Sessions

The Sessions view shows end-to-end session timelines for agent interactions. A session captures everything that happened between an agent and the systems it interacted with — gateway requests, guardrail evaluations, tool calls, and any associated browser activity.

Navigate to Highflame StudioObservatorySessions.


Session list

The session list shows recent sessions in reverse chronological order:

Column
Description

Session ID

Unique identifier for the session

Started

When the session began

Duration

Total session duration

Agent

The agent identity (ZeroID) associated with the session

User

The human user who initiated the session, if known

Turns

Number of conversation turns in the session

Events

Number of recorded events in the session

Violations

Number of threat events within the session

Max severity

Highest severity threat seen in the session

Status

Active or completed

Filter sessions by time range, agent identity, user, source product, or violation count. Sessions with violations are highlighted.


Session timeline

Click any session to expand its timeline. The timeline shows all events in the session in chronological order:

Event types in the timeline

Event type
Source
Description

Gateway request

Agent Gateway

An inbound request received by the gateway

Gateway response

Agent Gateway

The response returned to the caller

Guardrail evaluation

Shield SDK / Gateway

A guardrail policy evaluated against a prompt, tool call, or response

Tool call

Shield SDK

A tool invoked by the agent

Tool response

Shield SDK

The result returned from a tool

Browser violation

Browser Security

A browser-side event correlated into this session by user identity

Threat event

Any

A detection that fired during this session

Each event in the timeline shows its timestamp, type, and a brief summary. Click any event to open the event detail panel.

Guardrail results in the timeline

Guardrail evaluation events show:

  • The policy evaluated

  • The input evaluated (prompt, tool call, or response)

  • The outcome: allow, block, or monitor

  • Which rules matched, if any

Timeline annotations

The timeline is annotated with:

  • Threat markers — red indicators on events that contain or triggered a detection

  • Policy gap markers — yellow indicators on Monitor-mode events that would not have been blocked


Session detail panel

The session detail panel (opened by clicking a session row) shows:

  • Summary — session metadata, agent identity, user, duration, products involved, source IDE

  • Agentic metrics — turn count, cumulative risk score, loop detection flag, token budget status, total tokens in/out, total cost (USD), unique tools used

  • Timeline — the full event sequence (described above)

  • Traces — links to OpenTelemetry traces associated with this session. Each agent LLM call that was instrumented with the Shield SDK or passed through the gateway will appear as a trace entry.

  • Violations — a filtered list of all threat events in this session, with links to the Threats view

  • Entity panel — recent sessions and risk score for the agent and user involved

The agentic metrics are particularly useful for identifying runaway sessions: a high cumulative risk score with many turns, or a loop_detected flag, indicates a session that may be stuck or manipulated. budget_exceeded means the session consumed more tokens than its configured limit.


Cross-product session correlation

Sessions are correlated across products using a shared session ID that flows through:

  1. Agent Gateway — assigns a session ID to each inbound connection and propagates it in response headers and webhook payloads

  2. Shield SDK — reads the session ID from the gateway response header (or generates one if the agent is not using the gateway) and attaches it to all SDK-emitted events

  3. Browser Security — links violations to sessions by user identity when a violation occurs within the session's active time window

This means a single session can include gateway events, SDK guardrail evaluations, and browser-side violations from the same user, all visible in a unified timeline.

Last updated