Trace

The Trace tab provides post-mortem analysis after a query completes. It adds the inner three observation layers – Prompt, Thinking, and Raw Response – that are not visible during the Live stream. The tab is hidden during execution and only rendered after the done event arrives.

Inner Layers

For each LLM call, the Trace tab exposes three layers exclusive to post-mortem inspection:

• Prompt – the full prompt sent to the model, with [System] / [Instructions] path-format separators
• Thinking – chain-of-thought reasoning, when thinking is enabled for that role
• Raw Response – the unprocessed model output before any parsing

Snapshot Consistency

The Trace tab uses HTML snapshots captured from the Live tab rather than re-rendering from data. During execution, each completed stage is snapshot into a stageSnapshots dictionary keyed by role (react, grader, judge, fallback, done). Multi-round pipelines produce multiple snapshots per role – one per round. This guarantees that what you see in Trace is exactly what streamed in Live.

Collapsible Layers

Each layer is collapsible. By default, Prompt, Thinking, and Raw Response are collapsed while Parsed Output is expanded. Click any layer header to toggle it. The Prompt layer shows path-format separators (e.g. [System] / [Instructions]) for quick scanning before expanding the full text.

Trace Cards

Each stage renders as a trace card with a dark background (#0d0d0d), subtle border, and a left-side color indicator matching the role – cyan for ReAct, yellow for Grader, green for Judge, pink for Fallback. Cards stack vertically and include the Live snapshot followed by the LLM layers below it.