Claude provider: event_callback not used — web dashboard shows no intermediate activity

[franklixuefei]

Feature Request / Bug

The Claude provider's execute() method accepts an event_callback parameter but never invokes it. This means the web dashboard (--web / --web-bg) shows only agent input and output, with no visibility into intermediate activity (tool calls, reasoning, MCP interactions).

The Copilot provider correctly forwards events via _forward_event(), emitting agent_reasoning, agent_tool_start, agent_tool_complete, agent_turn_start, and agent_message events that the web dashboard renders in real-time.

Root Cause

In src/conductor/providers/claude.py:

1. execute() (line ~395) accepts event_callback: EventCallback | None = None
2. But _execute_with_retry() (line ~420) is called without forwarding event_callback
3. The agentic loop in _execute_agentic_loop() has all the information needed (tool names, tool results, iteration counts, API responses) but never calls the callback

Expected Behavior

The web dashboard should show real-time intermediate activity for Claude provider agents:

• Tool call start — which MCP tool is being called and with what arguments
• Tool call complete — tool result (truncated)
• Turn/iteration progress — which iteration of the agentic loop
• Text content — any text blocks in intermediate responses

This matches the Copilot provider's behavior (see _forward_event() at line ~1161 of copilot.py).

Actual Behavior

• Web dashboard shows agent name, input prompt, and final output only
• No intermediate activity visible — tool calls, file reads, MCP interactions are invisible
• Users have to tail -f the log file to see what's happening

Suggested Fix

1. Pass event_callback through execute() → _execute_with_retry() → _execute_with_retry_inner() → _execute_agentic_loop()

2. In _execute_agentic_loop(), emit events at key points:

# At start of each iteration:
if event_callback:
    event_callback("agent_turn_start", {"turn": iteration})

# When MCP tool calls are identified:
for tool_use in mcp_tool_uses:
    if event_callback:
        event_callback("agent_tool_start", {
            "tool_name": tool_use.name,
            "arguments": str(dict(tool_use.input))[:500] if hasattr(tool_use, "input") else None,
        })

# After each MCP tool completes:
if event_callback:
    event_callback("agent_tool_complete", {
        "tool_name": tool_use.name,
        "result": str(result)[:500] if result else None,
    })

# For text content in responses:
for block in response.content:
    if hasattr(block, "type") and block.type == "text" and block.text:
        if event_callback:
            event_callback("agent_message", {"content": block.text})

Relationship to #37 and #38

This is independent of the tool_filter (#37) and parse recovery (#38) bugs, but becomes visible only after those are fixed — with MCP tools actually working, the lack of intermediate events is noticeable.

Environment

• Conductor v0.1.0 (commit 391ddaf)
• Python 3.12.13

Impact

Medium — The web dashboard is significantly less useful with the Claude provider compared to Copilot. Users cannot monitor agent progress in real-time and must rely on log files instead.