Managed Agents: emit ContextTraceEmitter events so langextract/langfuse traces are non-empty

[MervinPraison]

Problem

Managed agents emit no trace events. The langextract + langfuse bridges landed in #1420 listen to ContextTraceEmitter (agent_start / tool_call_* / llm_response / agent_end), but both AnthropicManagedAgent._execute_sync (src/praisonai/praisonai/integrations/managed_agents.py:377-406) and LocalManagedAgent._execute_sync (src/praisonai/praisonai/integrations/managed_local.py:548-577) bypass this pipeline entirely. Consequence: a run with praisonai --observe langextract managed run "…" produces an empty HTML.

Acceptance criteria

• AnthropicManagedAgent._execute_sync emits agent_start before SSE stream, tool_call_start/end around each agent.tool_use event, llm_response when aggregated text is available, agent_end on session.status_idle.
• LocalManagedAgent._execute_sync emits the same five event types around agent.chat(...) (already partially covered via inner Agent but managed-level agent_start/end must be explicit).
• praisonai --observe langextract managed run "Write a haiku" (or --observe langfuse) produces a non-empty HTML / Langfuse trace with the managed agent visible.
• Zero-overhead when no emitter is installed (get_context_emitter() returns the disabled singleton — see praisonaiagents/trace/context_events.py).
• Unit tests with a fake ContextTraceSinkProtocol confirm event sequence and counts.
• Real agentic test (gated): Agent(backend=ManagedAgent()).start("Say hi") with an installed ContextListSink shows ≥ 2 events.

Implementation plan

1. In both _execute_sync methods: emitter = get_context_emitter() at the top (cheap), then emitter.agent_start(self._cfg.get("name","Agent"), {"input": prompt, "goal": ...}) before sending the user message; matching emitter.agent_end(...) in finally:.
2. In _process_events of AnthropicManagedAgent: on agent.tool_use call emitter.tool_call_start(agent_name, tool_name, input); no direct "end" event from Anthropic, so emit tool_call_end synthetically right after (or on the next event) with duration_ms from a local timer.
3. On agent.message aggregated text: call emitter.llm_response(agent_name, response_content=joined_text) once per turn.
4. For LocalManagedAgent, the inner Agent.chat() already emits context events via chat_mixin; still add managed-level agent_start/end to mark the boundary.
5. No core changes needed — all emission happens from the wrapper.

Files

Modify:

• src/praisonai/praisonai/integrations/managed_agents.py — _execute_sync, _process_events.
• src/praisonai/praisonai/integrations/managed_local.py — _execute_sync.

Tests:

• src/praisonai-agents/tests/managed/test_managed_trace_events.py (new) — uses ContextListSink + trace_context.

Invariants

• Protocol-driven core untouched.
• No new deps; emission is opt-in via get_context_emitter().
• Backward-compatible.

References

• Langextract context bridge (analogous wrapper-side wiring): fix(langextract): bridge ContextTraceEmitter so real agent events produce a non-empty trace #1420.
• ContextTraceEmitter API: src/praisonai-agents/praisonaiagents/trace/context_events.py:352-900.
• Analysis: meta issue Meta: Managed Agents follow-ups — safety, observability, tests, CLI parity #1425.

cc @claude — per .windsurf/workflows/e2e-analysis-issue-pr-merge.md.

[praisonai-triage-agent]

Claude finished @MervinPraison's task —— View job • claude/issue-1427-20260416-2118 • [Create PR ➔](https://github.com/MervinPraison/PraisonAI/compare/main...claude/issue-1427-20260416-2118?quick_pull=1&title=Issue%20%231427%3A%20Changes%20from%20Claude&body=This%20PR%20addresses%20issue%20%231427%0A%0AGenerated%20with%20%5BClaude%20Code%5D(https%3A%2F%2Fclaude.ai%2Fcode)

---

Implementing Managed Agents Trace Events ✅

Successfully implemented ContextTraceEmitter events for Managed Agents so that langextract/langfuse traces are non-empty.

Completed Tasks:

• Read AGENTS.md for architecture guidelines
• Understand current trace event system implementation
• Review managed agent implementations
• Implement trace events in AnthropicManagedAgent._execute_sync
• Implement trace events in LocalManagedAgent._execute_sync
• Create unit tests for trace events
• Run unit tests to verify implementation
• Test real agentic example
• Commit and push changes
• Create pull request

Summary

Problem: Managed agents bypassed the ContextTraceEmitter pipeline entirely, causing praisonai --observe langextract managed run "..." to produce empty HTML traces.

Solution: Added trace event emission to both AnthropicManagedAgent._execute_sync and LocalManagedAgent._execute_sync:

Changes Made

AnthropicManagedAgent:

• ✅ agent_start before SSE stream
• ✅ tool_call_start/tool_call_end around each agent.tool_use event
• ✅ llm_response when aggregated text is available
• ✅ agent_end on session.status_idle

LocalManagedAgent:

• ✅ agent_start/agent_end around agent.chat() calls
• ✅ llm_response for response content
• ✅ Inner Agent.chat() events preserved via existing chat_mixin

Zero-overhead: Uses get_context_emitter() which returns disabled singleton when no emitter is installed.

Comprehensive testing: Unit tests verify event sequence and counts. Real agentic test confirms ≥ 2 events are emitted.

Verification

✅ Both managed agent implementations now emit proper trace events
✅ Zero performance impact when no emitter is installed
✅ Backward compatible - existing code unchanged
✅ Unit tests pass - correct event sequence verified
✅ Integration tested with mock and real scenarios

Pull Request: #1433

The issue is now resolved - managed agents will produce non-empty traces for langextract/langfuse observability tools.

--- | Branch: claude/issue-1427-20260416-2118 | Pull Request #1433