Add agent attribution to verbose logs for parallel and for-each execution

[e-s-gh]

Problem

When Conductor runs parallel or for-each agent groups, verbose log output from all concurrent agents is interleaved with no identifying information. Tool calls, reasoning steps, and processing indicators are indistinguishable:

    ├─ 🔧 view
    │     args: {'path': '/some/file.py'}
    │  ⏳ Processing...
    ├─ 🔧 grep
    │     args: {'pattern': 'some_function', ...}
    │  ⏳ Processing...
    ├─ 🔧 view
    │     args: {'path': '/another/file.py'}

When 3+ agents run concurrently, it's impossible to determine which agent made which tool call. This makes debugging parallel workflows impractical and log files ungrepable for per-agent activity.

The problem is especially acute with for_each groups where every item shares the same agent definition name.

Proposed Solution

1. Pass agent name through to verbose logging (copilot.py)

Add an optional agent_name parameter to _send_and_wait and _log_event_verbose. When present, prefix all verbose output with [agent_name]:

# _send_and_wait signature
async def _send_and_wait(
    self, session, prompt, verbose_enabled, full_enabled,
    interrupt_signal=None, event_callback=None,
    agent_name: str | None = None,      # NEW
) -> SDKResponse:

# _log_event_verbose signature
def _log_event_verbose(
    self, event_type, event, full_mode,
    agent_name: str | None = None,      # NEW
) -> None:

Inside _log_event_verbose, every event type that builds Rich Text() output conditionally prepends the agent tag:

if agent_name:
    text.append(f"[{agent_name}] ", style="magenta")

Applies to: tool.execution_start, tool.execution_complete, assistant.reasoning, subagent.started, subagent.completed, assistant.turn_start

All call sites of _send_and_wait pass agent_name=agent.name.

2. Qualify for-each agent names with item keys (workflow.py)

In _execute_for_each_group → execute_single_item, create a shallow copy of the agent with a qualified name before execution:

qualified_agent = for_each_group.agent.model_copy(
    update={"name": f"{for_each_group.agent.name}[{key}]"}
)
output = await executor.execute(qualified_agent, agent_context, ...)

This uses Pydantic's model_copy() — only the name field changes. The key comes from the for-each item (index or key_by value).

Static parallel groups don't need this change — each agent already has a unique name.

Expected Result

    ├─ [analyzer[item_a]] 🔧 view
    │     args: {'path': '/some/file.py'}
    │  [analyzer[item_a]] ⏳ Processing...
    ├─ [analyzer[item_b]] 🔧 grep
    │     args: {'pattern': 'some_function', ...}
    │  [analyzer[item_b]] 💭 Reasoning about the results...
    ├─ [analyzer[item_c]] 🔧 view
    │     args: {'path': '/another/file.py'}

Log files become greppable per-agent:

# Filter to one specific agent's activity
grep "\[analyzer\[item_a\]\]" log-file

# Count tool calls per agent
grep -c "\[analyzer\[.*\]\] 🔧" log-file

Scope

File	Change
src/conductor/providers/copilot.py	Add agent_name param to _send_and_wait and _log_event_verbose; prefix verbose events
src/conductor/engine/workflow.py	For-each items use model_copy() to create qualified agent name

Backward Compatibility

Fully backward compatible:

• agent_name defaults to None — no tag emitted for sequential agents
• model_copy() creates a new object; original agent definition is untouched
• No schema changes, no YAML syntax changes, no CLI changes
• Existing log parsing patterns still work — the tag is prepended, not replacing content