Missing Human-in-the-Loop Tracing Support

[prajwalun]

Title: Missing tracing support for human-in-the-loop agent workflows

Description:
My resume agent pauses for human input when it detects potential fabrication risks or needs clarification. The current tracing API doesn't handle these pause/resume scenarios properly, making it difficult to track the complete workflow.

Problem:
When my agent pauses for user clarification, the traces break or become disconnected:

if fabrication_risk['needs_clarification']:
    # How do we trace this pause for user input?
    request_clarification()
    continue  # Resume processing after user input

This creates fragmented traces that are hard to debug and monitor.

Expected Behavior:
Tracing should support pause/resume patterns for human-in-the-loop workflows:

# Option 1: Pause/resume context managers
with judgment.pause_for_human_input(reason="clarification_needed"):
    user_response = await get_user_clarification()
    
with judgment.resume_from_human_input(previous_span_id=span_id):
    continue_processing(user_response)

# Option 2: Session continuation
session_id = judgment.get_current_session_id()
# ... pause for user input ...
judgment.resume_session(session_id, user_input=user_response)

Current Behavior:

• Traces become fragmented when agent execution pauses for human input
• No way to correlate spans before and after human interaction
• Difficult to debug and monitor complete workflows
• Session continuity is lost

Impact:
This makes it impossible to properly monitor and debug human-in-the-loop agent workflows, which are common in production systems.

[linear]

JUD-1443 Missing Human-in-the-Loop Tracing Support

[hegu-1]

This issue feels like it needs a first-class continuity model, not only a span-linking patch.

For HITL agent workflows, the pause is not just an observability gap. It is a state transition where the agent is intentionally handing judgment back to a human and later resuming with new authority. I would probably model three explicit records:

• pause snapshot: reason, active task/thread, risk class, pending decision, current trace/span id
• human return: input/provenance, whether it changes the task boundary, whether it authorizes continuation
• resume edge: previous span/session id, resumed state hash, and the next agent step allowed after the human response

That gives you a trace tree that can answer more than “what happened?” It can answer “what was unresolved when the agent stopped, what did the human decide, and what is the resumed execution allowed to rely on?”

Related reference if useful: I have been experimenting with this as continuity / snapshot / trace in a Personal AI OS skeleton: https://github.com/hegu-1/personal-ai-os. The repo is early, but the framing maps pretty directly to HITL pause/resume tracing.