On Site Agent · Logs

What this is

This is an internal log viewer for the On Site Agent project on Braintrust. It lets you browse user conversations, inspect AI responses, and review quality and safety evaluations.

Navigating traces

• Left sidebar — lists recent user prompts. Click any to load the full conversation in the center panel.
• Arrow keys ↑ ↓ — once a trace is selected, navigate up and down the list without using the mouse.
• Color coding — the left border alternates orange / blue to visually group messages from the same session (conversation). Siblings are always kept together regardless of date order.
• 👍 / 👎 — shown on traces where the user submitted explicit feedback.

Filters

• Prompt — filter by system prompt ID (populated from loaded data).
• Is Staff — hide or show messages from staff accounts. Defaults to False (real users only).
• Has Feedback — show only traces where a user left a thumbs up or down.

Right panel

Selecting a trace loads full detail on demand: user feedback, quality and safety eval scores, referenced texts, and metadata (model, origin, session, duration). This loads separately after the conversation text.

Performance & data

• The viewer loads 200 traces at a time. Use Load more at the bottom of the sidebar to fetch additional pages.
• Data is fetched live from the Braintrust API via a proxy — it reflects the current state of the project logs.
• The API enforces rate limits. If you see a "Rate limited" status, the viewer will automatically retry with exponential backoff. Loading may take 10–30 seconds on a cold start.
• Response bodies are loaded on demand when you click a trace, not up front — this keeps initial load fast.