Plan audit entry per step

Plan Audit Entry per Loop Step

When to use

Your agent has a planner that runs multiple times (one per loop iteration). When it misbehaves, you want to know exactly what was decided, with what budget, and whether the planner rerouted to a different source family. Logging via print makes this lossy; you want a typed dict in state.

How it works

• A PlanAudit TypedDict declares the audit shape: loop, tool_budget, planned_count, rerouted, optional reroute_reason.
• Every planning node populates it before returning.
• Downstream nodes (e.g. investigate) read the audit so summaries can include "loop 2: planner re-routed from grafana to datadog because Grafana returned no logs".

Example

class PlanAudit(TypedDict, total=False):
    loop: int
    tool_budget: int
    planned_count: int
    rerouted: bool
    reroute_reason: str

def node_plan_actions(state):
    loop_count = state.get("investigation_loop_count", 0)
    plan, available_sources, available_action_names, _, rerouted, reroute_reason = build_plan_actions(...)
    audit_entry: PlanAudit = {
        "loop": loop_count,
        "tool_budget": input_data.tool_budget,
        "planned_count": len(plan.actions if plan else []),
        "rerouted": rerouted,
    }
    if rerouted:
        audit_entry["reroute_reason"] = reroute_reason
    return {
        "planned_actions": plan.actions if plan else [],
        "plan_rationale": plan.rationale if plan else "",
        "available_sources": available_sources,
        "available_action_names": available_action_names,
        "plan_audit": audit_entry,
    }

Gotchas

• Use TypedDict(total=False) so you can omit optional fields without forcing a sentinel value.
• Don't accumulate the entire history in state — keep only the most recent plan_audit. The full history can be reconstructed from LangSmith traces or LangGraph's checkpoint table.
• Surface the audit in the final report so users can self-debug ("agent looped 3 times because Grafana returned no logs" is more actionable than "couldn't find root cause").