+
 @dataclass
 class AgentResponse:
     """Response from agent execution."""
@@ -202,7 +237,7 @@ class WebberClient:
         prompt: str,
         working_dir: str = ".",
         mode: PermissionMode = PermissionMode.default,
-    ) -> AsyncIterator[str]:
+    ) -> AsyncIterator[StreamEvent]:
         """
         Run an agent with streaming response.
 
@@ -213,7 +248,7 @@ class WebberClient:
             mode: Permission mode controlling tool access
 
         Yields:
-            Text chunks as they arrive
+            StreamEvent objects as they arrive
         """
         # Use a fresh client for streaming with longer timeout
         async with httpx.AsyncClient(
@@ -235,13 +270,30 @@ class WebberClient:
                     if line.startswith("data: "):
                         try:
                             data = json.loads(line[6:])
-                            event = data.get("event")
-                            if event == "chunk":
-                                yield data.get("data", "")
-                            elif event == "error":
-                                raise Exception(data.get("data", "Unknown error"))
-                            elif event == "done":
+                            event_type = data.get("event")
+
+                            # Parse event type
+                            try:
+                                evt_type = StreamEventType(event_type)
+                            except ValueError:
+                                continue  # Unknown event type
+
+                            # Build StreamEvent from response data
+                            yield StreamEvent(
+                                event=evt_type,
+                                tool=data.get("tool"),
+                                args=data.get("args"),
+                                result_summary=data.get("result_summary"),
+                                message=data.get("message"),
+                                text=data.get("text") or data.get("data"),  # 'data' for legacy chunk
+                                error_message=data.get("error_message") or data.get("data"),
+                                mode=data.get("mode"),
+                            )
+
+                            # Stop on done or error
+                            if evt_type in (StreamEventType.done, StreamEventType.error):
                                 break
+
                         except json.JSONDecodeError:
                             continue
 

webber-cli/webber_cli/main.py

@@ -21,7 +21,7 @@ from rich.markdown import Markdown
 from rich.panel import Panel
 from rich.prompt import Confirm
 
-from webber_cli.client import WebberClient, PermissionMode
+from webber_cli.client import WebberClient, PermissionMode, StreamEvent, StreamEventType
 from webber_cli.config import get_config, init_config, get_config_path, CONFIG_FILE
 from webber_cli.theme import get_console, get_theme
 
@@ -471,15 +471,56 @@ async def _chat_loop(
                 console.print()
 
                 if stream:
-                    # Stream response in real-time, buffering for persistence
+                    # Stream response with structured events
                     response_chunks: list[str] = []
+                    current_tool: str | None = None
                     try:
-                        async for chunk in client.run_agent_stream(
+                        async for event in client.run_agent_stream(
                             agent_type, user_input, working_dir, current_mode
                         ):
-                            sys.stdout.write(chunk)
+                            if event.event == StreamEventType.thinking:
+                                # Show thinking status
+                                console.print(f"[dim]{event.message or 'Thinking...'}[/]")
+
+                            elif event.event == StreamEventType.tool_start:
+                                # Show tool starting
+                                current_tool = event.tool
+                                args_display = ""
+                                if event.args:
+                                    # Format key args for display
+                                    key_args = []
+                                    for k, v in list(event.args.items())[:2]:
+                                        v_str = str(v)[:40] + "..." if len(str(v)) > 40 else str(v)
+                                        key_args.append(f"{k}={v_str}")
+                                    args_display = f" ({', '.join(key_args)})"
+                                console.print(f"[info]→ {event.tool}[/]{args_display}", end="")
+
+                            elif event.event == StreamEventType.tool_done:
+                                # Show tool completed
+                                result = event.result_summary or "done"
+                                console.print(f" [success]✓[/] [dim]{result}[/]")
+                                current_tool = None
+
+                            elif event.event == StreamEventType.response:
+                                # Stream response text
+                                if event.text:
+                                    sys.stdout.write(event.text)
                                     sys.stdout.flush()
-                            response_chunks.append(chunk)
+                                    response_chunks.append(event.text)
+
+                            elif event.event == StreamEventType.chunk:
+                                # Legacy text chunk (for other agents)
+                                if event.text:
+                                    sys.stdout.write(event.text)
+                                    sys.stdout.flush()
+                                    response_chunks.append(event.text)
+
+                            elif event.event == StreamEventType.error:
+                                console.print(f"\n[error]Error:[/] {event.error_message}")
+
+                            elif event.event == StreamEventType.done:
+                                pass  # Stream complete
+
                         console.print()  # Newline after streaming
 
                         # Save messages to conversation if we have a session
@@ -672,13 +713,23 @@ async def _explore(
         console.print()
 
         if stream:
-            # Stream response in real-time
+            # Stream response with structured events
             try:
-                async for chunk in client.run_agent_stream(
+                async for event in client.run_agent_stream(
                     "task", query, working_dir, mode
                 ):
-                    sys.stdout.write(chunk)
+                    if event.event == StreamEventType.thinking:
+                        console.print(f"[dim]{event.message or 'Thinking...'}[/]")
+                    elif event.event == StreamEventType.tool_start:
+                        console.print(f"[info]→ {event.tool}[/]", end="")
+                    elif event.event == StreamEventType.tool_done:
+                        console.print(f" [success]✓[/] [dim]{event.result_summary or 'done'}[/]")
+                    elif event.event in (StreamEventType.response, StreamEventType.chunk):
+                        if event.text:
+                            sys.stdout.write(event.text)
                             sys.stdout.flush()
+                    elif event.event == StreamEventType.error:
+                        console.print(f"\n[error]Error:[/] {event.error_message}")
                 console.print()  # Newline after streaming
             except Exception as e:
                 console.print(f"\n[error]Stream error:[/] {e}")

webber-sandbox/README.md

@@ -0,0 +1 @@
+This directory contains an application to analyse for the webber coding agent to see what it reports this directory is for. It is to test the llm.